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);
160 EXPORT_SYMBOL(put_pages_list);
163 * get_kernel_pages() - pin kernel pages in memory
164 * @kiov: An array of struct kvec structures
165 * @nr_segs: number of segments to pin
166 * @write: pinning for read/write, currently ignored
167 * @pages: array that receives pointers to the pages pinned.
168 * Should be at least nr_segs long.
170 * Returns number of pages pinned. This may be fewer than the number
171 * requested. If nr_pages is 0 or negative, returns 0. If no pages
172 * were pinned, returns -errno. Each page returned must be released
173 * with a put_page() call when it is finished with.
175 int get_kernel_pages(const struct kvec *kiov, int nr_segs, int write,
180 for (seg = 0; seg < nr_segs; seg++) {
181 if (WARN_ON(kiov[seg].iov_len != PAGE_SIZE))
184 pages[seg] = kmap_to_page(kiov[seg].iov_base);
185 get_page(pages[seg]);
190 EXPORT_SYMBOL_GPL(get_kernel_pages);
192 static void pagevec_lru_move_fn(struct pagevec *pvec,
193 void (*move_fn)(struct page *page, struct lruvec *lruvec))
196 struct lruvec *lruvec = NULL;
197 unsigned long flags = 0;
199 for (i = 0; i < pagevec_count(pvec); i++) {
200 struct page *page = pvec->pages[i];
201 struct folio *folio = page_folio(page);
203 /* block memcg migration during page moving between lru */
204 if (!TestClearPageLRU(page))
207 lruvec = folio_lruvec_relock_irqsave(folio, lruvec, &flags);
208 (*move_fn)(page, lruvec);
213 unlock_page_lruvec_irqrestore(lruvec, flags);
214 release_pages(pvec->pages, pvec->nr);
215 pagevec_reinit(pvec);
218 static void pagevec_move_tail_fn(struct page *page, struct lruvec *lruvec)
220 struct folio *folio = page_folio(page);
222 if (!folio_test_unevictable(folio)) {
223 lruvec_del_folio(lruvec, folio);
224 folio_clear_active(folio);
225 lruvec_add_folio_tail(lruvec, folio);
226 __count_vm_events(PGROTATED, folio_nr_pages(folio));
230 /* return true if pagevec needs to drain */
231 static bool pagevec_add_and_need_flush(struct pagevec *pvec, struct page *page)
235 if (!pagevec_add(pvec, page) || PageCompound(page) ||
236 lru_cache_disabled())
243 * Writeback is about to end against a folio which has been marked for
244 * immediate reclaim. If it still appears to be reclaimable, move it
245 * to the tail of the inactive list.
247 * folio_rotate_reclaimable() must disable IRQs, to prevent nasty races.
249 void folio_rotate_reclaimable(struct folio *folio)
251 if (!folio_test_locked(folio) && !folio_test_dirty(folio) &&
252 !folio_test_unevictable(folio) && folio_test_lru(folio)) {
253 struct pagevec *pvec;
257 local_lock_irqsave(&lru_rotate.lock, flags);
258 pvec = this_cpu_ptr(&lru_rotate.pvec);
259 if (pagevec_add_and_need_flush(pvec, &folio->page))
260 pagevec_lru_move_fn(pvec, pagevec_move_tail_fn);
261 local_unlock_irqrestore(&lru_rotate.lock, flags);
265 void lru_note_cost(struct lruvec *lruvec, bool file, unsigned int nr_pages)
268 unsigned long lrusize;
271 * Hold lruvec->lru_lock is safe here, since
272 * 1) The pinned lruvec in reclaim, or
273 * 2) From a pre-LRU page during refault (which also holds the
274 * rcu lock, so would be safe even if the page was on the LRU
275 * and could move simultaneously to a new lruvec).
277 spin_lock_irq(&lruvec->lru_lock);
278 /* Record cost event */
280 lruvec->file_cost += nr_pages;
282 lruvec->anon_cost += nr_pages;
285 * Decay previous events
287 * Because workloads change over time (and to avoid
288 * overflow) we keep these statistics as a floating
289 * average, which ends up weighing recent refaults
290 * more than old ones.
292 lrusize = lruvec_page_state(lruvec, NR_INACTIVE_ANON) +
293 lruvec_page_state(lruvec, NR_ACTIVE_ANON) +
294 lruvec_page_state(lruvec, NR_INACTIVE_FILE) +
295 lruvec_page_state(lruvec, NR_ACTIVE_FILE);
297 if (lruvec->file_cost + lruvec->anon_cost > lrusize / 4) {
298 lruvec->file_cost /= 2;
299 lruvec->anon_cost /= 2;
301 spin_unlock_irq(&lruvec->lru_lock);
302 } while ((lruvec = parent_lruvec(lruvec)));
305 void lru_note_cost_folio(struct folio *folio)
307 lru_note_cost(folio_lruvec(folio), folio_is_file_lru(folio),
308 folio_nr_pages(folio));
311 static void __folio_activate(struct folio *folio, struct lruvec *lruvec)
313 if (!folio_test_active(folio) && !folio_test_unevictable(folio)) {
314 long nr_pages = folio_nr_pages(folio);
316 lruvec_del_folio(lruvec, folio);
317 folio_set_active(folio);
318 lruvec_add_folio(lruvec, folio);
319 trace_mm_lru_activate(folio);
321 __count_vm_events(PGACTIVATE, nr_pages);
322 __count_memcg_events(lruvec_memcg(lruvec), PGACTIVATE,
328 static void __activate_page(struct page *page, struct lruvec *lruvec)
330 return __folio_activate(page_folio(page), lruvec);
333 static void activate_page_drain(int cpu)
335 struct pagevec *pvec = &per_cpu(lru_pvecs.activate_page, cpu);
337 if (pagevec_count(pvec))
338 pagevec_lru_move_fn(pvec, __activate_page);
341 static bool need_activate_page_drain(int cpu)
343 return pagevec_count(&per_cpu(lru_pvecs.activate_page, cpu)) != 0;
346 static void folio_activate(struct folio *folio)
348 if (folio_test_lru(folio) && !folio_test_active(folio) &&
349 !folio_test_unevictable(folio)) {
350 struct pagevec *pvec;
353 local_lock(&lru_pvecs.lock);
354 pvec = this_cpu_ptr(&lru_pvecs.activate_page);
355 if (pagevec_add_and_need_flush(pvec, &folio->page))
356 pagevec_lru_move_fn(pvec, __activate_page);
357 local_unlock(&lru_pvecs.lock);
362 static inline void activate_page_drain(int cpu)
366 static void folio_activate(struct folio *folio)
368 struct lruvec *lruvec;
370 if (folio_test_clear_lru(folio)) {
371 lruvec = folio_lruvec_lock_irq(folio);
372 __folio_activate(folio, lruvec);
373 unlock_page_lruvec_irq(lruvec);
374 folio_set_lru(folio);
379 static void __lru_cache_activate_folio(struct folio *folio)
381 struct pagevec *pvec;
384 local_lock(&lru_pvecs.lock);
385 pvec = this_cpu_ptr(&lru_pvecs.lru_add);
388 * Search backwards on the optimistic assumption that the page being
389 * activated has just been added to this pagevec. Note that only
390 * the local pagevec is examined as a !PageLRU page could be in the
391 * process of being released, reclaimed, migrated or on a remote
392 * pagevec that is currently being drained. Furthermore, marking
393 * a remote pagevec's page PageActive potentially hits a race where
394 * a page is marked PageActive just after it is added to the inactive
395 * list causing accounting errors and BUG_ON checks to trigger.
397 for (i = pagevec_count(pvec) - 1; i >= 0; i--) {
398 struct page *pagevec_page = pvec->pages[i];
400 if (pagevec_page == &folio->page) {
401 folio_set_active(folio);
406 local_unlock(&lru_pvecs.lock);
410 * Mark a page as having seen activity.
412 * inactive,unreferenced -> inactive,referenced
413 * inactive,referenced -> active,unreferenced
414 * active,unreferenced -> active,referenced
416 * When a newly allocated page is not yet visible, so safe for non-atomic ops,
417 * __SetPageReferenced(page) may be substituted for mark_page_accessed(page).
419 void folio_mark_accessed(struct folio *folio)
421 if (!folio_test_referenced(folio)) {
422 folio_set_referenced(folio);
423 } else if (folio_test_unevictable(folio)) {
425 * Unevictable pages are on the "LRU_UNEVICTABLE" list. But,
426 * this list is never rotated or maintained, so marking an
427 * evictable page accessed has no effect.
429 } else if (!folio_test_active(folio)) {
431 * If the page is on the LRU, queue it for activation via
432 * lru_pvecs.activate_page. Otherwise, assume the page is on a
433 * pagevec, mark it active and it'll be moved to the active
434 * LRU on the next drain.
436 if (folio_test_lru(folio))
437 folio_activate(folio);
439 __lru_cache_activate_folio(folio);
440 folio_clear_referenced(folio);
441 workingset_activation(folio);
443 if (folio_test_idle(folio))
444 folio_clear_idle(folio);
446 EXPORT_SYMBOL(folio_mark_accessed);
449 * folio_add_lru - Add a folio to an LRU list.
450 * @folio: The folio to be added to the LRU.
452 * Queue the folio for addition to the LRU. The decision on whether
453 * to add the page to the [in]active [file|anon] list is deferred until the
454 * pagevec is drained. This gives a chance for the caller of folio_add_lru()
455 * have the folio added to the active list using folio_mark_accessed().
457 void folio_add_lru(struct folio *folio)
459 struct pagevec *pvec;
461 VM_BUG_ON_FOLIO(folio_test_active(folio) && folio_test_unevictable(folio), folio);
462 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
465 local_lock(&lru_pvecs.lock);
466 pvec = this_cpu_ptr(&lru_pvecs.lru_add);
467 if (pagevec_add_and_need_flush(pvec, &folio->page))
468 __pagevec_lru_add(pvec);
469 local_unlock(&lru_pvecs.lock);
471 EXPORT_SYMBOL(folio_add_lru);
474 * lru_cache_add_inactive_or_unevictable
475 * @page: the page to be added to LRU
476 * @vma: vma in which page is mapped for determining reclaimability
478 * Place @page on the inactive or unevictable LRU list, depending on its
481 void lru_cache_add_inactive_or_unevictable(struct page *page,
482 struct vm_area_struct *vma)
486 VM_BUG_ON_PAGE(PageLRU(page), page);
488 unevictable = (vma->vm_flags & (VM_LOCKED | VM_SPECIAL)) == VM_LOCKED;
489 if (unlikely(unevictable) && !TestSetPageMlocked(page)) {
490 int nr_pages = thp_nr_pages(page);
492 * We use the irq-unsafe __mod_zone_page_state because this
493 * counter is not modified from interrupt context, and the pte
494 * lock is held(spinlock), which implies preemption disabled.
496 __mod_zone_page_state(page_zone(page), NR_MLOCK, nr_pages);
497 count_vm_events(UNEVICTABLE_PGMLOCKED, nr_pages);
503 * If the page can not be invalidated, it is moved to the
504 * inactive list to speed up its reclaim. It is moved to the
505 * head of the list, rather than the tail, to give the flusher
506 * threads some time to write it out, as this is much more
507 * effective than the single-page writeout from reclaim.
509 * If the page isn't page_mapped and dirty/writeback, the page
510 * could reclaim asap using PG_reclaim.
512 * 1. active, mapped page -> none
513 * 2. active, dirty/writeback page -> inactive, head, PG_reclaim
514 * 3. inactive, mapped page -> none
515 * 4. inactive, dirty/writeback page -> inactive, head, PG_reclaim
516 * 5. inactive, clean -> inactive, tail
519 * In 4, why it moves inactive's head, the VM expects the page would
520 * be write it out by flusher threads as this is much more effective
521 * than the single-page writeout from reclaim.
523 static void lru_deactivate_file_fn(struct page *page, struct lruvec *lruvec)
525 bool active = PageActive(page);
526 int nr_pages = thp_nr_pages(page);
528 if (PageUnevictable(page))
531 /* Some processes are using the page */
532 if (page_mapped(page))
535 del_page_from_lru_list(page, lruvec);
536 ClearPageActive(page);
537 ClearPageReferenced(page);
539 if (PageWriteback(page) || PageDirty(page)) {
541 * PG_reclaim could be raced with end_page_writeback
542 * It can make readahead confusing. But race window
543 * is _really_ small and it's non-critical problem.
545 add_page_to_lru_list(page, lruvec);
546 SetPageReclaim(page);
549 * The page's writeback ends up during pagevec
550 * We move that page into tail of inactive.
552 add_page_to_lru_list_tail(page, lruvec);
553 __count_vm_events(PGROTATED, nr_pages);
557 __count_vm_events(PGDEACTIVATE, nr_pages);
558 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
563 static void lru_deactivate_fn(struct page *page, struct lruvec *lruvec)
565 if (PageActive(page) && !PageUnevictable(page)) {
566 int nr_pages = thp_nr_pages(page);
568 del_page_from_lru_list(page, lruvec);
569 ClearPageActive(page);
570 ClearPageReferenced(page);
571 add_page_to_lru_list(page, lruvec);
573 __count_vm_events(PGDEACTIVATE, nr_pages);
574 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
579 static void lru_lazyfree_fn(struct page *page, struct lruvec *lruvec)
581 if (PageAnon(page) && PageSwapBacked(page) &&
582 !PageSwapCache(page) && !PageUnevictable(page)) {
583 int nr_pages = thp_nr_pages(page);
585 del_page_from_lru_list(page, lruvec);
586 ClearPageActive(page);
587 ClearPageReferenced(page);
589 * Lazyfree pages are clean anonymous pages. They have
590 * PG_swapbacked flag cleared, to distinguish them from normal
593 ClearPageSwapBacked(page);
594 add_page_to_lru_list(page, lruvec);
596 __count_vm_events(PGLAZYFREE, nr_pages);
597 __count_memcg_events(lruvec_memcg(lruvec), PGLAZYFREE,
603 * Drain pages out of the cpu's pagevecs.
604 * Either "cpu" is the current CPU, and preemption has already been
605 * disabled; or "cpu" is being hot-unplugged, and is already dead.
607 void lru_add_drain_cpu(int cpu)
609 struct pagevec *pvec = &per_cpu(lru_pvecs.lru_add, cpu);
611 if (pagevec_count(pvec))
612 __pagevec_lru_add(pvec);
614 pvec = &per_cpu(lru_rotate.pvec, cpu);
615 /* Disabling interrupts below acts as a compiler barrier. */
616 if (data_race(pagevec_count(pvec))) {
619 /* No harm done if a racing interrupt already did this */
620 local_lock_irqsave(&lru_rotate.lock, flags);
621 pagevec_lru_move_fn(pvec, pagevec_move_tail_fn);
622 local_unlock_irqrestore(&lru_rotate.lock, flags);
625 pvec = &per_cpu(lru_pvecs.lru_deactivate_file, cpu);
626 if (pagevec_count(pvec))
627 pagevec_lru_move_fn(pvec, lru_deactivate_file_fn);
629 pvec = &per_cpu(lru_pvecs.lru_deactivate, cpu);
630 if (pagevec_count(pvec))
631 pagevec_lru_move_fn(pvec, lru_deactivate_fn);
633 pvec = &per_cpu(lru_pvecs.lru_lazyfree, cpu);
634 if (pagevec_count(pvec))
635 pagevec_lru_move_fn(pvec, lru_lazyfree_fn);
637 activate_page_drain(cpu);
641 * deactivate_file_page - forcefully deactivate a file page
642 * @page: page to deactivate
644 * This function hints the VM that @page is a good reclaim candidate,
645 * for example if its invalidation fails due to the page being dirty
646 * or under writeback.
648 void deactivate_file_page(struct page *page)
651 * In a workload with many unevictable page such as mprotect,
652 * unevictable page deactivation for accelerating reclaim is pointless.
654 if (PageUnevictable(page))
657 if (likely(get_page_unless_zero(page))) {
658 struct pagevec *pvec;
660 local_lock(&lru_pvecs.lock);
661 pvec = this_cpu_ptr(&lru_pvecs.lru_deactivate_file);
663 if (pagevec_add_and_need_flush(pvec, page))
664 pagevec_lru_move_fn(pvec, lru_deactivate_file_fn);
665 local_unlock(&lru_pvecs.lock);
670 * deactivate_page - deactivate a page
671 * @page: page to deactivate
673 * deactivate_page() moves @page to the inactive list if @page was on the active
674 * list and was not an unevictable page. This is done to accelerate the reclaim
677 void deactivate_page(struct page *page)
679 if (PageLRU(page) && PageActive(page) && !PageUnevictable(page)) {
680 struct pagevec *pvec;
682 local_lock(&lru_pvecs.lock);
683 pvec = this_cpu_ptr(&lru_pvecs.lru_deactivate);
685 if (pagevec_add_and_need_flush(pvec, page))
686 pagevec_lru_move_fn(pvec, lru_deactivate_fn);
687 local_unlock(&lru_pvecs.lock);
692 * mark_page_lazyfree - make an anon page lazyfree
693 * @page: page to deactivate
695 * mark_page_lazyfree() moves @page to the inactive file list.
696 * This is done to accelerate the reclaim of @page.
698 void mark_page_lazyfree(struct page *page)
700 if (PageLRU(page) && PageAnon(page) && PageSwapBacked(page) &&
701 !PageSwapCache(page) && !PageUnevictable(page)) {
702 struct pagevec *pvec;
704 local_lock(&lru_pvecs.lock);
705 pvec = this_cpu_ptr(&lru_pvecs.lru_lazyfree);
707 if (pagevec_add_and_need_flush(pvec, page))
708 pagevec_lru_move_fn(pvec, lru_lazyfree_fn);
709 local_unlock(&lru_pvecs.lock);
713 void lru_add_drain(void)
715 local_lock(&lru_pvecs.lock);
716 lru_add_drain_cpu(smp_processor_id());
717 local_unlock(&lru_pvecs.lock);
721 * It's called from per-cpu workqueue context in SMP case so
722 * lru_add_drain_cpu and invalidate_bh_lrus_cpu should run on
723 * the same cpu. It shouldn't be a problem in !SMP case since
724 * the core is only one and the locks will disable preemption.
726 static void lru_add_and_bh_lrus_drain(void)
728 local_lock(&lru_pvecs.lock);
729 lru_add_drain_cpu(smp_processor_id());
730 local_unlock(&lru_pvecs.lock);
731 invalidate_bh_lrus_cpu();
734 void lru_add_drain_cpu_zone(struct zone *zone)
736 local_lock(&lru_pvecs.lock);
737 lru_add_drain_cpu(smp_processor_id());
738 drain_local_pages(zone);
739 local_unlock(&lru_pvecs.lock);
744 static DEFINE_PER_CPU(struct work_struct, lru_add_drain_work);
746 static void lru_add_drain_per_cpu(struct work_struct *dummy)
748 lru_add_and_bh_lrus_drain();
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.
758 inline void __lru_add_drain_all(bool force_all_cpus)
761 * lru_drain_gen - Global pages generation number
763 * (A) Definition: global lru_drain_gen = x implies that all generations
764 * 0 < n <= x are already *scheduled* for draining.
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
770 static unsigned int lru_drain_gen;
771 static struct cpumask has_work;
772 static DEFINE_MUTEX(lock);
773 unsigned cpu, this_gen;
776 * Make sure nobody triggers this path before mm_percpu_wq is fully
779 if (WARN_ON(!mm_percpu_wq))
783 * Guarantee pagevec counter stores visible by this CPU are visible to
784 * other CPUs before loading the current drain generation.
789 * (B) Locally cache global LRU draining generation number
791 * The read barrier ensures that the counter is loaded before the mutex
792 * is taken. It pairs with smp_mb() inside the mutex critical section
795 this_gen = smp_load_acquire(&lru_drain_gen);
800 * (C) Exit the draining operation if a newer generation, from another
801 * lru_add_drain_all(), was already scheduled for draining. Check (A).
803 if (unlikely(this_gen != lru_drain_gen && !force_all_cpus))
807 * (D) Increment global generation number
809 * Pairs with smp_load_acquire() at (B), outside of the critical
810 * section. Use a full memory barrier to guarantee that the new global
811 * drain generation number is stored before loading pagevec counters.
813 * This pairing must be done here, before the for_each_online_cpu loop
814 * below which drains the page vectors.
816 * Let x, y, and z represent some system CPU numbers, where x < y < z.
817 * Assume CPU #z is in the middle of the for_each_online_cpu loop
818 * below and has already reached CPU #y's per-cpu data. CPU #x comes
819 * along, adds some pages to its per-cpu vectors, then calls
820 * lru_add_drain_all().
822 * If the paired barrier is done at any later step, e.g. after the
823 * loop, CPU #x will just exit at (C) and miss flushing out all of its
826 WRITE_ONCE(lru_drain_gen, lru_drain_gen + 1);
829 cpumask_clear(&has_work);
830 for_each_online_cpu(cpu) {
831 struct work_struct *work = &per_cpu(lru_add_drain_work, cpu);
833 if (force_all_cpus ||
834 pagevec_count(&per_cpu(lru_pvecs.lru_add, cpu)) ||
835 data_race(pagevec_count(&per_cpu(lru_rotate.pvec, cpu))) ||
836 pagevec_count(&per_cpu(lru_pvecs.lru_deactivate_file, cpu)) ||
837 pagevec_count(&per_cpu(lru_pvecs.lru_deactivate, cpu)) ||
838 pagevec_count(&per_cpu(lru_pvecs.lru_lazyfree, cpu)) ||
839 need_activate_page_drain(cpu) ||
840 has_bh_in_lru(cpu, NULL)) {
841 INIT_WORK(work, lru_add_drain_per_cpu);
842 queue_work_on(cpu, mm_percpu_wq, work);
843 __cpumask_set_cpu(cpu, &has_work);
847 for_each_cpu(cpu, &has_work)
848 flush_work(&per_cpu(lru_add_drain_work, cpu));
854 void lru_add_drain_all(void)
856 __lru_add_drain_all(false);
859 void lru_add_drain_all(void)
863 #endif /* CONFIG_SMP */
865 atomic_t lru_disable_count = ATOMIC_INIT(0);
868 * lru_cache_disable() needs to be called before we start compiling
869 * a list of pages to be migrated using isolate_lru_page().
870 * It drains pages on LRU cache and then disable on all cpus until
871 * lru_cache_enable is called.
873 * Must be paired with a call to lru_cache_enable().
875 void lru_cache_disable(void)
877 atomic_inc(&lru_disable_count);
880 * lru_add_drain_all in the force mode will schedule draining on
881 * all online CPUs so any calls of lru_cache_disabled wrapped by
882 * local_lock or preemption disabled would be ordered by that.
883 * The atomic operation doesn't need to have stronger ordering
884 * requirements because that is enforeced by the scheduling
887 __lru_add_drain_all(true);
889 lru_add_and_bh_lrus_drain();
894 * release_pages - batched put_page()
895 * @pages: array of pages to release
896 * @nr: number of pages
898 * Decrement the reference count on all the pages in @pages. If it
899 * fell to zero, remove the page from the LRU and free it.
901 void release_pages(struct page **pages, int nr)
904 LIST_HEAD(pages_to_free);
905 struct lruvec *lruvec = NULL;
906 unsigned long flags = 0;
907 unsigned int lock_batch;
909 for (i = 0; i < nr; i++) {
910 struct page *page = pages[i];
911 struct folio *folio = page_folio(page);
914 * Make sure the IRQ-safe lock-holding time does not get
915 * excessive with a continuous string of pages from the
916 * same lruvec. The lock is held only if lruvec != NULL.
918 if (lruvec && ++lock_batch == SWAP_CLUSTER_MAX) {
919 unlock_page_lruvec_irqrestore(lruvec, flags);
924 if (is_huge_zero_page(page))
927 if (is_zone_device_page(page)) {
929 unlock_page_lruvec_irqrestore(lruvec, flags);
933 * ZONE_DEVICE pages that return 'false' from
934 * page_is_devmap_managed() do not require special
935 * processing, and instead, expect a call to
936 * put_page_testzero().
938 if (page_is_devmap_managed(page)) {
939 put_devmap_managed_page(page);
942 if (put_page_testzero(page))
943 put_dev_pagemap(page->pgmap);
947 if (!put_page_testzero(page))
950 if (PageCompound(page)) {
952 unlock_page_lruvec_irqrestore(lruvec, flags);
955 __put_compound_page(page);
960 struct lruvec *prev_lruvec = lruvec;
962 lruvec = folio_lruvec_relock_irqsave(folio, lruvec,
964 if (prev_lruvec != lruvec)
967 del_page_from_lru_list(page, lruvec);
968 __clear_page_lru_flags(page);
971 __ClearPageWaiters(page);
973 list_add(&page->lru, &pages_to_free);
976 unlock_page_lruvec_irqrestore(lruvec, flags);
978 mem_cgroup_uncharge_list(&pages_to_free);
979 free_unref_page_list(&pages_to_free);
981 EXPORT_SYMBOL(release_pages);
984 * The pages which we're about to release may be in the deferred lru-addition
985 * queues. That would prevent them from really being freed right now. That's
986 * OK from a correctness point of view but is inefficient - those pages may be
987 * cache-warm and we want to give them back to the page allocator ASAP.
989 * So __pagevec_release() will drain those queues here. __pagevec_lru_add()
990 * and __pagevec_lru_add_active() call release_pages() directly to avoid
993 void __pagevec_release(struct pagevec *pvec)
995 if (!pvec->percpu_pvec_drained) {
997 pvec->percpu_pvec_drained = true;
999 release_pages(pvec->pages, pagevec_count(pvec));
1000 pagevec_reinit(pvec);
1002 EXPORT_SYMBOL(__pagevec_release);
1004 static void __pagevec_lru_add_fn(struct folio *folio, struct lruvec *lruvec)
1006 int was_unevictable = folio_test_clear_unevictable(folio);
1007 long nr_pages = folio_nr_pages(folio);
1009 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
1012 * A folio becomes evictable in two ways:
1013 * 1) Within LRU lock [munlock_vma_page() and __munlock_pagevec()].
1014 * 2) Before acquiring LRU lock to put the folio on the correct LRU
1016 * a) do PageLRU check with lock [check_move_unevictable_pages]
1017 * b) do PageLRU check before lock [clear_page_mlock]
1019 * (1) & (2a) are ok as LRU lock will serialize them. For (2b), we need
1020 * following strict ordering:
1022 * #0: __pagevec_lru_add_fn #1: clear_page_mlock
1024 * folio_set_lru() folio_test_clear_mlocked()
1025 * smp_mb() // explicit ordering // above provides strict
1027 * folio_test_mlocked() folio_test_lru()
1030 * if '#1' does not observe setting of PG_lru by '#0' and
1031 * fails isolation, the explicit barrier will make sure that
1032 * folio_evictable check will put the folio on the correct
1033 * LRU. Without smp_mb(), folio_set_lru() can be reordered
1034 * after folio_test_mlocked() check and can make '#1' fail the
1035 * isolation of the folio whose mlocked bit is cleared (#0 is
1036 * also looking at the same folio) and the evictable folio will
1037 * be stranded on an unevictable LRU.
1039 folio_set_lru(folio);
1040 smp_mb__after_atomic();
1042 if (folio_evictable(folio)) {
1043 if (was_unevictable)
1044 __count_vm_events(UNEVICTABLE_PGRESCUED, nr_pages);
1046 folio_clear_active(folio);
1047 folio_set_unevictable(folio);
1048 if (!was_unevictable)
1049 __count_vm_events(UNEVICTABLE_PGCULLED, nr_pages);
1052 lruvec_add_folio(lruvec, folio);
1053 trace_mm_lru_insertion(folio);
1057 * Add the passed pages to the LRU, then drop the caller's refcount
1058 * on them. Reinitialises the caller's pagevec.
1060 void __pagevec_lru_add(struct pagevec *pvec)
1063 struct lruvec *lruvec = NULL;
1064 unsigned long flags = 0;
1066 for (i = 0; i < pagevec_count(pvec); i++) {
1067 struct folio *folio = page_folio(pvec->pages[i]);
1069 lruvec = folio_lruvec_relock_irqsave(folio, lruvec, &flags);
1070 __pagevec_lru_add_fn(folio, lruvec);
1073 unlock_page_lruvec_irqrestore(lruvec, flags);
1074 release_pages(pvec->pages, pvec->nr);
1075 pagevec_reinit(pvec);
1079 * pagevec_remove_exceptionals - pagevec exceptionals pruning
1080 * @pvec: The pagevec to prune
1082 * find_get_entries() fills both pages and XArray value entries (aka
1083 * exceptional entries) into the pagevec. This function prunes all
1084 * exceptionals from @pvec without leaving holes, so that it can be
1085 * passed on to page-only pagevec operations.
1087 void pagevec_remove_exceptionals(struct pagevec *pvec)
1091 for (i = 0, j = 0; i < pagevec_count(pvec); i++) {
1092 struct page *page = pvec->pages[i];
1093 if (!xa_is_value(page))
1094 pvec->pages[j++] = page;
1100 * pagevec_lookup_range - gang pagecache lookup
1101 * @pvec: Where the resulting pages are placed
1102 * @mapping: The address_space to search
1103 * @start: The starting page index
1104 * @end: The final page index
1106 * pagevec_lookup_range() will search for & return a group of up to PAGEVEC_SIZE
1107 * pages in the mapping starting from index @start and upto index @end
1108 * (inclusive). The pages are placed in @pvec. pagevec_lookup() takes a
1109 * reference against the pages in @pvec.
1111 * The search returns a group of mapping-contiguous pages with ascending
1112 * indexes. There may be holes in the indices due to not-present pages. We
1113 * also update @start to index the next page for the traversal.
1115 * pagevec_lookup_range() returns the number of pages which were found. If this
1116 * number is smaller than PAGEVEC_SIZE, the end of specified range has been
1119 unsigned pagevec_lookup_range(struct pagevec *pvec,
1120 struct address_space *mapping, pgoff_t *start, pgoff_t end)
1122 pvec->nr = find_get_pages_range(mapping, start, end, PAGEVEC_SIZE,
1124 return pagevec_count(pvec);
1126 EXPORT_SYMBOL(pagevec_lookup_range);
1128 unsigned pagevec_lookup_range_tag(struct pagevec *pvec,
1129 struct address_space *mapping, pgoff_t *index, pgoff_t end,
1132 pvec->nr = find_get_pages_range_tag(mapping, index, end, tag,
1133 PAGEVEC_SIZE, pvec->pages);
1134 return pagevec_count(pvec);
1136 EXPORT_SYMBOL(pagevec_lookup_range_tag);
1139 * Perform any setup for the swap system
1141 void __init swap_setup(void)
1143 unsigned long megs = totalram_pages() >> (20 - PAGE_SHIFT);
1145 /* Use a smaller cluster for small-memory machines */
1151 * Right now other parts of the system means that we
1152 * _really_ don't want to cluster much more
1156 #ifdef CONFIG_DEV_PAGEMAP_OPS
1157 void put_devmap_managed_page(struct page *page)
1161 if (WARN_ON_ONCE(!page_is_devmap_managed(page)))
1164 count = page_ref_dec_return(page);
1167 * devmap page refcounts are 1-based, rather than 0-based: if
1168 * refcount is 1, then the page is free and the refcount is
1169 * stable because nobody holds a reference on the page.
1172 free_devmap_managed_page(page);
1176 EXPORT_SYMBOL(put_devmap_managed_page);