1 // SPDX-License-Identifier: GPL-2.0-only
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
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.
13 * Swap aging added 23.2.95, Stephen Tweedie.
14 * Buffermem limits added 12.3.98, Rik van Riel.
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>
43 #define CREATE_TRACE_POINTS
44 #include <trace/events/pagemap.h>
46 /* How many pages do we try to swap or page in/out together? */
49 /* Protecting only lru_rotate.pvec which requires disabling interrupts */
54 static DEFINE_PER_CPU(struct lru_rotate, lru_rotate) = {
55 .lock = INIT_LOCAL_LOCK(lock),
59 * The following struct pagevec are grouped together because they are protected
60 * by disabling preemption (and interrupts remain enabled).
64 struct pagevec lru_add;
65 struct pagevec lru_deactivate_file;
66 struct pagevec lru_deactivate;
67 struct pagevec lru_lazyfree;
69 struct pagevec activate_page;
72 static DEFINE_PER_CPU(struct lru_pvecs, lru_pvecs) = {
73 .lock = INIT_LOCAL_LOCK(lock),
77 * This path almost never happens for VM activity - pages are normally
78 * freed via pagevecs. But it gets used by networking.
80 static void __page_cache_release(struct page *page)
83 struct folio *folio = page_folio(page);
84 struct lruvec *lruvec;
87 lruvec = folio_lruvec_lock_irqsave(folio, &flags);
88 del_page_from_lru_list(page, lruvec);
89 __clear_page_lru_flags(page);
90 unlock_page_lruvec_irqrestore(lruvec, flags);
92 __ClearPageWaiters(page);
95 static void __put_single_page(struct page *page)
97 __page_cache_release(page);
98 mem_cgroup_uncharge(page_folio(page));
99 free_unref_page(page, 0);
102 static void __put_compound_page(struct page *page)
105 * __page_cache_release() is supposed to be called for thp, not for
106 * hugetlb. This is because hugetlb page does never have PageLRU set
107 * (it's never listed to any LRU lists) and no memcg routines should
108 * be called for hugetlb (it has a separate hugetlb_cgroup.)
111 __page_cache_release(page);
112 destroy_compound_page(page);
115 void __put_page(struct page *page)
117 if (is_zone_device_page(page)) {
118 put_dev_pagemap(page->pgmap);
121 * The page belongs to the device that created pgmap. Do
122 * not return it to page allocator.
127 if (unlikely(PageCompound(page)))
128 __put_compound_page(page);
130 __put_single_page(page);
132 EXPORT_SYMBOL(__put_page);
135 * put_pages_list() - release a list of pages
136 * @pages: list of pages threaded on page->lru
138 * Release a list of pages which are strung together on page.lru.
140 void put_pages_list(struct list_head *pages)
142 struct page *page, *next;
144 list_for_each_entry_safe(page, next, pages, lru) {
145 if (!put_page_testzero(page)) {
146 list_del(&page->lru);
149 if (PageHead(page)) {
150 list_del(&page->lru);
151 __put_compound_page(page);
154 /* Cannot be PageLRU because it's passed to us using the lru */
155 __ClearPageWaiters(page);
158 free_unref_page_list(pages);
159 INIT_LIST_HEAD(pages);
161 EXPORT_SYMBOL(put_pages_list);
164 * get_kernel_pages() - pin kernel pages in memory
165 * @kiov: An array of struct kvec structures
166 * @nr_segs: number of segments to pin
167 * @write: pinning for read/write, currently ignored
168 * @pages: array that receives pointers to the pages pinned.
169 * Should be at least nr_segs long.
171 * Returns number of pages pinned. This may be fewer than the number
172 * requested. If nr_pages is 0 or negative, returns 0. If no pages
173 * were pinned, returns -errno. Each page returned must be released
174 * with a put_page() call when it is finished with.
176 int get_kernel_pages(const struct kvec *kiov, int nr_segs, int write,
181 for (seg = 0; seg < nr_segs; seg++) {
182 if (WARN_ON(kiov[seg].iov_len != PAGE_SIZE))
185 pages[seg] = kmap_to_page(kiov[seg].iov_base);
186 get_page(pages[seg]);
191 EXPORT_SYMBOL_GPL(get_kernel_pages);
193 static void pagevec_lru_move_fn(struct pagevec *pvec,
194 void (*move_fn)(struct page *page, struct lruvec *lruvec))
197 struct lruvec *lruvec = NULL;
198 unsigned long flags = 0;
200 for (i = 0; i < pagevec_count(pvec); i++) {
201 struct page *page = pvec->pages[i];
202 struct folio *folio = page_folio(page);
204 /* block memcg migration during page moving between lru */
205 if (!TestClearPageLRU(page))
208 lruvec = folio_lruvec_relock_irqsave(folio, lruvec, &flags);
209 (*move_fn)(page, lruvec);
214 unlock_page_lruvec_irqrestore(lruvec, flags);
215 release_pages(pvec->pages, pvec->nr);
216 pagevec_reinit(pvec);
219 static void pagevec_move_tail_fn(struct page *page, struct lruvec *lruvec)
221 struct folio *folio = page_folio(page);
223 if (!folio_test_unevictable(folio)) {
224 lruvec_del_folio(lruvec, folio);
225 folio_clear_active(folio);
226 lruvec_add_folio_tail(lruvec, folio);
227 __count_vm_events(PGROTATED, folio_nr_pages(folio));
231 /* return true if pagevec needs to drain */
232 static bool pagevec_add_and_need_flush(struct pagevec *pvec, struct page *page)
236 if (!pagevec_add(pvec, page) || PageCompound(page) ||
237 lru_cache_disabled())
244 * Writeback is about to end against a folio which has been marked for
245 * immediate reclaim. If it still appears to be reclaimable, move it
246 * to the tail of the inactive list.
248 * folio_rotate_reclaimable() must disable IRQs, to prevent nasty races.
250 void folio_rotate_reclaimable(struct folio *folio)
252 if (!folio_test_locked(folio) && !folio_test_dirty(folio) &&
253 !folio_test_unevictable(folio) && folio_test_lru(folio)) {
254 struct pagevec *pvec;
258 local_lock_irqsave(&lru_rotate.lock, flags);
259 pvec = this_cpu_ptr(&lru_rotate.pvec);
260 if (pagevec_add_and_need_flush(pvec, &folio->page))
261 pagevec_lru_move_fn(pvec, pagevec_move_tail_fn);
262 local_unlock_irqrestore(&lru_rotate.lock, flags);
266 void lru_note_cost(struct lruvec *lruvec, bool file, unsigned int nr_pages)
269 unsigned long lrusize;
272 * Hold lruvec->lru_lock is safe here, since
273 * 1) The pinned lruvec in reclaim, or
274 * 2) From a pre-LRU page during refault (which also holds the
275 * rcu lock, so would be safe even if the page was on the LRU
276 * and could move simultaneously to a new lruvec).
278 spin_lock_irq(&lruvec->lru_lock);
279 /* Record cost event */
281 lruvec->file_cost += nr_pages;
283 lruvec->anon_cost += nr_pages;
286 * Decay previous events
288 * Because workloads change over time (and to avoid
289 * overflow) we keep these statistics as a floating
290 * average, which ends up weighing recent refaults
291 * more than old ones.
293 lrusize = lruvec_page_state(lruvec, NR_INACTIVE_ANON) +
294 lruvec_page_state(lruvec, NR_ACTIVE_ANON) +
295 lruvec_page_state(lruvec, NR_INACTIVE_FILE) +
296 lruvec_page_state(lruvec, NR_ACTIVE_FILE);
298 if (lruvec->file_cost + lruvec->anon_cost > lrusize / 4) {
299 lruvec->file_cost /= 2;
300 lruvec->anon_cost /= 2;
302 spin_unlock_irq(&lruvec->lru_lock);
303 } while ((lruvec = parent_lruvec(lruvec)));
306 void lru_note_cost_folio(struct folio *folio)
308 lru_note_cost(folio_lruvec(folio), folio_is_file_lru(folio),
309 folio_nr_pages(folio));
312 static void __folio_activate(struct folio *folio, struct lruvec *lruvec)
314 if (!folio_test_active(folio) && !folio_test_unevictable(folio)) {
315 long nr_pages = folio_nr_pages(folio);
317 lruvec_del_folio(lruvec, folio);
318 folio_set_active(folio);
319 lruvec_add_folio(lruvec, folio);
320 trace_mm_lru_activate(folio);
322 __count_vm_events(PGACTIVATE, nr_pages);
323 __count_memcg_events(lruvec_memcg(lruvec), PGACTIVATE,
329 static void __activate_page(struct page *page, struct lruvec *lruvec)
331 return __folio_activate(page_folio(page), lruvec);
334 static void activate_page_drain(int cpu)
336 struct pagevec *pvec = &per_cpu(lru_pvecs.activate_page, cpu);
338 if (pagevec_count(pvec))
339 pagevec_lru_move_fn(pvec, __activate_page);
342 static bool need_activate_page_drain(int cpu)
344 return pagevec_count(&per_cpu(lru_pvecs.activate_page, cpu)) != 0;
347 static void folio_activate(struct folio *folio)
349 if (folio_test_lru(folio) && !folio_test_active(folio) &&
350 !folio_test_unevictable(folio)) {
351 struct pagevec *pvec;
354 local_lock(&lru_pvecs.lock);
355 pvec = this_cpu_ptr(&lru_pvecs.activate_page);
356 if (pagevec_add_and_need_flush(pvec, &folio->page))
357 pagevec_lru_move_fn(pvec, __activate_page);
358 local_unlock(&lru_pvecs.lock);
363 static inline void activate_page_drain(int cpu)
367 static void folio_activate(struct folio *folio)
369 struct lruvec *lruvec;
371 if (folio_test_clear_lru(folio)) {
372 lruvec = folio_lruvec_lock_irq(folio);
373 __folio_activate(folio, lruvec);
374 unlock_page_lruvec_irq(lruvec);
375 folio_set_lru(folio);
380 static void __lru_cache_activate_folio(struct folio *folio)
382 struct pagevec *pvec;
385 local_lock(&lru_pvecs.lock);
386 pvec = this_cpu_ptr(&lru_pvecs.lru_add);
389 * Search backwards on the optimistic assumption that the page being
390 * activated has just been added to this pagevec. Note that only
391 * the local pagevec is examined as a !PageLRU page could be in the
392 * process of being released, reclaimed, migrated or on a remote
393 * pagevec that is currently being drained. Furthermore, marking
394 * a remote pagevec's page PageActive potentially hits a race where
395 * a page is marked PageActive just after it is added to the inactive
396 * list causing accounting errors and BUG_ON checks to trigger.
398 for (i = pagevec_count(pvec) - 1; i >= 0; i--) {
399 struct page *pagevec_page = pvec->pages[i];
401 if (pagevec_page == &folio->page) {
402 folio_set_active(folio);
407 local_unlock(&lru_pvecs.lock);
411 * Mark a page as having seen activity.
413 * inactive,unreferenced -> inactive,referenced
414 * inactive,referenced -> active,unreferenced
415 * active,unreferenced -> active,referenced
417 * When a newly allocated page is not yet visible, so safe for non-atomic ops,
418 * __SetPageReferenced(page) may be substituted for mark_page_accessed(page).
420 void folio_mark_accessed(struct folio *folio)
422 if (!folio_test_referenced(folio)) {
423 folio_set_referenced(folio);
424 } else if (folio_test_unevictable(folio)) {
426 * Unevictable pages are on the "LRU_UNEVICTABLE" list. But,
427 * this list is never rotated or maintained, so marking an
428 * unevictable page accessed has no effect.
430 } else if (!folio_test_active(folio)) {
432 * If the page is on the LRU, queue it for activation via
433 * lru_pvecs.activate_page. Otherwise, assume the page is on a
434 * pagevec, mark it active and it'll be moved to the active
435 * LRU on the next drain.
437 if (folio_test_lru(folio))
438 folio_activate(folio);
440 __lru_cache_activate_folio(folio);
441 folio_clear_referenced(folio);
442 workingset_activation(folio);
444 if (folio_test_idle(folio))
445 folio_clear_idle(folio);
447 EXPORT_SYMBOL(folio_mark_accessed);
450 * folio_add_lru - Add a folio to an LRU list.
451 * @folio: The folio to be added to the LRU.
453 * Queue the folio for addition to the LRU. The decision on whether
454 * to add the page to the [in]active [file|anon] list is deferred until the
455 * pagevec is drained. This gives a chance for the caller of folio_add_lru()
456 * have the folio added to the active list using folio_mark_accessed().
458 void folio_add_lru(struct folio *folio)
460 struct pagevec *pvec;
462 VM_BUG_ON_FOLIO(folio_test_active(folio) && folio_test_unevictable(folio), folio);
463 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
466 local_lock(&lru_pvecs.lock);
467 pvec = this_cpu_ptr(&lru_pvecs.lru_add);
468 if (pagevec_add_and_need_flush(pvec, &folio->page))
469 __pagevec_lru_add(pvec);
470 local_unlock(&lru_pvecs.lock);
472 EXPORT_SYMBOL(folio_add_lru);
475 * lru_cache_add_inactive_or_unevictable
476 * @page: the page to be added to LRU
477 * @vma: vma in which page is mapped for determining reclaimability
479 * Place @page on the inactive or unevictable LRU list, depending on its
482 void lru_cache_add_inactive_or_unevictable(struct page *page,
483 struct vm_area_struct *vma)
487 VM_BUG_ON_PAGE(PageLRU(page), page);
489 unevictable = (vma->vm_flags & (VM_LOCKED | VM_SPECIAL)) == VM_LOCKED;
490 if (unlikely(unevictable) && !TestSetPageMlocked(page)) {
491 int nr_pages = thp_nr_pages(page);
493 * We use the irq-unsafe __mod_zone_page_state because this
494 * counter is not modified from interrupt context, and the pte
495 * lock is held(spinlock), which implies preemption disabled.
497 __mod_zone_page_state(page_zone(page), NR_MLOCK, nr_pages);
498 count_vm_events(UNEVICTABLE_PGMLOCKED, nr_pages);
504 * If the page can not be invalidated, it is moved to the
505 * inactive list to speed up its reclaim. It is moved to the
506 * head of the list, rather than the tail, to give the flusher
507 * threads some time to write it out, as this is much more
508 * effective than the single-page writeout from reclaim.
510 * If the page isn't page_mapped and dirty/writeback, the page
511 * could reclaim asap using PG_reclaim.
513 * 1. active, mapped page -> none
514 * 2. active, dirty/writeback page -> inactive, head, PG_reclaim
515 * 3. inactive, mapped page -> none
516 * 4. inactive, dirty/writeback page -> inactive, head, PG_reclaim
517 * 5. inactive, clean -> inactive, tail
520 * In 4, why it moves inactive's head, the VM expects the page would
521 * be write it out by flusher threads as this is much more effective
522 * than the single-page writeout from reclaim.
524 static void lru_deactivate_file_fn(struct page *page, struct lruvec *lruvec)
526 bool active = PageActive(page);
527 int nr_pages = thp_nr_pages(page);
529 if (PageUnevictable(page))
532 /* Some processes are using the page */
533 if (page_mapped(page))
536 del_page_from_lru_list(page, lruvec);
537 ClearPageActive(page);
538 ClearPageReferenced(page);
540 if (PageWriteback(page) || PageDirty(page)) {
542 * PG_reclaim could be raced with end_page_writeback
543 * It can make readahead confusing. But race window
544 * is _really_ small and it's non-critical problem.
546 add_page_to_lru_list(page, lruvec);
547 SetPageReclaim(page);
550 * The page's writeback ends up during pagevec
551 * We move that page into tail of inactive.
553 add_page_to_lru_list_tail(page, lruvec);
554 __count_vm_events(PGROTATED, nr_pages);
558 __count_vm_events(PGDEACTIVATE, nr_pages);
559 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
564 static void lru_deactivate_fn(struct page *page, struct lruvec *lruvec)
566 if (PageActive(page) && !PageUnevictable(page)) {
567 int nr_pages = thp_nr_pages(page);
569 del_page_from_lru_list(page, lruvec);
570 ClearPageActive(page);
571 ClearPageReferenced(page);
572 add_page_to_lru_list(page, lruvec);
574 __count_vm_events(PGDEACTIVATE, nr_pages);
575 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
580 static void lru_lazyfree_fn(struct page *page, struct lruvec *lruvec)
582 if (PageAnon(page) && PageSwapBacked(page) &&
583 !PageSwapCache(page) && !PageUnevictable(page)) {
584 int nr_pages = thp_nr_pages(page);
586 del_page_from_lru_list(page, lruvec);
587 ClearPageActive(page);
588 ClearPageReferenced(page);
590 * Lazyfree pages are clean anonymous pages. They have
591 * PG_swapbacked flag cleared, to distinguish them from normal
594 ClearPageSwapBacked(page);
595 add_page_to_lru_list(page, lruvec);
597 __count_vm_events(PGLAZYFREE, nr_pages);
598 __count_memcg_events(lruvec_memcg(lruvec), PGLAZYFREE,
604 * Drain pages out of the cpu's pagevecs.
605 * Either "cpu" is the current CPU, and preemption has already been
606 * disabled; or "cpu" is being hot-unplugged, and is already dead.
608 void lru_add_drain_cpu(int cpu)
610 struct pagevec *pvec = &per_cpu(lru_pvecs.lru_add, cpu);
612 if (pagevec_count(pvec))
613 __pagevec_lru_add(pvec);
615 pvec = &per_cpu(lru_rotate.pvec, cpu);
616 /* Disabling interrupts below acts as a compiler barrier. */
617 if (data_race(pagevec_count(pvec))) {
620 /* No harm done if a racing interrupt already did this */
621 local_lock_irqsave(&lru_rotate.lock, flags);
622 pagevec_lru_move_fn(pvec, pagevec_move_tail_fn);
623 local_unlock_irqrestore(&lru_rotate.lock, flags);
626 pvec = &per_cpu(lru_pvecs.lru_deactivate_file, cpu);
627 if (pagevec_count(pvec))
628 pagevec_lru_move_fn(pvec, lru_deactivate_file_fn);
630 pvec = &per_cpu(lru_pvecs.lru_deactivate, cpu);
631 if (pagevec_count(pvec))
632 pagevec_lru_move_fn(pvec, lru_deactivate_fn);
634 pvec = &per_cpu(lru_pvecs.lru_lazyfree, cpu);
635 if (pagevec_count(pvec))
636 pagevec_lru_move_fn(pvec, lru_lazyfree_fn);
638 activate_page_drain(cpu);
642 * deactivate_file_page - forcefully deactivate a file page
643 * @page: page to deactivate
645 * This function hints the VM that @page is a good reclaim candidate,
646 * for example if its invalidation fails due to the page being dirty
647 * or under writeback.
649 void deactivate_file_page(struct page *page)
652 * In a workload with many unevictable page such as mprotect,
653 * unevictable page deactivation for accelerating reclaim is pointless.
655 if (PageUnevictable(page))
658 if (likely(get_page_unless_zero(page))) {
659 struct pagevec *pvec;
661 local_lock(&lru_pvecs.lock);
662 pvec = this_cpu_ptr(&lru_pvecs.lru_deactivate_file);
664 if (pagevec_add_and_need_flush(pvec, page))
665 pagevec_lru_move_fn(pvec, lru_deactivate_file_fn);
666 local_unlock(&lru_pvecs.lock);
671 * deactivate_page - deactivate a page
672 * @page: page to deactivate
674 * deactivate_page() moves @page to the inactive list if @page was on the active
675 * list and was not an unevictable page. This is done to accelerate the reclaim
678 void deactivate_page(struct page *page)
680 if (PageLRU(page) && PageActive(page) && !PageUnevictable(page)) {
681 struct pagevec *pvec;
683 local_lock(&lru_pvecs.lock);
684 pvec = this_cpu_ptr(&lru_pvecs.lru_deactivate);
686 if (pagevec_add_and_need_flush(pvec, page))
687 pagevec_lru_move_fn(pvec, lru_deactivate_fn);
688 local_unlock(&lru_pvecs.lock);
693 * mark_page_lazyfree - make an anon page lazyfree
694 * @page: page to deactivate
696 * mark_page_lazyfree() moves @page to the inactive file list.
697 * This is done to accelerate the reclaim of @page.
699 void mark_page_lazyfree(struct page *page)
701 if (PageLRU(page) && PageAnon(page) && PageSwapBacked(page) &&
702 !PageSwapCache(page) && !PageUnevictable(page)) {
703 struct pagevec *pvec;
705 local_lock(&lru_pvecs.lock);
706 pvec = this_cpu_ptr(&lru_pvecs.lru_lazyfree);
708 if (pagevec_add_and_need_flush(pvec, page))
709 pagevec_lru_move_fn(pvec, lru_lazyfree_fn);
710 local_unlock(&lru_pvecs.lock);
714 void lru_add_drain(void)
716 local_lock(&lru_pvecs.lock);
717 lru_add_drain_cpu(smp_processor_id());
718 local_unlock(&lru_pvecs.lock);
722 * It's called from per-cpu workqueue context in SMP case so
723 * lru_add_drain_cpu and invalidate_bh_lrus_cpu should run on
724 * the same cpu. It shouldn't be a problem in !SMP case since
725 * the core is only one and the locks will disable preemption.
727 static void lru_add_and_bh_lrus_drain(void)
729 local_lock(&lru_pvecs.lock);
730 lru_add_drain_cpu(smp_processor_id());
731 local_unlock(&lru_pvecs.lock);
732 invalidate_bh_lrus_cpu();
735 void lru_add_drain_cpu_zone(struct zone *zone)
737 local_lock(&lru_pvecs.lock);
738 lru_add_drain_cpu(smp_processor_id());
739 drain_local_pages(zone);
740 local_unlock(&lru_pvecs.lock);
745 static DEFINE_PER_CPU(struct work_struct, lru_add_drain_work);
747 static void lru_add_drain_per_cpu(struct work_struct *dummy)
749 lru_add_and_bh_lrus_drain();
753 * Doesn't need any cpu hotplug locking because we do rely on per-cpu
754 * kworkers being shut down before our page_alloc_cpu_dead callback is
755 * executed on the offlined cpu.
756 * Calling this function with cpu hotplug locks held can actually lead
757 * to obscure indirect dependencies via WQ context.
759 inline void __lru_add_drain_all(bool force_all_cpus)
762 * lru_drain_gen - Global pages generation number
764 * (A) Definition: global lru_drain_gen = x implies that all generations
765 * 0 < n <= x are already *scheduled* for draining.
767 * This is an optimization for the highly-contended use case where a
768 * user space workload keeps constantly generating a flow of pages for
771 static unsigned int lru_drain_gen;
772 static struct cpumask has_work;
773 static DEFINE_MUTEX(lock);
774 unsigned cpu, this_gen;
777 * Make sure nobody triggers this path before mm_percpu_wq is fully
780 if (WARN_ON(!mm_percpu_wq))
784 * Guarantee pagevec counter stores visible by this CPU are visible to
785 * other CPUs before loading the current drain generation.
790 * (B) Locally cache global LRU draining generation number
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
796 this_gen = smp_load_acquire(&lru_drain_gen);
801 * (C) Exit the draining operation if a newer generation, from another
802 * lru_add_drain_all(), was already scheduled for draining. Check (A).
804 if (unlikely(this_gen != lru_drain_gen && !force_all_cpus))
808 * (D) Increment global generation number
810 * Pairs with smp_load_acquire() at (B), outside of the critical
811 * section. Use a full memory barrier to guarantee that the new global
812 * drain generation number is stored before loading pagevec counters.
814 * This pairing must be done here, before the for_each_online_cpu loop
815 * below which drains the page vectors.
817 * Let x, y, and z represent some system CPU numbers, where x < y < z.
818 * Assume CPU #z is in the middle of the for_each_online_cpu loop
819 * below and has already reached CPU #y's per-cpu data. CPU #x comes
820 * along, adds some pages to its per-cpu vectors, then calls
821 * lru_add_drain_all().
823 * If the paired barrier is done at any later step, e.g. after the
824 * loop, CPU #x will just exit at (C) and miss flushing out all of its
827 WRITE_ONCE(lru_drain_gen, lru_drain_gen + 1);
830 cpumask_clear(&has_work);
831 for_each_online_cpu(cpu) {
832 struct work_struct *work = &per_cpu(lru_add_drain_work, cpu);
834 if (force_all_cpus ||
835 pagevec_count(&per_cpu(lru_pvecs.lru_add, cpu)) ||
836 data_race(pagevec_count(&per_cpu(lru_rotate.pvec, cpu))) ||
837 pagevec_count(&per_cpu(lru_pvecs.lru_deactivate_file, cpu)) ||
838 pagevec_count(&per_cpu(lru_pvecs.lru_deactivate, cpu)) ||
839 pagevec_count(&per_cpu(lru_pvecs.lru_lazyfree, cpu)) ||
840 need_activate_page_drain(cpu) ||
841 has_bh_in_lru(cpu, NULL)) {
842 INIT_WORK(work, lru_add_drain_per_cpu);
843 queue_work_on(cpu, mm_percpu_wq, work);
844 __cpumask_set_cpu(cpu, &has_work);
848 for_each_cpu(cpu, &has_work)
849 flush_work(&per_cpu(lru_add_drain_work, cpu));
855 void lru_add_drain_all(void)
857 __lru_add_drain_all(false);
860 void lru_add_drain_all(void)
864 #endif /* CONFIG_SMP */
866 atomic_t lru_disable_count = ATOMIC_INIT(0);
869 * lru_cache_disable() needs to be called before we start compiling
870 * a list of pages to be migrated using isolate_lru_page().
871 * It drains pages on LRU cache and then disable on all cpus until
872 * lru_cache_enable is called.
874 * Must be paired with a call to lru_cache_enable().
876 void lru_cache_disable(void)
878 atomic_inc(&lru_disable_count);
881 * lru_add_drain_all in the force mode will schedule draining on
882 * all online CPUs so any calls of lru_cache_disabled wrapped by
883 * local_lock or preemption disabled would be ordered by that.
884 * The atomic operation doesn't need to have stronger ordering
885 * requirements because that is enforced by the scheduling
888 __lru_add_drain_all(true);
890 lru_add_and_bh_lrus_drain();
895 * release_pages - batched put_page()
896 * @pages: array of pages to release
897 * @nr: number of pages
899 * Decrement the reference count on all the pages in @pages. If it
900 * fell to zero, remove the page from the LRU and free it.
902 void release_pages(struct page **pages, int nr)
905 LIST_HEAD(pages_to_free);
906 struct lruvec *lruvec = NULL;
907 unsigned long flags = 0;
908 unsigned int lock_batch;
910 for (i = 0; i < nr; i++) {
911 struct page *page = pages[i];
912 struct folio *folio = page_folio(page);
915 * Make sure the IRQ-safe lock-holding time does not get
916 * excessive with a continuous string of pages from the
917 * same lruvec. The lock is held only if lruvec != NULL.
919 if (lruvec && ++lock_batch == SWAP_CLUSTER_MAX) {
920 unlock_page_lruvec_irqrestore(lruvec, flags);
925 if (is_huge_zero_page(page))
928 if (is_zone_device_page(page)) {
930 unlock_page_lruvec_irqrestore(lruvec, flags);
934 * ZONE_DEVICE pages that return 'false' from
935 * page_is_devmap_managed() do not require special
936 * processing, and instead, expect a call to
937 * put_page_testzero().
939 if (page_is_devmap_managed(page)) {
940 put_devmap_managed_page(page);
943 if (put_page_testzero(page))
944 put_dev_pagemap(page->pgmap);
948 if (!put_page_testzero(page))
951 if (PageCompound(page)) {
953 unlock_page_lruvec_irqrestore(lruvec, flags);
956 __put_compound_page(page);
961 struct lruvec *prev_lruvec = lruvec;
963 lruvec = folio_lruvec_relock_irqsave(folio, lruvec,
965 if (prev_lruvec != lruvec)
968 del_page_from_lru_list(page, lruvec);
969 __clear_page_lru_flags(page);
972 __ClearPageWaiters(page);
974 list_add(&page->lru, &pages_to_free);
977 unlock_page_lruvec_irqrestore(lruvec, flags);
979 mem_cgroup_uncharge_list(&pages_to_free);
980 free_unref_page_list(&pages_to_free);
982 EXPORT_SYMBOL(release_pages);
985 * The pages which we're about to release may be in the deferred lru-addition
986 * queues. That would prevent them from really being freed right now. That's
987 * OK from a correctness point of view but is inefficient - those pages may be
988 * cache-warm and we want to give them back to the page allocator ASAP.
990 * So __pagevec_release() will drain those queues here. __pagevec_lru_add()
991 * and __pagevec_lru_add_active() call release_pages() directly to avoid
994 void __pagevec_release(struct pagevec *pvec)
996 if (!pvec->percpu_pvec_drained) {
998 pvec->percpu_pvec_drained = true;
1000 release_pages(pvec->pages, pagevec_count(pvec));
1001 pagevec_reinit(pvec);
1003 EXPORT_SYMBOL(__pagevec_release);
1005 static void __pagevec_lru_add_fn(struct folio *folio, struct lruvec *lruvec)
1007 int was_unevictable = folio_test_clear_unevictable(folio);
1008 long nr_pages = folio_nr_pages(folio);
1010 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
1013 * A folio becomes evictable in two ways:
1014 * 1) Within LRU lock [munlock_vma_page() and __munlock_pagevec()].
1015 * 2) Before acquiring LRU lock to put the folio on the correct LRU
1017 * a) do PageLRU check with lock [check_move_unevictable_pages]
1018 * b) do PageLRU check before lock [clear_page_mlock]
1020 * (1) & (2a) are ok as LRU lock will serialize them. For (2b), we need
1021 * following strict ordering:
1023 * #0: __pagevec_lru_add_fn #1: clear_page_mlock
1025 * folio_set_lru() folio_test_clear_mlocked()
1026 * smp_mb() // explicit ordering // above provides strict
1028 * folio_test_mlocked() folio_test_lru()
1031 * if '#1' does not observe setting of PG_lru by '#0' and
1032 * fails isolation, the explicit barrier will make sure that
1033 * folio_evictable check will put the folio on the correct
1034 * LRU. Without smp_mb(), folio_set_lru() can be reordered
1035 * after folio_test_mlocked() check and can make '#1' fail the
1036 * isolation of the folio whose mlocked bit is cleared (#0 is
1037 * also looking at the same folio) and the evictable folio will
1038 * be stranded on an unevictable LRU.
1040 folio_set_lru(folio);
1041 smp_mb__after_atomic();
1043 if (folio_evictable(folio)) {
1044 if (was_unevictable)
1045 __count_vm_events(UNEVICTABLE_PGRESCUED, nr_pages);
1047 folio_clear_active(folio);
1048 folio_set_unevictable(folio);
1049 if (!was_unevictable)
1050 __count_vm_events(UNEVICTABLE_PGCULLED, nr_pages);
1053 lruvec_add_folio(lruvec, folio);
1054 trace_mm_lru_insertion(folio);
1058 * Add the passed pages to the LRU, then drop the caller's refcount
1059 * on them. Reinitialises the caller's pagevec.
1061 void __pagevec_lru_add(struct pagevec *pvec)
1064 struct lruvec *lruvec = NULL;
1065 unsigned long flags = 0;
1067 for (i = 0; i < pagevec_count(pvec); i++) {
1068 struct folio *folio = page_folio(pvec->pages[i]);
1070 lruvec = folio_lruvec_relock_irqsave(folio, lruvec, &flags);
1071 __pagevec_lru_add_fn(folio, lruvec);
1074 unlock_page_lruvec_irqrestore(lruvec, flags);
1075 release_pages(pvec->pages, pvec->nr);
1076 pagevec_reinit(pvec);
1080 * folio_batch_remove_exceptionals() - Prune non-folios from a batch.
1081 * @fbatch: The batch to prune
1083 * find_get_entries() fills a batch with both folios and shadow/swap/DAX
1084 * entries. This function prunes all the non-folio entries from @fbatch
1085 * without leaving holes, so that it can be passed on to folio-only batch
1088 void folio_batch_remove_exceptionals(struct folio_batch *fbatch)
1092 for (i = 0, j = 0; i < folio_batch_count(fbatch); i++) {
1093 struct folio *folio = fbatch->folios[i];
1094 if (!xa_is_value(folio))
1095 fbatch->folios[j++] = folio;
1101 * pagevec_lookup_range - gang pagecache lookup
1102 * @pvec: Where the resulting pages are placed
1103 * @mapping: The address_space to search
1104 * @start: The starting page index
1105 * @end: The final page index
1107 * pagevec_lookup_range() will search for & return a group of up to PAGEVEC_SIZE
1108 * pages in the mapping starting from index @start and upto index @end
1109 * (inclusive). The pages are placed in @pvec. pagevec_lookup() takes a
1110 * reference against the pages in @pvec.
1112 * The search returns a group of mapping-contiguous pages with ascending
1113 * indexes. There may be holes in the indices due to not-present pages. We
1114 * also update @start to index the next page for the traversal.
1116 * pagevec_lookup_range() returns the number of pages which were found. If this
1117 * number is smaller than PAGEVEC_SIZE, the end of specified range has been
1120 unsigned pagevec_lookup_range(struct pagevec *pvec,
1121 struct address_space *mapping, pgoff_t *start, pgoff_t end)
1123 pvec->nr = find_get_pages_range(mapping, start, end, PAGEVEC_SIZE,
1125 return pagevec_count(pvec);
1127 EXPORT_SYMBOL(pagevec_lookup_range);
1129 unsigned pagevec_lookup_range_tag(struct pagevec *pvec,
1130 struct address_space *mapping, pgoff_t *index, pgoff_t end,
1133 pvec->nr = find_get_pages_range_tag(mapping, index, end, tag,
1134 PAGEVEC_SIZE, pvec->pages);
1135 return pagevec_count(pvec);
1137 EXPORT_SYMBOL(pagevec_lookup_range_tag);
1140 * Perform any setup for the swap system
1142 void __init swap_setup(void)
1144 unsigned long megs = totalram_pages() >> (20 - PAGE_SHIFT);
1146 /* Use a smaller cluster for small-memory machines */
1152 * Right now other parts of the system means that we
1153 * _really_ don't want to cluster much more
1157 #ifdef CONFIG_DEV_PAGEMAP_OPS
1158 void put_devmap_managed_page(struct page *page)
1162 if (WARN_ON_ONCE(!page_is_devmap_managed(page)))
1165 count = page_ref_dec_return(page);
1168 * devmap page refcounts are 1-based, rather than 0-based: if
1169 * refcount is 1, then the page is free and the refcount is
1170 * stable because nobody holds a reference on the page.
1173 free_devmap_managed_page(page);
1177 EXPORT_SYMBOL(put_devmap_managed_page);