Line data Source code
1 : // SPDX-License-Identifier: GPL-2.0
2 : /*
3 : * Workingset detection
4 : *
5 : * Copyright (C) 2013 Red Hat, Inc., Johannes Weiner
6 : */
7 :
8 : #include <linux/memcontrol.h>
9 : #include <linux/mm_inline.h>
10 : #include <linux/writeback.h>
11 : #include <linux/shmem_fs.h>
12 : #include <linux/pagemap.h>
13 : #include <linux/atomic.h>
14 : #include <linux/module.h>
15 : #include <linux/swap.h>
16 : #include <linux/dax.h>
17 : #include <linux/fs.h>
18 : #include <linux/mm.h>
19 :
20 : /*
21 : * Double CLOCK lists
22 : *
23 : * Per node, two clock lists are maintained for file pages: the
24 : * inactive and the active list. Freshly faulted pages start out at
25 : * the head of the inactive list and page reclaim scans pages from the
26 : * tail. Pages that are accessed multiple times on the inactive list
27 : * are promoted to the active list, to protect them from reclaim,
28 : * whereas active pages are demoted to the inactive list when the
29 : * active list grows too big.
30 : *
31 : * fault ------------------------+
32 : * |
33 : * +--------------+ | +-------------+
34 : * reclaim <- | inactive | <-+-- demotion | active | <--+
35 : * +--------------+ +-------------+ |
36 : * | |
37 : * +-------------- promotion ------------------+
38 : *
39 : *
40 : * Access frequency and refault distance
41 : *
42 : * A workload is thrashing when its pages are frequently used but they
43 : * are evicted from the inactive list every time before another access
44 : * would have promoted them to the active list.
45 : *
46 : * In cases where the average access distance between thrashing pages
47 : * is bigger than the size of memory there is nothing that can be
48 : * done - the thrashing set could never fit into memory under any
49 : * circumstance.
50 : *
51 : * However, the average access distance could be bigger than the
52 : * inactive list, yet smaller than the size of memory. In this case,
53 : * the set could fit into memory if it weren't for the currently
54 : * active pages - which may be used more, hopefully less frequently:
55 : *
56 : * +-memory available to cache-+
57 : * | |
58 : * +-inactive------+-active----+
59 : * a b | c d e f g h i | J K L M N |
60 : * +---------------+-----------+
61 : *
62 : * It is prohibitively expensive to accurately track access frequency
63 : * of pages. But a reasonable approximation can be made to measure
64 : * thrashing on the inactive list, after which refaulting pages can be
65 : * activated optimistically to compete with the existing active pages.
66 : *
67 : * Approximating inactive page access frequency - Observations:
68 : *
69 : * 1. When a page is accessed for the first time, it is added to the
70 : * head of the inactive list, slides every existing inactive page
71 : * towards the tail by one slot, and pushes the current tail page
72 : * out of memory.
73 : *
74 : * 2. When a page is accessed for the second time, it is promoted to
75 : * the active list, shrinking the inactive list by one slot. This
76 : * also slides all inactive pages that were faulted into the cache
77 : * more recently than the activated page towards the tail of the
78 : * inactive list.
79 : *
80 : * Thus:
81 : *
82 : * 1. The sum of evictions and activations between any two points in
83 : * time indicate the minimum number of inactive pages accessed in
84 : * between.
85 : *
86 : * 2. Moving one inactive page N page slots towards the tail of the
87 : * list requires at least N inactive page accesses.
88 : *
89 : * Combining these:
90 : *
91 : * 1. When a page is finally evicted from memory, the number of
92 : * inactive pages accessed while the page was in cache is at least
93 : * the number of page slots on the inactive list.
94 : *
95 : * 2. In addition, measuring the sum of evictions and activations (E)
96 : * at the time of a page's eviction, and comparing it to another
97 : * reading (R) at the time the page faults back into memory tells
98 : * the minimum number of accesses while the page was not cached.
99 : * This is called the refault distance.
100 : *
101 : * Because the first access of the page was the fault and the second
102 : * access the refault, we combine the in-cache distance with the
103 : * out-of-cache distance to get the complete minimum access distance
104 : * of this page:
105 : *
106 : * NR_inactive + (R - E)
107 : *
108 : * And knowing the minimum access distance of a page, we can easily
109 : * tell if the page would be able to stay in cache assuming all page
110 : * slots in the cache were available:
111 : *
112 : * NR_inactive + (R - E) <= NR_inactive + NR_active
113 : *
114 : * If we have swap we should consider about NR_inactive_anon and
115 : * NR_active_anon, so for page cache and anonymous respectively:
116 : *
117 : * NR_inactive_file + (R - E) <= NR_inactive_file + NR_active_file
118 : * + NR_inactive_anon + NR_active_anon
119 : *
120 : * NR_inactive_anon + (R - E) <= NR_inactive_anon + NR_active_anon
121 : * + NR_inactive_file + NR_active_file
122 : *
123 : * Which can be further simplified to:
124 : *
125 : * (R - E) <= NR_active_file + NR_inactive_anon + NR_active_anon
126 : *
127 : * (R - E) <= NR_active_anon + NR_inactive_file + NR_active_file
128 : *
129 : * Put into words, the refault distance (out-of-cache) can be seen as
130 : * a deficit in inactive list space (in-cache). If the inactive list
131 : * had (R - E) more page slots, the page would not have been evicted
132 : * in between accesses, but activated instead. And on a full system,
133 : * the only thing eating into inactive list space is active pages.
134 : *
135 : *
136 : * Refaulting inactive pages
137 : *
138 : * All that is known about the active list is that the pages have been
139 : * accessed more than once in the past. This means that at any given
140 : * time there is actually a good chance that pages on the active list
141 : * are no longer in active use.
142 : *
143 : * So when a refault distance of (R - E) is observed and there are at
144 : * least (R - E) pages in the userspace workingset, the refaulting page
145 : * is activated optimistically in the hope that (R - E) pages are actually
146 : * used less frequently than the refaulting page - or even not used at
147 : * all anymore.
148 : *
149 : * That means if inactive cache is refaulting with a suitable refault
150 : * distance, we assume the cache workingset is transitioning and put
151 : * pressure on the current workingset.
152 : *
153 : * If this is wrong and demotion kicks in, the pages which are truly
154 : * used more frequently will be reactivated while the less frequently
155 : * used once will be evicted from memory.
156 : *
157 : * But if this is right, the stale pages will be pushed out of memory
158 : * and the used pages get to stay in cache.
159 : *
160 : * Refaulting active pages
161 : *
162 : * If on the other hand the refaulting pages have recently been
163 : * deactivated, it means that the active list is no longer protecting
164 : * actively used cache from reclaim. The cache is NOT transitioning to
165 : * a different workingset; the existing workingset is thrashing in the
166 : * space allocated to the page cache.
167 : *
168 : *
169 : * Implementation
170 : *
171 : * For each node's LRU lists, a counter for inactive evictions and
172 : * activations is maintained (node->nonresident_age).
173 : *
174 : * On eviction, a snapshot of this counter (along with some bits to
175 : * identify the node) is stored in the now empty page cache
176 : * slot of the evicted page. This is called a shadow entry.
177 : *
178 : * On cache misses for which there are shadow entries, an eligible
179 : * refault distance will immediately activate the refaulting page.
180 : */
181 :
182 : #define WORKINGSET_SHIFT 1
183 : #define EVICTION_SHIFT ((BITS_PER_LONG - BITS_PER_XA_VALUE) + \
184 : WORKINGSET_SHIFT + NODES_SHIFT + \
185 : MEM_CGROUP_ID_SHIFT)
186 : #define EVICTION_MASK (~0UL >> EVICTION_SHIFT)
187 :
188 : /*
189 : * Eviction timestamps need to be able to cover the full range of
190 : * actionable refaults. However, bits are tight in the xarray
191 : * entry, and after storing the identifier for the lruvec there might
192 : * not be enough left to represent every single actionable refault. In
193 : * that case, we have to sacrifice granularity for distance, and group
194 : * evictions into coarser buckets by shaving off lower timestamp bits.
195 : */
196 : static unsigned int bucket_order __read_mostly;
197 :
198 0 : static void *pack_shadow(int memcgid, pg_data_t *pgdat, unsigned long eviction,
199 : bool workingset)
200 : {
201 0 : eviction &= EVICTION_MASK;
202 0 : eviction = (eviction << MEM_CGROUP_ID_SHIFT) | memcgid;
203 0 : eviction = (eviction << NODES_SHIFT) | pgdat->node_id;
204 0 : eviction = (eviction << WORKINGSET_SHIFT) | workingset;
205 :
206 0 : return xa_mk_value(eviction);
207 : }
208 :
209 : static void unpack_shadow(void *shadow, int *memcgidp, pg_data_t **pgdat,
210 : unsigned long *evictionp, bool *workingsetp)
211 : {
212 0 : unsigned long entry = xa_to_value(shadow);
213 : int memcgid, nid;
214 : bool workingset;
215 :
216 0 : workingset = entry & ((1UL << WORKINGSET_SHIFT) - 1);
217 0 : entry >>= WORKINGSET_SHIFT;
218 0 : nid = entry & ((1UL << NODES_SHIFT) - 1);
219 0 : entry >>= NODES_SHIFT;
220 0 : memcgid = entry & ((1UL << MEM_CGROUP_ID_SHIFT) - 1);
221 0 : entry >>= MEM_CGROUP_ID_SHIFT;
222 :
223 0 : *memcgidp = memcgid;
224 0 : *pgdat = NODE_DATA(nid);
225 0 : *evictionp = entry;
226 0 : *workingsetp = workingset;
227 : }
228 :
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;
237 : struct lru_gen_folio *lrugen;
238 : int type = folio_is_file_lru(folio);
239 : int delta = folio_nr_pages(folio);
240 : int refs = folio_lru_refs(folio);
241 : int tier = lru_tier_from_refs(refs);
242 : struct mem_cgroup *memcg = folio_memcg(folio);
243 : struct pglist_data *pgdat = folio_pgdat(folio);
244 :
245 : BUILD_BUG_ON(LRU_GEN_WIDTH + LRU_REFS_WIDTH > BITS_PER_LONG - EVICTION_SHIFT);
246 :
247 : lruvec = mem_cgroup_lruvec(memcg, pgdat);
248 : lrugen = &lruvec->lrugen;
249 : min_seq = READ_ONCE(lrugen->min_seq[type]);
250 : token = (min_seq << LRU_REFS_WIDTH) | max(refs - 1, 0);
251 :
252 : hist = lru_hist_from_seq(min_seq);
253 : atomic_long_add(delta, &lrugen->evicted[hist][type][tier]);
254 :
255 : return pack_shadow(mem_cgroup_id(memcg), pgdat, token, refs);
256 : }
257 :
258 : 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;
266 : struct lru_gen_folio *lrugen;
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 :
326 : /**
327 : * workingset_age_nonresident - age non-resident entries as LRU ages
328 : * @lruvec: the lruvec that was aged
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 0 : void workingset_age_nonresident(struct lruvec *lruvec, unsigned long nr_pages)
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 {
350 0 : atomic_long_add(nr_pages, &lruvec->nonresident_age);
351 0 : } while ((lruvec = parent_lruvec(lruvec)));
352 0 : }
353 :
354 : /**
355 : * workingset_eviction - note the eviction of a folio from memory
356 : * @target_memcg: the cgroup that is causing the reclaim
357 : * @folio: the folio being evicted
358 : *
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.
361 : */
362 0 : void *workingset_eviction(struct folio *folio, struct mem_cgroup *target_memcg)
363 : {
364 0 : struct pglist_data *pgdat = folio_pgdat(folio);
365 : unsigned long eviction;
366 : struct lruvec *lruvec;
367 : int memcgid;
368 :
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);
373 :
374 : if (lru_gen_enabled())
375 : return lru_gen_eviction(folio);
376 :
377 0 : lruvec = mem_cgroup_lruvec(target_memcg, pgdat);
378 : /* XXX: target_memcg can be NULL, go through lruvec */
379 0 : memcgid = mem_cgroup_id(lruvec_memcg(lruvec));
380 0 : eviction = atomic_long_read(&lruvec->nonresident_age);
381 0 : eviction >>= bucket_order;
382 0 : workingset_age_nonresident(lruvec, folio_nr_pages(folio));
383 0 : return pack_shadow(memcgid, pgdat, eviction,
384 0 : folio_test_workingset(folio));
385 : }
386 :
387 : /**
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.
391 : *
392 : * Calculates and evaluates the refault distance of the previously
393 : * evicted folio in the context of the node and the memcg whose memory
394 : * pressure caused the eviction.
395 : */
396 0 : void workingset_refault(struct folio *folio, void *shadow)
397 : {
398 0 : bool file = folio_is_file_lru(folio);
399 : struct mem_cgroup *eviction_memcg;
400 : struct lruvec *eviction_lruvec;
401 : unsigned long refault_distance;
402 : unsigned long workingset_size;
403 : struct pglist_data *pgdat;
404 : struct mem_cgroup *memcg;
405 : unsigned long eviction;
406 : struct lruvec *lruvec;
407 : unsigned long refault;
408 : bool workingset;
409 : int memcgid;
410 : long nr;
411 :
412 : if (lru_gen_enabled()) {
413 : lru_gen_refault(folio, shadow);
414 : return;
415 : }
416 :
417 0 : unpack_shadow(shadow, &memcgid, &pgdat, &eviction, &workingset);
418 0 : eviction <<= bucket_order;
419 :
420 : /* Flush stats (and potentially sleep) before holding RCU read lock */
421 : mem_cgroup_flush_stats_ratelimited();
422 :
423 : rcu_read_lock();
424 : /*
425 : * Look up the memcg associated with the stored ID. It might
426 : * have been deleted since the folio's eviction.
427 : *
428 : * Note that in rare events the ID could have been recycled
429 : * for a new cgroup that refaults a shared folio. This is
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 : */
440 0 : eviction_memcg = mem_cgroup_from_id(memcgid);
441 : if (!mem_cgroup_disabled() && !eviction_memcg)
442 : goto out;
443 0 : eviction_lruvec = mem_cgroup_lruvec(eviction_memcg, pgdat);
444 0 : refault = atomic_long_read(&eviction_lruvec->nonresident_age);
445 :
446 : /*
447 : * Calculate the refault distance
448 : *
449 : * The unsigned subtraction here gives an accurate distance
450 : * across nonresident_age overflows in most cases. There is a
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
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
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.
461 : */
462 0 : refault_distance = (refault - eviction) & EVICTION_MASK;
463 :
464 : /*
465 : * The activation decision for this folio is made at the level
466 : * where the eviction occurred, as that is where the LRU order
467 : * during folio reclaim is being determined.
468 : *
469 : * However, the cgroup that will own the folio is the one that
470 : * is actually experiencing the refault event.
471 : */
472 0 : nr = folio_nr_pages(folio);
473 0 : memcg = folio_memcg(folio);
474 0 : pgdat = folio_pgdat(folio);
475 0 : lruvec = mem_cgroup_lruvec(memcg, pgdat);
476 :
477 0 : mod_lruvec_state(lruvec, WORKINGSET_REFAULT_BASE + file, nr);
478 : /*
479 : * Compare the distance to the existing workingset size. We
480 : * don't activate pages that couldn't stay resident even if
481 : * all the memory was available to the workingset. Whether
482 : * workingset competition needs to consider anon or not depends
483 : * on having free swap space.
484 : */
485 0 : workingset_size = lruvec_page_state(eviction_lruvec, NR_ACTIVE_FILE);
486 0 : if (!file) {
487 0 : workingset_size += lruvec_page_state(eviction_lruvec,
488 : NR_INACTIVE_FILE);
489 : }
490 0 : if (mem_cgroup_get_nr_swap_pages(eviction_memcg) > 0) {
491 0 : workingset_size += lruvec_page_state(eviction_lruvec,
492 : NR_ACTIVE_ANON);
493 0 : if (file) {
494 0 : workingset_size += lruvec_page_state(eviction_lruvec,
495 : NR_INACTIVE_ANON);
496 : }
497 : }
498 0 : if (refault_distance > workingset_size)
499 : goto out;
500 :
501 0 : folio_set_active(folio);
502 0 : workingset_age_nonresident(lruvec, nr);
503 0 : mod_lruvec_state(lruvec, WORKINGSET_ACTIVATE_BASE + file, nr);
504 :
505 : /* Folio was active prior to eviction */
506 0 : if (workingset) {
507 0 : folio_set_workingset(folio);
508 : /*
509 : * XXX: Move to folio_add_lru() when it supports new vs
510 : * putback
511 : */
512 0 : lru_note_cost_refault(folio);
513 0 : mod_lruvec_state(lruvec, WORKINGSET_RESTORE_BASE + file, nr);
514 : }
515 : out:
516 : rcu_read_unlock();
517 : }
518 :
519 : /**
520 : * workingset_activation - note a page activation
521 : * @folio: Folio that is being activated.
522 : */
523 0 : void workingset_activation(struct folio *folio)
524 : {
525 : struct mem_cgroup *memcg;
526 :
527 : rcu_read_lock();
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 : */
535 0 : memcg = folio_memcg_rcu(folio);
536 : if (!mem_cgroup_disabled() && !memcg)
537 : goto out;
538 0 : workingset_age_nonresident(folio_lruvec(folio), folio_nr_pages(folio));
539 : out:
540 : rcu_read_unlock();
541 0 : }
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 :
555 : struct list_lru shadow_nodes;
556 :
557 0 : void workingset_update_node(struct xa_node *node)
558 : {
559 : struct address_space *mapping;
560 :
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
567 : * as node->private_list is protected by the i_pages lock.
568 : */
569 0 : mapping = container_of(node->array, struct address_space, i_pages);
570 : lockdep_assert_held(&mapping->i_pages.xa_lock);
571 :
572 0 : if (node->count && node->count == node->nr_values) {
573 0 : if (list_empty(&node->private_list)) {
574 0 : list_lru_add(&shadow_nodes, &node->private_list);
575 : __inc_lruvec_kmem_state(node, WORKINGSET_NODES);
576 : }
577 : } else {
578 0 : if (!list_empty(&node->private_list)) {
579 0 : list_lru_del(&shadow_nodes, &node->private_list);
580 : __dec_lruvec_kmem_state(node, WORKINGSET_NODES);
581 : }
582 : }
583 0 : }
584 :
585 0 : static unsigned long count_shadow_nodes(struct shrinker *shrinker,
586 : struct shrink_control *sc)
587 : {
588 : unsigned long max_nodes;
589 : unsigned long nodes;
590 : unsigned long pages;
591 :
592 0 : nodes = list_lru_shrink_count(&shadow_nodes, sc);
593 0 : if (!nodes)
594 : return SHRINK_EMPTY;
595 :
596 : /*
597 : * Approximate a reasonable limit for the nodes
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.
611 : *
612 : * On 64-bit with 7 xa_nodes per page and 64 slots
613 : * each, this will reclaim shadow entries when they consume
614 : * ~1.8% of available memory:
615 : *
616 : * PAGE_SIZE / xa_nodes / node_entries * 8 / PAGE_SIZE
617 : */
618 : #ifdef CONFIG_MEMCG
619 : if (sc->memcg) {
620 : struct lruvec *lruvec;
621 : int i;
622 :
623 : lruvec = mem_cgroup_lruvec(sc->memcg, NODE_DATA(sc->nid));
624 : for (pages = 0, i = 0; i < NR_LRU_LISTS; i++)
625 : pages += lruvec_page_state_local(lruvec,
626 : NR_LRU_BASE + i);
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;
631 : } else
632 : #endif
633 0 : pages = node_present_pages(sc->nid);
634 :
635 0 : max_nodes = pages >> (XA_CHUNK_SHIFT - 3);
636 :
637 0 : if (nodes <= max_nodes)
638 : return 0;
639 0 : return nodes - max_nodes;
640 : }
641 :
642 0 : static enum lru_status shadow_lru_isolate(struct list_head *item,
643 : struct list_lru_one *lru,
644 : spinlock_t *lru_lock,
645 : void *arg) __must_hold(lru_lock)
646 : {
647 0 : struct xa_node *node = container_of(item, struct xa_node, private_list);
648 : struct address_space *mapping;
649 : int ret;
650 :
651 : /*
652 : * Page cache insertions and deletions synchronously maintain
653 : * the shadow node LRU under the i_pages lock and the
654 : * lru_lock. Because the page cache tree is emptied before
655 : * the inode can be destroyed, holding the lru_lock pins any
656 : * address_space that has nodes on the LRU.
657 : *
658 : * We can then safely transition to the i_pages lock to
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 :
663 0 : mapping = container_of(node->array, struct address_space, i_pages);
664 :
665 : /* Coming from the list, invert the lock order */
666 0 : if (!xa_trylock(&mapping->i_pages)) {
667 : spin_unlock_irq(lru_lock);
668 : ret = LRU_RETRY;
669 : goto out;
670 : }
671 :
672 : /* For page cache we need to hold i_lock */
673 0 : if (mapping->host != NULL) {
674 0 : 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 : }
680 : }
681 :
682 0 : list_lru_isolate(lru, item);
683 0 : __dec_lruvec_kmem_state(node, WORKINGSET_NODES);
684 :
685 0 : 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 : */
692 0 : if (WARN_ON_ONCE(!node->nr_values))
693 : goto out_invalid;
694 0 : if (WARN_ON_ONCE(node->count != node->nr_values))
695 : goto out_invalid;
696 0 : xa_delete_node(node, workingset_update_node);
697 : __inc_lruvec_kmem_state(node, WORKINGSET_NODERECLAIM);
698 :
699 : out_invalid:
700 0 : xa_unlock_irq(&mapping->i_pages);
701 0 : if (mapping->host != NULL) {
702 0 : if (mapping_shrinkable(mapping))
703 0 : inode_add_lru(mapping->host);
704 0 : spin_unlock(&mapping->host->i_lock);
705 : }
706 0 : ret = LRU_REMOVED_RETRY;
707 : out:
708 0 : cond_resched();
709 0 : spin_lock_irq(lru_lock);
710 0 : return ret;
711 : }
712 :
713 0 : static unsigned long scan_shadow_nodes(struct shrinker *shrinker,
714 : struct shrink_control *sc)
715 : {
716 : /* list_lru lock nests inside the IRQ-safe i_pages lock */
717 0 : return list_lru_shrink_walk_irq(&shadow_nodes, sc, shadow_lru_isolate,
718 : NULL);
719 : }
720 :
721 : static struct shrinker workingset_shadow_shrinker = {
722 : .count_objects = count_shadow_nodes,
723 : .scan_objects = scan_shadow_nodes,
724 : .seeks = 0, /* ->count reports only fully expendable nodes */
725 : .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE,
726 : };
727 :
728 : /*
729 : * Our list_lru->lock is IRQ-safe as it nests inside the IRQ-safe
730 : * i_pages lock.
731 : */
732 : static struct lock_class_key shadow_nodes_key;
733 :
734 1 : static int __init workingset_init(void)
735 : {
736 : unsigned int timestamp_bits;
737 : unsigned int max_order;
738 : int ret;
739 :
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 1 : timestamp_bits = BITS_PER_LONG - EVICTION_SHIFT;
749 2 : max_order = fls_long(totalram_pages() - 1);
750 1 : if (max_order > timestamp_bits)
751 0 : bucket_order = max_order - timestamp_bits;
752 1 : pr_info("workingset: timestamp_bits=%d max_order=%d bucket_order=%u\n",
753 : timestamp_bits, max_order, bucket_order);
754 :
755 1 : ret = prealloc_shrinker(&workingset_shadow_shrinker, "mm-shadow");
756 1 : if (ret)
757 : goto err;
758 1 : ret = __list_lru_init(&shadow_nodes, true, &shadow_nodes_key,
759 : &workingset_shadow_shrinker);
760 1 : if (ret)
761 : goto err_list_lru;
762 1 : register_shrinker_prepared(&workingset_shadow_shrinker);
763 1 : return 0;
764 : err_list_lru:
765 0 : free_prealloced_shrinker(&workingset_shadow_shrinker);
766 : err:
767 : return ret;
768 : }
769 : module_init(workingset_init);
|
