Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/mm/swap.c
4 *
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 */
7
8/*
9 * This file contains the default values for the operation of the
10 * Linux VM subsystem. Fine-tuning documentation can be found in
11 * Documentation/admin-guide/sysctl/vm.rst.
12 * Started 18.12.91
13 * Swap aging added 23.2.95, Stephen Tweedie.
14 * Buffermem limits added 12.3.98, Rik van Riel.
15 */
16
17#include <linux/mm.h>
18#include <linux/sched.h>
19#include <linux/kernel_stat.h>
20#include <linux/swap.h>
21#include <linux/mman.h>
22#include <linux/pagemap.h>
23#include <linux/pagevec.h>
24#include <linux/init.h>
25#include <linux/export.h>
26#include <linux/mm_inline.h>
27#include <linux/percpu_counter.h>
28#include <linux/memremap.h>
29#include <linux/percpu.h>
30#include <linux/cpu.h>
31#include <linux/notifier.h>
32#include <linux/backing-dev.h>
33#include <linux/memcontrol.h>
34#include <linux/gfp.h>
35#include <linux/uio.h>
36#include <linux/hugetlb.h>
37#include <linux/page_idle.h>
38#include <linux/local_lock.h>
39#include <linux/buffer_head.h>
40
41#include "internal.h"
42
43#define CREATE_TRACE_POINTS
44#include <trace/events/pagemap.h>
45
46/* How many pages do we try to swap or page in/out together? As a power of 2 */
47int page_cluster;
48const int page_cluster_max = 31;
49
50/* Protecting only lru_rotate.fbatch which requires disabling interrupts */
51struct lru_rotate {
52 local_lock_t lock;
53 struct folio_batch fbatch;
54};
55static DEFINE_PER_CPU(struct lru_rotate, lru_rotate) = {
56 .lock = INIT_LOCAL_LOCK(lock),
57};
58
59/*
60 * The following folio batches are grouped together because they are protected
61 * by disabling preemption (and interrupts remain enabled).
62 */
63struct cpu_fbatches {
64 local_lock_t lock;
65 struct folio_batch lru_add;
66 struct folio_batch lru_deactivate_file;
67 struct folio_batch lru_deactivate;
68 struct folio_batch lru_lazyfree;
69#ifdef CONFIG_SMP
70 struct folio_batch activate;
71#endif
72};
73static DEFINE_PER_CPU(struct cpu_fbatches, cpu_fbatches) = {
74 .lock = INIT_LOCAL_LOCK(lock),
75};
76
77/*
78 * This path almost never happens for VM activity - pages are normally freed
79 * via pagevecs. But it gets used by networking - and for compound pages.
80 */
81static void __page_cache_release(struct folio *folio)
82{
83 if (folio_test_lru(folio)) {
84 struct lruvec *lruvec;
85 unsigned long flags;
86
87 lruvec = folio_lruvec_lock_irqsave(folio, &flags);
88 lruvec_del_folio(lruvec, folio);
89 __folio_clear_lru_flags(folio);
90 unlock_page_lruvec_irqrestore(lruvec, flags);
91 }
92 /* See comment on folio_test_mlocked in release_pages() */
93 if (unlikely(folio_test_mlocked(folio))) {
94 long nr_pages = folio_nr_pages(folio);
95
96 __folio_clear_mlocked(folio);
97 zone_stat_mod_folio(folio, NR_MLOCK, -nr_pages);
98 count_vm_events(UNEVICTABLE_PGCLEARED, nr_pages);
99 }
100}
101
102static void __folio_put_small(struct folio *folio)
103{
104 __page_cache_release(folio);
105 mem_cgroup_uncharge(folio);
106 free_unref_page(&folio->page, 0);
107}
108
109static void __folio_put_large(struct folio *folio)
110{
111 /*
112 * __page_cache_release() is supposed to be called for thp, not for
113 * hugetlb. This is because hugetlb page does never have PageLRU set
114 * (it's never listed to any LRU lists) and no memcg routines should
115 * be called for hugetlb (it has a separate hugetlb_cgroup.)
116 */
117 if (!folio_test_hugetlb(folio))
118 __page_cache_release(folio);
119 destroy_large_folio(folio);
120}
121
122void __folio_put(struct folio *folio)
123{
124 if (unlikely(folio_is_zone_device(folio)))
125 free_zone_device_page(&folio->page);
126 else if (unlikely(folio_test_large(folio)))
127 __folio_put_large(folio);
128 else
129 __folio_put_small(folio);
130}
131EXPORT_SYMBOL(__folio_put);
132
133/**
134 * put_pages_list() - release a list of pages
135 * @pages: list of pages threaded on page->lru
136 *
137 * Release a list of pages which are strung together on page.lru.
138 */
139void put_pages_list(struct list_head *pages)
140{
141 struct folio *folio, *next;
142
143 list_for_each_entry_safe(folio, next, pages, lru) {
144 if (!folio_put_testzero(folio)) {
145 list_del(&folio->lru);
146 continue;
147 }
148 if (folio_test_large(folio)) {
149 list_del(&folio->lru);
150 __folio_put_large(folio);
151 continue;
152 }
153 /* LRU flag must be clear because it's passed using the lru */
154 }
155
156 free_unref_page_list(pages);
157 INIT_LIST_HEAD(pages);
158}
159EXPORT_SYMBOL(put_pages_list);
160
161/*
162 * get_kernel_pages() - pin kernel pages in memory
163 * @kiov: An array of struct kvec structures
164 * @nr_segs: number of segments to pin
165 * @write: pinning for read/write, currently ignored
166 * @pages: array that receives pointers to the pages pinned.
167 * Should be at least nr_segs long.
168 *
169 * Returns number of pages pinned. This may be fewer than the number requested.
170 * If nr_segs is 0 or negative, returns 0. If no pages were pinned, returns 0.
171 * Each page returned must be released with a put_page() call when it is
172 * finished with.
173 */
174int get_kernel_pages(const struct kvec *kiov, int nr_segs, int write,
175 struct page **pages)
176{
177 int seg;
178
179 for (seg = 0; seg < nr_segs; seg++) {
180 if (WARN_ON(kiov[seg].iov_len != PAGE_SIZE))
181 return seg;
182
183 pages[seg] = kmap_to_page(kiov[seg].iov_base);
184 get_page(pages[seg]);
185 }
186
187 return seg;
188}
189EXPORT_SYMBOL_GPL(get_kernel_pages);
190
191typedef void (*move_fn_t)(struct lruvec *lruvec, struct folio *folio);
192
193static void lru_add_fn(struct lruvec *lruvec, struct folio *folio)
194{
195 int was_unevictable = folio_test_clear_unevictable(folio);
196 long nr_pages = folio_nr_pages(folio);
197
198 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
199
200 /*
201 * Is an smp_mb__after_atomic() still required here, before
202 * folio_evictable() tests the mlocked flag, to rule out the possibility
203 * of stranding an evictable folio on an unevictable LRU? I think
204 * not, because __munlock_page() only clears the mlocked flag
205 * while the LRU lock is held.
206 *
207 * (That is not true of __page_cache_release(), and not necessarily
208 * true of release_pages(): but those only clear the mlocked flag after
209 * folio_put_testzero() has excluded any other users of the folio.)
210 */
211 if (folio_evictable(folio)) {
212 if (was_unevictable)
213 __count_vm_events(UNEVICTABLE_PGRESCUED, nr_pages);
214 } else {
215 folio_clear_active(folio);
216 folio_set_unevictable(folio);
217 /*
218 * folio->mlock_count = !!folio_test_mlocked(folio)?
219 * But that leaves __mlock_page() in doubt whether another
220 * actor has already counted the mlock or not. Err on the
221 * safe side, underestimate, let page reclaim fix it, rather
222 * than leaving a page on the unevictable LRU indefinitely.
223 */
224 folio->mlock_count = 0;
225 if (!was_unevictable)
226 __count_vm_events(UNEVICTABLE_PGCULLED, nr_pages);
227 }
228
229 lruvec_add_folio(lruvec, folio);
230 trace_mm_lru_insertion(folio);
231}
232
233static void folio_batch_move_lru(struct folio_batch *fbatch, move_fn_t move_fn)
234{
235 int i;
236 struct lruvec *lruvec = NULL;
237 unsigned long flags = 0;
238
239 for (i = 0; i < folio_batch_count(fbatch); i++) {
240 struct folio *folio = fbatch->folios[i];
241
242 /* block memcg migration while the folio moves between lru */
243 if (move_fn != lru_add_fn && !folio_test_clear_lru(folio))
244 continue;
245
246 lruvec = folio_lruvec_relock_irqsave(folio, lruvec, &flags);
247 move_fn(lruvec, folio);
248
249 folio_set_lru(folio);
250 }
251
252 if (lruvec)
253 unlock_page_lruvec_irqrestore(lruvec, flags);
254 folios_put(fbatch->folios, folio_batch_count(fbatch));
255 folio_batch_init(fbatch);
256}
257
258static void folio_batch_add_and_move(struct folio_batch *fbatch,
259 struct folio *folio, move_fn_t move_fn)
260{
261 if (folio_batch_add(fbatch, folio) && !folio_test_large(folio) &&
262 !lru_cache_disabled())
263 return;
264 folio_batch_move_lru(fbatch, move_fn);
265}
266
267static void lru_move_tail_fn(struct lruvec *lruvec, struct folio *folio)
268{
269 if (!folio_test_unevictable(folio)) {
270 lruvec_del_folio(lruvec, folio);
271 folio_clear_active(folio);
272 lruvec_add_folio_tail(lruvec, folio);
273 __count_vm_events(PGROTATED, folio_nr_pages(folio));
274 }
275}
276
277/*
278 * Writeback is about to end against a folio which has been marked for
279 * immediate reclaim. If it still appears to be reclaimable, move it
280 * to the tail of the inactive list.
281 *
282 * folio_rotate_reclaimable() must disable IRQs, to prevent nasty races.
283 */
284void folio_rotate_reclaimable(struct folio *folio)
285{
286 if (!folio_test_locked(folio) && !folio_test_dirty(folio) &&
287 !folio_test_unevictable(folio) && folio_test_lru(folio)) {
288 struct folio_batch *fbatch;
289 unsigned long flags;
290
291 folio_get(folio);
292 local_lock_irqsave(&lru_rotate.lock, flags);
293 fbatch = this_cpu_ptr(&lru_rotate.fbatch);
294 folio_batch_add_and_move(fbatch, folio, lru_move_tail_fn);
295 local_unlock_irqrestore(&lru_rotate.lock, flags);
296 }
297}
298
299void lru_note_cost(struct lruvec *lruvec, bool file,
300 unsigned int nr_io, unsigned int nr_rotated)
301{
302 unsigned long cost;
303
304 /*
305 * Reflect the relative cost of incurring IO and spending CPU
306 * time on rotations. This doesn't attempt to make a precise
307 * comparison, it just says: if reloads are about comparable
308 * between the LRU lists, or rotations are overwhelmingly
309 * different between them, adjust scan balance for CPU work.
310 */
311 cost = nr_io * SWAP_CLUSTER_MAX + nr_rotated;
312
313 do {
314 unsigned long lrusize;
315
316 /*
317 * Hold lruvec->lru_lock is safe here, since
318 * 1) The pinned lruvec in reclaim, or
319 * 2) From a pre-LRU page during refault (which also holds the
320 * rcu lock, so would be safe even if the page was on the LRU
321 * and could move simultaneously to a new lruvec).
322 */
323 spin_lock_irq(&lruvec->lru_lock);
324 /* Record cost event */
325 if (file)
326 lruvec->file_cost += cost;
327 else
328 lruvec->anon_cost += cost;
329
330 /*
331 * Decay previous events
332 *
333 * Because workloads change over time (and to avoid
334 * overflow) we keep these statistics as a floating
335 * average, which ends up weighing recent refaults
336 * more than old ones.
337 */
338 lrusize = lruvec_page_state(lruvec, NR_INACTIVE_ANON) +
339 lruvec_page_state(lruvec, NR_ACTIVE_ANON) +
340 lruvec_page_state(lruvec, NR_INACTIVE_FILE) +
341 lruvec_page_state(lruvec, NR_ACTIVE_FILE);
342
343 if (lruvec->file_cost + lruvec->anon_cost > lrusize / 4) {
344 lruvec->file_cost /= 2;
345 lruvec->anon_cost /= 2;
346 }
347 spin_unlock_irq(&lruvec->lru_lock);
348 } while ((lruvec = parent_lruvec(lruvec)));
349}
350
351void lru_note_cost_refault(struct folio *folio)
352{
353 lru_note_cost(folio_lruvec(folio), folio_is_file_lru(folio),
354 folio_nr_pages(folio), 0);
355}
356
357static void folio_activate_fn(struct lruvec *lruvec, struct folio *folio)
358{
359 if (!folio_test_active(folio) && !folio_test_unevictable(folio)) {
360 long nr_pages = folio_nr_pages(folio);
361
362 lruvec_del_folio(lruvec, folio);
363 folio_set_active(folio);
364 lruvec_add_folio(lruvec, folio);
365 trace_mm_lru_activate(folio);
366
367 __count_vm_events(PGACTIVATE, nr_pages);
368 __count_memcg_events(lruvec_memcg(lruvec), PGACTIVATE,
369 nr_pages);
370 }
371}
372
373#ifdef CONFIG_SMP
374static void folio_activate_drain(int cpu)
375{
376 struct folio_batch *fbatch = &per_cpu(cpu_fbatches.activate, cpu);
377
378 if (folio_batch_count(fbatch))
379 folio_batch_move_lru(fbatch, folio_activate_fn);
380}
381
382void folio_activate(struct folio *folio)
383{
384 if (folio_test_lru(folio) && !folio_test_active(folio) &&
385 !folio_test_unevictable(folio)) {
386 struct folio_batch *fbatch;
387
388 folio_get(folio);
389 local_lock(&cpu_fbatches.lock);
390 fbatch = this_cpu_ptr(&cpu_fbatches.activate);
391 folio_batch_add_and_move(fbatch, folio, folio_activate_fn);
392 local_unlock(&cpu_fbatches.lock);
393 }
394}
395
396#else
397static inline void folio_activate_drain(int cpu)
398{
399}
400
401void folio_activate(struct folio *folio)
402{
403 struct lruvec *lruvec;
404
405 if (folio_test_clear_lru(folio)) {
406 lruvec = folio_lruvec_lock_irq(folio);
407 folio_activate_fn(lruvec, folio);
408 unlock_page_lruvec_irq(lruvec);
409 folio_set_lru(folio);
410 }
411}
412#endif
413
414static void __lru_cache_activate_folio(struct folio *folio)
415{
416 struct folio_batch *fbatch;
417 int i;
418
419 local_lock(&cpu_fbatches.lock);
420 fbatch = this_cpu_ptr(&cpu_fbatches.lru_add);
421
422 /*
423 * Search backwards on the optimistic assumption that the folio being
424 * activated has just been added to this batch. Note that only
425 * the local batch is examined as a !LRU folio could be in the
426 * process of being released, reclaimed, migrated or on a remote
427 * batch that is currently being drained. Furthermore, marking
428 * a remote batch's folio active potentially hits a race where
429 * a folio is marked active just after it is added to the inactive
430 * list causing accounting errors and BUG_ON checks to trigger.
431 */
432 for (i = folio_batch_count(fbatch) - 1; i >= 0; i--) {
433 struct folio *batch_folio = fbatch->folios[i];
434
435 if (batch_folio == folio) {
436 folio_set_active(folio);
437 break;
438 }
439 }
440
441 local_unlock(&cpu_fbatches.lock);
442}
443
444#ifdef CONFIG_LRU_GEN
445static void folio_inc_refs(struct folio *folio)
446{
447 unsigned long new_flags, old_flags = READ_ONCE(folio->flags);
448
449 if (folio_test_unevictable(folio))
450 return;
451
452 if (!folio_test_referenced(folio)) {
453 folio_set_referenced(folio);
454 return;
455 }
456
457 if (!folio_test_workingset(folio)) {
458 folio_set_workingset(folio);
459 return;
460 }
461
462 /* see the comment on MAX_NR_TIERS */
463 do {
464 new_flags = old_flags & LRU_REFS_MASK;
465 if (new_flags == LRU_REFS_MASK)
466 break;
467
468 new_flags += BIT(LRU_REFS_PGOFF);
469 new_flags |= old_flags & ~LRU_REFS_MASK;
470 } while (!try_cmpxchg(&folio->flags, &old_flags, new_flags));
471}
472#else
473static void folio_inc_refs(struct folio *folio)
474{
475}
476#endif /* CONFIG_LRU_GEN */
477
478/*
479 * Mark a page as having seen activity.
480 *
481 * inactive,unreferenced -> inactive,referenced
482 * inactive,referenced -> active,unreferenced
483 * active,unreferenced -> active,referenced
484 *
485 * When a newly allocated page is not yet visible, so safe for non-atomic ops,
486 * __SetPageReferenced(page) may be substituted for mark_page_accessed(page).
487 */
488void folio_mark_accessed(struct folio *folio)
489{
490 if (lru_gen_enabled()) {
491 folio_inc_refs(folio);
492 return;
493 }
494
495 if (!folio_test_referenced(folio)) {
496 folio_set_referenced(folio);
497 } else if (folio_test_unevictable(folio)) {
498 /*
499 * Unevictable pages are on the "LRU_UNEVICTABLE" list. But,
500 * this list is never rotated or maintained, so marking an
501 * unevictable page accessed has no effect.
502 */
503 } else if (!folio_test_active(folio)) {
504 /*
505 * If the folio is on the LRU, queue it for activation via
506 * cpu_fbatches.activate. Otherwise, assume the folio is in a
507 * folio_batch, mark it active and it'll be moved to the active
508 * LRU on the next drain.
509 */
510 if (folio_test_lru(folio))
511 folio_activate(folio);
512 else
513 __lru_cache_activate_folio(folio);
514 folio_clear_referenced(folio);
515 workingset_activation(folio);
516 }
517 if (folio_test_idle(folio))
518 folio_clear_idle(folio);
519}
520EXPORT_SYMBOL(folio_mark_accessed);
521
522/**
523 * folio_add_lru - Add a folio to an LRU list.
524 * @folio: The folio to be added to the LRU.
525 *
526 * Queue the folio for addition to the LRU. The decision on whether
527 * to add the page to the [in]active [file|anon] list is deferred until the
528 * folio_batch is drained. This gives a chance for the caller of folio_add_lru()
529 * have the folio added to the active list using folio_mark_accessed().
530 */
531void folio_add_lru(struct folio *folio)
532{
533 struct folio_batch *fbatch;
534
535 VM_BUG_ON_FOLIO(folio_test_active(folio) &&
536 folio_test_unevictable(folio), folio);
537 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
538
539 /* see the comment in lru_gen_add_folio() */
540 if (lru_gen_enabled() && !folio_test_unevictable(folio) &&
541 lru_gen_in_fault() && !(current->flags & PF_MEMALLOC))
542 folio_set_active(folio);
543
544 folio_get(folio);
545 local_lock(&cpu_fbatches.lock);
546 fbatch = this_cpu_ptr(&cpu_fbatches.lru_add);
547 folio_batch_add_and_move(fbatch, folio, lru_add_fn);
548 local_unlock(&cpu_fbatches.lock);
549}
550EXPORT_SYMBOL(folio_add_lru);
551
552/**
553 * folio_add_lru_vma() - Add a folio to the appropate LRU list for this VMA.
554 * @folio: The folio to be added to the LRU.
555 * @vma: VMA in which the folio is mapped.
556 *
557 * If the VMA is mlocked, @folio is added to the unevictable list.
558 * Otherwise, it is treated the same way as folio_add_lru().
559 */
560void folio_add_lru_vma(struct folio *folio, struct vm_area_struct *vma)
561{
562 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
563
564 if (unlikely((vma->vm_flags & (VM_LOCKED | VM_SPECIAL)) == VM_LOCKED))
565 mlock_new_page(&folio->page);
566 else
567 folio_add_lru(folio);
568}
569
570/*
571 * If the folio cannot be invalidated, it is moved to the
572 * inactive list to speed up its reclaim. It is moved to the
573 * head of the list, rather than the tail, to give the flusher
574 * threads some time to write it out, as this is much more
575 * effective than the single-page writeout from reclaim.
576 *
577 * If the folio isn't mapped and dirty/writeback, the folio
578 * could be reclaimed asap using the reclaim flag.
579 *
580 * 1. active, mapped folio -> none
581 * 2. active, dirty/writeback folio -> inactive, head, reclaim
582 * 3. inactive, mapped folio -> none
583 * 4. inactive, dirty/writeback folio -> inactive, head, reclaim
584 * 5. inactive, clean -> inactive, tail
585 * 6. Others -> none
586 *
587 * In 4, it moves to the head of the inactive list so the folio is
588 * written out by flusher threads as this is much more efficient
589 * than the single-page writeout from reclaim.
590 */
591static void lru_deactivate_file_fn(struct lruvec *lruvec, struct folio *folio)
592{
593 bool active = folio_test_active(folio);
594 long nr_pages = folio_nr_pages(folio);
595
596 if (folio_test_unevictable(folio))
597 return;
598
599 /* Some processes are using the folio */
600 if (folio_mapped(folio))
601 return;
602
603 lruvec_del_folio(lruvec, folio);
604 folio_clear_active(folio);
605 folio_clear_referenced(folio);
606
607 if (folio_test_writeback(folio) || folio_test_dirty(folio)) {
608 /*
609 * Setting the reclaim flag could race with
610 * folio_end_writeback() and confuse readahead. But the
611 * race window is _really_ small and it's not a critical
612 * problem.
613 */
614 lruvec_add_folio(lruvec, folio);
615 folio_set_reclaim(folio);
616 } else {
617 /*
618 * The folio's writeback ended while it was in the batch.
619 * We move that folio to the tail of the inactive list.
620 */
621 lruvec_add_folio_tail(lruvec, folio);
622 __count_vm_events(PGROTATED, nr_pages);
623 }
624
625 if (active) {
626 __count_vm_events(PGDEACTIVATE, nr_pages);
627 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
628 nr_pages);
629 }
630}
631
632static void lru_deactivate_fn(struct lruvec *lruvec, struct folio *folio)
633{
634 if (!folio_test_unevictable(folio) && (folio_test_active(folio) || lru_gen_enabled())) {
635 long nr_pages = folio_nr_pages(folio);
636
637 lruvec_del_folio(lruvec, folio);
638 folio_clear_active(folio);
639 folio_clear_referenced(folio);
640 lruvec_add_folio(lruvec, folio);
641
642 __count_vm_events(PGDEACTIVATE, nr_pages);
643 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
644 nr_pages);
645 }
646}
647
648static void lru_lazyfree_fn(struct lruvec *lruvec, struct folio *folio)
649{
650 if (folio_test_anon(folio) && folio_test_swapbacked(folio) &&
651 !folio_test_swapcache(folio) && !folio_test_unevictable(folio)) {
652 long nr_pages = folio_nr_pages(folio);
653
654 lruvec_del_folio(lruvec, folio);
655 folio_clear_active(folio);
656 folio_clear_referenced(folio);
657 /*
658 * Lazyfree folios are clean anonymous folios. They have
659 * the swapbacked flag cleared, to distinguish them from normal
660 * anonymous folios
661 */
662 folio_clear_swapbacked(folio);
663 lruvec_add_folio(lruvec, folio);
664
665 __count_vm_events(PGLAZYFREE, nr_pages);
666 __count_memcg_events(lruvec_memcg(lruvec), PGLAZYFREE,
667 nr_pages);
668 }
669}
670
671/*
672 * Drain pages out of the cpu's folio_batch.
673 * Either "cpu" is the current CPU, and preemption has already been
674 * disabled; or "cpu" is being hot-unplugged, and is already dead.
675 */
676void lru_add_drain_cpu(int cpu)
677{
678 struct cpu_fbatches *fbatches = &per_cpu(cpu_fbatches, cpu);
679 struct folio_batch *fbatch = &fbatches->lru_add;
680
681 if (folio_batch_count(fbatch))
682 folio_batch_move_lru(fbatch, lru_add_fn);
683
684 fbatch = &per_cpu(lru_rotate.fbatch, cpu);
685 /* Disabling interrupts below acts as a compiler barrier. */
686 if (data_race(folio_batch_count(fbatch))) {
687 unsigned long flags;
688
689 /* No harm done if a racing interrupt already did this */
690 local_lock_irqsave(&lru_rotate.lock, flags);
691 folio_batch_move_lru(fbatch, lru_move_tail_fn);
692 local_unlock_irqrestore(&lru_rotate.lock, flags);
693 }
694
695 fbatch = &fbatches->lru_deactivate_file;
696 if (folio_batch_count(fbatch))
697 folio_batch_move_lru(fbatch, lru_deactivate_file_fn);
698
699 fbatch = &fbatches->lru_deactivate;
700 if (folio_batch_count(fbatch))
701 folio_batch_move_lru(fbatch, lru_deactivate_fn);
702
703 fbatch = &fbatches->lru_lazyfree;
704 if (folio_batch_count(fbatch))
705 folio_batch_move_lru(fbatch, lru_lazyfree_fn);
706
707 folio_activate_drain(cpu);
708}
709
710/**
711 * deactivate_file_folio() - Deactivate a file folio.
712 * @folio: Folio to deactivate.
713 *
714 * This function hints to the VM that @folio is a good reclaim candidate,
715 * for example if its invalidation fails due to the folio being dirty
716 * or under writeback.
717 *
718 * Context: Caller holds a reference on the folio.
719 */
720void deactivate_file_folio(struct folio *folio)
721{
722 struct folio_batch *fbatch;
723
724 /* Deactivating an unevictable folio will not accelerate reclaim */
725 if (folio_test_unevictable(folio))
726 return;
727
728 folio_get(folio);
729 local_lock(&cpu_fbatches.lock);
730 fbatch = this_cpu_ptr(&cpu_fbatches.lru_deactivate_file);
731 folio_batch_add_and_move(fbatch, folio, lru_deactivate_file_fn);
732 local_unlock(&cpu_fbatches.lock);
733}
734
735/*
736 * deactivate_page - deactivate a page
737 * @page: page to deactivate
738 *
739 * deactivate_page() moves @page to the inactive list if @page was on the active
740 * list and was not an unevictable page. This is done to accelerate the reclaim
741 * of @page.
742 */
743void deactivate_page(struct page *page)
744{
745 struct folio *folio = page_folio(page);
746
747 if (folio_test_lru(folio) && !folio_test_unevictable(folio) &&
748 (folio_test_active(folio) || lru_gen_enabled())) {
749 struct folio_batch *fbatch;
750
751 folio_get(folio);
752 local_lock(&cpu_fbatches.lock);
753 fbatch = this_cpu_ptr(&cpu_fbatches.lru_deactivate);
754 folio_batch_add_and_move(fbatch, folio, lru_deactivate_fn);
755 local_unlock(&cpu_fbatches.lock);
756 }
757}
758
759/**
760 * mark_page_lazyfree - make an anon page lazyfree
761 * @page: page to deactivate
762 *
763 * mark_page_lazyfree() moves @page to the inactive file list.
764 * This is done to accelerate the reclaim of @page.
765 */
766void mark_page_lazyfree(struct page *page)
767{
768 struct folio *folio = page_folio(page);
769
770 if (folio_test_lru(folio) && folio_test_anon(folio) &&
771 folio_test_swapbacked(folio) && !folio_test_swapcache(folio) &&
772 !folio_test_unevictable(folio)) {
773 struct folio_batch *fbatch;
774
775 folio_get(folio);
776 local_lock(&cpu_fbatches.lock);
777 fbatch = this_cpu_ptr(&cpu_fbatches.lru_lazyfree);
778 folio_batch_add_and_move(fbatch, folio, lru_lazyfree_fn);
779 local_unlock(&cpu_fbatches.lock);
780 }
781}
782
783void lru_add_drain(void)
784{
785 local_lock(&cpu_fbatches.lock);
786 lru_add_drain_cpu(smp_processor_id());
787 local_unlock(&cpu_fbatches.lock);
788 mlock_page_drain_local();
789}
790
791/*
792 * It's called from per-cpu workqueue context in SMP case so
793 * lru_add_drain_cpu and invalidate_bh_lrus_cpu should run on
794 * the same cpu. It shouldn't be a problem in !SMP case since
795 * the core is only one and the locks will disable preemption.
796 */
797static void lru_add_and_bh_lrus_drain(void)
798{
799 local_lock(&cpu_fbatches.lock);
800 lru_add_drain_cpu(smp_processor_id());
801 local_unlock(&cpu_fbatches.lock);
802 invalidate_bh_lrus_cpu();
803 mlock_page_drain_local();
804}
805
806void lru_add_drain_cpu_zone(struct zone *zone)
807{
808 local_lock(&cpu_fbatches.lock);
809 lru_add_drain_cpu(smp_processor_id());
810 drain_local_pages(zone);
811 local_unlock(&cpu_fbatches.lock);
812 mlock_page_drain_local();
813}
814
815#ifdef CONFIG_SMP
816
817static DEFINE_PER_CPU(struct work_struct, lru_add_drain_work);
818
819static void lru_add_drain_per_cpu(struct work_struct *dummy)
820{
821 lru_add_and_bh_lrus_drain();
822}
823
824static bool cpu_needs_drain(unsigned int cpu)
825{
826 struct cpu_fbatches *fbatches = &per_cpu(cpu_fbatches, cpu);
827
828 /* Check these in order of likelihood that they're not zero */
829 return folio_batch_count(&fbatches->lru_add) ||
830 data_race(folio_batch_count(&per_cpu(lru_rotate.fbatch, cpu))) ||
831 folio_batch_count(&fbatches->lru_deactivate_file) ||
832 folio_batch_count(&fbatches->lru_deactivate) ||
833 folio_batch_count(&fbatches->lru_lazyfree) ||
834 folio_batch_count(&fbatches->activate) ||
835 need_mlock_page_drain(cpu) ||
836 has_bh_in_lru(cpu, NULL);
837}
838
839/*
840 * Doesn't need any cpu hotplug locking because we do rely on per-cpu
841 * kworkers being shut down before our page_alloc_cpu_dead callback is
842 * executed on the offlined cpu.
843 * Calling this function with cpu hotplug locks held can actually lead
844 * to obscure indirect dependencies via WQ context.
845 */
846static inline void __lru_add_drain_all(bool force_all_cpus)
847{
848 /*
849 * lru_drain_gen - Global pages generation number
850 *
851 * (A) Definition: global lru_drain_gen = x implies that all generations
852 * 0 < n <= x are already *scheduled* for draining.
853 *
854 * This is an optimization for the highly-contended use case where a
855 * user space workload keeps constantly generating a flow of pages for
856 * each CPU.
857 */
858 static unsigned int lru_drain_gen;
859 static struct cpumask has_work;
860 static DEFINE_MUTEX(lock);
861 unsigned cpu, this_gen;
862
863 /*
864 * Make sure nobody triggers this path before mm_percpu_wq is fully
865 * initialized.
866 */
867 if (WARN_ON(!mm_percpu_wq))
868 return;
869
870 /*
871 * Guarantee folio_batch counter stores visible by this CPU
872 * are visible to other CPUs before loading the current drain
873 * generation.
874 */
875 smp_mb();
876
877 /*
878 * (B) Locally cache global LRU draining generation number
879 *
880 * The read barrier ensures that the counter is loaded before the mutex
881 * is taken. It pairs with smp_mb() inside the mutex critical section
882 * at (D).
883 */
884 this_gen = smp_load_acquire(&lru_drain_gen);
885
886 mutex_lock(&lock);
887
888 /*
889 * (C) Exit the draining operation if a newer generation, from another
890 * lru_add_drain_all(), was already scheduled for draining. Check (A).
891 */
892 if (unlikely(this_gen != lru_drain_gen && !force_all_cpus))
893 goto done;
894
895 /*
896 * (D) Increment global generation number
897 *
898 * Pairs with smp_load_acquire() at (B), outside of the critical
899 * section. Use a full memory barrier to guarantee that the
900 * new global drain generation number is stored before loading
901 * folio_batch counters.
902 *
903 * This pairing must be done here, before the for_each_online_cpu loop
904 * below which drains the page vectors.
905 *
906 * Let x, y, and z represent some system CPU numbers, where x < y < z.
907 * Assume CPU #z is in the middle of the for_each_online_cpu loop
908 * below and has already reached CPU #y's per-cpu data. CPU #x comes
909 * along, adds some pages to its per-cpu vectors, then calls
910 * lru_add_drain_all().
911 *
912 * If the paired barrier is done at any later step, e.g. after the
913 * loop, CPU #x will just exit at (C) and miss flushing out all of its
914 * added pages.
915 */
916 WRITE_ONCE(lru_drain_gen, lru_drain_gen + 1);
917 smp_mb();
918
919 cpumask_clear(&has_work);
920 for_each_online_cpu(cpu) {
921 struct work_struct *work = &per_cpu(lru_add_drain_work, cpu);
922
923 if (cpu_needs_drain(cpu)) {
924 INIT_WORK(work, lru_add_drain_per_cpu);
925 queue_work_on(cpu, mm_percpu_wq, work);
926 __cpumask_set_cpu(cpu, &has_work);
927 }
928 }
929
930 for_each_cpu(cpu, &has_work)
931 flush_work(&per_cpu(lru_add_drain_work, cpu));
932
933done:
934 mutex_unlock(&lock);
935}
936
937void lru_add_drain_all(void)
938{
939 __lru_add_drain_all(false);
940}
941#else
942void lru_add_drain_all(void)
943{
944 lru_add_drain();
945}
946#endif /* CONFIG_SMP */
947
948atomic_t lru_disable_count = ATOMIC_INIT(0);
949
950/*
951 * lru_cache_disable() needs to be called before we start compiling
952 * a list of pages to be migrated using isolate_lru_page().
953 * It drains pages on LRU cache and then disable on all cpus until
954 * lru_cache_enable is called.
955 *
956 * Must be paired with a call to lru_cache_enable().
957 */
958void lru_cache_disable(void)
959{
960 atomic_inc(&lru_disable_count);
961 /*
962 * Readers of lru_disable_count are protected by either disabling
963 * preemption or rcu_read_lock:
964 *
965 * preempt_disable, local_irq_disable [bh_lru_lock()]
966 * rcu_read_lock [rt_spin_lock CONFIG_PREEMPT_RT]
967 * preempt_disable [local_lock !CONFIG_PREEMPT_RT]
968 *
969 * Since v5.1 kernel, synchronize_rcu() is guaranteed to wait on
970 * preempt_disable() regions of code. So any CPU which sees
971 * lru_disable_count = 0 will have exited the critical
972 * section when synchronize_rcu() returns.
973 */
974 synchronize_rcu_expedited();
975#ifdef CONFIG_SMP
976 __lru_add_drain_all(true);
977#else
978 lru_add_and_bh_lrus_drain();
979#endif
980}
981
982/**
983 * release_pages - batched put_page()
984 * @arg: array of pages to release
985 * @nr: number of pages
986 *
987 * Decrement the reference count on all the pages in @arg. If it
988 * fell to zero, remove the page from the LRU and free it.
989 *
990 * Note that the argument can be an array of pages, encoded pages,
991 * or folio pointers. We ignore any encoded bits, and turn any of
992 * them into just a folio that gets free'd.
993 */
994void release_pages(release_pages_arg arg, int nr)
995{
996 int i;
997 struct encoded_page **encoded = arg.encoded_pages;
998 LIST_HEAD(pages_to_free);
999 struct lruvec *lruvec = NULL;
1000 unsigned long flags = 0;
1001 unsigned int lock_batch;
1002
1003 for (i = 0; i < nr; i++) {
1004 struct folio *folio;
1005
1006 /* Turn any of the argument types into a folio */
1007 folio = page_folio(encoded_page_ptr(encoded[i]));
1008
1009 /*
1010 * Make sure the IRQ-safe lock-holding time does not get
1011 * excessive with a continuous string of pages from the
1012 * same lruvec. The lock is held only if lruvec != NULL.
1013 */
1014 if (lruvec && ++lock_batch == SWAP_CLUSTER_MAX) {
1015 unlock_page_lruvec_irqrestore(lruvec, flags);
1016 lruvec = NULL;
1017 }
1018
1019 if (is_huge_zero_page(&folio->page))
1020 continue;
1021
1022 if (folio_is_zone_device(folio)) {
1023 if (lruvec) {
1024 unlock_page_lruvec_irqrestore(lruvec, flags);
1025 lruvec = NULL;
1026 }
1027 if (put_devmap_managed_page(&folio->page))
1028 continue;
1029 if (folio_put_testzero(folio))
1030 free_zone_device_page(&folio->page);
1031 continue;
1032 }
1033
1034 if (!folio_put_testzero(folio))
1035 continue;
1036
1037 if (folio_test_large(folio)) {
1038 if (lruvec) {
1039 unlock_page_lruvec_irqrestore(lruvec, flags);
1040 lruvec = NULL;
1041 }
1042 __folio_put_large(folio);
1043 continue;
1044 }
1045
1046 if (folio_test_lru(folio)) {
1047 struct lruvec *prev_lruvec = lruvec;
1048
1049 lruvec = folio_lruvec_relock_irqsave(folio, lruvec,
1050 &flags);
1051 if (prev_lruvec != lruvec)
1052 lock_batch = 0;
1053
1054 lruvec_del_folio(lruvec, folio);
1055 __folio_clear_lru_flags(folio);
1056 }
1057
1058 /*
1059 * In rare cases, when truncation or holepunching raced with
1060 * munlock after VM_LOCKED was cleared, Mlocked may still be
1061 * found set here. This does not indicate a problem, unless
1062 * "unevictable_pgs_cleared" appears worryingly large.
1063 */
1064 if (unlikely(folio_test_mlocked(folio))) {
1065 __folio_clear_mlocked(folio);
1066 zone_stat_sub_folio(folio, NR_MLOCK);
1067 count_vm_event(UNEVICTABLE_PGCLEARED);
1068 }
1069
1070 list_add(&folio->lru, &pages_to_free);
1071 }
1072 if (lruvec)
1073 unlock_page_lruvec_irqrestore(lruvec, flags);
1074
1075 mem_cgroup_uncharge_list(&pages_to_free);
1076 free_unref_page_list(&pages_to_free);
1077}
1078EXPORT_SYMBOL(release_pages);
1079
1080/*
1081 * The pages which we're about to release may be in the deferred lru-addition
1082 * queues. That would prevent them from really being freed right now. That's
1083 * OK from a correctness point of view but is inefficient - those pages may be
1084 * cache-warm and we want to give them back to the page allocator ASAP.
1085 *
1086 * So __pagevec_release() will drain those queues here.
1087 * folio_batch_move_lru() calls folios_put() directly to avoid
1088 * mutual recursion.
1089 */
1090void __pagevec_release(struct pagevec *pvec)
1091{
1092 if (!pvec->percpu_pvec_drained) {
1093 lru_add_drain();
1094 pvec->percpu_pvec_drained = true;
1095 }
1096 release_pages(pvec->pages, pagevec_count(pvec));
1097 pagevec_reinit(pvec);
1098}
1099EXPORT_SYMBOL(__pagevec_release);
1100
1101/**
1102 * folio_batch_remove_exceptionals() - Prune non-folios from a batch.
1103 * @fbatch: The batch to prune
1104 *
1105 * find_get_entries() fills a batch with both folios and shadow/swap/DAX
1106 * entries. This function prunes all the non-folio entries from @fbatch
1107 * without leaving holes, so that it can be passed on to folio-only batch
1108 * operations.
1109 */
1110void folio_batch_remove_exceptionals(struct folio_batch *fbatch)
1111{
1112 unsigned int i, j;
1113
1114 for (i = 0, j = 0; i < folio_batch_count(fbatch); i++) {
1115 struct folio *folio = fbatch->folios[i];
1116 if (!xa_is_value(folio))
1117 fbatch->folios[j++] = folio;
1118 }
1119 fbatch->nr = j;
1120}
1121
1122unsigned pagevec_lookup_range_tag(struct pagevec *pvec,
1123 struct address_space *mapping, pgoff_t *index, pgoff_t end,
1124 xa_mark_t tag)
1125{
1126 pvec->nr = find_get_pages_range_tag(mapping, index, end, tag,
1127 PAGEVEC_SIZE, pvec->pages);
1128 return pagevec_count(pvec);
1129}
1130EXPORT_SYMBOL(pagevec_lookup_range_tag);
1131
1132/*
1133 * Perform any setup for the swap system
1134 */
1135void __init swap_setup(void)
1136{
1137 unsigned long megs = totalram_pages() >> (20 - PAGE_SHIFT);
1138
1139 /* Use a smaller cluster for small-memory machines */
1140 if (megs < 16)
1141 page_cluster = 2;
1142 else
1143 page_cluster = 3;
1144 /*
1145 * Right now other parts of the system means that we
1146 * _really_ don't want to cluster much more
1147 */
1148}
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/mm/swap.c
4 *
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 */
7
8/*
9 * This file contains the default values for the operation of the
10 * Linux VM subsystem. Fine-tuning documentation can be found in
11 * Documentation/admin-guide/sysctl/vm.rst.
12 * Started 18.12.91
13 * Swap aging added 23.2.95, Stephen Tweedie.
14 * Buffermem limits added 12.3.98, Rik van Riel.
15 */
16
17#include <linux/mm.h>
18#include <linux/sched.h>
19#include <linux/kernel_stat.h>
20#include <linux/swap.h>
21#include <linux/mman.h>
22#include <linux/pagemap.h>
23#include <linux/pagevec.h>
24#include <linux/init.h>
25#include <linux/export.h>
26#include <linux/mm_inline.h>
27#include <linux/percpu_counter.h>
28#include <linux/memremap.h>
29#include <linux/percpu.h>
30#include <linux/cpu.h>
31#include <linux/notifier.h>
32#include <linux/backing-dev.h>
33#include <linux/memcontrol.h>
34#include <linux/gfp.h>
35#include <linux/uio.h>
36#include <linux/hugetlb.h>
37#include <linux/page_idle.h>
38#include <linux/local_lock.h>
39#include <linux/buffer_head.h>
40
41#include "internal.h"
42
43#define CREATE_TRACE_POINTS
44#include <trace/events/pagemap.h>
45
46/* How many pages do we try to swap or page in/out together? As a power of 2 */
47int page_cluster;
48const int page_cluster_max = 31;
49
50/* Protecting only lru_rotate.fbatch which requires disabling interrupts */
51struct lru_rotate {
52 local_lock_t lock;
53 struct folio_batch fbatch;
54};
55static DEFINE_PER_CPU(struct lru_rotate, lru_rotate) = {
56 .lock = INIT_LOCAL_LOCK(lock),
57};
58
59/*
60 * The following folio batches are grouped together because they are protected
61 * by disabling preemption (and interrupts remain enabled).
62 */
63struct cpu_fbatches {
64 local_lock_t lock;
65 struct folio_batch lru_add;
66 struct folio_batch lru_deactivate_file;
67 struct folio_batch lru_deactivate;
68 struct folio_batch lru_lazyfree;
69#ifdef CONFIG_SMP
70 struct folio_batch activate;
71#endif
72};
73static DEFINE_PER_CPU(struct cpu_fbatches, cpu_fbatches) = {
74 .lock = INIT_LOCAL_LOCK(lock),
75};
76
77static void __page_cache_release(struct folio *folio, struct lruvec **lruvecp,
78 unsigned long *flagsp)
79{
80 if (folio_test_lru(folio)) {
81 folio_lruvec_relock_irqsave(folio, lruvecp, flagsp);
82 lruvec_del_folio(*lruvecp, folio);
83 __folio_clear_lru_flags(folio);
84 }
85
86 /*
87 * In rare cases, when truncation or holepunching raced with
88 * munlock after VM_LOCKED was cleared, Mlocked may still be
89 * found set here. This does not indicate a problem, unless
90 * "unevictable_pgs_cleared" appears worryingly large.
91 */
92 if (unlikely(folio_test_mlocked(folio))) {
93 long nr_pages = folio_nr_pages(folio);
94
95 __folio_clear_mlocked(folio);
96 zone_stat_mod_folio(folio, NR_MLOCK, -nr_pages);
97 count_vm_events(UNEVICTABLE_PGCLEARED, nr_pages);
98 }
99}
100
101/*
102 * This path almost never happens for VM activity - pages are normally freed
103 * in batches. But it gets used by networking - and for compound pages.
104 */
105static void page_cache_release(struct folio *folio)
106{
107 struct lruvec *lruvec = NULL;
108 unsigned long flags;
109
110 __page_cache_release(folio, &lruvec, &flags);
111 if (lruvec)
112 unlock_page_lruvec_irqrestore(lruvec, flags);
113}
114
115static void __folio_put_small(struct folio *folio)
116{
117 page_cache_release(folio);
118 mem_cgroup_uncharge(folio);
119 free_unref_page(&folio->page, 0);
120}
121
122static void __folio_put_large(struct folio *folio)
123{
124 /*
125 * __page_cache_release() is supposed to be called for thp, not for
126 * hugetlb. This is because hugetlb page does never have PageLRU set
127 * (it's never listed to any LRU lists) and no memcg routines should
128 * be called for hugetlb (it has a separate hugetlb_cgroup.)
129 */
130 if (!folio_test_hugetlb(folio))
131 page_cache_release(folio);
132 destroy_large_folio(folio);
133}
134
135void __folio_put(struct folio *folio)
136{
137 if (unlikely(folio_is_zone_device(folio)))
138 free_zone_device_page(&folio->page);
139 else if (unlikely(folio_test_large(folio)))
140 __folio_put_large(folio);
141 else
142 __folio_put_small(folio);
143}
144EXPORT_SYMBOL(__folio_put);
145
146/**
147 * put_pages_list() - release a list of pages
148 * @pages: list of pages threaded on page->lru
149 *
150 * Release a list of pages which are strung together on page.lru.
151 */
152void put_pages_list(struct list_head *pages)
153{
154 struct folio_batch fbatch;
155 struct folio *folio, *next;
156
157 folio_batch_init(&fbatch);
158 list_for_each_entry_safe(folio, next, pages, lru) {
159 if (!folio_put_testzero(folio))
160 continue;
161 if (folio_test_large(folio)) {
162 __folio_put_large(folio);
163 continue;
164 }
165 /* LRU flag must be clear because it's passed using the lru */
166 if (folio_batch_add(&fbatch, folio) > 0)
167 continue;
168 free_unref_folios(&fbatch);
169 }
170
171 if (fbatch.nr)
172 free_unref_folios(&fbatch);
173 INIT_LIST_HEAD(pages);
174}
175EXPORT_SYMBOL(put_pages_list);
176
177typedef void (*move_fn_t)(struct lruvec *lruvec, struct folio *folio);
178
179static void lru_add_fn(struct lruvec *lruvec, struct folio *folio)
180{
181 int was_unevictable = folio_test_clear_unevictable(folio);
182 long nr_pages = folio_nr_pages(folio);
183
184 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
185
186 /*
187 * Is an smp_mb__after_atomic() still required here, before
188 * folio_evictable() tests the mlocked flag, to rule out the possibility
189 * of stranding an evictable folio on an unevictable LRU? I think
190 * not, because __munlock_folio() only clears the mlocked flag
191 * while the LRU lock is held.
192 *
193 * (That is not true of __page_cache_release(), and not necessarily
194 * true of folios_put(): but those only clear the mlocked flag after
195 * folio_put_testzero() has excluded any other users of the folio.)
196 */
197 if (folio_evictable(folio)) {
198 if (was_unevictable)
199 __count_vm_events(UNEVICTABLE_PGRESCUED, nr_pages);
200 } else {
201 folio_clear_active(folio);
202 folio_set_unevictable(folio);
203 /*
204 * folio->mlock_count = !!folio_test_mlocked(folio)?
205 * But that leaves __mlock_folio() in doubt whether another
206 * actor has already counted the mlock or not. Err on the
207 * safe side, underestimate, let page reclaim fix it, rather
208 * than leaving a page on the unevictable LRU indefinitely.
209 */
210 folio->mlock_count = 0;
211 if (!was_unevictable)
212 __count_vm_events(UNEVICTABLE_PGCULLED, nr_pages);
213 }
214
215 lruvec_add_folio(lruvec, folio);
216 trace_mm_lru_insertion(folio);
217}
218
219static void folio_batch_move_lru(struct folio_batch *fbatch, move_fn_t move_fn)
220{
221 int i;
222 struct lruvec *lruvec = NULL;
223 unsigned long flags = 0;
224
225 for (i = 0; i < folio_batch_count(fbatch); i++) {
226 struct folio *folio = fbatch->folios[i];
227
228 /* block memcg migration while the folio moves between lru */
229 if (move_fn != lru_add_fn && !folio_test_clear_lru(folio))
230 continue;
231
232 folio_lruvec_relock_irqsave(folio, &lruvec, &flags);
233 move_fn(lruvec, folio);
234
235 folio_set_lru(folio);
236 }
237
238 if (lruvec)
239 unlock_page_lruvec_irqrestore(lruvec, flags);
240 folios_put(fbatch);
241}
242
243static void folio_batch_add_and_move(struct folio_batch *fbatch,
244 struct folio *folio, move_fn_t move_fn)
245{
246 if (folio_batch_add(fbatch, folio) && !folio_test_large(folio) &&
247 !lru_cache_disabled())
248 return;
249 folio_batch_move_lru(fbatch, move_fn);
250}
251
252static void lru_move_tail_fn(struct lruvec *lruvec, struct folio *folio)
253{
254 if (!folio_test_unevictable(folio)) {
255 lruvec_del_folio(lruvec, folio);
256 folio_clear_active(folio);
257 lruvec_add_folio_tail(lruvec, folio);
258 __count_vm_events(PGROTATED, folio_nr_pages(folio));
259 }
260}
261
262/*
263 * Writeback is about to end against a folio which has been marked for
264 * immediate reclaim. If it still appears to be reclaimable, move it
265 * to the tail of the inactive list.
266 *
267 * folio_rotate_reclaimable() must disable IRQs, to prevent nasty races.
268 */
269void folio_rotate_reclaimable(struct folio *folio)
270{
271 if (!folio_test_locked(folio) && !folio_test_dirty(folio) &&
272 !folio_test_unevictable(folio) && folio_test_lru(folio)) {
273 struct folio_batch *fbatch;
274 unsigned long flags;
275
276 folio_get(folio);
277 local_lock_irqsave(&lru_rotate.lock, flags);
278 fbatch = this_cpu_ptr(&lru_rotate.fbatch);
279 folio_batch_add_and_move(fbatch, folio, lru_move_tail_fn);
280 local_unlock_irqrestore(&lru_rotate.lock, flags);
281 }
282}
283
284void lru_note_cost(struct lruvec *lruvec, bool file,
285 unsigned int nr_io, unsigned int nr_rotated)
286{
287 unsigned long cost;
288
289 /*
290 * Reflect the relative cost of incurring IO and spending CPU
291 * time on rotations. This doesn't attempt to make a precise
292 * comparison, it just says: if reloads are about comparable
293 * between the LRU lists, or rotations are overwhelmingly
294 * different between them, adjust scan balance for CPU work.
295 */
296 cost = nr_io * SWAP_CLUSTER_MAX + nr_rotated;
297
298 do {
299 unsigned long lrusize;
300
301 /*
302 * Hold lruvec->lru_lock is safe here, since
303 * 1) The pinned lruvec in reclaim, or
304 * 2) From a pre-LRU page during refault (which also holds the
305 * rcu lock, so would be safe even if the page was on the LRU
306 * and could move simultaneously to a new lruvec).
307 */
308 spin_lock_irq(&lruvec->lru_lock);
309 /* Record cost event */
310 if (file)
311 lruvec->file_cost += cost;
312 else
313 lruvec->anon_cost += cost;
314
315 /*
316 * Decay previous events
317 *
318 * Because workloads change over time (and to avoid
319 * overflow) we keep these statistics as a floating
320 * average, which ends up weighing recent refaults
321 * more than old ones.
322 */
323 lrusize = lruvec_page_state(lruvec, NR_INACTIVE_ANON) +
324 lruvec_page_state(lruvec, NR_ACTIVE_ANON) +
325 lruvec_page_state(lruvec, NR_INACTIVE_FILE) +
326 lruvec_page_state(lruvec, NR_ACTIVE_FILE);
327
328 if (lruvec->file_cost + lruvec->anon_cost > lrusize / 4) {
329 lruvec->file_cost /= 2;
330 lruvec->anon_cost /= 2;
331 }
332 spin_unlock_irq(&lruvec->lru_lock);
333 } while ((lruvec = parent_lruvec(lruvec)));
334}
335
336void lru_note_cost_refault(struct folio *folio)
337{
338 lru_note_cost(folio_lruvec(folio), folio_is_file_lru(folio),
339 folio_nr_pages(folio), 0);
340}
341
342static void folio_activate_fn(struct lruvec *lruvec, struct folio *folio)
343{
344 if (!folio_test_active(folio) && !folio_test_unevictable(folio)) {
345 long nr_pages = folio_nr_pages(folio);
346
347 lruvec_del_folio(lruvec, folio);
348 folio_set_active(folio);
349 lruvec_add_folio(lruvec, folio);
350 trace_mm_lru_activate(folio);
351
352 __count_vm_events(PGACTIVATE, nr_pages);
353 __count_memcg_events(lruvec_memcg(lruvec), PGACTIVATE,
354 nr_pages);
355 }
356}
357
358#ifdef CONFIG_SMP
359static void folio_activate_drain(int cpu)
360{
361 struct folio_batch *fbatch = &per_cpu(cpu_fbatches.activate, cpu);
362
363 if (folio_batch_count(fbatch))
364 folio_batch_move_lru(fbatch, folio_activate_fn);
365}
366
367void folio_activate(struct folio *folio)
368{
369 if (folio_test_lru(folio) && !folio_test_active(folio) &&
370 !folio_test_unevictable(folio)) {
371 struct folio_batch *fbatch;
372
373 folio_get(folio);
374 local_lock(&cpu_fbatches.lock);
375 fbatch = this_cpu_ptr(&cpu_fbatches.activate);
376 folio_batch_add_and_move(fbatch, folio, folio_activate_fn);
377 local_unlock(&cpu_fbatches.lock);
378 }
379}
380
381#else
382static inline void folio_activate_drain(int cpu)
383{
384}
385
386void folio_activate(struct folio *folio)
387{
388 struct lruvec *lruvec;
389
390 if (folio_test_clear_lru(folio)) {
391 lruvec = folio_lruvec_lock_irq(folio);
392 folio_activate_fn(lruvec, folio);
393 unlock_page_lruvec_irq(lruvec);
394 folio_set_lru(folio);
395 }
396}
397#endif
398
399static void __lru_cache_activate_folio(struct folio *folio)
400{
401 struct folio_batch *fbatch;
402 int i;
403
404 local_lock(&cpu_fbatches.lock);
405 fbatch = this_cpu_ptr(&cpu_fbatches.lru_add);
406
407 /*
408 * Search backwards on the optimistic assumption that the folio being
409 * activated has just been added to this batch. Note that only
410 * the local batch is examined as a !LRU folio could be in the
411 * process of being released, reclaimed, migrated or on a remote
412 * batch that is currently being drained. Furthermore, marking
413 * a remote batch's folio active potentially hits a race where
414 * a folio is marked active just after it is added to the inactive
415 * list causing accounting errors and BUG_ON checks to trigger.
416 */
417 for (i = folio_batch_count(fbatch) - 1; i >= 0; i--) {
418 struct folio *batch_folio = fbatch->folios[i];
419
420 if (batch_folio == folio) {
421 folio_set_active(folio);
422 break;
423 }
424 }
425
426 local_unlock(&cpu_fbatches.lock);
427}
428
429#ifdef CONFIG_LRU_GEN
430static void folio_inc_refs(struct folio *folio)
431{
432 unsigned long new_flags, old_flags = READ_ONCE(folio->flags);
433
434 if (folio_test_unevictable(folio))
435 return;
436
437 if (!folio_test_referenced(folio)) {
438 folio_set_referenced(folio);
439 return;
440 }
441
442 if (!folio_test_workingset(folio)) {
443 folio_set_workingset(folio);
444 return;
445 }
446
447 /* see the comment on MAX_NR_TIERS */
448 do {
449 new_flags = old_flags & LRU_REFS_MASK;
450 if (new_flags == LRU_REFS_MASK)
451 break;
452
453 new_flags += BIT(LRU_REFS_PGOFF);
454 new_flags |= old_flags & ~LRU_REFS_MASK;
455 } while (!try_cmpxchg(&folio->flags, &old_flags, new_flags));
456}
457#else
458static void folio_inc_refs(struct folio *folio)
459{
460}
461#endif /* CONFIG_LRU_GEN */
462
463/*
464 * Mark a page as having seen activity.
465 *
466 * inactive,unreferenced -> inactive,referenced
467 * inactive,referenced -> active,unreferenced
468 * active,unreferenced -> active,referenced
469 *
470 * When a newly allocated page is not yet visible, so safe for non-atomic ops,
471 * __SetPageReferenced(page) may be substituted for mark_page_accessed(page).
472 */
473void folio_mark_accessed(struct folio *folio)
474{
475 if (lru_gen_enabled()) {
476 folio_inc_refs(folio);
477 return;
478 }
479
480 if (!folio_test_referenced(folio)) {
481 folio_set_referenced(folio);
482 } else if (folio_test_unevictable(folio)) {
483 /*
484 * Unevictable pages are on the "LRU_UNEVICTABLE" list. But,
485 * this list is never rotated or maintained, so marking an
486 * unevictable page accessed has no effect.
487 */
488 } else if (!folio_test_active(folio)) {
489 /*
490 * If the folio is on the LRU, queue it for activation via
491 * cpu_fbatches.activate. Otherwise, assume the folio is in a
492 * folio_batch, mark it active and it'll be moved to the active
493 * LRU on the next drain.
494 */
495 if (folio_test_lru(folio))
496 folio_activate(folio);
497 else
498 __lru_cache_activate_folio(folio);
499 folio_clear_referenced(folio);
500 workingset_activation(folio);
501 }
502 if (folio_test_idle(folio))
503 folio_clear_idle(folio);
504}
505EXPORT_SYMBOL(folio_mark_accessed);
506
507/**
508 * folio_add_lru - Add a folio to an LRU list.
509 * @folio: The folio to be added to the LRU.
510 *
511 * Queue the folio for addition to the LRU. The decision on whether
512 * to add the page to the [in]active [file|anon] list is deferred until the
513 * folio_batch is drained. This gives a chance for the caller of folio_add_lru()
514 * have the folio added to the active list using folio_mark_accessed().
515 */
516void folio_add_lru(struct folio *folio)
517{
518 struct folio_batch *fbatch;
519
520 VM_BUG_ON_FOLIO(folio_test_active(folio) &&
521 folio_test_unevictable(folio), folio);
522 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
523
524 /* see the comment in lru_gen_add_folio() */
525 if (lru_gen_enabled() && !folio_test_unevictable(folio) &&
526 lru_gen_in_fault() && !(current->flags & PF_MEMALLOC))
527 folio_set_active(folio);
528
529 folio_get(folio);
530 local_lock(&cpu_fbatches.lock);
531 fbatch = this_cpu_ptr(&cpu_fbatches.lru_add);
532 folio_batch_add_and_move(fbatch, folio, lru_add_fn);
533 local_unlock(&cpu_fbatches.lock);
534}
535EXPORT_SYMBOL(folio_add_lru);
536
537/**
538 * folio_add_lru_vma() - Add a folio to the appropate LRU list for this VMA.
539 * @folio: The folio to be added to the LRU.
540 * @vma: VMA in which the folio is mapped.
541 *
542 * If the VMA is mlocked, @folio is added to the unevictable list.
543 * Otherwise, it is treated the same way as folio_add_lru().
544 */
545void folio_add_lru_vma(struct folio *folio, struct vm_area_struct *vma)
546{
547 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
548
549 if (unlikely((vma->vm_flags & (VM_LOCKED | VM_SPECIAL)) == VM_LOCKED))
550 mlock_new_folio(folio);
551 else
552 folio_add_lru(folio);
553}
554
555/*
556 * If the folio cannot be invalidated, it is moved to the
557 * inactive list to speed up its reclaim. It is moved to the
558 * head of the list, rather than the tail, to give the flusher
559 * threads some time to write it out, as this is much more
560 * effective than the single-page writeout from reclaim.
561 *
562 * If the folio isn't mapped and dirty/writeback, the folio
563 * could be reclaimed asap using the reclaim flag.
564 *
565 * 1. active, mapped folio -> none
566 * 2. active, dirty/writeback folio -> inactive, head, reclaim
567 * 3. inactive, mapped folio -> none
568 * 4. inactive, dirty/writeback folio -> inactive, head, reclaim
569 * 5. inactive, clean -> inactive, tail
570 * 6. Others -> none
571 *
572 * In 4, it moves to the head of the inactive list so the folio is
573 * written out by flusher threads as this is much more efficient
574 * than the single-page writeout from reclaim.
575 */
576static void lru_deactivate_file_fn(struct lruvec *lruvec, struct folio *folio)
577{
578 bool active = folio_test_active(folio);
579 long nr_pages = folio_nr_pages(folio);
580
581 if (folio_test_unevictable(folio))
582 return;
583
584 /* Some processes are using the folio */
585 if (folio_mapped(folio))
586 return;
587
588 lruvec_del_folio(lruvec, folio);
589 folio_clear_active(folio);
590 folio_clear_referenced(folio);
591
592 if (folio_test_writeback(folio) || folio_test_dirty(folio)) {
593 /*
594 * Setting the reclaim flag could race with
595 * folio_end_writeback() and confuse readahead. But the
596 * race window is _really_ small and it's not a critical
597 * problem.
598 */
599 lruvec_add_folio(lruvec, folio);
600 folio_set_reclaim(folio);
601 } else {
602 /*
603 * The folio's writeback ended while it was in the batch.
604 * We move that folio to the tail of the inactive list.
605 */
606 lruvec_add_folio_tail(lruvec, folio);
607 __count_vm_events(PGROTATED, nr_pages);
608 }
609
610 if (active) {
611 __count_vm_events(PGDEACTIVATE, nr_pages);
612 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
613 nr_pages);
614 }
615}
616
617static void lru_deactivate_fn(struct lruvec *lruvec, struct folio *folio)
618{
619 if (!folio_test_unevictable(folio) && (folio_test_active(folio) || lru_gen_enabled())) {
620 long nr_pages = folio_nr_pages(folio);
621
622 lruvec_del_folio(lruvec, folio);
623 folio_clear_active(folio);
624 folio_clear_referenced(folio);
625 lruvec_add_folio(lruvec, folio);
626
627 __count_vm_events(PGDEACTIVATE, nr_pages);
628 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
629 nr_pages);
630 }
631}
632
633static void lru_lazyfree_fn(struct lruvec *lruvec, struct folio *folio)
634{
635 if (folio_test_anon(folio) && folio_test_swapbacked(folio) &&
636 !folio_test_swapcache(folio) && !folio_test_unevictable(folio)) {
637 long nr_pages = folio_nr_pages(folio);
638
639 lruvec_del_folio(lruvec, folio);
640 folio_clear_active(folio);
641 folio_clear_referenced(folio);
642 /*
643 * Lazyfree folios are clean anonymous folios. They have
644 * the swapbacked flag cleared, to distinguish them from normal
645 * anonymous folios
646 */
647 folio_clear_swapbacked(folio);
648 lruvec_add_folio(lruvec, folio);
649
650 __count_vm_events(PGLAZYFREE, nr_pages);
651 __count_memcg_events(lruvec_memcg(lruvec), PGLAZYFREE,
652 nr_pages);
653 }
654}
655
656/*
657 * Drain pages out of the cpu's folio_batch.
658 * Either "cpu" is the current CPU, and preemption has already been
659 * disabled; or "cpu" is being hot-unplugged, and is already dead.
660 */
661void lru_add_drain_cpu(int cpu)
662{
663 struct cpu_fbatches *fbatches = &per_cpu(cpu_fbatches, cpu);
664 struct folio_batch *fbatch = &fbatches->lru_add;
665
666 if (folio_batch_count(fbatch))
667 folio_batch_move_lru(fbatch, lru_add_fn);
668
669 fbatch = &per_cpu(lru_rotate.fbatch, cpu);
670 /* Disabling interrupts below acts as a compiler barrier. */
671 if (data_race(folio_batch_count(fbatch))) {
672 unsigned long flags;
673
674 /* No harm done if a racing interrupt already did this */
675 local_lock_irqsave(&lru_rotate.lock, flags);
676 folio_batch_move_lru(fbatch, lru_move_tail_fn);
677 local_unlock_irqrestore(&lru_rotate.lock, flags);
678 }
679
680 fbatch = &fbatches->lru_deactivate_file;
681 if (folio_batch_count(fbatch))
682 folio_batch_move_lru(fbatch, lru_deactivate_file_fn);
683
684 fbatch = &fbatches->lru_deactivate;
685 if (folio_batch_count(fbatch))
686 folio_batch_move_lru(fbatch, lru_deactivate_fn);
687
688 fbatch = &fbatches->lru_lazyfree;
689 if (folio_batch_count(fbatch))
690 folio_batch_move_lru(fbatch, lru_lazyfree_fn);
691
692 folio_activate_drain(cpu);
693}
694
695/**
696 * deactivate_file_folio() - Deactivate a file folio.
697 * @folio: Folio to deactivate.
698 *
699 * This function hints to the VM that @folio is a good reclaim candidate,
700 * for example if its invalidation fails due to the folio being dirty
701 * or under writeback.
702 *
703 * Context: Caller holds a reference on the folio.
704 */
705void deactivate_file_folio(struct folio *folio)
706{
707 struct folio_batch *fbatch;
708
709 /* Deactivating an unevictable folio will not accelerate reclaim */
710 if (folio_test_unevictable(folio))
711 return;
712
713 folio_get(folio);
714 local_lock(&cpu_fbatches.lock);
715 fbatch = this_cpu_ptr(&cpu_fbatches.lru_deactivate_file);
716 folio_batch_add_and_move(fbatch, folio, lru_deactivate_file_fn);
717 local_unlock(&cpu_fbatches.lock);
718}
719
720/*
721 * folio_deactivate - deactivate a folio
722 * @folio: folio to deactivate
723 *
724 * folio_deactivate() moves @folio to the inactive list if @folio was on the
725 * active list and was not unevictable. This is done to accelerate the
726 * reclaim of @folio.
727 */
728void folio_deactivate(struct folio *folio)
729{
730 if (folio_test_lru(folio) && !folio_test_unevictable(folio) &&
731 (folio_test_active(folio) || lru_gen_enabled())) {
732 struct folio_batch *fbatch;
733
734 folio_get(folio);
735 local_lock(&cpu_fbatches.lock);
736 fbatch = this_cpu_ptr(&cpu_fbatches.lru_deactivate);
737 folio_batch_add_and_move(fbatch, folio, lru_deactivate_fn);
738 local_unlock(&cpu_fbatches.lock);
739 }
740}
741
742/**
743 * folio_mark_lazyfree - make an anon folio lazyfree
744 * @folio: folio to deactivate
745 *
746 * folio_mark_lazyfree() moves @folio to the inactive file list.
747 * This is done to accelerate the reclaim of @folio.
748 */
749void folio_mark_lazyfree(struct folio *folio)
750{
751 if (folio_test_lru(folio) && folio_test_anon(folio) &&
752 folio_test_swapbacked(folio) && !folio_test_swapcache(folio) &&
753 !folio_test_unevictable(folio)) {
754 struct folio_batch *fbatch;
755
756 folio_get(folio);
757 local_lock(&cpu_fbatches.lock);
758 fbatch = this_cpu_ptr(&cpu_fbatches.lru_lazyfree);
759 folio_batch_add_and_move(fbatch, folio, lru_lazyfree_fn);
760 local_unlock(&cpu_fbatches.lock);
761 }
762}
763
764void lru_add_drain(void)
765{
766 local_lock(&cpu_fbatches.lock);
767 lru_add_drain_cpu(smp_processor_id());
768 local_unlock(&cpu_fbatches.lock);
769 mlock_drain_local();
770}
771
772/*
773 * It's called from per-cpu workqueue context in SMP case so
774 * lru_add_drain_cpu and invalidate_bh_lrus_cpu should run on
775 * the same cpu. It shouldn't be a problem in !SMP case since
776 * the core is only one and the locks will disable preemption.
777 */
778static void lru_add_and_bh_lrus_drain(void)
779{
780 local_lock(&cpu_fbatches.lock);
781 lru_add_drain_cpu(smp_processor_id());
782 local_unlock(&cpu_fbatches.lock);
783 invalidate_bh_lrus_cpu();
784 mlock_drain_local();
785}
786
787void lru_add_drain_cpu_zone(struct zone *zone)
788{
789 local_lock(&cpu_fbatches.lock);
790 lru_add_drain_cpu(smp_processor_id());
791 drain_local_pages(zone);
792 local_unlock(&cpu_fbatches.lock);
793 mlock_drain_local();
794}
795
796#ifdef CONFIG_SMP
797
798static DEFINE_PER_CPU(struct work_struct, lru_add_drain_work);
799
800static void lru_add_drain_per_cpu(struct work_struct *dummy)
801{
802 lru_add_and_bh_lrus_drain();
803}
804
805static bool cpu_needs_drain(unsigned int cpu)
806{
807 struct cpu_fbatches *fbatches = &per_cpu(cpu_fbatches, cpu);
808
809 /* Check these in order of likelihood that they're not zero */
810 return folio_batch_count(&fbatches->lru_add) ||
811 data_race(folio_batch_count(&per_cpu(lru_rotate.fbatch, cpu))) ||
812 folio_batch_count(&fbatches->lru_deactivate_file) ||
813 folio_batch_count(&fbatches->lru_deactivate) ||
814 folio_batch_count(&fbatches->lru_lazyfree) ||
815 folio_batch_count(&fbatches->activate) ||
816 need_mlock_drain(cpu) ||
817 has_bh_in_lru(cpu, NULL);
818}
819
820/*
821 * Doesn't need any cpu hotplug locking because we do rely on per-cpu
822 * kworkers being shut down before our page_alloc_cpu_dead callback is
823 * executed on the offlined cpu.
824 * Calling this function with cpu hotplug locks held can actually lead
825 * to obscure indirect dependencies via WQ context.
826 */
827static inline void __lru_add_drain_all(bool force_all_cpus)
828{
829 /*
830 * lru_drain_gen - Global pages generation number
831 *
832 * (A) Definition: global lru_drain_gen = x implies that all generations
833 * 0 < n <= x are already *scheduled* for draining.
834 *
835 * This is an optimization for the highly-contended use case where a
836 * user space workload keeps constantly generating a flow of pages for
837 * each CPU.
838 */
839 static unsigned int lru_drain_gen;
840 static struct cpumask has_work;
841 static DEFINE_MUTEX(lock);
842 unsigned cpu, this_gen;
843
844 /*
845 * Make sure nobody triggers this path before mm_percpu_wq is fully
846 * initialized.
847 */
848 if (WARN_ON(!mm_percpu_wq))
849 return;
850
851 /*
852 * Guarantee folio_batch counter stores visible by this CPU
853 * are visible to other CPUs before loading the current drain
854 * generation.
855 */
856 smp_mb();
857
858 /*
859 * (B) Locally cache global LRU draining generation number
860 *
861 * The read barrier ensures that the counter is loaded before the mutex
862 * is taken. It pairs with smp_mb() inside the mutex critical section
863 * at (D).
864 */
865 this_gen = smp_load_acquire(&lru_drain_gen);
866
867 mutex_lock(&lock);
868
869 /*
870 * (C) Exit the draining operation if a newer generation, from another
871 * lru_add_drain_all(), was already scheduled for draining. Check (A).
872 */
873 if (unlikely(this_gen != lru_drain_gen && !force_all_cpus))
874 goto done;
875
876 /*
877 * (D) Increment global generation number
878 *
879 * Pairs with smp_load_acquire() at (B), outside of the critical
880 * section. Use a full memory barrier to guarantee that the
881 * new global drain generation number is stored before loading
882 * folio_batch counters.
883 *
884 * This pairing must be done here, before the for_each_online_cpu loop
885 * below which drains the page vectors.
886 *
887 * Let x, y, and z represent some system CPU numbers, where x < y < z.
888 * Assume CPU #z is in the middle of the for_each_online_cpu loop
889 * below and has already reached CPU #y's per-cpu data. CPU #x comes
890 * along, adds some pages to its per-cpu vectors, then calls
891 * lru_add_drain_all().
892 *
893 * If the paired barrier is done at any later step, e.g. after the
894 * loop, CPU #x will just exit at (C) and miss flushing out all of its
895 * added pages.
896 */
897 WRITE_ONCE(lru_drain_gen, lru_drain_gen + 1);
898 smp_mb();
899
900 cpumask_clear(&has_work);
901 for_each_online_cpu(cpu) {
902 struct work_struct *work = &per_cpu(lru_add_drain_work, cpu);
903
904 if (cpu_needs_drain(cpu)) {
905 INIT_WORK(work, lru_add_drain_per_cpu);
906 queue_work_on(cpu, mm_percpu_wq, work);
907 __cpumask_set_cpu(cpu, &has_work);
908 }
909 }
910
911 for_each_cpu(cpu, &has_work)
912 flush_work(&per_cpu(lru_add_drain_work, cpu));
913
914done:
915 mutex_unlock(&lock);
916}
917
918void lru_add_drain_all(void)
919{
920 __lru_add_drain_all(false);
921}
922#else
923void lru_add_drain_all(void)
924{
925 lru_add_drain();
926}
927#endif /* CONFIG_SMP */
928
929atomic_t lru_disable_count = ATOMIC_INIT(0);
930
931/*
932 * lru_cache_disable() needs to be called before we start compiling
933 * a list of pages to be migrated using isolate_lru_page().
934 * It drains pages on LRU cache and then disable on all cpus until
935 * lru_cache_enable is called.
936 *
937 * Must be paired with a call to lru_cache_enable().
938 */
939void lru_cache_disable(void)
940{
941 atomic_inc(&lru_disable_count);
942 /*
943 * Readers of lru_disable_count are protected by either disabling
944 * preemption or rcu_read_lock:
945 *
946 * preempt_disable, local_irq_disable [bh_lru_lock()]
947 * rcu_read_lock [rt_spin_lock CONFIG_PREEMPT_RT]
948 * preempt_disable [local_lock !CONFIG_PREEMPT_RT]
949 *
950 * Since v5.1 kernel, synchronize_rcu() is guaranteed to wait on
951 * preempt_disable() regions of code. So any CPU which sees
952 * lru_disable_count = 0 will have exited the critical
953 * section when synchronize_rcu() returns.
954 */
955 synchronize_rcu_expedited();
956#ifdef CONFIG_SMP
957 __lru_add_drain_all(true);
958#else
959 lru_add_and_bh_lrus_drain();
960#endif
961}
962
963/**
964 * folios_put_refs - Reduce the reference count on a batch of folios.
965 * @folios: The folios.
966 * @refs: The number of refs to subtract from each folio.
967 *
968 * Like folio_put(), but for a batch of folios. This is more efficient
969 * than writing the loop yourself as it will optimise the locks which need
970 * to be taken if the folios are freed. The folios batch is returned
971 * empty and ready to be reused for another batch; there is no need
972 * to reinitialise it. If @refs is NULL, we subtract one from each
973 * folio refcount.
974 *
975 * Context: May be called in process or interrupt context, but not in NMI
976 * context. May be called while holding a spinlock.
977 */
978void folios_put_refs(struct folio_batch *folios, unsigned int *refs)
979{
980 int i, j;
981 struct lruvec *lruvec = NULL;
982 unsigned long flags = 0;
983
984 for (i = 0, j = 0; i < folios->nr; i++) {
985 struct folio *folio = folios->folios[i];
986 unsigned int nr_refs = refs ? refs[i] : 1;
987
988 if (is_huge_zero_page(&folio->page))
989 continue;
990
991 if (folio_is_zone_device(folio)) {
992 if (lruvec) {
993 unlock_page_lruvec_irqrestore(lruvec, flags);
994 lruvec = NULL;
995 }
996 if (put_devmap_managed_page_refs(&folio->page, nr_refs))
997 continue;
998 if (folio_ref_sub_and_test(folio, nr_refs))
999 free_zone_device_page(&folio->page);
1000 continue;
1001 }
1002
1003 if (!folio_ref_sub_and_test(folio, nr_refs))
1004 continue;
1005
1006 /* hugetlb has its own memcg */
1007 if (folio_test_hugetlb(folio)) {
1008 if (lruvec) {
1009 unlock_page_lruvec_irqrestore(lruvec, flags);
1010 lruvec = NULL;
1011 }
1012 free_huge_folio(folio);
1013 continue;
1014 }
1015 if (folio_test_large(folio) &&
1016 folio_test_large_rmappable(folio))
1017 folio_undo_large_rmappable(folio);
1018
1019 __page_cache_release(folio, &lruvec, &flags);
1020
1021 if (j != i)
1022 folios->folios[j] = folio;
1023 j++;
1024 }
1025 if (lruvec)
1026 unlock_page_lruvec_irqrestore(lruvec, flags);
1027 if (!j) {
1028 folio_batch_reinit(folios);
1029 return;
1030 }
1031
1032 folios->nr = j;
1033 mem_cgroup_uncharge_folios(folios);
1034 free_unref_folios(folios);
1035}
1036EXPORT_SYMBOL(folios_put_refs);
1037
1038/**
1039 * release_pages - batched put_page()
1040 * @arg: array of pages to release
1041 * @nr: number of pages
1042 *
1043 * Decrement the reference count on all the pages in @arg. If it
1044 * fell to zero, remove the page from the LRU and free it.
1045 *
1046 * Note that the argument can be an array of pages, encoded pages,
1047 * or folio pointers. We ignore any encoded bits, and turn any of
1048 * them into just a folio that gets free'd.
1049 */
1050void release_pages(release_pages_arg arg, int nr)
1051{
1052 struct folio_batch fbatch;
1053 int refs[PAGEVEC_SIZE];
1054 struct encoded_page **encoded = arg.encoded_pages;
1055 int i;
1056
1057 folio_batch_init(&fbatch);
1058 for (i = 0; i < nr; i++) {
1059 /* Turn any of the argument types into a folio */
1060 struct folio *folio = page_folio(encoded_page_ptr(encoded[i]));
1061
1062 /* Is our next entry actually "nr_pages" -> "nr_refs" ? */
1063 refs[fbatch.nr] = 1;
1064 if (unlikely(encoded_page_flags(encoded[i]) &
1065 ENCODED_PAGE_BIT_NR_PAGES_NEXT))
1066 refs[fbatch.nr] = encoded_nr_pages(encoded[++i]);
1067
1068 if (folio_batch_add(&fbatch, folio) > 0)
1069 continue;
1070 folios_put_refs(&fbatch, refs);
1071 }
1072
1073 if (fbatch.nr)
1074 folios_put_refs(&fbatch, refs);
1075}
1076EXPORT_SYMBOL(release_pages);
1077
1078/*
1079 * The folios which we're about to release may be in the deferred lru-addition
1080 * queues. That would prevent them from really being freed right now. That's
1081 * OK from a correctness point of view but is inefficient - those folios may be
1082 * cache-warm and we want to give them back to the page allocator ASAP.
1083 *
1084 * So __folio_batch_release() will drain those queues here.
1085 * folio_batch_move_lru() calls folios_put() directly to avoid
1086 * mutual recursion.
1087 */
1088void __folio_batch_release(struct folio_batch *fbatch)
1089{
1090 if (!fbatch->percpu_pvec_drained) {
1091 lru_add_drain();
1092 fbatch->percpu_pvec_drained = true;
1093 }
1094 folios_put(fbatch);
1095}
1096EXPORT_SYMBOL(__folio_batch_release);
1097
1098/**
1099 * folio_batch_remove_exceptionals() - Prune non-folios from a batch.
1100 * @fbatch: The batch to prune
1101 *
1102 * find_get_entries() fills a batch with both folios and shadow/swap/DAX
1103 * entries. This function prunes all the non-folio entries from @fbatch
1104 * without leaving holes, so that it can be passed on to folio-only batch
1105 * operations.
1106 */
1107void folio_batch_remove_exceptionals(struct folio_batch *fbatch)
1108{
1109 unsigned int i, j;
1110
1111 for (i = 0, j = 0; i < folio_batch_count(fbatch); i++) {
1112 struct folio *folio = fbatch->folios[i];
1113 if (!xa_is_value(folio))
1114 fbatch->folios[j++] = folio;
1115 }
1116 fbatch->nr = j;
1117}
1118
1119/*
1120 * Perform any setup for the swap system
1121 */
1122void __init swap_setup(void)
1123{
1124 unsigned long megs = totalram_pages() >> (20 - PAGE_SHIFT);
1125
1126 /* Use a smaller cluster for small-memory machines */
1127 if (megs < 16)
1128 page_cluster = 2;
1129 else
1130 page_cluster = 3;
1131 /*
1132 * Right now other parts of the system means that we
1133 * _really_ don't want to cluster much more
1134 */
1135}