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 freed
78 * via pagevecs. But it gets used by networking - and for compound pages.
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 /* See comment on PageMlocked in release_pages() */
93 if (unlikely(PageMlocked(page))) {
94 int nr_pages = thp_nr_pages(page);
96 __ClearPageMlocked(page);
97 mod_zone_page_state(page_zone(page), NR_MLOCK, -nr_pages);
98 count_vm_events(UNEVICTABLE_PGCLEARED, nr_pages);
102 static void __put_single_page(struct page *page)
104 __page_cache_release(page);
105 mem_cgroup_uncharge(page_folio(page));
106 free_unref_page(page, 0);
109 static void __put_compound_page(struct page *page)
112 * __page_cache_release() is supposed to be called for thp, not for
113 * hugetlb. This is because hugetlb page does never have PageLRU set
114 * (it's never listed to any LRU lists) and no memcg routines should
115 * be called for hugetlb (it has a separate hugetlb_cgroup.)
118 __page_cache_release(page);
119 destroy_compound_page(page);
122 void __put_page(struct page *page)
124 if (unlikely(is_zone_device_page(page)))
125 free_zone_device_page(page);
126 else if (unlikely(PageCompound(page)))
127 __put_compound_page(page);
129 __put_single_page(page);
131 EXPORT_SYMBOL(__put_page);
134 * put_pages_list() - release a list of pages
135 * @pages: list of pages threaded on page->lru
137 * Release a list of pages which are strung together on page.lru.
139 void put_pages_list(struct list_head *pages)
141 struct page *page, *next;
143 list_for_each_entry_safe(page, next, pages, lru) {
144 if (!put_page_testzero(page)) {
145 list_del(&page->lru);
148 if (PageHead(page)) {
149 list_del(&page->lru);
150 __put_compound_page(page);
153 /* Cannot be PageLRU because it's passed to us using the lru */
156 free_unref_page_list(pages);
157 INIT_LIST_HEAD(pages);
159 EXPORT_SYMBOL(put_pages_list);
162 * get_kernel_pages() - pin kernel pages in memory
163 * @kiov: An array of struct kvec structures
164 * @nr_segs: number of segments to pin
165 * @write: pinning for read/write, currently ignored
166 * @pages: array that receives pointers to the pages pinned.
167 * Should be at least nr_segs long.
169 * Returns number of pages pinned. This may be fewer than the number requested.
170 * If nr_segs is 0 or negative, returns 0. If no pages were pinned, returns 0.
171 * Each page returned must be released with a put_page() call when it is
174 int get_kernel_pages(const struct kvec *kiov, int nr_segs, int write,
179 for (seg = 0; seg < nr_segs; seg++) {
180 if (WARN_ON(kiov[seg].iov_len != PAGE_SIZE))
183 pages[seg] = kmap_to_page(kiov[seg].iov_base);
184 get_page(pages[seg]);
189 EXPORT_SYMBOL_GPL(get_kernel_pages);
191 static void pagevec_lru_move_fn(struct pagevec *pvec,
192 void (*move_fn)(struct page *page, struct lruvec *lruvec))
195 struct lruvec *lruvec = NULL;
196 unsigned long flags = 0;
198 for (i = 0; i < pagevec_count(pvec); i++) {
199 struct page *page = pvec->pages[i];
200 struct folio *folio = page_folio(page);
202 /* block memcg migration during page moving between lru */
203 if (!TestClearPageLRU(page))
206 lruvec = folio_lruvec_relock_irqsave(folio, lruvec, &flags);
207 (*move_fn)(page, lruvec);
212 unlock_page_lruvec_irqrestore(lruvec, flags);
213 release_pages(pvec->pages, pvec->nr);
214 pagevec_reinit(pvec);
217 static void pagevec_move_tail_fn(struct page *page, struct lruvec *lruvec)
219 struct folio *folio = page_folio(page);
221 if (!folio_test_unevictable(folio)) {
222 lruvec_del_folio(lruvec, folio);
223 folio_clear_active(folio);
224 lruvec_add_folio_tail(lruvec, folio);
225 __count_vm_events(PGROTATED, folio_nr_pages(folio));
229 /* return true if pagevec needs to drain */
230 static bool pagevec_add_and_need_flush(struct pagevec *pvec, struct page *page)
234 if (!pagevec_add(pvec, page) || PageCompound(page) ||
235 lru_cache_disabled())
242 * Writeback is about to end against a folio which has been marked for
243 * immediate reclaim. If it still appears to be reclaimable, move it
244 * to the tail of the inactive list.
246 * folio_rotate_reclaimable() must disable IRQs, to prevent nasty races.
248 void folio_rotate_reclaimable(struct folio *folio)
250 if (!folio_test_locked(folio) && !folio_test_dirty(folio) &&
251 !folio_test_unevictable(folio) && folio_test_lru(folio)) {
252 struct pagevec *pvec;
256 local_lock_irqsave(&lru_rotate.lock, flags);
257 pvec = this_cpu_ptr(&lru_rotate.pvec);
258 if (pagevec_add_and_need_flush(pvec, &folio->page))
259 pagevec_lru_move_fn(pvec, pagevec_move_tail_fn);
260 local_unlock_irqrestore(&lru_rotate.lock, flags);
264 void lru_note_cost(struct lruvec *lruvec, bool file, unsigned int nr_pages)
267 unsigned long lrusize;
270 * Hold lruvec->lru_lock is safe here, since
271 * 1) The pinned lruvec in reclaim, or
272 * 2) From a pre-LRU page during refault (which also holds the
273 * rcu lock, so would be safe even if the page was on the LRU
274 * and could move simultaneously to a new lruvec).
276 spin_lock_irq(&lruvec->lru_lock);
277 /* Record cost event */
279 lruvec->file_cost += nr_pages;
281 lruvec->anon_cost += nr_pages;
284 * Decay previous events
286 * Because workloads change over time (and to avoid
287 * overflow) we keep these statistics as a floating
288 * average, which ends up weighing recent refaults
289 * more than old ones.
291 lrusize = lruvec_page_state(lruvec, NR_INACTIVE_ANON) +
292 lruvec_page_state(lruvec, NR_ACTIVE_ANON) +
293 lruvec_page_state(lruvec, NR_INACTIVE_FILE) +
294 lruvec_page_state(lruvec, NR_ACTIVE_FILE);
296 if (lruvec->file_cost + lruvec->anon_cost > lrusize / 4) {
297 lruvec->file_cost /= 2;
298 lruvec->anon_cost /= 2;
300 spin_unlock_irq(&lruvec->lru_lock);
301 } while ((lruvec = parent_lruvec(lruvec)));
304 void lru_note_cost_folio(struct folio *folio)
306 lru_note_cost(folio_lruvec(folio), folio_is_file_lru(folio),
307 folio_nr_pages(folio));
310 static void __folio_activate(struct folio *folio, struct lruvec *lruvec)
312 if (!folio_test_active(folio) && !folio_test_unevictable(folio)) {
313 long nr_pages = folio_nr_pages(folio);
315 lruvec_del_folio(lruvec, folio);
316 folio_set_active(folio);
317 lruvec_add_folio(lruvec, folio);
318 trace_mm_lru_activate(folio);
320 __count_vm_events(PGACTIVATE, nr_pages);
321 __count_memcg_events(lruvec_memcg(lruvec), PGACTIVATE,
327 static void __activate_page(struct page *page, struct lruvec *lruvec)
329 return __folio_activate(page_folio(page), lruvec);
332 static void activate_page_drain(int cpu)
334 struct pagevec *pvec = &per_cpu(lru_pvecs.activate_page, cpu);
336 if (pagevec_count(pvec))
337 pagevec_lru_move_fn(pvec, __activate_page);
340 static bool need_activate_page_drain(int cpu)
342 return pagevec_count(&per_cpu(lru_pvecs.activate_page, cpu)) != 0;
345 static void folio_activate(struct folio *folio)
347 if (folio_test_lru(folio) && !folio_test_active(folio) &&
348 !folio_test_unevictable(folio)) {
349 struct pagevec *pvec;
352 local_lock(&lru_pvecs.lock);
353 pvec = this_cpu_ptr(&lru_pvecs.activate_page);
354 if (pagevec_add_and_need_flush(pvec, &folio->page))
355 pagevec_lru_move_fn(pvec, __activate_page);
356 local_unlock(&lru_pvecs.lock);
361 static inline void activate_page_drain(int cpu)
365 static void folio_activate(struct folio *folio)
367 struct lruvec *lruvec;
369 if (folio_test_clear_lru(folio)) {
370 lruvec = folio_lruvec_lock_irq(folio);
371 __folio_activate(folio, lruvec);
372 unlock_page_lruvec_irq(lruvec);
373 folio_set_lru(folio);
378 static void __lru_cache_activate_folio(struct folio *folio)
380 struct pagevec *pvec;
383 local_lock(&lru_pvecs.lock);
384 pvec = this_cpu_ptr(&lru_pvecs.lru_add);
387 * Search backwards on the optimistic assumption that the page being
388 * activated has just been added to this pagevec. Note that only
389 * the local pagevec is examined as a !PageLRU page could be in the
390 * process of being released, reclaimed, migrated or on a remote
391 * pagevec that is currently being drained. Furthermore, marking
392 * a remote pagevec's page PageActive potentially hits a race where
393 * a page is marked PageActive just after it is added to the inactive
394 * list causing accounting errors and BUG_ON checks to trigger.
396 for (i = pagevec_count(pvec) - 1; i >= 0; i--) {
397 struct page *pagevec_page = pvec->pages[i];
399 if (pagevec_page == &folio->page) {
400 folio_set_active(folio);
405 local_unlock(&lru_pvecs.lock);
409 * Mark a page as having seen activity.
411 * inactive,unreferenced -> inactive,referenced
412 * inactive,referenced -> active,unreferenced
413 * active,unreferenced -> active,referenced
415 * When a newly allocated page is not yet visible, so safe for non-atomic ops,
416 * __SetPageReferenced(page) may be substituted for mark_page_accessed(page).
418 void folio_mark_accessed(struct folio *folio)
420 if (!folio_test_referenced(folio)) {
421 folio_set_referenced(folio);
422 } else if (folio_test_unevictable(folio)) {
424 * Unevictable pages are on the "LRU_UNEVICTABLE" list. But,
425 * this list is never rotated or maintained, so marking an
426 * unevictable page accessed has no effect.
428 } else if (!folio_test_active(folio)) {
430 * If the page is on the LRU, queue it for activation via
431 * lru_pvecs.activate_page. Otherwise, assume the page is on a
432 * pagevec, mark it active and it'll be moved to the active
433 * LRU on the next drain.
435 if (folio_test_lru(folio))
436 folio_activate(folio);
438 __lru_cache_activate_folio(folio);
439 folio_clear_referenced(folio);
440 workingset_activation(folio);
442 if (folio_test_idle(folio))
443 folio_clear_idle(folio);
445 EXPORT_SYMBOL(folio_mark_accessed);
448 * folio_add_lru - Add a folio to an LRU list.
449 * @folio: The folio to be added to the LRU.
451 * Queue the folio for addition to the LRU. The decision on whether
452 * to add the page to the [in]active [file|anon] list is deferred until the
453 * pagevec is drained. This gives a chance for the caller of folio_add_lru()
454 * have the folio added to the active list using folio_mark_accessed().
456 void folio_add_lru(struct folio *folio)
458 struct pagevec *pvec;
460 VM_BUG_ON_FOLIO(folio_test_active(folio) && folio_test_unevictable(folio), folio);
461 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
464 local_lock(&lru_pvecs.lock);
465 pvec = this_cpu_ptr(&lru_pvecs.lru_add);
466 if (pagevec_add_and_need_flush(pvec, &folio->page))
467 __pagevec_lru_add(pvec);
468 local_unlock(&lru_pvecs.lock);
470 EXPORT_SYMBOL(folio_add_lru);
473 * lru_cache_add_inactive_or_unevictable
474 * @page: the page to be added to LRU
475 * @vma: vma in which page is mapped for determining reclaimability
477 * Place @page on the inactive or unevictable LRU list, depending on its
480 void lru_cache_add_inactive_or_unevictable(struct page *page,
481 struct vm_area_struct *vma)
483 VM_BUG_ON_PAGE(PageLRU(page), page);
485 if (unlikely((vma->vm_flags & (VM_LOCKED | VM_SPECIAL)) == VM_LOCKED))
486 mlock_new_page(page);
492 * If the page can not be invalidated, it is moved to the
493 * inactive list to speed up its reclaim. It is moved to the
494 * head of the list, rather than the tail, to give the flusher
495 * threads some time to write it out, as this is much more
496 * effective than the single-page writeout from reclaim.
498 * If the page isn't page_mapped and dirty/writeback, the page
499 * could reclaim asap using PG_reclaim.
501 * 1. active, mapped page -> none
502 * 2. active, dirty/writeback page -> inactive, head, PG_reclaim
503 * 3. inactive, mapped page -> none
504 * 4. inactive, dirty/writeback page -> inactive, head, PG_reclaim
505 * 5. inactive, clean -> inactive, tail
508 * In 4, why it moves inactive's head, the VM expects the page would
509 * be write it out by flusher threads as this is much more effective
510 * than the single-page writeout from reclaim.
512 static void lru_deactivate_file_fn(struct page *page, struct lruvec *lruvec)
514 bool active = PageActive(page);
515 int nr_pages = thp_nr_pages(page);
517 if (PageUnevictable(page))
520 /* Some processes are using the page */
521 if (page_mapped(page))
524 del_page_from_lru_list(page, lruvec);
525 ClearPageActive(page);
526 ClearPageReferenced(page);
528 if (PageWriteback(page) || PageDirty(page)) {
530 * PG_reclaim could be raced with end_page_writeback
531 * It can make readahead confusing. But race window
532 * is _really_ small and it's non-critical problem.
534 add_page_to_lru_list(page, lruvec);
535 SetPageReclaim(page);
538 * The page's writeback ends up during pagevec
539 * We move that page into tail of inactive.
541 add_page_to_lru_list_tail(page, lruvec);
542 __count_vm_events(PGROTATED, nr_pages);
546 __count_vm_events(PGDEACTIVATE, nr_pages);
547 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
552 static void lru_deactivate_fn(struct page *page, struct lruvec *lruvec)
554 if (PageActive(page) && !PageUnevictable(page)) {
555 int nr_pages = thp_nr_pages(page);
557 del_page_from_lru_list(page, lruvec);
558 ClearPageActive(page);
559 ClearPageReferenced(page);
560 add_page_to_lru_list(page, lruvec);
562 __count_vm_events(PGDEACTIVATE, nr_pages);
563 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
568 static void lru_lazyfree_fn(struct page *page, struct lruvec *lruvec)
570 if (PageAnon(page) && PageSwapBacked(page) &&
571 !PageSwapCache(page) && !PageUnevictable(page)) {
572 int nr_pages = thp_nr_pages(page);
574 del_page_from_lru_list(page, lruvec);
575 ClearPageActive(page);
576 ClearPageReferenced(page);
578 * Lazyfree pages are clean anonymous pages. They have
579 * PG_swapbacked flag cleared, to distinguish them from normal
582 ClearPageSwapBacked(page);
583 add_page_to_lru_list(page, lruvec);
585 __count_vm_events(PGLAZYFREE, nr_pages);
586 __count_memcg_events(lruvec_memcg(lruvec), PGLAZYFREE,
592 * Drain pages out of the cpu's pagevecs.
593 * Either "cpu" is the current CPU, and preemption has already been
594 * disabled; or "cpu" is being hot-unplugged, and is already dead.
596 void lru_add_drain_cpu(int cpu)
598 struct pagevec *pvec = &per_cpu(lru_pvecs.lru_add, cpu);
600 if (pagevec_count(pvec))
601 __pagevec_lru_add(pvec);
603 pvec = &per_cpu(lru_rotate.pvec, cpu);
604 /* Disabling interrupts below acts as a compiler barrier. */
605 if (data_race(pagevec_count(pvec))) {
608 /* No harm done if a racing interrupt already did this */
609 local_lock_irqsave(&lru_rotate.lock, flags);
610 pagevec_lru_move_fn(pvec, pagevec_move_tail_fn);
611 local_unlock_irqrestore(&lru_rotate.lock, flags);
614 pvec = &per_cpu(lru_pvecs.lru_deactivate_file, cpu);
615 if (pagevec_count(pvec))
616 pagevec_lru_move_fn(pvec, lru_deactivate_file_fn);
618 pvec = &per_cpu(lru_pvecs.lru_deactivate, cpu);
619 if (pagevec_count(pvec))
620 pagevec_lru_move_fn(pvec, lru_deactivate_fn);
622 pvec = &per_cpu(lru_pvecs.lru_lazyfree, cpu);
623 if (pagevec_count(pvec))
624 pagevec_lru_move_fn(pvec, lru_lazyfree_fn);
626 activate_page_drain(cpu);
630 * deactivate_file_folio() - Forcefully deactivate a file folio.
631 * @folio: Folio to deactivate.
633 * This function hints to the VM that @folio is a good reclaim candidate,
634 * for example if its invalidation fails due to the folio being dirty
635 * or under writeback.
637 * Context: Caller holds a reference on the page.
639 void deactivate_file_folio(struct folio *folio)
641 struct pagevec *pvec;
644 * In a workload with many unevictable pages such as mprotect,
645 * unevictable folio deactivation for accelerating reclaim is pointless.
647 if (folio_test_unevictable(folio))
651 local_lock(&lru_pvecs.lock);
652 pvec = this_cpu_ptr(&lru_pvecs.lru_deactivate_file);
654 if (pagevec_add_and_need_flush(pvec, &folio->page))
655 pagevec_lru_move_fn(pvec, lru_deactivate_file_fn);
656 local_unlock(&lru_pvecs.lock);
660 * deactivate_page - deactivate a page
661 * @page: page to deactivate
663 * deactivate_page() moves @page to the inactive list if @page was on the active
664 * list and was not an unevictable page. This is done to accelerate the reclaim
667 void deactivate_page(struct page *page)
669 if (PageLRU(page) && PageActive(page) && !PageUnevictable(page)) {
670 struct pagevec *pvec;
672 local_lock(&lru_pvecs.lock);
673 pvec = this_cpu_ptr(&lru_pvecs.lru_deactivate);
675 if (pagevec_add_and_need_flush(pvec, page))
676 pagevec_lru_move_fn(pvec, lru_deactivate_fn);
677 local_unlock(&lru_pvecs.lock);
682 * mark_page_lazyfree - make an anon page lazyfree
683 * @page: page to deactivate
685 * mark_page_lazyfree() moves @page to the inactive file list.
686 * This is done to accelerate the reclaim of @page.
688 void mark_page_lazyfree(struct page *page)
690 if (PageLRU(page) && PageAnon(page) && PageSwapBacked(page) &&
691 !PageSwapCache(page) && !PageUnevictable(page)) {
692 struct pagevec *pvec;
694 local_lock(&lru_pvecs.lock);
695 pvec = this_cpu_ptr(&lru_pvecs.lru_lazyfree);
697 if (pagevec_add_and_need_flush(pvec, page))
698 pagevec_lru_move_fn(pvec, lru_lazyfree_fn);
699 local_unlock(&lru_pvecs.lock);
703 void lru_add_drain(void)
705 local_lock(&lru_pvecs.lock);
706 lru_add_drain_cpu(smp_processor_id());
707 local_unlock(&lru_pvecs.lock);
708 mlock_page_drain_local();
712 * It's called from per-cpu workqueue context in SMP case so
713 * lru_add_drain_cpu and invalidate_bh_lrus_cpu should run on
714 * the same cpu. It shouldn't be a problem in !SMP case since
715 * the core is only one and the locks will disable preemption.
717 static void lru_add_and_bh_lrus_drain(void)
719 local_lock(&lru_pvecs.lock);
720 lru_add_drain_cpu(smp_processor_id());
721 local_unlock(&lru_pvecs.lock);
722 invalidate_bh_lrus_cpu();
723 mlock_page_drain_local();
726 void lru_add_drain_cpu_zone(struct zone *zone)
728 local_lock(&lru_pvecs.lock);
729 lru_add_drain_cpu(smp_processor_id());
730 drain_local_pages(zone);
731 local_unlock(&lru_pvecs.lock);
732 mlock_page_drain_local();
737 static DEFINE_PER_CPU(struct work_struct, lru_add_drain_work);
739 static void lru_add_drain_per_cpu(struct work_struct *dummy)
741 lru_add_and_bh_lrus_drain();
745 * Doesn't need any cpu hotplug locking because we do rely on per-cpu
746 * kworkers being shut down before our page_alloc_cpu_dead callback is
747 * executed on the offlined cpu.
748 * Calling this function with cpu hotplug locks held can actually lead
749 * to obscure indirect dependencies via WQ context.
751 static inline void __lru_add_drain_all(bool force_all_cpus)
754 * lru_drain_gen - Global pages generation number
756 * (A) Definition: global lru_drain_gen = x implies that all generations
757 * 0 < n <= x are already *scheduled* for draining.
759 * This is an optimization for the highly-contended use case where a
760 * user space workload keeps constantly generating a flow of pages for
763 static unsigned int lru_drain_gen;
764 static struct cpumask has_work;
765 static DEFINE_MUTEX(lock);
766 unsigned cpu, this_gen;
769 * Make sure nobody triggers this path before mm_percpu_wq is fully
772 if (WARN_ON(!mm_percpu_wq))
776 * Guarantee pagevec counter stores visible by this CPU are visible to
777 * other CPUs before loading the current drain generation.
782 * (B) Locally cache global LRU draining generation number
784 * The read barrier ensures that the counter is loaded before the mutex
785 * is taken. It pairs with smp_mb() inside the mutex critical section
788 this_gen = smp_load_acquire(&lru_drain_gen);
793 * (C) Exit the draining operation if a newer generation, from another
794 * lru_add_drain_all(), was already scheduled for draining. Check (A).
796 if (unlikely(this_gen != lru_drain_gen && !force_all_cpus))
800 * (D) Increment global generation number
802 * Pairs with smp_load_acquire() at (B), outside of the critical
803 * section. Use a full memory barrier to guarantee that the new global
804 * drain generation number is stored before loading pagevec counters.
806 * This pairing must be done here, before the for_each_online_cpu loop
807 * below which drains the page vectors.
809 * Let x, y, and z represent some system CPU numbers, where x < y < z.
810 * Assume CPU #z is in the middle of the for_each_online_cpu loop
811 * below and has already reached CPU #y's per-cpu data. CPU #x comes
812 * along, adds some pages to its per-cpu vectors, then calls
813 * lru_add_drain_all().
815 * If the paired barrier is done at any later step, e.g. after the
816 * loop, CPU #x will just exit at (C) and miss flushing out all of its
819 WRITE_ONCE(lru_drain_gen, lru_drain_gen + 1);
822 cpumask_clear(&has_work);
823 for_each_online_cpu(cpu) {
824 struct work_struct *work = &per_cpu(lru_add_drain_work, cpu);
826 if (pagevec_count(&per_cpu(lru_pvecs.lru_add, cpu)) ||
827 data_race(pagevec_count(&per_cpu(lru_rotate.pvec, cpu))) ||
828 pagevec_count(&per_cpu(lru_pvecs.lru_deactivate_file, cpu)) ||
829 pagevec_count(&per_cpu(lru_pvecs.lru_deactivate, cpu)) ||
830 pagevec_count(&per_cpu(lru_pvecs.lru_lazyfree, cpu)) ||
831 need_activate_page_drain(cpu) ||
832 need_mlock_page_drain(cpu) ||
833 has_bh_in_lru(cpu, NULL)) {
834 INIT_WORK(work, lru_add_drain_per_cpu);
835 queue_work_on(cpu, mm_percpu_wq, work);
836 __cpumask_set_cpu(cpu, &has_work);
840 for_each_cpu(cpu, &has_work)
841 flush_work(&per_cpu(lru_add_drain_work, cpu));
847 void lru_add_drain_all(void)
849 __lru_add_drain_all(false);
852 void lru_add_drain_all(void)
856 #endif /* CONFIG_SMP */
858 atomic_t lru_disable_count = ATOMIC_INIT(0);
861 * lru_cache_disable() needs to be called before we start compiling
862 * a list of pages to be migrated using isolate_lru_page().
863 * It drains pages on LRU cache and then disable on all cpus until
864 * lru_cache_enable is called.
866 * Must be paired with a call to lru_cache_enable().
868 void lru_cache_disable(void)
870 atomic_inc(&lru_disable_count);
872 * Readers of lru_disable_count are protected by either disabling
873 * preemption or rcu_read_lock:
875 * preempt_disable, local_irq_disable [bh_lru_lock()]
876 * rcu_read_lock [rt_spin_lock CONFIG_PREEMPT_RT]
877 * preempt_disable [local_lock !CONFIG_PREEMPT_RT]
879 * Since v5.1 kernel, synchronize_rcu() is guaranteed to wait on
880 * preempt_disable() regions of code. So any CPU which sees
881 * lru_disable_count = 0 will have exited the critical
882 * section when synchronize_rcu() returns.
884 synchronize_rcu_expedited();
886 __lru_add_drain_all(true);
888 lru_add_and_bh_lrus_drain();
893 * release_pages - batched put_page()
894 * @pages: array of pages to release
895 * @nr: number of pages
897 * Decrement the reference count on all the pages in @pages. If it
898 * fell to zero, remove the page from the LRU and free it.
900 void release_pages(struct page **pages, int nr)
903 LIST_HEAD(pages_to_free);
904 struct lruvec *lruvec = NULL;
905 unsigned long flags = 0;
906 unsigned int lock_batch;
908 for (i = 0; i < nr; i++) {
909 struct page *page = pages[i];
910 struct folio *folio = page_folio(page);
913 * Make sure the IRQ-safe lock-holding time does not get
914 * excessive with a continuous string of pages from the
915 * same lruvec. The lock is held only if lruvec != NULL.
917 if (lruvec && ++lock_batch == SWAP_CLUSTER_MAX) {
918 unlock_page_lruvec_irqrestore(lruvec, flags);
923 if (is_huge_zero_page(page))
926 if (is_zone_device_page(page)) {
928 unlock_page_lruvec_irqrestore(lruvec, flags);
931 if (put_devmap_managed_page(page))
933 if (put_page_testzero(page))
934 free_zone_device_page(page);
938 if (!put_page_testzero(page))
941 if (PageCompound(page)) {
943 unlock_page_lruvec_irqrestore(lruvec, flags);
946 __put_compound_page(page);
951 struct lruvec *prev_lruvec = lruvec;
953 lruvec = folio_lruvec_relock_irqsave(folio, lruvec,
955 if (prev_lruvec != lruvec)
958 del_page_from_lru_list(page, lruvec);
959 __clear_page_lru_flags(page);
963 * In rare cases, when truncation or holepunching raced with
964 * munlock after VM_LOCKED was cleared, Mlocked may still be
965 * found set here. This does not indicate a problem, unless
966 * "unevictable_pgs_cleared" appears worryingly large.
968 if (unlikely(PageMlocked(page))) {
969 __ClearPageMlocked(page);
970 dec_zone_page_state(page, NR_MLOCK);
971 count_vm_event(UNEVICTABLE_PGCLEARED);
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);
1012 folio_set_lru(folio);
1014 * Is an smp_mb__after_atomic() still required here, before
1015 * folio_evictable() tests PageMlocked, to rule out the possibility
1016 * of stranding an evictable folio on an unevictable LRU? I think
1017 * not, because __munlock_page() only clears PageMlocked while the LRU
1020 * (That is not true of __page_cache_release(), and not necessarily
1021 * true of release_pages(): but those only clear PageMlocked after
1022 * put_page_testzero() has excluded any other users of the page.)
1024 if (folio_evictable(folio)) {
1025 if (was_unevictable)
1026 __count_vm_events(UNEVICTABLE_PGRESCUED, nr_pages);
1028 folio_clear_active(folio);
1029 folio_set_unevictable(folio);
1031 * folio->mlock_count = !!folio_test_mlocked(folio)?
1032 * But that leaves __mlock_page() in doubt whether another
1033 * actor has already counted the mlock or not. Err on the
1034 * safe side, underestimate, let page reclaim fix it, rather
1035 * than leaving a page on the unevictable LRU indefinitely.
1037 folio->mlock_count = 0;
1038 if (!was_unevictable)
1039 __count_vm_events(UNEVICTABLE_PGCULLED, nr_pages);
1042 lruvec_add_folio(lruvec, folio);
1043 trace_mm_lru_insertion(folio);
1047 * Add the passed pages to the LRU, then drop the caller's refcount
1048 * on them. Reinitialises the caller's pagevec.
1050 void __pagevec_lru_add(struct pagevec *pvec)
1053 struct lruvec *lruvec = NULL;
1054 unsigned long flags = 0;
1056 for (i = 0; i < pagevec_count(pvec); i++) {
1057 struct folio *folio = page_folio(pvec->pages[i]);
1059 lruvec = folio_lruvec_relock_irqsave(folio, lruvec, &flags);
1060 __pagevec_lru_add_fn(folio, lruvec);
1063 unlock_page_lruvec_irqrestore(lruvec, flags);
1064 release_pages(pvec->pages, pvec->nr);
1065 pagevec_reinit(pvec);
1069 * folio_batch_remove_exceptionals() - Prune non-folios from a batch.
1070 * @fbatch: The batch to prune
1072 * find_get_entries() fills a batch with both folios and shadow/swap/DAX
1073 * entries. This function prunes all the non-folio entries from @fbatch
1074 * without leaving holes, so that it can be passed on to folio-only batch
1077 void folio_batch_remove_exceptionals(struct folio_batch *fbatch)
1081 for (i = 0, j = 0; i < folio_batch_count(fbatch); i++) {
1082 struct folio *folio = fbatch->folios[i];
1083 if (!xa_is_value(folio))
1084 fbatch->folios[j++] = folio;
1090 * pagevec_lookup_range - gang pagecache lookup
1091 * @pvec: Where the resulting pages are placed
1092 * @mapping: The address_space to search
1093 * @start: The starting page index
1094 * @end: The final page index
1096 * pagevec_lookup_range() will search for & return a group of up to PAGEVEC_SIZE
1097 * pages in the mapping starting from index @start and upto index @end
1098 * (inclusive). The pages are placed in @pvec. pagevec_lookup() takes a
1099 * reference against the pages in @pvec.
1101 * The search returns a group of mapping-contiguous pages with ascending
1102 * indexes. There may be holes in the indices due to not-present pages. We
1103 * also update @start to index the next page for the traversal.
1105 * pagevec_lookup_range() returns the number of pages which were found. If this
1106 * number is smaller than PAGEVEC_SIZE, the end of specified range has been
1109 unsigned pagevec_lookup_range(struct pagevec *pvec,
1110 struct address_space *mapping, pgoff_t *start, pgoff_t end)
1112 pvec->nr = find_get_pages_range(mapping, start, end, PAGEVEC_SIZE,
1114 return pagevec_count(pvec);
1116 EXPORT_SYMBOL(pagevec_lookup_range);
1118 unsigned pagevec_lookup_range_tag(struct pagevec *pvec,
1119 struct address_space *mapping, pgoff_t *index, pgoff_t end,
1122 pvec->nr = find_get_pages_range_tag(mapping, index, end, tag,
1123 PAGEVEC_SIZE, pvec->pages);
1124 return pagevec_count(pvec);
1126 EXPORT_SYMBOL(pagevec_lookup_range_tag);
1129 * Perform any setup for the swap system
1131 void __init swap_setup(void)
1133 unsigned long megs = totalram_pages() >> (20 - PAGE_SHIFT);
1135 /* Use a smaller cluster for small-memory machines */
1141 * Right now other parts of the system means that we
1142 * _really_ don't want to cluster much more