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. Currently
139 * used by read_cache_pages() and related error recovery code.
141 void put_pages_list(struct list_head *pages)
143 while (!list_empty(pages)) {
146 victim = lru_to_page(pages);
147 list_del(&victim->lru);
151 EXPORT_SYMBOL(put_pages_list);
154 * get_kernel_pages() - pin kernel pages in memory
155 * @kiov: An array of struct kvec structures
156 * @nr_segs: number of segments to pin
157 * @write: pinning for read/write, currently ignored
158 * @pages: array that receives pointers to the pages pinned.
159 * Should be at least nr_segs long.
161 * Returns number of pages pinned. This may be fewer than the number
162 * requested. If nr_pages is 0 or negative, returns 0. If no pages
163 * were pinned, returns -errno. Each page returned must be released
164 * with a put_page() call when it is finished with.
166 int get_kernel_pages(const struct kvec *kiov, int nr_segs, int write,
171 for (seg = 0; seg < nr_segs; seg++) {
172 if (WARN_ON(kiov[seg].iov_len != PAGE_SIZE))
175 pages[seg] = kmap_to_page(kiov[seg].iov_base);
176 get_page(pages[seg]);
181 EXPORT_SYMBOL_GPL(get_kernel_pages);
183 static void pagevec_lru_move_fn(struct pagevec *pvec,
184 void (*move_fn)(struct page *page, struct lruvec *lruvec))
187 struct lruvec *lruvec = NULL;
188 unsigned long flags = 0;
190 for (i = 0; i < pagevec_count(pvec); i++) {
191 struct page *page = pvec->pages[i];
192 struct folio *folio = page_folio(page);
194 /* block memcg migration during page moving between lru */
195 if (!TestClearPageLRU(page))
198 lruvec = folio_lruvec_relock_irqsave(folio, lruvec, &flags);
199 (*move_fn)(page, lruvec);
204 unlock_page_lruvec_irqrestore(lruvec, flags);
205 release_pages(pvec->pages, pvec->nr);
206 pagevec_reinit(pvec);
209 static void pagevec_move_tail_fn(struct page *page, struct lruvec *lruvec)
211 struct folio *folio = page_folio(page);
213 if (!folio_test_unevictable(folio)) {
214 lruvec_del_folio(lruvec, folio);
215 folio_clear_active(folio);
216 lruvec_add_folio_tail(lruvec, folio);
217 __count_vm_events(PGROTATED, folio_nr_pages(folio));
221 /* return true if pagevec needs to drain */
222 static bool pagevec_add_and_need_flush(struct pagevec *pvec, struct page *page)
226 if (!pagevec_add(pvec, page) || PageCompound(page) ||
227 lru_cache_disabled())
234 * Writeback is about to end against a folio which has been marked for
235 * immediate reclaim. If it still appears to be reclaimable, move it
236 * to the tail of the inactive list.
238 * folio_rotate_reclaimable() must disable IRQs, to prevent nasty races.
240 void folio_rotate_reclaimable(struct folio *folio)
242 if (!folio_test_locked(folio) && !folio_test_dirty(folio) &&
243 !folio_test_unevictable(folio) && folio_test_lru(folio)) {
244 struct pagevec *pvec;
248 local_lock_irqsave(&lru_rotate.lock, flags);
249 pvec = this_cpu_ptr(&lru_rotate.pvec);
250 if (pagevec_add_and_need_flush(pvec, &folio->page))
251 pagevec_lru_move_fn(pvec, pagevec_move_tail_fn);
252 local_unlock_irqrestore(&lru_rotate.lock, flags);
256 void lru_note_cost(struct lruvec *lruvec, bool file, unsigned int nr_pages)
259 unsigned long lrusize;
262 * Hold lruvec->lru_lock is safe here, since
263 * 1) The pinned lruvec in reclaim, or
264 * 2) From a pre-LRU page during refault (which also holds the
265 * rcu lock, so would be safe even if the page was on the LRU
266 * and could move simultaneously to a new lruvec).
268 spin_lock_irq(&lruvec->lru_lock);
269 /* Record cost event */
271 lruvec->file_cost += nr_pages;
273 lruvec->anon_cost += nr_pages;
276 * Decay previous events
278 * Because workloads change over time (and to avoid
279 * overflow) we keep these statistics as a floating
280 * average, which ends up weighing recent refaults
281 * more than old ones.
283 lrusize = lruvec_page_state(lruvec, NR_INACTIVE_ANON) +
284 lruvec_page_state(lruvec, NR_ACTIVE_ANON) +
285 lruvec_page_state(lruvec, NR_INACTIVE_FILE) +
286 lruvec_page_state(lruvec, NR_ACTIVE_FILE);
288 if (lruvec->file_cost + lruvec->anon_cost > lrusize / 4) {
289 lruvec->file_cost /= 2;
290 lruvec->anon_cost /= 2;
292 spin_unlock_irq(&lruvec->lru_lock);
293 } while ((lruvec = parent_lruvec(lruvec)));
296 void lru_note_cost_folio(struct folio *folio)
298 lru_note_cost(folio_lruvec(folio), folio_is_file_lru(folio),
299 folio_nr_pages(folio));
302 static void __folio_activate(struct folio *folio, struct lruvec *lruvec)
304 if (!folio_test_active(folio) && !folio_test_unevictable(folio)) {
305 long nr_pages = folio_nr_pages(folio);
307 lruvec_del_folio(lruvec, folio);
308 folio_set_active(folio);
309 lruvec_add_folio(lruvec, folio);
310 trace_mm_lru_activate(folio);
312 __count_vm_events(PGACTIVATE, nr_pages);
313 __count_memcg_events(lruvec_memcg(lruvec), PGACTIVATE,
319 static void __activate_page(struct page *page, struct lruvec *lruvec)
321 return __folio_activate(page_folio(page), lruvec);
324 static void activate_page_drain(int cpu)
326 struct pagevec *pvec = &per_cpu(lru_pvecs.activate_page, cpu);
328 if (pagevec_count(pvec))
329 pagevec_lru_move_fn(pvec, __activate_page);
332 static bool need_activate_page_drain(int cpu)
334 return pagevec_count(&per_cpu(lru_pvecs.activate_page, cpu)) != 0;
337 static void folio_activate(struct folio *folio)
339 if (folio_test_lru(folio) && !folio_test_active(folio) &&
340 !folio_test_unevictable(folio)) {
341 struct pagevec *pvec;
344 local_lock(&lru_pvecs.lock);
345 pvec = this_cpu_ptr(&lru_pvecs.activate_page);
346 if (pagevec_add_and_need_flush(pvec, &folio->page))
347 pagevec_lru_move_fn(pvec, __activate_page);
348 local_unlock(&lru_pvecs.lock);
353 static inline void activate_page_drain(int cpu)
357 static void folio_activate(struct folio *folio)
359 struct lruvec *lruvec;
361 if (folio_test_clear_lru(folio)) {
362 lruvec = folio_lruvec_lock_irq(folio);
363 __folio_activate(folio, lruvec);
364 unlock_page_lruvec_irq(lruvec);
365 folio_set_lru(folio);
370 static void __lru_cache_activate_folio(struct folio *folio)
372 struct pagevec *pvec;
375 local_lock(&lru_pvecs.lock);
376 pvec = this_cpu_ptr(&lru_pvecs.lru_add);
379 * Search backwards on the optimistic assumption that the page being
380 * activated has just been added to this pagevec. Note that only
381 * the local pagevec is examined as a !PageLRU page could be in the
382 * process of being released, reclaimed, migrated or on a remote
383 * pagevec that is currently being drained. Furthermore, marking
384 * a remote pagevec's page PageActive potentially hits a race where
385 * a page is marked PageActive just after it is added to the inactive
386 * list causing accounting errors and BUG_ON checks to trigger.
388 for (i = pagevec_count(pvec) - 1; i >= 0; i--) {
389 struct page *pagevec_page = pvec->pages[i];
391 if (pagevec_page == &folio->page) {
392 folio_set_active(folio);
397 local_unlock(&lru_pvecs.lock);
401 * Mark a page as having seen activity.
403 * inactive,unreferenced -> inactive,referenced
404 * inactive,referenced -> active,unreferenced
405 * active,unreferenced -> active,referenced
407 * When a newly allocated page is not yet visible, so safe for non-atomic ops,
408 * __SetPageReferenced(page) may be substituted for mark_page_accessed(page).
410 void folio_mark_accessed(struct folio *folio)
412 if (!folio_test_referenced(folio)) {
413 folio_set_referenced(folio);
414 } else if (folio_test_unevictable(folio)) {
416 * Unevictable pages are on the "LRU_UNEVICTABLE" list. But,
417 * this list is never rotated or maintained, so marking an
418 * evictable page accessed has no effect.
420 } else if (!folio_test_active(folio)) {
422 * If the page is on the LRU, queue it for activation via
423 * lru_pvecs.activate_page. Otherwise, assume the page is on a
424 * pagevec, mark it active and it'll be moved to the active
425 * LRU on the next drain.
427 if (folio_test_lru(folio))
428 folio_activate(folio);
430 __lru_cache_activate_folio(folio);
431 folio_clear_referenced(folio);
432 workingset_activation(folio);
434 if (folio_test_idle(folio))
435 folio_clear_idle(folio);
437 EXPORT_SYMBOL(folio_mark_accessed);
440 * folio_add_lru - Add a folio to an LRU list.
441 * @folio: The folio to be added to the LRU.
443 * Queue the folio for addition to the LRU. The decision on whether
444 * to add the page to the [in]active [file|anon] list is deferred until the
445 * pagevec is drained. This gives a chance for the caller of folio_add_lru()
446 * have the folio added to the active list using folio_mark_accessed().
448 void folio_add_lru(struct folio *folio)
450 struct pagevec *pvec;
452 VM_BUG_ON_FOLIO(folio_test_active(folio) && folio_test_unevictable(folio), folio);
453 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
456 local_lock(&lru_pvecs.lock);
457 pvec = this_cpu_ptr(&lru_pvecs.lru_add);
458 if (pagevec_add_and_need_flush(pvec, &folio->page))
459 __pagevec_lru_add(pvec);
460 local_unlock(&lru_pvecs.lock);
462 EXPORT_SYMBOL(folio_add_lru);
465 * lru_cache_add_inactive_or_unevictable
466 * @page: the page to be added to LRU
467 * @vma: vma in which page is mapped for determining reclaimability
469 * Place @page on the inactive or unevictable LRU list, depending on its
472 void lru_cache_add_inactive_or_unevictable(struct page *page,
473 struct vm_area_struct *vma)
477 VM_BUG_ON_PAGE(PageLRU(page), page);
479 unevictable = (vma->vm_flags & (VM_LOCKED | VM_SPECIAL)) == VM_LOCKED;
480 if (unlikely(unevictable) && !TestSetPageMlocked(page)) {
481 int nr_pages = thp_nr_pages(page);
483 * We use the irq-unsafe __mod_zone_page_state because this
484 * counter is not modified from interrupt context, and the pte
485 * lock is held(spinlock), which implies preemption disabled.
487 __mod_zone_page_state(page_zone(page), NR_MLOCK, nr_pages);
488 count_vm_events(UNEVICTABLE_PGMLOCKED, nr_pages);
494 * If the page can not be invalidated, it is moved to the
495 * inactive list to speed up its reclaim. It is moved to the
496 * head of the list, rather than the tail, to give the flusher
497 * threads some time to write it out, as this is much more
498 * effective than the single-page writeout from reclaim.
500 * If the page isn't page_mapped and dirty/writeback, the page
501 * could reclaim asap using PG_reclaim.
503 * 1. active, mapped page -> none
504 * 2. active, dirty/writeback page -> inactive, head, PG_reclaim
505 * 3. inactive, mapped page -> none
506 * 4. inactive, dirty/writeback page -> inactive, head, PG_reclaim
507 * 5. inactive, clean -> inactive, tail
510 * In 4, why it moves inactive's head, the VM expects the page would
511 * be write it out by flusher threads as this is much more effective
512 * than the single-page writeout from reclaim.
514 static void lru_deactivate_file_fn(struct page *page, struct lruvec *lruvec)
516 bool active = PageActive(page);
517 int nr_pages = thp_nr_pages(page);
519 if (PageUnevictable(page))
522 /* Some processes are using the page */
523 if (page_mapped(page))
526 del_page_from_lru_list(page, lruvec);
527 ClearPageActive(page);
528 ClearPageReferenced(page);
530 if (PageWriteback(page) || PageDirty(page)) {
532 * PG_reclaim could be raced with end_page_writeback
533 * It can make readahead confusing. But race window
534 * is _really_ small and it's non-critical problem.
536 add_page_to_lru_list(page, lruvec);
537 SetPageReclaim(page);
540 * The page's writeback ends up during pagevec
541 * We move that page into tail of inactive.
543 add_page_to_lru_list_tail(page, lruvec);
544 __count_vm_events(PGROTATED, nr_pages);
548 __count_vm_events(PGDEACTIVATE, nr_pages);
549 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
554 static void lru_deactivate_fn(struct page *page, struct lruvec *lruvec)
556 if (PageActive(page) && !PageUnevictable(page)) {
557 int nr_pages = thp_nr_pages(page);
559 del_page_from_lru_list(page, lruvec);
560 ClearPageActive(page);
561 ClearPageReferenced(page);
562 add_page_to_lru_list(page, lruvec);
564 __count_vm_events(PGDEACTIVATE, nr_pages);
565 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
570 static void lru_lazyfree_fn(struct page *page, struct lruvec *lruvec)
572 if (PageAnon(page) && PageSwapBacked(page) &&
573 !PageSwapCache(page) && !PageUnevictable(page)) {
574 int nr_pages = thp_nr_pages(page);
576 del_page_from_lru_list(page, lruvec);
577 ClearPageActive(page);
578 ClearPageReferenced(page);
580 * Lazyfree pages are clean anonymous pages. They have
581 * PG_swapbacked flag cleared, to distinguish them from normal
584 ClearPageSwapBacked(page);
585 add_page_to_lru_list(page, lruvec);
587 __count_vm_events(PGLAZYFREE, nr_pages);
588 __count_memcg_events(lruvec_memcg(lruvec), PGLAZYFREE,
594 * Drain pages out of the cpu's pagevecs.
595 * Either "cpu" is the current CPU, and preemption has already been
596 * disabled; or "cpu" is being hot-unplugged, and is already dead.
598 void lru_add_drain_cpu(int cpu)
600 struct pagevec *pvec = &per_cpu(lru_pvecs.lru_add, cpu);
602 if (pagevec_count(pvec))
603 __pagevec_lru_add(pvec);
605 pvec = &per_cpu(lru_rotate.pvec, cpu);
606 /* Disabling interrupts below acts as a compiler barrier. */
607 if (data_race(pagevec_count(pvec))) {
610 /* No harm done if a racing interrupt already did this */
611 local_lock_irqsave(&lru_rotate.lock, flags);
612 pagevec_lru_move_fn(pvec, pagevec_move_tail_fn);
613 local_unlock_irqrestore(&lru_rotate.lock, flags);
616 pvec = &per_cpu(lru_pvecs.lru_deactivate_file, cpu);
617 if (pagevec_count(pvec))
618 pagevec_lru_move_fn(pvec, lru_deactivate_file_fn);
620 pvec = &per_cpu(lru_pvecs.lru_deactivate, cpu);
621 if (pagevec_count(pvec))
622 pagevec_lru_move_fn(pvec, lru_deactivate_fn);
624 pvec = &per_cpu(lru_pvecs.lru_lazyfree, cpu);
625 if (pagevec_count(pvec))
626 pagevec_lru_move_fn(pvec, lru_lazyfree_fn);
628 activate_page_drain(cpu);
632 * deactivate_file_page - forcefully deactivate a file page
633 * @page: page to deactivate
635 * This function hints the VM that @page is a good reclaim candidate,
636 * for example if its invalidation fails due to the page being dirty
637 * or under writeback.
639 void deactivate_file_page(struct page *page)
642 * In a workload with many unevictable page such as mprotect,
643 * unevictable page deactivation for accelerating reclaim is pointless.
645 if (PageUnevictable(page))
648 if (likely(get_page_unless_zero(page))) {
649 struct pagevec *pvec;
651 local_lock(&lru_pvecs.lock);
652 pvec = this_cpu_ptr(&lru_pvecs.lru_deactivate_file);
654 if (pagevec_add_and_need_flush(pvec, page))
655 pagevec_lru_move_fn(pvec, lru_deactivate_file_fn);
656 local_unlock(&lru_pvecs.lock);
661 * deactivate_page - deactivate a page
662 * @page: page to deactivate
664 * deactivate_page() moves @page to the inactive list if @page was on the active
665 * list and was not an unevictable page. This is done to accelerate the reclaim
668 void deactivate_page(struct page *page)
670 if (PageLRU(page) && PageActive(page) && !PageUnevictable(page)) {
671 struct pagevec *pvec;
673 local_lock(&lru_pvecs.lock);
674 pvec = this_cpu_ptr(&lru_pvecs.lru_deactivate);
676 if (pagevec_add_and_need_flush(pvec, page))
677 pagevec_lru_move_fn(pvec, lru_deactivate_fn);
678 local_unlock(&lru_pvecs.lock);
683 * mark_page_lazyfree - make an anon page lazyfree
684 * @page: page to deactivate
686 * mark_page_lazyfree() moves @page to the inactive file list.
687 * This is done to accelerate the reclaim of @page.
689 void mark_page_lazyfree(struct page *page)
691 if (PageLRU(page) && PageAnon(page) && PageSwapBacked(page) &&
692 !PageSwapCache(page) && !PageUnevictable(page)) {
693 struct pagevec *pvec;
695 local_lock(&lru_pvecs.lock);
696 pvec = this_cpu_ptr(&lru_pvecs.lru_lazyfree);
698 if (pagevec_add_and_need_flush(pvec, page))
699 pagevec_lru_move_fn(pvec, lru_lazyfree_fn);
700 local_unlock(&lru_pvecs.lock);
704 void lru_add_drain(void)
706 local_lock(&lru_pvecs.lock);
707 lru_add_drain_cpu(smp_processor_id());
708 local_unlock(&lru_pvecs.lock);
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();
725 void lru_add_drain_cpu_zone(struct zone *zone)
727 local_lock(&lru_pvecs.lock);
728 lru_add_drain_cpu(smp_processor_id());
729 drain_local_pages(zone);
730 local_unlock(&lru_pvecs.lock);
735 static DEFINE_PER_CPU(struct work_struct, lru_add_drain_work);
737 static void lru_add_drain_per_cpu(struct work_struct *dummy)
739 lru_add_and_bh_lrus_drain();
743 * Doesn't need any cpu hotplug locking because we do rely on per-cpu
744 * kworkers being shut down before our page_alloc_cpu_dead callback is
745 * executed on the offlined cpu.
746 * Calling this function with cpu hotplug locks held can actually lead
747 * to obscure indirect dependencies via WQ context.
749 inline void __lru_add_drain_all(bool force_all_cpus)
752 * lru_drain_gen - Global pages generation number
754 * (A) Definition: global lru_drain_gen = x implies that all generations
755 * 0 < n <= x are already *scheduled* for draining.
757 * This is an optimization for the highly-contended use case where a
758 * user space workload keeps constantly generating a flow of pages for
761 static unsigned int lru_drain_gen;
762 static struct cpumask has_work;
763 static DEFINE_MUTEX(lock);
764 unsigned cpu, this_gen;
767 * Make sure nobody triggers this path before mm_percpu_wq is fully
770 if (WARN_ON(!mm_percpu_wq))
774 * Guarantee pagevec counter stores visible by this CPU are visible to
775 * other CPUs before loading the current drain generation.
780 * (B) Locally cache global LRU draining generation number
782 * The read barrier ensures that the counter is loaded before the mutex
783 * is taken. It pairs with smp_mb() inside the mutex critical section
786 this_gen = smp_load_acquire(&lru_drain_gen);
791 * (C) Exit the draining operation if a newer generation, from another
792 * lru_add_drain_all(), was already scheduled for draining. Check (A).
794 if (unlikely(this_gen != lru_drain_gen && !force_all_cpus))
798 * (D) Increment global generation number
800 * Pairs with smp_load_acquire() at (B), outside of the critical
801 * section. Use a full memory barrier to guarantee that the new global
802 * drain generation number is stored before loading pagevec counters.
804 * This pairing must be done here, before the for_each_online_cpu loop
805 * below which drains the page vectors.
807 * Let x, y, and z represent some system CPU numbers, where x < y < z.
808 * Assume CPU #z is in the middle of the for_each_online_cpu loop
809 * below and has already reached CPU #y's per-cpu data. CPU #x comes
810 * along, adds some pages to its per-cpu vectors, then calls
811 * lru_add_drain_all().
813 * If the paired barrier is done at any later step, e.g. after the
814 * loop, CPU #x will just exit at (C) and miss flushing out all of its
817 WRITE_ONCE(lru_drain_gen, lru_drain_gen + 1);
820 cpumask_clear(&has_work);
821 for_each_online_cpu(cpu) {
822 struct work_struct *work = &per_cpu(lru_add_drain_work, cpu);
824 if (force_all_cpus ||
825 pagevec_count(&per_cpu(lru_pvecs.lru_add, cpu)) ||
826 data_race(pagevec_count(&per_cpu(lru_rotate.pvec, cpu))) ||
827 pagevec_count(&per_cpu(lru_pvecs.lru_deactivate_file, cpu)) ||
828 pagevec_count(&per_cpu(lru_pvecs.lru_deactivate, cpu)) ||
829 pagevec_count(&per_cpu(lru_pvecs.lru_lazyfree, cpu)) ||
830 need_activate_page_drain(cpu) ||
831 has_bh_in_lru(cpu, NULL)) {
832 INIT_WORK(work, lru_add_drain_per_cpu);
833 queue_work_on(cpu, mm_percpu_wq, work);
834 __cpumask_set_cpu(cpu, &has_work);
838 for_each_cpu(cpu, &has_work)
839 flush_work(&per_cpu(lru_add_drain_work, cpu));
845 void lru_add_drain_all(void)
847 __lru_add_drain_all(false);
850 void lru_add_drain_all(void)
854 #endif /* CONFIG_SMP */
856 atomic_t lru_disable_count = ATOMIC_INIT(0);
859 * lru_cache_disable() needs to be called before we start compiling
860 * a list of pages to be migrated using isolate_lru_page().
861 * It drains pages on LRU cache and then disable on all cpus until
862 * lru_cache_enable is called.
864 * Must be paired with a call to lru_cache_enable().
866 void lru_cache_disable(void)
868 atomic_inc(&lru_disable_count);
871 * lru_add_drain_all in the force mode will schedule draining on
872 * all online CPUs so any calls of lru_cache_disabled wrapped by
873 * local_lock or preemption disabled would be ordered by that.
874 * The atomic operation doesn't need to have stronger ordering
875 * requirements because that is enforeced by the scheduling
878 __lru_add_drain_all(true);
880 lru_add_and_bh_lrus_drain();
885 * release_pages - batched put_page()
886 * @pages: array of pages to release
887 * @nr: number of pages
889 * Decrement the reference count on all the pages in @pages. If it
890 * fell to zero, remove the page from the LRU and free it.
892 void release_pages(struct page **pages, int nr)
895 LIST_HEAD(pages_to_free);
896 struct lruvec *lruvec = NULL;
897 unsigned long flags = 0;
898 unsigned int lock_batch;
900 for (i = 0; i < nr; i++) {
901 struct page *page = pages[i];
902 struct folio *folio = page_folio(page);
905 * Make sure the IRQ-safe lock-holding time does not get
906 * excessive with a continuous string of pages from the
907 * same lruvec. The lock is held only if lruvec != NULL.
909 if (lruvec && ++lock_batch == SWAP_CLUSTER_MAX) {
910 unlock_page_lruvec_irqrestore(lruvec, flags);
915 if (is_huge_zero_page(page))
918 if (is_zone_device_page(page)) {
920 unlock_page_lruvec_irqrestore(lruvec, flags);
924 * ZONE_DEVICE pages that return 'false' from
925 * page_is_devmap_managed() do not require special
926 * processing, and instead, expect a call to
927 * put_page_testzero().
929 if (page_is_devmap_managed(page)) {
930 put_devmap_managed_page(page);
933 if (put_page_testzero(page))
934 put_dev_pagemap(page->pgmap);
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);
962 __ClearPageWaiters(page);
964 list_add(&page->lru, &pages_to_free);
967 unlock_page_lruvec_irqrestore(lruvec, flags);
969 mem_cgroup_uncharge_list(&pages_to_free);
970 free_unref_page_list(&pages_to_free);
972 EXPORT_SYMBOL(release_pages);
975 * The pages which we're about to release may be in the deferred lru-addition
976 * queues. That would prevent them from really being freed right now. That's
977 * OK from a correctness point of view but is inefficient - those pages may be
978 * cache-warm and we want to give them back to the page allocator ASAP.
980 * So __pagevec_release() will drain those queues here. __pagevec_lru_add()
981 * and __pagevec_lru_add_active() call release_pages() directly to avoid
984 void __pagevec_release(struct pagevec *pvec)
986 if (!pvec->percpu_pvec_drained) {
988 pvec->percpu_pvec_drained = true;
990 release_pages(pvec->pages, pagevec_count(pvec));
991 pagevec_reinit(pvec);
993 EXPORT_SYMBOL(__pagevec_release);
995 static void __pagevec_lru_add_fn(struct folio *folio, struct lruvec *lruvec)
997 int was_unevictable = folio_test_clear_unevictable(folio);
998 long nr_pages = folio_nr_pages(folio);
1000 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
1003 * A folio becomes evictable in two ways:
1004 * 1) Within LRU lock [munlock_vma_page() and __munlock_pagevec()].
1005 * 2) Before acquiring LRU lock to put the folio on the correct LRU
1007 * a) do PageLRU check with lock [check_move_unevictable_pages]
1008 * b) do PageLRU check before lock [clear_page_mlock]
1010 * (1) & (2a) are ok as LRU lock will serialize them. For (2b), we need
1011 * following strict ordering:
1013 * #0: __pagevec_lru_add_fn #1: clear_page_mlock
1015 * folio_set_lru() folio_test_clear_mlocked()
1016 * smp_mb() // explicit ordering // above provides strict
1018 * folio_test_mlocked() folio_test_lru()
1021 * if '#1' does not observe setting of PG_lru by '#0' and
1022 * fails isolation, the explicit barrier will make sure that
1023 * folio_evictable check will put the folio on the correct
1024 * LRU. Without smp_mb(), folio_set_lru() can be reordered
1025 * after folio_test_mlocked() check and can make '#1' fail the
1026 * isolation of the folio whose mlocked bit is cleared (#0 is
1027 * also looking at the same folio) and the evictable folio will
1028 * be stranded on an unevictable LRU.
1030 folio_set_lru(folio);
1031 smp_mb__after_atomic();
1033 if (folio_evictable(folio)) {
1034 if (was_unevictable)
1035 __count_vm_events(UNEVICTABLE_PGRESCUED, nr_pages);
1037 folio_clear_active(folio);
1038 folio_set_unevictable(folio);
1039 if (!was_unevictable)
1040 __count_vm_events(UNEVICTABLE_PGCULLED, nr_pages);
1043 lruvec_add_folio(lruvec, folio);
1044 trace_mm_lru_insertion(folio);
1048 * Add the passed pages to the LRU, then drop the caller's refcount
1049 * on them. Reinitialises the caller's pagevec.
1051 void __pagevec_lru_add(struct pagevec *pvec)
1054 struct lruvec *lruvec = NULL;
1055 unsigned long flags = 0;
1057 for (i = 0; i < pagevec_count(pvec); i++) {
1058 struct folio *folio = page_folio(pvec->pages[i]);
1060 lruvec = folio_lruvec_relock_irqsave(folio, lruvec, &flags);
1061 __pagevec_lru_add_fn(folio, lruvec);
1064 unlock_page_lruvec_irqrestore(lruvec, flags);
1065 release_pages(pvec->pages, pvec->nr);
1066 pagevec_reinit(pvec);
1070 * pagevec_remove_exceptionals - pagevec exceptionals pruning
1071 * @pvec: The pagevec to prune
1073 * find_get_entries() fills both pages and XArray value entries (aka
1074 * exceptional entries) into the pagevec. This function prunes all
1075 * exceptionals from @pvec without leaving holes, so that it can be
1076 * passed on to page-only pagevec operations.
1078 void pagevec_remove_exceptionals(struct pagevec *pvec)
1082 for (i = 0, j = 0; i < pagevec_count(pvec); i++) {
1083 struct page *page = pvec->pages[i];
1084 if (!xa_is_value(page))
1085 pvec->pages[j++] = page;
1091 * pagevec_lookup_range - gang pagecache lookup
1092 * @pvec: Where the resulting pages are placed
1093 * @mapping: The address_space to search
1094 * @start: The starting page index
1095 * @end: The final page index
1097 * pagevec_lookup_range() will search for & return a group of up to PAGEVEC_SIZE
1098 * pages in the mapping starting from index @start and upto index @end
1099 * (inclusive). The pages are placed in @pvec. pagevec_lookup() takes a
1100 * reference against the pages in @pvec.
1102 * The search returns a group of mapping-contiguous pages with ascending
1103 * indexes. There may be holes in the indices due to not-present pages. We
1104 * also update @start to index the next page for the traversal.
1106 * pagevec_lookup_range() returns the number of pages which were found. If this
1107 * number is smaller than PAGEVEC_SIZE, the end of specified range has been
1110 unsigned pagevec_lookup_range(struct pagevec *pvec,
1111 struct address_space *mapping, pgoff_t *start, pgoff_t end)
1113 pvec->nr = find_get_pages_range(mapping, start, end, PAGEVEC_SIZE,
1115 return pagevec_count(pvec);
1117 EXPORT_SYMBOL(pagevec_lookup_range);
1119 unsigned pagevec_lookup_range_tag(struct pagevec *pvec,
1120 struct address_space *mapping, pgoff_t *index, pgoff_t end,
1123 pvec->nr = find_get_pages_range_tag(mapping, index, end, tag,
1124 PAGEVEC_SIZE, pvec->pages);
1125 return pagevec_count(pvec);
1127 EXPORT_SYMBOL(pagevec_lookup_range_tag);
1130 * Perform any setup for the swap system
1132 void __init swap_setup(void)
1134 unsigned long megs = totalram_pages() >> (20 - PAGE_SHIFT);
1136 /* Use a smaller cluster for small-memory machines */
1142 * Right now other parts of the system means that we
1143 * _really_ don't want to cluster much more
1147 #ifdef CONFIG_DEV_PAGEMAP_OPS
1148 void put_devmap_managed_page(struct page *page)
1152 if (WARN_ON_ONCE(!page_is_devmap_managed(page)))
1155 count = page_ref_dec_return(page);
1158 * devmap page refcounts are 1-based, rather than 0-based: if
1159 * refcount is 1, then the page is free and the refcount is
1160 * stable because nobody holds a reference on the page.
1163 free_devmap_managed_page(page);
1167 EXPORT_SYMBOL(put_devmap_managed_page);