1 | // SPDX-License-Identifier: GPL-2.0 |
2 | /* |
3 | * Workingset detection |
4 | * |
5 | * Copyright (C) 2013 Red Hat, Inc., Johannes Weiner |
6 | */ |
7 | |
8 | #include <linux/memcontrol.h> |
9 | #include <linux/mm_inline.h> |
10 | #include <linux/writeback.h> |
11 | #include <linux/shmem_fs.h> |
12 | #include <linux/pagemap.h> |
13 | #include <linux/atomic.h> |
14 | #include <linux/module.h> |
15 | #include <linux/swap.h> |
16 | #include <linux/dax.h> |
17 | #include <linux/fs.h> |
18 | #include <linux/mm.h> |
19 | #include "internal.h" |
20 | |
21 | /* |
22 | * Double CLOCK lists |
23 | * |
24 | * Per node, two clock lists are maintained for file pages: the |
25 | * inactive and the active list. Freshly faulted pages start out at |
26 | * the head of the inactive list and page reclaim scans pages from the |
27 | * tail. Pages that are accessed multiple times on the inactive list |
28 | * are promoted to the active list, to protect them from reclaim, |
29 | * whereas active pages are demoted to the inactive list when the |
30 | * active list grows too big. |
31 | * |
32 | * fault ------------------------+ |
33 | * | |
34 | * +--------------+ | +-------------+ |
35 | * reclaim <- | inactive | <-+-- demotion | active | <--+ |
36 | * +--------------+ +-------------+ | |
37 | * | | |
38 | * +-------------- promotion ------------------+ |
39 | * |
40 | * |
41 | * Access frequency and refault distance |
42 | * |
43 | * A workload is thrashing when its pages are frequently used but they |
44 | * are evicted from the inactive list every time before another access |
45 | * would have promoted them to the active list. |
46 | * |
47 | * In cases where the average access distance between thrashing pages |
48 | * is bigger than the size of memory there is nothing that can be |
49 | * done - the thrashing set could never fit into memory under any |
50 | * circumstance. |
51 | * |
52 | * However, the average access distance could be bigger than the |
53 | * inactive list, yet smaller than the size of memory. In this case, |
54 | * the set could fit into memory if it weren't for the currently |
55 | * active pages - which may be used more, hopefully less frequently: |
56 | * |
57 | * +-memory available to cache-+ |
58 | * | | |
59 | * +-inactive------+-active----+ |
60 | * a b | c d e f g h i | J K L M N | |
61 | * +---------------+-----------+ |
62 | * |
63 | * It is prohibitively expensive to accurately track access frequency |
64 | * of pages. But a reasonable approximation can be made to measure |
65 | * thrashing on the inactive list, after which refaulting pages can be |
66 | * activated optimistically to compete with the existing active pages. |
67 | * |
68 | * Approximating inactive page access frequency - Observations: |
69 | * |
70 | * 1. When a page is accessed for the first time, it is added to the |
71 | * head of the inactive list, slides every existing inactive page |
72 | * towards the tail by one slot, and pushes the current tail page |
73 | * out of memory. |
74 | * |
75 | * 2. When a page is accessed for the second time, it is promoted to |
76 | * the active list, shrinking the inactive list by one slot. This |
77 | * also slides all inactive pages that were faulted into the cache |
78 | * more recently than the activated page towards the tail of the |
79 | * inactive list. |
80 | * |
81 | * Thus: |
82 | * |
83 | * 1. The sum of evictions and activations between any two points in |
84 | * time indicate the minimum number of inactive pages accessed in |
85 | * between. |
86 | * |
87 | * 2. Moving one inactive page N page slots towards the tail of the |
88 | * list requires at least N inactive page accesses. |
89 | * |
90 | * Combining these: |
91 | * |
92 | * 1. When a page is finally evicted from memory, the number of |
93 | * inactive pages accessed while the page was in cache is at least |
94 | * the number of page slots on the inactive list. |
95 | * |
96 | * 2. In addition, measuring the sum of evictions and activations (E) |
97 | * at the time of a page's eviction, and comparing it to another |
98 | * reading (R) at the time the page faults back into memory tells |
99 | * the minimum number of accesses while the page was not cached. |
100 | * This is called the refault distance. |
101 | * |
102 | * Because the first access of the page was the fault and the second |
103 | * access the refault, we combine the in-cache distance with the |
104 | * out-of-cache distance to get the complete minimum access distance |
105 | * of this page: |
106 | * |
107 | * NR_inactive + (R - E) |
108 | * |
109 | * And knowing the minimum access distance of a page, we can easily |
110 | * tell if the page would be able to stay in cache assuming all page |
111 | * slots in the cache were available: |
112 | * |
113 | * NR_inactive + (R - E) <= NR_inactive + NR_active |
114 | * |
115 | * If we have swap we should consider about NR_inactive_anon and |
116 | * NR_active_anon, so for page cache and anonymous respectively: |
117 | * |
118 | * NR_inactive_file + (R - E) <= NR_inactive_file + NR_active_file |
119 | * + NR_inactive_anon + NR_active_anon |
120 | * |
121 | * NR_inactive_anon + (R - E) <= NR_inactive_anon + NR_active_anon |
122 | * + NR_inactive_file + NR_active_file |
123 | * |
124 | * Which can be further simplified to: |
125 | * |
126 | * (R - E) <= NR_active_file + NR_inactive_anon + NR_active_anon |
127 | * |
128 | * (R - E) <= NR_active_anon + NR_inactive_file + NR_active_file |
129 | * |
130 | * Put into words, the refault distance (out-of-cache) can be seen as |
131 | * a deficit in inactive list space (in-cache). If the inactive list |
132 | * had (R - E) more page slots, the page would not have been evicted |
133 | * in between accesses, but activated instead. And on a full system, |
134 | * the only thing eating into inactive list space is active pages. |
135 | * |
136 | * |
137 | * Refaulting inactive pages |
138 | * |
139 | * All that is known about the active list is that the pages have been |
140 | * accessed more than once in the past. This means that at any given |
141 | * time there is actually a good chance that pages on the active list |
142 | * are no longer in active use. |
143 | * |
144 | * So when a refault distance of (R - E) is observed and there are at |
145 | * least (R - E) pages in the userspace workingset, the refaulting page |
146 | * is activated optimistically in the hope that (R - E) pages are actually |
147 | * used less frequently than the refaulting page - or even not used at |
148 | * all anymore. |
149 | * |
150 | * That means if inactive cache is refaulting with a suitable refault |
151 | * distance, we assume the cache workingset is transitioning and put |
152 | * pressure on the current workingset. |
153 | * |
154 | * If this is wrong and demotion kicks in, the pages which are truly |
155 | * used more frequently will be reactivated while the less frequently |
156 | * used once will be evicted from memory. |
157 | * |
158 | * But if this is right, the stale pages will be pushed out of memory |
159 | * and the used pages get to stay in cache. |
160 | * |
161 | * Refaulting active pages |
162 | * |
163 | * If on the other hand the refaulting pages have recently been |
164 | * deactivated, it means that the active list is no longer protecting |
165 | * actively used cache from reclaim. The cache is NOT transitioning to |
166 | * a different workingset; the existing workingset is thrashing in the |
167 | * space allocated to the page cache. |
168 | * |
169 | * |
170 | * Implementation |
171 | * |
172 | * For each node's LRU lists, a counter for inactive evictions and |
173 | * activations is maintained (node->nonresident_age). |
174 | * |
175 | * On eviction, a snapshot of this counter (along with some bits to |
176 | * identify the node) is stored in the now empty page cache |
177 | * slot of the evicted page. This is called a shadow entry. |
178 | * |
179 | * On cache misses for which there are shadow entries, an eligible |
180 | * refault distance will immediately activate the refaulting page. |
181 | */ |
182 | |
183 | #define WORKINGSET_SHIFT 1 |
184 | #define EVICTION_SHIFT ((BITS_PER_LONG - BITS_PER_XA_VALUE) + \ |
185 | WORKINGSET_SHIFT + NODES_SHIFT + \ |
186 | MEM_CGROUP_ID_SHIFT) |
187 | #define EVICTION_MASK (~0UL >> EVICTION_SHIFT) |
188 | |
189 | /* |
190 | * Eviction timestamps need to be able to cover the full range of |
191 | * actionable refaults. However, bits are tight in the xarray |
192 | * entry, and after storing the identifier for the lruvec there might |
193 | * not be enough left to represent every single actionable refault. In |
194 | * that case, we have to sacrifice granularity for distance, and group |
195 | * evictions into coarser buckets by shaving off lower timestamp bits. |
196 | */ |
197 | static unsigned int bucket_order __read_mostly; |
198 | |
199 | static void *pack_shadow(int memcgid, pg_data_t *pgdat, unsigned long eviction, |
200 | bool workingset) |
201 | { |
202 | eviction &= EVICTION_MASK; |
203 | eviction = (eviction << MEM_CGROUP_ID_SHIFT) | memcgid; |
204 | eviction = (eviction << NODES_SHIFT) | pgdat->node_id; |
205 | eviction = (eviction << WORKINGSET_SHIFT) | workingset; |
206 | |
207 | return xa_mk_value(v: eviction); |
208 | } |
209 | |
210 | static void unpack_shadow(void *shadow, int *memcgidp, pg_data_t **pgdat, |
211 | unsigned long *evictionp, bool *workingsetp) |
212 | { |
213 | unsigned long entry = xa_to_value(entry: shadow); |
214 | int memcgid, nid; |
215 | bool workingset; |
216 | |
217 | workingset = entry & ((1UL << WORKINGSET_SHIFT) - 1); |
218 | entry >>= WORKINGSET_SHIFT; |
219 | nid = entry & ((1UL << NODES_SHIFT) - 1); |
220 | entry >>= NODES_SHIFT; |
221 | memcgid = entry & ((1UL << MEM_CGROUP_ID_SHIFT) - 1); |
222 | entry >>= MEM_CGROUP_ID_SHIFT; |
223 | |
224 | *memcgidp = memcgid; |
225 | *pgdat = NODE_DATA(nid); |
226 | *evictionp = entry; |
227 | *workingsetp = workingset; |
228 | } |
229 | |
230 | #ifdef CONFIG_LRU_GEN |
231 | |
232 | static void *lru_gen_eviction(struct folio *folio) |
233 | { |
234 | int hist; |
235 | unsigned long token; |
236 | unsigned long min_seq; |
237 | struct lruvec *lruvec; |
238 | struct lru_gen_folio *lrugen; |
239 | int type = folio_is_file_lru(folio); |
240 | int delta = folio_nr_pages(folio); |
241 | int refs = folio_lru_refs(folio); |
242 | 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(seq: min_seq); |
254 | atomic_long_add(i: delta, v: &lrugen->evicted[hist][type][tier]); |
255 | |
256 | return pack_shadow(memcgid: mem_cgroup_id(memcg), pgdat, eviction: token, workingset: refs); |
257 | } |
258 | |
259 | /* |
260 | * Tests if the shadow entry is for a folio that was recently evicted. |
261 | * Fills in @lruvec, @token, @workingset with the values unpacked from shadow. |
262 | */ |
263 | static bool lru_gen_test_recent(void *shadow, bool file, struct lruvec **lruvec, |
264 | unsigned long *token, bool *workingset) |
265 | { |
266 | int memcg_id; |
267 | unsigned long min_seq; |
268 | struct mem_cgroup *memcg; |
269 | struct pglist_data *pgdat; |
270 | |
271 | unpack_shadow(shadow, memcgidp: &memcg_id, pgdat: &pgdat, evictionp: token, workingsetp: workingset); |
272 | |
273 | memcg = mem_cgroup_from_id(id: memcg_id); |
274 | *lruvec = mem_cgroup_lruvec(memcg, pgdat); |
275 | |
276 | min_seq = READ_ONCE((*lruvec)->lrugen.min_seq[file]); |
277 | return (*token >> LRU_REFS_WIDTH) == (min_seq & (EVICTION_MASK >> LRU_REFS_WIDTH)); |
278 | } |
279 | |
280 | static void lru_gen_refault(struct folio *folio, void *shadow) |
281 | { |
282 | bool recent; |
283 | int hist, tier, refs; |
284 | bool workingset; |
285 | unsigned long token; |
286 | struct lruvec *lruvec; |
287 | struct lru_gen_folio *lrugen; |
288 | int type = folio_is_file_lru(folio); |
289 | int delta = folio_nr_pages(folio); |
290 | |
291 | rcu_read_lock(); |
292 | |
293 | recent = lru_gen_test_recent(shadow, file: type, lruvec: &lruvec, token: &token, workingset: &workingset); |
294 | if (lruvec != folio_lruvec(folio)) |
295 | goto unlock; |
296 | |
297 | mod_lruvec_state(lruvec, idx: WORKINGSET_REFAULT_BASE + type, val: delta); |
298 | |
299 | if (!recent) |
300 | goto unlock; |
301 | |
302 | lrugen = &lruvec->lrugen; |
303 | |
304 | hist = lru_hist_from_seq(READ_ONCE(lrugen->min_seq[type])); |
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(i: delta, v: &lrugen->refaulted[hist][type][tier]); |
310 | mod_lruvec_state(lruvec, idx: WORKINGSET_ACTIVATE_BASE + type, val: delta); |
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, |
317 | * they would have been protected by sort_folio(). |
318 | */ |
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)); |
321 | mod_lruvec_state(lruvec, idx: WORKINGSET_RESTORE_BASE + type, val: 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 | |
334 | static bool lru_gen_test_recent(void *shadow, bool file, struct lruvec **lruvec, |
335 | unsigned long *token, bool *workingset) |
336 | { |
337 | return false; |
338 | } |
339 | |
340 | static void lru_gen_refault(struct folio *folio, void *shadow) |
341 | { |
342 | } |
343 | |
344 | #endif /* CONFIG_LRU_GEN */ |
345 | |
346 | /** |
347 | * workingset_age_nonresident - age non-resident entries as LRU ages |
348 | * @lruvec: the lruvec that was aged |
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) |
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 { |
370 | atomic_long_add(i: nr_pages, v: &lruvec->nonresident_age); |
371 | } while ((lruvec = parent_lruvec(lruvec))); |
372 | } |
373 | |
374 | /** |
375 | * workingset_eviction - note the eviction of a folio from memory |
376 | * @target_memcg: the cgroup that is causing the reclaim |
377 | * @folio: the folio being evicted |
378 | * |
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. |
381 | */ |
382 | void *workingset_eviction(struct folio *folio, struct mem_cgroup *target_memcg) |
383 | { |
384 | struct pglist_data *pgdat = folio_pgdat(folio); |
385 | unsigned long eviction; |
386 | struct lruvec *lruvec; |
387 | int memcgid; |
388 | |
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); |
393 | |
394 | if (lru_gen_enabled()) |
395 | return lru_gen_eviction(folio); |
396 | |
397 | lruvec = mem_cgroup_lruvec(memcg: target_memcg, pgdat); |
398 | /* XXX: target_memcg can be NULL, go through lruvec */ |
399 | memcgid = mem_cgroup_id(memcg: lruvec_memcg(lruvec)); |
400 | eviction = atomic_long_read(v: &lruvec->nonresident_age); |
401 | eviction >>= bucket_order; |
402 | workingset_age_nonresident(lruvec, nr_pages: folio_nr_pages(folio)); |
403 | return pack_shadow(memcgid, pgdat, eviction, |
404 | workingset: folio_test_workingset(folio)); |
405 | } |
406 | |
407 | /** |
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. |
417 | */ |
418 | bool workingset_test_recent(void *shadow, bool file, bool *workingset) |
419 | { |
420 | struct mem_cgroup *eviction_memcg; |
421 | struct lruvec *eviction_lruvec; |
422 | unsigned long refault_distance; |
423 | unsigned long workingset_size; |
424 | unsigned long refault; |
425 | int memcgid; |
426 | struct pglist_data *pgdat; |
427 | unsigned long eviction; |
428 | |
429 | rcu_read_lock(); |
430 | |
431 | if (lru_gen_enabled()) { |
432 | bool recent = lru_gen_test_recent(shadow, file, |
433 | lruvec: &eviction_lruvec, token: &eviction, workingset); |
434 | |
435 | rcu_read_unlock(); |
436 | return recent; |
437 | } |
438 | |
439 | |
440 | unpack_shadow(shadow, memcgidp: &memcgid, pgdat: &pgdat, evictionp: &eviction, workingsetp: workingset); |
441 | eviction <<= bucket_order; |
442 | |
443 | /* |
444 | * Look up the memcg associated with the stored ID. It might |
445 | * have been deleted since the folio's eviction. |
446 | * |
447 | * Note that in rare events the ID could have been recycled |
448 | * for a new cgroup that refaults a shared folio. This is |
449 | * impossible to tell from the available data. However, this |
450 | * should be a rare and limited disturbance, and activations |
451 | * are always speculative anyway. Ultimately, it's the aging |
452 | * algorithm's job to shake out the minimum access frequency |
453 | * for the active cache. |
454 | * |
455 | * XXX: On !CONFIG_MEMCG, this will always return NULL; it |
456 | * would be better if the root_mem_cgroup existed in all |
457 | * configurations instead. |
458 | */ |
459 | eviction_memcg = mem_cgroup_from_id(id: memcgid); |
460 | if (!mem_cgroup_disabled() && |
461 | (!eviction_memcg || !mem_cgroup_tryget(memcg: eviction_memcg))) { |
462 | rcu_read_unlock(); |
463 | return false; |
464 | } |
465 | |
466 | rcu_read_unlock(); |
467 | |
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(memcg: eviction_memcg); |
474 | |
475 | eviction_lruvec = mem_cgroup_lruvec(memcg: eviction_memcg, pgdat); |
476 | refault = atomic_long_read(v: &eviction_lruvec->nonresident_age); |
477 | |
478 | /* |
479 | * Calculate the refault distance |
480 | * |
481 | * The unsigned subtraction here gives an accurate distance |
482 | * across nonresident_age overflows in most cases. There is a |
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 |
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 |
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. |
493 | */ |
494 | refault_distance = (refault - eviction) & EVICTION_MASK; |
495 | |
496 | /* |
497 | * Compare the distance to the existing workingset size. We |
498 | * don't activate pages that couldn't stay resident even if |
499 | * all the memory was available to the workingset. Whether |
500 | * workingset competition needs to consider anon or not depends |
501 | * on having free swap space. |
502 | */ |
503 | workingset_size = lruvec_page_state(lruvec: eviction_lruvec, idx: NR_ACTIVE_FILE); |
504 | if (!file) { |
505 | workingset_size += lruvec_page_state(lruvec: eviction_lruvec, |
506 | idx: NR_INACTIVE_FILE); |
507 | } |
508 | if (mem_cgroup_get_nr_swap_pages(memcg: eviction_memcg) > 0) { |
509 | workingset_size += lruvec_page_state(lruvec: eviction_lruvec, |
510 | idx: NR_ACTIVE_ANON); |
511 | if (file) { |
512 | workingset_size += lruvec_page_state(lruvec: eviction_lruvec, |
513 | idx: NR_INACTIVE_ANON); |
514 | } |
515 | } |
516 | |
517 | mem_cgroup_put(memcg: eviction_memcg); |
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 | |
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 |
550 | * is actually experiencing the refault event. Make sure the folio is |
551 | * locked to guarantee folio_memcg() stability throughout. |
552 | */ |
553 | VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); |
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, idx: WORKINGSET_REFAULT_BASE + file, val: nr); |
560 | |
561 | if (!workingset_test_recent(shadow, file, workingset: &workingset)) |
562 | return; |
563 | |
564 | folio_set_active(folio); |
565 | workingset_age_nonresident(lruvec, nr_pages: nr); |
566 | mod_lruvec_state(lruvec, idx: WORKINGSET_ACTIVATE_BASE + file, val: nr); |
567 | |
568 | /* Folio was active prior to eviction */ |
569 | if (workingset) { |
570 | folio_set_workingset(folio); |
571 | /* |
572 | * XXX: Move to folio_add_lru() when it supports new vs |
573 | * putback |
574 | */ |
575 | lru_note_cost_refault(folio); |
576 | mod_lruvec_state(lruvec, idx: WORKINGSET_RESTORE_BASE + file, val: nr); |
577 | } |
578 | } |
579 | |
580 | /** |
581 | * workingset_activation - note a page activation |
582 | * @folio: Folio that is being activated. |
583 | */ |
584 | void workingset_activation(struct folio *folio) |
585 | { |
586 | struct mem_cgroup *memcg; |
587 | |
588 | rcu_read_lock(); |
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 | */ |
596 | memcg = folio_memcg_rcu(folio); |
597 | if (!mem_cgroup_disabled() && !memcg) |
598 | goto out; |
599 | workingset_age_nonresident(lruvec: folio_lruvec(folio), nr_pages: folio_nr_pages(folio)); |
600 | out: |
601 | rcu_read_unlock(); |
602 | } |
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 | |
616 | struct list_lru shadow_nodes; |
617 | |
618 | void workingset_update_node(struct xa_node *node) |
619 | { |
620 | struct address_space *mapping; |
621 | |
622 | /* |
623 | * Track non-empty nodes that contain only shadow entries; |
624 | * unlink those that contain pages or are being freed. |
625 | * |
626 | * Avoid acquiring the list_lru lock when the nodes are |
627 | * already where they should be. The list_empty() test is safe |
628 | * as node->private_list is protected by the i_pages lock. |
629 | */ |
630 | mapping = container_of(node->array, struct address_space, i_pages); |
631 | lockdep_assert_held(&mapping->i_pages.xa_lock); |
632 | |
633 | if (node->count && node->count == node->nr_values) { |
634 | if (list_empty(head: &node->private_list)) { |
635 | list_lru_add_obj(lru: &shadow_nodes, item: &node->private_list); |
636 | __inc_lruvec_kmem_state(p: node, idx: WORKINGSET_NODES); |
637 | } |
638 | } else { |
639 | if (!list_empty(head: &node->private_list)) { |
640 | list_lru_del_obj(lru: &shadow_nodes, item: &node->private_list); |
641 | __dec_lruvec_kmem_state(p: node, idx: WORKINGSET_NODES); |
642 | } |
643 | } |
644 | } |
645 | |
646 | static unsigned long count_shadow_nodes(struct shrinker *shrinker, |
647 | struct shrink_control *sc) |
648 | { |
649 | unsigned long max_nodes; |
650 | unsigned long nodes; |
651 | unsigned long pages; |
652 | |
653 | nodes = list_lru_shrink_count(lru: &shadow_nodes, sc); |
654 | if (!nodes) |
655 | return SHRINK_EMPTY; |
656 | |
657 | /* |
658 | * Approximate a reasonable limit for the nodes |
659 | * containing shadow entries. We don't need to keep more |
660 | * shadow entries than possible pages on the active list, |
661 | * since refault distances bigger than that are dismissed. |
662 | * |
663 | * The size of the active list converges toward 100% of |
664 | * overall page cache as memory grows, with only a tiny |
665 | * inactive list. Assume the total cache size for that. |
666 | * |
667 | * Nodes might be sparsely populated, with only one shadow |
668 | * entry in the extreme case. Obviously, we cannot keep one |
669 | * node for every eligible shadow entry, so compromise on a |
670 | * worst-case density of 1/8th. Below that, not all eligible |
671 | * refaults can be detected anymore. |
672 | * |
673 | * On 64-bit with 7 xa_nodes per page and 64 slots |
674 | * each, this will reclaim shadow entries when they consume |
675 | * ~1.8% of available memory: |
676 | * |
677 | * PAGE_SIZE / xa_nodes / node_entries * 8 / PAGE_SIZE |
678 | */ |
679 | #ifdef CONFIG_MEMCG |
680 | if (sc->memcg) { |
681 | struct lruvec *lruvec; |
682 | int i; |
683 | |
684 | mem_cgroup_flush_stats_ratelimited(memcg: sc->memcg); |
685 | lruvec = mem_cgroup_lruvec(memcg: sc->memcg, NODE_DATA(sc->nid)); |
686 | for (pages = 0, i = 0; i < NR_LRU_LISTS; i++) |
687 | pages += lruvec_page_state_local(lruvec, |
688 | idx: NR_LRU_BASE + i); |
689 | pages += lruvec_page_state_local( |
690 | lruvec, idx: NR_SLAB_RECLAIMABLE_B) >> PAGE_SHIFT; |
691 | pages += lruvec_page_state_local( |
692 | lruvec, idx: NR_SLAB_UNRECLAIMABLE_B) >> PAGE_SHIFT; |
693 | } else |
694 | #endif |
695 | pages = node_present_pages(sc->nid); |
696 | |
697 | max_nodes = pages >> (XA_CHUNK_SHIFT - 3); |
698 | |
699 | if (nodes <= max_nodes) |
700 | return 0; |
701 | return nodes - max_nodes; |
702 | } |
703 | |
704 | static enum lru_status shadow_lru_isolate(struct list_head *item, |
705 | struct list_lru_one *lru, |
706 | spinlock_t *lru_lock, |
707 | void *arg) __must_hold(lru_lock) |
708 | { |
709 | struct xa_node *node = container_of(item, struct xa_node, private_list); |
710 | struct address_space *mapping; |
711 | int ret; |
712 | |
713 | /* |
714 | * Page cache insertions and deletions synchronously maintain |
715 | * the shadow node LRU under the i_pages lock and the |
716 | * lru_lock. Because the page cache tree is emptied before |
717 | * the inode can be destroyed, holding the lru_lock pins any |
718 | * address_space that has nodes on the LRU. |
719 | * |
720 | * We can then safely transition to the i_pages lock to |
721 | * pin only the address_space of the particular node we want |
722 | * to reclaim, take the node off-LRU, and drop the lru_lock. |
723 | */ |
724 | |
725 | mapping = container_of(node->array, struct address_space, i_pages); |
726 | |
727 | /* Coming from the list, invert the lock order */ |
728 | if (!xa_trylock(&mapping->i_pages)) { |
729 | spin_unlock_irq(lock: lru_lock); |
730 | ret = LRU_RETRY; |
731 | goto out; |
732 | } |
733 | |
734 | /* For page cache we need to hold i_lock */ |
735 | if (mapping->host != NULL) { |
736 | if (!spin_trylock(lock: &mapping->host->i_lock)) { |
737 | xa_unlock(&mapping->i_pages); |
738 | spin_unlock_irq(lock: lru_lock); |
739 | ret = LRU_RETRY; |
740 | goto out; |
741 | } |
742 | } |
743 | |
744 | list_lru_isolate(list: lru, item); |
745 | __dec_lruvec_kmem_state(p: node, idx: WORKINGSET_NODES); |
746 | |
747 | spin_unlock(lock: lru_lock); |
748 | |
749 | /* |
750 | * The nodes should only contain one or more shadow entries, |
751 | * no pages, so we expect to be able to remove them all and |
752 | * delete and free the empty node afterwards. |
753 | */ |
754 | if (WARN_ON_ONCE(!node->nr_values)) |
755 | goto out_invalid; |
756 | if (WARN_ON_ONCE(node->count != node->nr_values)) |
757 | goto out_invalid; |
758 | xa_delete_node(node, workingset_update_node); |
759 | __inc_lruvec_kmem_state(p: node, idx: WORKINGSET_NODERECLAIM); |
760 | |
761 | out_invalid: |
762 | xa_unlock_irq(&mapping->i_pages); |
763 | if (mapping->host != NULL) { |
764 | if (mapping_shrinkable(mapping)) |
765 | inode_add_lru(inode: mapping->host); |
766 | spin_unlock(lock: &mapping->host->i_lock); |
767 | } |
768 | ret = LRU_REMOVED_RETRY; |
769 | out: |
770 | cond_resched(); |
771 | spin_lock_irq(lock: lru_lock); |
772 | return ret; |
773 | } |
774 | |
775 | static unsigned long scan_shadow_nodes(struct shrinker *shrinker, |
776 | struct shrink_control *sc) |
777 | { |
778 | /* list_lru lock nests inside the IRQ-safe i_pages lock */ |
779 | return list_lru_shrink_walk_irq(lru: &shadow_nodes, sc, isolate: shadow_lru_isolate, |
780 | NULL); |
781 | } |
782 | |
783 | /* |
784 | * Our list_lru->lock is IRQ-safe as it nests inside the IRQ-safe |
785 | * i_pages lock. |
786 | */ |
787 | static struct lock_class_key shadow_nodes_key; |
788 | |
789 | static int __init workingset_init(void) |
790 | { |
791 | struct shrinker *workingset_shadow_shrinker; |
792 | unsigned int timestamp_bits; |
793 | unsigned int max_order; |
794 | int ret = -ENOMEM; |
795 | |
796 | BUILD_BUG_ON(BITS_PER_LONG < EVICTION_SHIFT); |
797 | /* |
798 | * Calculate the eviction bucket size to cover the longest |
799 | * actionable refault distance, which is currently half of |
800 | * memory (totalram_pages/2). However, memory hotplug may add |
801 | * some more pages at runtime, so keep working with up to |
802 | * double the initial memory by using totalram_pages as-is. |
803 | */ |
804 | timestamp_bits = BITS_PER_LONG - EVICTION_SHIFT; |
805 | max_order = fls_long(l: totalram_pages() - 1); |
806 | if (max_order > timestamp_bits) |
807 | bucket_order = max_order - timestamp_bits; |
808 | pr_info("workingset: timestamp_bits=%d max_order=%d bucket_order=%u\n" , |
809 | timestamp_bits, max_order, bucket_order); |
810 | |
811 | workingset_shadow_shrinker = shrinker_alloc(SHRINKER_NUMA_AWARE | |
812 | SHRINKER_MEMCG_AWARE, |
813 | fmt: "mm-shadow" ); |
814 | if (!workingset_shadow_shrinker) |
815 | goto err; |
816 | |
817 | ret = __list_lru_init(lru: &shadow_nodes, memcg_aware: true, key: &shadow_nodes_key, |
818 | shrinker: workingset_shadow_shrinker); |
819 | if (ret) |
820 | goto err_list_lru; |
821 | |
822 | workingset_shadow_shrinker->count_objects = count_shadow_nodes; |
823 | workingset_shadow_shrinker->scan_objects = scan_shadow_nodes; |
824 | /* ->count reports only fully expendable nodes */ |
825 | workingset_shadow_shrinker->seeks = 0; |
826 | |
827 | shrinker_register(shrinker: workingset_shadow_shrinker); |
828 | return 0; |
829 | err_list_lru: |
830 | shrinker_free(shrinker: workingset_shadow_shrinker); |
831 | err: |
832 | return ret; |
833 | } |
834 | module_init(workingset_init); |
835 | |