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