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? As a power of 2 */
48 const int page_cluster_max = 31;
50 /* Protecting only lru_rotate.fbatch which requires disabling interrupts */
53 struct folio_batch fbatch;
55 static DEFINE_PER_CPU(struct lru_rotate, lru_rotate) = {
56 .lock = INIT_LOCAL_LOCK(lock),
60 * The following folio batches are grouped together because they are protected
61 * by disabling preemption (and interrupts remain enabled).
65 struct folio_batch lru_add;
66 struct folio_batch lru_deactivate_file;
67 struct folio_batch lru_deactivate;
68 struct folio_batch lru_lazyfree;
70 struct folio_batch activate;
73 static DEFINE_PER_CPU(struct cpu_fbatches, cpu_fbatches) = {
74 .lock = INIT_LOCAL_LOCK(lock),
78 * This path almost never happens for VM activity - pages are normally freed
79 * via pagevecs. But it gets used by networking - and for compound pages.
81 static void __page_cache_release(struct folio *folio)
83 if (folio_test_lru(folio)) {
84 struct lruvec *lruvec;
87 lruvec = folio_lruvec_lock_irqsave(folio, &flags);
88 lruvec_del_folio(lruvec, folio);
89 __folio_clear_lru_flags(folio);
90 unlock_page_lruvec_irqrestore(lruvec, flags);
92 /* See comment on folio_test_mlocked in release_pages() */
93 if (unlikely(folio_test_mlocked(folio))) {
94 long nr_pages = folio_nr_pages(folio);
96 __folio_clear_mlocked(folio);
97 zone_stat_mod_folio(folio, NR_MLOCK, -nr_pages);
98 count_vm_events(UNEVICTABLE_PGCLEARED, nr_pages);
102 static void __folio_put_small(struct folio *folio)
104 __page_cache_release(folio);
105 mem_cgroup_uncharge(folio);
106 free_unref_page(&folio->page, 0);
109 static void __folio_put_large(struct folio *folio)
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.)
117 if (!folio_test_hugetlb(folio))
118 __page_cache_release(folio);
119 destroy_large_folio(folio);
122 void __folio_put(struct folio *folio)
124 if (unlikely(folio_is_zone_device(folio)))
125 free_zone_device_page(&folio->page);
126 else if (unlikely(folio_test_large(folio)))
127 __folio_put_large(folio);
129 __folio_put_small(folio);
131 EXPORT_SYMBOL(__folio_put);
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 folio *folio, *next;
143 list_for_each_entry_safe(folio, next, pages, lru) {
144 if (!folio_put_testzero(folio)) {
145 list_del(&folio->lru);
148 if (folio_test_large(folio)) {
149 list_del(&folio->lru);
150 __folio_put_large(folio);
153 /* LRU flag must be clear because it's passed 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 typedef void (*move_fn_t)(struct lruvec *lruvec, struct folio *folio);
193 static void lru_add_fn(struct lruvec *lruvec, struct folio *folio)
195 int was_unevictable = folio_test_clear_unevictable(folio);
196 long nr_pages = folio_nr_pages(folio);
198 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
201 * Is an smp_mb__after_atomic() still required here, before
202 * folio_evictable() tests the mlocked flag, to rule out the possibility
203 * of stranding an evictable folio on an unevictable LRU? I think
204 * not, because __munlock_page() only clears the mlocked flag
205 * while the LRU lock is held.
207 * (That is not true of __page_cache_release(), and not necessarily
208 * true of release_pages(): but those only clear the mlocked flag after
209 * folio_put_testzero() has excluded any other users of the folio.)
211 if (folio_evictable(folio)) {
213 __count_vm_events(UNEVICTABLE_PGRESCUED, nr_pages);
215 folio_clear_active(folio);
216 folio_set_unevictable(folio);
218 * folio->mlock_count = !!folio_test_mlocked(folio)?
219 * But that leaves __mlock_page() in doubt whether another
220 * actor has already counted the mlock or not. Err on the
221 * safe side, underestimate, let page reclaim fix it, rather
222 * than leaving a page on the unevictable LRU indefinitely.
224 folio->mlock_count = 0;
225 if (!was_unevictable)
226 __count_vm_events(UNEVICTABLE_PGCULLED, nr_pages);
229 lruvec_add_folio(lruvec, folio);
230 trace_mm_lru_insertion(folio);
233 static void folio_batch_move_lru(struct folio_batch *fbatch, move_fn_t move_fn)
236 struct lruvec *lruvec = NULL;
237 unsigned long flags = 0;
239 for (i = 0; i < folio_batch_count(fbatch); i++) {
240 struct folio *folio = fbatch->folios[i];
242 /* block memcg migration while the folio moves between lru */
243 if (move_fn != lru_add_fn && !folio_test_clear_lru(folio))
246 lruvec = folio_lruvec_relock_irqsave(folio, lruvec, &flags);
247 move_fn(lruvec, folio);
249 folio_set_lru(folio);
253 unlock_page_lruvec_irqrestore(lruvec, flags);
254 folios_put(fbatch->folios, folio_batch_count(fbatch));
255 folio_batch_init(fbatch);
258 static void folio_batch_add_and_move(struct folio_batch *fbatch,
259 struct folio *folio, move_fn_t move_fn)
261 if (folio_batch_add(fbatch, folio) && !folio_test_large(folio) &&
262 !lru_cache_disabled())
264 folio_batch_move_lru(fbatch, move_fn);
267 static void lru_move_tail_fn(struct lruvec *lruvec, struct folio *folio)
269 if (!folio_test_unevictable(folio)) {
270 lruvec_del_folio(lruvec, folio);
271 folio_clear_active(folio);
272 lruvec_add_folio_tail(lruvec, folio);
273 __count_vm_events(PGROTATED, folio_nr_pages(folio));
278 * Writeback is about to end against a folio which has been marked for
279 * immediate reclaim. If it still appears to be reclaimable, move it
280 * to the tail of the inactive list.
282 * folio_rotate_reclaimable() must disable IRQs, to prevent nasty races.
284 void folio_rotate_reclaimable(struct folio *folio)
286 if (!folio_test_locked(folio) && !folio_test_dirty(folio) &&
287 !folio_test_unevictable(folio) && folio_test_lru(folio)) {
288 struct folio_batch *fbatch;
292 local_lock_irqsave(&lru_rotate.lock, flags);
293 fbatch = this_cpu_ptr(&lru_rotate.fbatch);
294 folio_batch_add_and_move(fbatch, folio, lru_move_tail_fn);
295 local_unlock_irqrestore(&lru_rotate.lock, flags);
299 void lru_note_cost(struct lruvec *lruvec, bool file,
300 unsigned int nr_io, unsigned int nr_rotated)
305 * Reflect the relative cost of incurring IO and spending CPU
306 * time on rotations. This doesn't attempt to make a precise
307 * comparison, it just says: if reloads are about comparable
308 * between the LRU lists, or rotations are overwhelmingly
309 * different between them, adjust scan balance for CPU work.
311 cost = nr_io * SWAP_CLUSTER_MAX + nr_rotated;
314 unsigned long lrusize;
317 * Hold lruvec->lru_lock is safe here, since
318 * 1) The pinned lruvec in reclaim, or
319 * 2) From a pre-LRU page during refault (which also holds the
320 * rcu lock, so would be safe even if the page was on the LRU
321 * and could move simultaneously to a new lruvec).
323 spin_lock_irq(&lruvec->lru_lock);
324 /* Record cost event */
326 lruvec->file_cost += cost;
328 lruvec->anon_cost += cost;
331 * Decay previous events
333 * Because workloads change over time (and to avoid
334 * overflow) we keep these statistics as a floating
335 * average, which ends up weighing recent refaults
336 * more than old ones.
338 lrusize = lruvec_page_state(lruvec, NR_INACTIVE_ANON) +
339 lruvec_page_state(lruvec, NR_ACTIVE_ANON) +
340 lruvec_page_state(lruvec, NR_INACTIVE_FILE) +
341 lruvec_page_state(lruvec, NR_ACTIVE_FILE);
343 if (lruvec->file_cost + lruvec->anon_cost > lrusize / 4) {
344 lruvec->file_cost /= 2;
345 lruvec->anon_cost /= 2;
347 spin_unlock_irq(&lruvec->lru_lock);
348 } while ((lruvec = parent_lruvec(lruvec)));
351 void lru_note_cost_refault(struct folio *folio)
353 lru_note_cost(folio_lruvec(folio), folio_is_file_lru(folio),
354 folio_nr_pages(folio), 0);
357 static void folio_activate_fn(struct lruvec *lruvec, struct folio *folio)
359 if (!folio_test_active(folio) && !folio_test_unevictable(folio)) {
360 long nr_pages = folio_nr_pages(folio);
362 lruvec_del_folio(lruvec, folio);
363 folio_set_active(folio);
364 lruvec_add_folio(lruvec, folio);
365 trace_mm_lru_activate(folio);
367 __count_vm_events(PGACTIVATE, nr_pages);
368 __count_memcg_events(lruvec_memcg(lruvec), PGACTIVATE,
374 static void folio_activate_drain(int cpu)
376 struct folio_batch *fbatch = &per_cpu(cpu_fbatches.activate, cpu);
378 if (folio_batch_count(fbatch))
379 folio_batch_move_lru(fbatch, folio_activate_fn);
382 void folio_activate(struct folio *folio)
384 if (folio_test_lru(folio) && !folio_test_active(folio) &&
385 !folio_test_unevictable(folio)) {
386 struct folio_batch *fbatch;
389 local_lock(&cpu_fbatches.lock);
390 fbatch = this_cpu_ptr(&cpu_fbatches.activate);
391 folio_batch_add_and_move(fbatch, folio, folio_activate_fn);
392 local_unlock(&cpu_fbatches.lock);
397 static inline void folio_activate_drain(int cpu)
401 void folio_activate(struct folio *folio)
403 struct lruvec *lruvec;
405 if (folio_test_clear_lru(folio)) {
406 lruvec = folio_lruvec_lock_irq(folio);
407 folio_activate_fn(lruvec, folio);
408 unlock_page_lruvec_irq(lruvec);
409 folio_set_lru(folio);
414 static void __lru_cache_activate_folio(struct folio *folio)
416 struct folio_batch *fbatch;
419 local_lock(&cpu_fbatches.lock);
420 fbatch = this_cpu_ptr(&cpu_fbatches.lru_add);
423 * Search backwards on the optimistic assumption that the folio being
424 * activated has just been added to this batch. Note that only
425 * the local batch is examined as a !LRU folio could be in the
426 * process of being released, reclaimed, migrated or on a remote
427 * batch that is currently being drained. Furthermore, marking
428 * a remote batch's folio active potentially hits a race where
429 * a folio is marked active just after it is added to the inactive
430 * list causing accounting errors and BUG_ON checks to trigger.
432 for (i = folio_batch_count(fbatch) - 1; i >= 0; i--) {
433 struct folio *batch_folio = fbatch->folios[i];
435 if (batch_folio == folio) {
436 folio_set_active(folio);
441 local_unlock(&cpu_fbatches.lock);
444 #ifdef CONFIG_LRU_GEN
445 static void folio_inc_refs(struct folio *folio)
447 unsigned long new_flags, old_flags = READ_ONCE(folio->flags);
449 if (folio_test_unevictable(folio))
452 if (!folio_test_referenced(folio)) {
453 folio_set_referenced(folio);
457 if (!folio_test_workingset(folio)) {
458 folio_set_workingset(folio);
462 /* see the comment on MAX_NR_TIERS */
464 new_flags = old_flags & LRU_REFS_MASK;
465 if (new_flags == LRU_REFS_MASK)
468 new_flags += BIT(LRU_REFS_PGOFF);
469 new_flags |= old_flags & ~LRU_REFS_MASK;
470 } while (!try_cmpxchg(&folio->flags, &old_flags, new_flags));
473 static void folio_inc_refs(struct folio *folio)
476 #endif /* CONFIG_LRU_GEN */
479 * Mark a page as having seen activity.
481 * inactive,unreferenced -> inactive,referenced
482 * inactive,referenced -> active,unreferenced
483 * active,unreferenced -> active,referenced
485 * When a newly allocated page is not yet visible, so safe for non-atomic ops,
486 * __SetPageReferenced(page) may be substituted for mark_page_accessed(page).
488 void folio_mark_accessed(struct folio *folio)
490 if (lru_gen_enabled()) {
491 folio_inc_refs(folio);
495 if (!folio_test_referenced(folio)) {
496 folio_set_referenced(folio);
497 } else if (folio_test_unevictable(folio)) {
499 * Unevictable pages are on the "LRU_UNEVICTABLE" list. But,
500 * this list is never rotated or maintained, so marking an
501 * unevictable page accessed has no effect.
503 } else if (!folio_test_active(folio)) {
505 * If the folio is on the LRU, queue it for activation via
506 * cpu_fbatches.activate. Otherwise, assume the folio is in a
507 * folio_batch, mark it active and it'll be moved to the active
508 * LRU on the next drain.
510 if (folio_test_lru(folio))
511 folio_activate(folio);
513 __lru_cache_activate_folio(folio);
514 folio_clear_referenced(folio);
515 workingset_activation(folio);
517 if (folio_test_idle(folio))
518 folio_clear_idle(folio);
520 EXPORT_SYMBOL(folio_mark_accessed);
523 * folio_add_lru - Add a folio to an LRU list.
524 * @folio: The folio to be added to the LRU.
526 * Queue the folio for addition to the LRU. The decision on whether
527 * to add the page to the [in]active [file|anon] list is deferred until the
528 * folio_batch is drained. This gives a chance for the caller of folio_add_lru()
529 * have the folio added to the active list using folio_mark_accessed().
531 void folio_add_lru(struct folio *folio)
533 struct folio_batch *fbatch;
535 VM_BUG_ON_FOLIO(folio_test_active(folio) &&
536 folio_test_unevictable(folio), folio);
537 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
539 /* see the comment in lru_gen_add_folio() */
540 if (lru_gen_enabled() && !folio_test_unevictable(folio) &&
541 lru_gen_in_fault() && !(current->flags & PF_MEMALLOC))
542 folio_set_active(folio);
545 local_lock(&cpu_fbatches.lock);
546 fbatch = this_cpu_ptr(&cpu_fbatches.lru_add);
547 folio_batch_add_and_move(fbatch, folio, lru_add_fn);
548 local_unlock(&cpu_fbatches.lock);
550 EXPORT_SYMBOL(folio_add_lru);
553 * folio_add_lru_vma() - Add a folio to the appropate LRU list for this VMA.
554 * @folio: The folio to be added to the LRU.
555 * @vma: VMA in which the folio is mapped.
557 * If the VMA is mlocked, @folio is added to the unevictable list.
558 * Otherwise, it is treated the same way as folio_add_lru().
560 void folio_add_lru_vma(struct folio *folio, struct vm_area_struct *vma)
562 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
564 if (unlikely((vma->vm_flags & (VM_LOCKED | VM_SPECIAL)) == VM_LOCKED))
565 mlock_new_page(&folio->page);
567 folio_add_lru(folio);
571 * If the folio cannot be invalidated, it is moved to the
572 * inactive list to speed up its reclaim. It is moved to the
573 * head of the list, rather than the tail, to give the flusher
574 * threads some time to write it out, as this is much more
575 * effective than the single-page writeout from reclaim.
577 * If the folio isn't mapped and dirty/writeback, the folio
578 * could be reclaimed asap using the reclaim flag.
580 * 1. active, mapped folio -> none
581 * 2. active, dirty/writeback folio -> inactive, head, reclaim
582 * 3. inactive, mapped folio -> none
583 * 4. inactive, dirty/writeback folio -> inactive, head, reclaim
584 * 5. inactive, clean -> inactive, tail
587 * In 4, it moves to the head of the inactive list so the folio is
588 * written out by flusher threads as this is much more efficient
589 * than the single-page writeout from reclaim.
591 static void lru_deactivate_file_fn(struct lruvec *lruvec, struct folio *folio)
593 bool active = folio_test_active(folio);
594 long nr_pages = folio_nr_pages(folio);
596 if (folio_test_unevictable(folio))
599 /* Some processes are using the folio */
600 if (folio_mapped(folio))
603 lruvec_del_folio(lruvec, folio);
604 folio_clear_active(folio);
605 folio_clear_referenced(folio);
607 if (folio_test_writeback(folio) || folio_test_dirty(folio)) {
609 * Setting the reclaim flag could race with
610 * folio_end_writeback() and confuse readahead. But the
611 * race window is _really_ small and it's not a critical
614 lruvec_add_folio(lruvec, folio);
615 folio_set_reclaim(folio);
618 * The folio's writeback ended while it was in the batch.
619 * We move that folio to the tail of the inactive list.
621 lruvec_add_folio_tail(lruvec, folio);
622 __count_vm_events(PGROTATED, nr_pages);
626 __count_vm_events(PGDEACTIVATE, nr_pages);
627 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
632 static void lru_deactivate_fn(struct lruvec *lruvec, struct folio *folio)
634 if (!folio_test_unevictable(folio) && (folio_test_active(folio) || lru_gen_enabled())) {
635 long nr_pages = folio_nr_pages(folio);
637 lruvec_del_folio(lruvec, folio);
638 folio_clear_active(folio);
639 folio_clear_referenced(folio);
640 lruvec_add_folio(lruvec, folio);
642 __count_vm_events(PGDEACTIVATE, nr_pages);
643 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
648 static void lru_lazyfree_fn(struct lruvec *lruvec, struct folio *folio)
650 if (folio_test_anon(folio) && folio_test_swapbacked(folio) &&
651 !folio_test_swapcache(folio) && !folio_test_unevictable(folio)) {
652 long nr_pages = folio_nr_pages(folio);
654 lruvec_del_folio(lruvec, folio);
655 folio_clear_active(folio);
656 folio_clear_referenced(folio);
658 * Lazyfree folios are clean anonymous folios. They have
659 * the swapbacked flag cleared, to distinguish them from normal
662 folio_clear_swapbacked(folio);
663 lruvec_add_folio(lruvec, folio);
665 __count_vm_events(PGLAZYFREE, nr_pages);
666 __count_memcg_events(lruvec_memcg(lruvec), PGLAZYFREE,
672 * Drain pages out of the cpu's folio_batch.
673 * Either "cpu" is the current CPU, and preemption has already been
674 * disabled; or "cpu" is being hot-unplugged, and is already dead.
676 void lru_add_drain_cpu(int cpu)
678 struct cpu_fbatches *fbatches = &per_cpu(cpu_fbatches, cpu);
679 struct folio_batch *fbatch = &fbatches->lru_add;
681 if (folio_batch_count(fbatch))
682 folio_batch_move_lru(fbatch, lru_add_fn);
684 fbatch = &per_cpu(lru_rotate.fbatch, cpu);
685 /* Disabling interrupts below acts as a compiler barrier. */
686 if (data_race(folio_batch_count(fbatch))) {
689 /* No harm done if a racing interrupt already did this */
690 local_lock_irqsave(&lru_rotate.lock, flags);
691 folio_batch_move_lru(fbatch, lru_move_tail_fn);
692 local_unlock_irqrestore(&lru_rotate.lock, flags);
695 fbatch = &fbatches->lru_deactivate_file;
696 if (folio_batch_count(fbatch))
697 folio_batch_move_lru(fbatch, lru_deactivate_file_fn);
699 fbatch = &fbatches->lru_deactivate;
700 if (folio_batch_count(fbatch))
701 folio_batch_move_lru(fbatch, lru_deactivate_fn);
703 fbatch = &fbatches->lru_lazyfree;
704 if (folio_batch_count(fbatch))
705 folio_batch_move_lru(fbatch, lru_lazyfree_fn);
707 folio_activate_drain(cpu);
711 * deactivate_file_folio() - Deactivate a file folio.
712 * @folio: Folio to deactivate.
714 * This function hints to the VM that @folio is a good reclaim candidate,
715 * for example if its invalidation fails due to the folio being dirty
716 * or under writeback.
718 * Context: Caller holds a reference on the folio.
720 void deactivate_file_folio(struct folio *folio)
722 struct folio_batch *fbatch;
724 /* Deactivating an unevictable folio will not accelerate reclaim */
725 if (folio_test_unevictable(folio))
729 local_lock(&cpu_fbatches.lock);
730 fbatch = this_cpu_ptr(&cpu_fbatches.lru_deactivate_file);
731 folio_batch_add_and_move(fbatch, folio, lru_deactivate_file_fn);
732 local_unlock(&cpu_fbatches.lock);
736 * deactivate_page - deactivate a page
737 * @page: page to deactivate
739 * deactivate_page() moves @page to the inactive list if @page was on the active
740 * list and was not an unevictable page. This is done to accelerate the reclaim
743 void deactivate_page(struct page *page)
745 struct folio *folio = page_folio(page);
747 if (folio_test_lru(folio) && !folio_test_unevictable(folio) &&
748 (folio_test_active(folio) || lru_gen_enabled())) {
749 struct folio_batch *fbatch;
752 local_lock(&cpu_fbatches.lock);
753 fbatch = this_cpu_ptr(&cpu_fbatches.lru_deactivate);
754 folio_batch_add_and_move(fbatch, folio, lru_deactivate_fn);
755 local_unlock(&cpu_fbatches.lock);
760 * mark_page_lazyfree - make an anon page lazyfree
761 * @page: page to deactivate
763 * mark_page_lazyfree() moves @page to the inactive file list.
764 * This is done to accelerate the reclaim of @page.
766 void mark_page_lazyfree(struct page *page)
768 struct folio *folio = page_folio(page);
770 if (folio_test_lru(folio) && folio_test_anon(folio) &&
771 folio_test_swapbacked(folio) && !folio_test_swapcache(folio) &&
772 !folio_test_unevictable(folio)) {
773 struct folio_batch *fbatch;
776 local_lock(&cpu_fbatches.lock);
777 fbatch = this_cpu_ptr(&cpu_fbatches.lru_lazyfree);
778 folio_batch_add_and_move(fbatch, folio, lru_lazyfree_fn);
779 local_unlock(&cpu_fbatches.lock);
783 void lru_add_drain(void)
785 local_lock(&cpu_fbatches.lock);
786 lru_add_drain_cpu(smp_processor_id());
787 local_unlock(&cpu_fbatches.lock);
788 mlock_page_drain_local();
792 * It's called from per-cpu workqueue context in SMP case so
793 * lru_add_drain_cpu and invalidate_bh_lrus_cpu should run on
794 * the same cpu. It shouldn't be a problem in !SMP case since
795 * the core is only one and the locks will disable preemption.
797 static void lru_add_and_bh_lrus_drain(void)
799 local_lock(&cpu_fbatches.lock);
800 lru_add_drain_cpu(smp_processor_id());
801 local_unlock(&cpu_fbatches.lock);
802 invalidate_bh_lrus_cpu();
803 mlock_page_drain_local();
806 void lru_add_drain_cpu_zone(struct zone *zone)
808 local_lock(&cpu_fbatches.lock);
809 lru_add_drain_cpu(smp_processor_id());
810 drain_local_pages(zone);
811 local_unlock(&cpu_fbatches.lock);
812 mlock_page_drain_local();
817 static DEFINE_PER_CPU(struct work_struct, lru_add_drain_work);
819 static void lru_add_drain_per_cpu(struct work_struct *dummy)
821 lru_add_and_bh_lrus_drain();
824 static bool cpu_needs_drain(unsigned int cpu)
826 struct cpu_fbatches *fbatches = &per_cpu(cpu_fbatches, cpu);
828 /* Check these in order of likelihood that they're not zero */
829 return folio_batch_count(&fbatches->lru_add) ||
830 data_race(folio_batch_count(&per_cpu(lru_rotate.fbatch, cpu))) ||
831 folio_batch_count(&fbatches->lru_deactivate_file) ||
832 folio_batch_count(&fbatches->lru_deactivate) ||
833 folio_batch_count(&fbatches->lru_lazyfree) ||
834 folio_batch_count(&fbatches->activate) ||
835 need_mlock_page_drain(cpu) ||
836 has_bh_in_lru(cpu, NULL);
840 * Doesn't need any cpu hotplug locking because we do rely on per-cpu
841 * kworkers being shut down before our page_alloc_cpu_dead callback is
842 * executed on the offlined cpu.
843 * Calling this function with cpu hotplug locks held can actually lead
844 * to obscure indirect dependencies via WQ context.
846 static inline void __lru_add_drain_all(bool force_all_cpus)
849 * lru_drain_gen - Global pages generation number
851 * (A) Definition: global lru_drain_gen = x implies that all generations
852 * 0 < n <= x are already *scheduled* for draining.
854 * This is an optimization for the highly-contended use case where a
855 * user space workload keeps constantly generating a flow of pages for
858 static unsigned int lru_drain_gen;
859 static struct cpumask has_work;
860 static DEFINE_MUTEX(lock);
861 unsigned cpu, this_gen;
864 * Make sure nobody triggers this path before mm_percpu_wq is fully
867 if (WARN_ON(!mm_percpu_wq))
871 * Guarantee folio_batch counter stores visible by this CPU
872 * are visible to other CPUs before loading the current drain
878 * (B) Locally cache global LRU draining generation number
880 * The read barrier ensures that the counter is loaded before the mutex
881 * is taken. It pairs with smp_mb() inside the mutex critical section
884 this_gen = smp_load_acquire(&lru_drain_gen);
889 * (C) Exit the draining operation if a newer generation, from another
890 * lru_add_drain_all(), was already scheduled for draining. Check (A).
892 if (unlikely(this_gen != lru_drain_gen && !force_all_cpus))
896 * (D) Increment global generation number
898 * Pairs with smp_load_acquire() at (B), outside of the critical
899 * section. Use a full memory barrier to guarantee that the
900 * new global drain generation number is stored before loading
901 * folio_batch counters.
903 * This pairing must be done here, before the for_each_online_cpu loop
904 * below which drains the page vectors.
906 * Let x, y, and z represent some system CPU numbers, where x < y < z.
907 * Assume CPU #z is in the middle of the for_each_online_cpu loop
908 * below and has already reached CPU #y's per-cpu data. CPU #x comes
909 * along, adds some pages to its per-cpu vectors, then calls
910 * lru_add_drain_all().
912 * If the paired barrier is done at any later step, e.g. after the
913 * loop, CPU #x will just exit at (C) and miss flushing out all of its
916 WRITE_ONCE(lru_drain_gen, lru_drain_gen + 1);
919 cpumask_clear(&has_work);
920 for_each_online_cpu(cpu) {
921 struct work_struct *work = &per_cpu(lru_add_drain_work, cpu);
923 if (cpu_needs_drain(cpu)) {
924 INIT_WORK(work, lru_add_drain_per_cpu);
925 queue_work_on(cpu, mm_percpu_wq, work);
926 __cpumask_set_cpu(cpu, &has_work);
930 for_each_cpu(cpu, &has_work)
931 flush_work(&per_cpu(lru_add_drain_work, cpu));
937 void lru_add_drain_all(void)
939 __lru_add_drain_all(false);
942 void lru_add_drain_all(void)
946 #endif /* CONFIG_SMP */
948 atomic_t lru_disable_count = ATOMIC_INIT(0);
951 * lru_cache_disable() needs to be called before we start compiling
952 * a list of pages to be migrated using isolate_lru_page().
953 * It drains pages on LRU cache and then disable on all cpus until
954 * lru_cache_enable is called.
956 * Must be paired with a call to lru_cache_enable().
958 void lru_cache_disable(void)
960 atomic_inc(&lru_disable_count);
962 * Readers of lru_disable_count are protected by either disabling
963 * preemption or rcu_read_lock:
965 * preempt_disable, local_irq_disable [bh_lru_lock()]
966 * rcu_read_lock [rt_spin_lock CONFIG_PREEMPT_RT]
967 * preempt_disable [local_lock !CONFIG_PREEMPT_RT]
969 * Since v5.1 kernel, synchronize_rcu() is guaranteed to wait on
970 * preempt_disable() regions of code. So any CPU which sees
971 * lru_disable_count = 0 will have exited the critical
972 * section when synchronize_rcu() returns.
974 synchronize_rcu_expedited();
976 __lru_add_drain_all(true);
978 lru_add_and_bh_lrus_drain();
983 * release_pages - batched put_page()
984 * @arg: array of pages to release
985 * @nr: number of pages
987 * Decrement the reference count on all the pages in @arg. If it
988 * fell to zero, remove the page from the LRU and free it.
990 * Note that the argument can be an array of pages, encoded pages,
991 * or folio pointers. We ignore any encoded bits, and turn any of
992 * them into just a folio that gets free'd.
994 void release_pages(release_pages_arg arg, int nr)
997 struct encoded_page **encoded = arg.encoded_pages;
998 LIST_HEAD(pages_to_free);
999 struct lruvec *lruvec = NULL;
1000 unsigned long flags = 0;
1001 unsigned int lock_batch;
1003 for (i = 0; i < nr; i++) {
1004 struct folio *folio;
1006 /* Turn any of the argument types into a folio */
1007 folio = page_folio(encoded_page_ptr(encoded[i]));
1010 * Make sure the IRQ-safe lock-holding time does not get
1011 * excessive with a continuous string of pages from the
1012 * same lruvec. The lock is held only if lruvec != NULL.
1014 if (lruvec && ++lock_batch == SWAP_CLUSTER_MAX) {
1015 unlock_page_lruvec_irqrestore(lruvec, flags);
1019 if (is_huge_zero_page(&folio->page))
1022 if (folio_is_zone_device(folio)) {
1024 unlock_page_lruvec_irqrestore(lruvec, flags);
1027 if (put_devmap_managed_page(&folio->page))
1029 if (folio_put_testzero(folio))
1030 free_zone_device_page(&folio->page);
1034 if (!folio_put_testzero(folio))
1037 if (folio_test_large(folio)) {
1039 unlock_page_lruvec_irqrestore(lruvec, flags);
1042 __folio_put_large(folio);
1046 if (folio_test_lru(folio)) {
1047 struct lruvec *prev_lruvec = lruvec;
1049 lruvec = folio_lruvec_relock_irqsave(folio, lruvec,
1051 if (prev_lruvec != lruvec)
1054 lruvec_del_folio(lruvec, folio);
1055 __folio_clear_lru_flags(folio);
1059 * In rare cases, when truncation or holepunching raced with
1060 * munlock after VM_LOCKED was cleared, Mlocked may still be
1061 * found set here. This does not indicate a problem, unless
1062 * "unevictable_pgs_cleared" appears worryingly large.
1064 if (unlikely(folio_test_mlocked(folio))) {
1065 __folio_clear_mlocked(folio);
1066 zone_stat_sub_folio(folio, NR_MLOCK);
1067 count_vm_event(UNEVICTABLE_PGCLEARED);
1070 list_add(&folio->lru, &pages_to_free);
1073 unlock_page_lruvec_irqrestore(lruvec, flags);
1075 mem_cgroup_uncharge_list(&pages_to_free);
1076 free_unref_page_list(&pages_to_free);
1078 EXPORT_SYMBOL(release_pages);
1081 * The pages which we're about to release may be in the deferred lru-addition
1082 * queues. That would prevent them from really being freed right now. That's
1083 * OK from a correctness point of view but is inefficient - those pages may be
1084 * cache-warm and we want to give them back to the page allocator ASAP.
1086 * So __pagevec_release() will drain those queues here.
1087 * folio_batch_move_lru() calls folios_put() directly to avoid
1090 void __pagevec_release(struct pagevec *pvec)
1092 if (!pvec->percpu_pvec_drained) {
1094 pvec->percpu_pvec_drained = true;
1096 release_pages(pvec->pages, pagevec_count(pvec));
1097 pagevec_reinit(pvec);
1099 EXPORT_SYMBOL(__pagevec_release);
1102 * folio_batch_remove_exceptionals() - Prune non-folios from a batch.
1103 * @fbatch: The batch to prune
1105 * find_get_entries() fills a batch with both folios and shadow/swap/DAX
1106 * entries. This function prunes all the non-folio entries from @fbatch
1107 * without leaving holes, so that it can be passed on to folio-only batch
1110 void folio_batch_remove_exceptionals(struct folio_batch *fbatch)
1114 for (i = 0, j = 0; i < folio_batch_count(fbatch); i++) {
1115 struct folio *folio = fbatch->folios[i];
1116 if (!xa_is_value(folio))
1117 fbatch->folios[j++] = folio;
1122 unsigned pagevec_lookup_range_tag(struct pagevec *pvec,
1123 struct address_space *mapping, pgoff_t *index, pgoff_t end,
1126 pvec->nr = find_get_pages_range_tag(mapping, index, end, tag,
1127 PAGEVEC_SIZE, pvec->pages);
1128 return pagevec_count(pvec);
1130 EXPORT_SYMBOL(pagevec_lookup_range_tag);
1133 * Perform any setup for the swap system
1135 void __init swap_setup(void)
1137 unsigned long megs = totalram_pages() >> (20 - PAGE_SHIFT);
1139 /* Use a smaller cluster for small-memory machines */
1145 * Right now other parts of the system means that we
1146 * _really_ don't want to cluster much more