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),
77 static void __page_cache_release(struct folio *folio, struct lruvec **lruvecp,
78 unsigned long *flagsp)
80 if (folio_test_lru(folio)) {
81 folio_lruvec_relock_irqsave(folio, lruvecp, flagsp);
82 lruvec_del_folio(*lruvecp, folio);
83 __folio_clear_lru_flags(folio);
87 * In rare cases, when truncation or holepunching raced with
88 * munlock after VM_LOCKED was cleared, Mlocked may still be
89 * found set here. This does not indicate a problem, unless
90 * "unevictable_pgs_cleared" appears worryingly large.
92 if (unlikely(folio_test_mlocked(folio))) {
93 long nr_pages = folio_nr_pages(folio);
95 __folio_clear_mlocked(folio);
96 zone_stat_mod_folio(folio, NR_MLOCK, -nr_pages);
97 count_vm_events(UNEVICTABLE_PGCLEARED, nr_pages);
102 * This path almost never happens for VM activity - pages are normally freed
103 * in batches. But it gets used by networking - and for compound pages.
105 static void page_cache_release(struct folio *folio)
107 struct lruvec *lruvec = NULL;
110 __page_cache_release(folio, &lruvec, &flags);
112 unlock_page_lruvec_irqrestore(lruvec, flags);
115 void __folio_put(struct folio *folio)
117 if (unlikely(folio_is_zone_device(folio))) {
118 free_zone_device_folio(folio);
120 } else if (folio_test_hugetlb(folio)) {
121 free_huge_folio(folio);
125 page_cache_release(folio);
126 if (folio_test_large(folio) && folio_test_large_rmappable(folio))
127 folio_undo_large_rmappable(folio);
128 mem_cgroup_uncharge(folio);
129 free_unref_page(&folio->page, folio_order(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_batch fbatch;
142 struct folio *folio, *next;
144 folio_batch_init(&fbatch);
145 list_for_each_entry_safe(folio, next, pages, lru) {
146 if (!folio_put_testzero(folio))
148 if (folio_test_hugetlb(folio)) {
149 free_huge_folio(folio);
152 /* LRU flag must be clear because it's passed using the lru */
153 if (folio_batch_add(&fbatch, folio) > 0)
155 free_unref_folios(&fbatch);
159 free_unref_folios(&fbatch);
160 INIT_LIST_HEAD(pages);
162 EXPORT_SYMBOL(put_pages_list);
164 typedef void (*move_fn_t)(struct lruvec *lruvec, struct folio *folio);
166 static void lru_add_fn(struct lruvec *lruvec, struct folio *folio)
168 int was_unevictable = folio_test_clear_unevictable(folio);
169 long nr_pages = folio_nr_pages(folio);
171 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
174 * Is an smp_mb__after_atomic() still required here, before
175 * folio_evictable() tests the mlocked flag, to rule out the possibility
176 * of stranding an evictable folio on an unevictable LRU? I think
177 * not, because __munlock_folio() only clears the mlocked flag
178 * while the LRU lock is held.
180 * (That is not true of __page_cache_release(), and not necessarily
181 * true of folios_put(): but those only clear the mlocked flag after
182 * folio_put_testzero() has excluded any other users of the folio.)
184 if (folio_evictable(folio)) {
186 __count_vm_events(UNEVICTABLE_PGRESCUED, nr_pages);
188 folio_clear_active(folio);
189 folio_set_unevictable(folio);
191 * folio->mlock_count = !!folio_test_mlocked(folio)?
192 * But that leaves __mlock_folio() in doubt whether another
193 * actor has already counted the mlock or not. Err on the
194 * safe side, underestimate, let page reclaim fix it, rather
195 * than leaving a page on the unevictable LRU indefinitely.
197 folio->mlock_count = 0;
198 if (!was_unevictable)
199 __count_vm_events(UNEVICTABLE_PGCULLED, nr_pages);
202 lruvec_add_folio(lruvec, folio);
203 trace_mm_lru_insertion(folio);
206 static void folio_batch_move_lru(struct folio_batch *fbatch, move_fn_t move_fn)
209 struct lruvec *lruvec = NULL;
210 unsigned long flags = 0;
212 for (i = 0; i < folio_batch_count(fbatch); i++) {
213 struct folio *folio = fbatch->folios[i];
215 /* block memcg migration while the folio moves between lru */
216 if (move_fn != lru_add_fn && !folio_test_clear_lru(folio))
219 folio_lruvec_relock_irqsave(folio, &lruvec, &flags);
220 move_fn(lruvec, folio);
222 folio_set_lru(folio);
226 unlock_page_lruvec_irqrestore(lruvec, flags);
230 static void folio_batch_add_and_move(struct folio_batch *fbatch,
231 struct folio *folio, move_fn_t move_fn)
233 if (folio_batch_add(fbatch, folio) && !folio_test_large(folio) &&
234 !lru_cache_disabled())
236 folio_batch_move_lru(fbatch, move_fn);
239 static void lru_move_tail_fn(struct lruvec *lruvec, struct folio *folio)
241 if (!folio_test_unevictable(folio)) {
242 lruvec_del_folio(lruvec, folio);
243 folio_clear_active(folio);
244 lruvec_add_folio_tail(lruvec, folio);
245 __count_vm_events(PGROTATED, folio_nr_pages(folio));
250 * Writeback is about to end against a folio which has been marked for
251 * immediate reclaim. If it still appears to be reclaimable, move it
252 * to the tail of the inactive list.
254 * folio_rotate_reclaimable() must disable IRQs, to prevent nasty races.
256 void folio_rotate_reclaimable(struct folio *folio)
258 if (!folio_test_locked(folio) && !folio_test_dirty(folio) &&
259 !folio_test_unevictable(folio) && folio_test_lru(folio)) {
260 struct folio_batch *fbatch;
264 local_lock_irqsave(&lru_rotate.lock, flags);
265 fbatch = this_cpu_ptr(&lru_rotate.fbatch);
266 folio_batch_add_and_move(fbatch, folio, lru_move_tail_fn);
267 local_unlock_irqrestore(&lru_rotate.lock, flags);
271 void lru_note_cost(struct lruvec *lruvec, bool file,
272 unsigned int nr_io, unsigned int nr_rotated)
277 * Reflect the relative cost of incurring IO and spending CPU
278 * time on rotations. This doesn't attempt to make a precise
279 * comparison, it just says: if reloads are about comparable
280 * between the LRU lists, or rotations are overwhelmingly
281 * different between them, adjust scan balance for CPU work.
283 cost = nr_io * SWAP_CLUSTER_MAX + nr_rotated;
286 unsigned long lrusize;
289 * Hold lruvec->lru_lock is safe here, since
290 * 1) The pinned lruvec in reclaim, or
291 * 2) From a pre-LRU page during refault (which also holds the
292 * rcu lock, so would be safe even if the page was on the LRU
293 * and could move simultaneously to a new lruvec).
295 spin_lock_irq(&lruvec->lru_lock);
296 /* Record cost event */
298 lruvec->file_cost += cost;
300 lruvec->anon_cost += cost;
303 * Decay previous events
305 * Because workloads change over time (and to avoid
306 * overflow) we keep these statistics as a floating
307 * average, which ends up weighing recent refaults
308 * more than old ones.
310 lrusize = lruvec_page_state(lruvec, NR_INACTIVE_ANON) +
311 lruvec_page_state(lruvec, NR_ACTIVE_ANON) +
312 lruvec_page_state(lruvec, NR_INACTIVE_FILE) +
313 lruvec_page_state(lruvec, NR_ACTIVE_FILE);
315 if (lruvec->file_cost + lruvec->anon_cost > lrusize / 4) {
316 lruvec->file_cost /= 2;
317 lruvec->anon_cost /= 2;
319 spin_unlock_irq(&lruvec->lru_lock);
320 } while ((lruvec = parent_lruvec(lruvec)));
323 void lru_note_cost_refault(struct folio *folio)
325 lru_note_cost(folio_lruvec(folio), folio_is_file_lru(folio),
326 folio_nr_pages(folio), 0);
329 static void folio_activate_fn(struct lruvec *lruvec, struct folio *folio)
331 if (!folio_test_active(folio) && !folio_test_unevictable(folio)) {
332 long nr_pages = folio_nr_pages(folio);
334 lruvec_del_folio(lruvec, folio);
335 folio_set_active(folio);
336 lruvec_add_folio(lruvec, folio);
337 trace_mm_lru_activate(folio);
339 __count_vm_events(PGACTIVATE, nr_pages);
340 __count_memcg_events(lruvec_memcg(lruvec), PGACTIVATE,
346 static void folio_activate_drain(int cpu)
348 struct folio_batch *fbatch = &per_cpu(cpu_fbatches.activate, cpu);
350 if (folio_batch_count(fbatch))
351 folio_batch_move_lru(fbatch, folio_activate_fn);
354 void folio_activate(struct folio *folio)
356 if (folio_test_lru(folio) && !folio_test_active(folio) &&
357 !folio_test_unevictable(folio)) {
358 struct folio_batch *fbatch;
361 local_lock(&cpu_fbatches.lock);
362 fbatch = this_cpu_ptr(&cpu_fbatches.activate);
363 folio_batch_add_and_move(fbatch, folio, folio_activate_fn);
364 local_unlock(&cpu_fbatches.lock);
369 static inline void folio_activate_drain(int cpu)
373 void folio_activate(struct folio *folio)
375 struct lruvec *lruvec;
377 if (folio_test_clear_lru(folio)) {
378 lruvec = folio_lruvec_lock_irq(folio);
379 folio_activate_fn(lruvec, folio);
380 unlock_page_lruvec_irq(lruvec);
381 folio_set_lru(folio);
386 static void __lru_cache_activate_folio(struct folio *folio)
388 struct folio_batch *fbatch;
391 local_lock(&cpu_fbatches.lock);
392 fbatch = this_cpu_ptr(&cpu_fbatches.lru_add);
395 * Search backwards on the optimistic assumption that the folio being
396 * activated has just been added to this batch. Note that only
397 * the local batch is examined as a !LRU folio could be in the
398 * process of being released, reclaimed, migrated or on a remote
399 * batch that is currently being drained. Furthermore, marking
400 * a remote batch's folio active potentially hits a race where
401 * a folio is marked active just after it is added to the inactive
402 * list causing accounting errors and BUG_ON checks to trigger.
404 for (i = folio_batch_count(fbatch) - 1; i >= 0; i--) {
405 struct folio *batch_folio = fbatch->folios[i];
407 if (batch_folio == folio) {
408 folio_set_active(folio);
413 local_unlock(&cpu_fbatches.lock);
416 #ifdef CONFIG_LRU_GEN
417 static void folio_inc_refs(struct folio *folio)
419 unsigned long new_flags, old_flags = READ_ONCE(folio->flags);
421 if (folio_test_unevictable(folio))
424 if (!folio_test_referenced(folio)) {
425 folio_set_referenced(folio);
429 if (!folio_test_workingset(folio)) {
430 folio_set_workingset(folio);
434 /* see the comment on MAX_NR_TIERS */
436 new_flags = old_flags & LRU_REFS_MASK;
437 if (new_flags == LRU_REFS_MASK)
440 new_flags += BIT(LRU_REFS_PGOFF);
441 new_flags |= old_flags & ~LRU_REFS_MASK;
442 } while (!try_cmpxchg(&folio->flags, &old_flags, new_flags));
445 static void folio_inc_refs(struct folio *folio)
448 #endif /* CONFIG_LRU_GEN */
451 * folio_mark_accessed - Mark a folio as having seen activity.
452 * @folio: The folio to mark.
454 * This function will perform one of the following transitions:
456 * * inactive,unreferenced -> inactive,referenced
457 * * inactive,referenced -> active,unreferenced
458 * * active,unreferenced -> active,referenced
460 * When a newly allocated folio is not yet visible, so safe for non-atomic ops,
461 * __folio_set_referenced() may be substituted for folio_mark_accessed().
463 void folio_mark_accessed(struct folio *folio)
465 if (lru_gen_enabled()) {
466 folio_inc_refs(folio);
470 if (!folio_test_referenced(folio)) {
471 folio_set_referenced(folio);
472 } else if (folio_test_unevictable(folio)) {
474 * Unevictable pages are on the "LRU_UNEVICTABLE" list. But,
475 * this list is never rotated or maintained, so marking an
476 * unevictable page accessed has no effect.
478 } else if (!folio_test_active(folio)) {
480 * If the folio is on the LRU, queue it for activation via
481 * cpu_fbatches.activate. Otherwise, assume the folio is in a
482 * folio_batch, mark it active and it'll be moved to the active
483 * LRU on the next drain.
485 if (folio_test_lru(folio))
486 folio_activate(folio);
488 __lru_cache_activate_folio(folio);
489 folio_clear_referenced(folio);
490 workingset_activation(folio);
492 if (folio_test_idle(folio))
493 folio_clear_idle(folio);
495 EXPORT_SYMBOL(folio_mark_accessed);
498 * folio_add_lru - Add a folio to an LRU list.
499 * @folio: The folio to be added to the LRU.
501 * Queue the folio for addition to the LRU. The decision on whether
502 * to add the page to the [in]active [file|anon] list is deferred until the
503 * folio_batch is drained. This gives a chance for the caller of folio_add_lru()
504 * have the folio added to the active list using folio_mark_accessed().
506 void folio_add_lru(struct folio *folio)
508 struct folio_batch *fbatch;
510 VM_BUG_ON_FOLIO(folio_test_active(folio) &&
511 folio_test_unevictable(folio), folio);
512 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
514 /* see the comment in lru_gen_add_folio() */
515 if (lru_gen_enabled() && !folio_test_unevictable(folio) &&
516 lru_gen_in_fault() && !(current->flags & PF_MEMALLOC))
517 folio_set_active(folio);
520 local_lock(&cpu_fbatches.lock);
521 fbatch = this_cpu_ptr(&cpu_fbatches.lru_add);
522 folio_batch_add_and_move(fbatch, folio, lru_add_fn);
523 local_unlock(&cpu_fbatches.lock);
525 EXPORT_SYMBOL(folio_add_lru);
528 * folio_add_lru_vma() - Add a folio to the appropate LRU list for this VMA.
529 * @folio: The folio to be added to the LRU.
530 * @vma: VMA in which the folio is mapped.
532 * If the VMA is mlocked, @folio is added to the unevictable list.
533 * Otherwise, it is treated the same way as folio_add_lru().
535 void folio_add_lru_vma(struct folio *folio, struct vm_area_struct *vma)
537 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
539 if (unlikely((vma->vm_flags & (VM_LOCKED | VM_SPECIAL)) == VM_LOCKED))
540 mlock_new_folio(folio);
542 folio_add_lru(folio);
546 * If the folio cannot be invalidated, it is moved to the
547 * inactive list to speed up its reclaim. It is moved to the
548 * head of the list, rather than the tail, to give the flusher
549 * threads some time to write it out, as this is much more
550 * effective than the single-page writeout from reclaim.
552 * If the folio isn't mapped and dirty/writeback, the folio
553 * could be reclaimed asap using the reclaim flag.
555 * 1. active, mapped folio -> none
556 * 2. active, dirty/writeback folio -> inactive, head, reclaim
557 * 3. inactive, mapped folio -> none
558 * 4. inactive, dirty/writeback folio -> inactive, head, reclaim
559 * 5. inactive, clean -> inactive, tail
562 * In 4, it moves to the head of the inactive list so the folio is
563 * written out by flusher threads as this is much more efficient
564 * than the single-page writeout from reclaim.
566 static void lru_deactivate_file_fn(struct lruvec *lruvec, struct folio *folio)
568 bool active = folio_test_active(folio);
569 long nr_pages = folio_nr_pages(folio);
571 if (folio_test_unevictable(folio))
574 /* Some processes are using the folio */
575 if (folio_mapped(folio))
578 lruvec_del_folio(lruvec, folio);
579 folio_clear_active(folio);
580 folio_clear_referenced(folio);
582 if (folio_test_writeback(folio) || folio_test_dirty(folio)) {
584 * Setting the reclaim flag could race with
585 * folio_end_writeback() and confuse readahead. But the
586 * race window is _really_ small and it's not a critical
589 lruvec_add_folio(lruvec, folio);
590 folio_set_reclaim(folio);
593 * The folio's writeback ended while it was in the batch.
594 * We move that folio to the tail of the inactive list.
596 lruvec_add_folio_tail(lruvec, folio);
597 __count_vm_events(PGROTATED, nr_pages);
601 __count_vm_events(PGDEACTIVATE, nr_pages);
602 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
607 static void lru_deactivate_fn(struct lruvec *lruvec, struct folio *folio)
609 if (!folio_test_unevictable(folio) && (folio_test_active(folio) || lru_gen_enabled())) {
610 long nr_pages = folio_nr_pages(folio);
612 lruvec_del_folio(lruvec, folio);
613 folio_clear_active(folio);
614 folio_clear_referenced(folio);
615 lruvec_add_folio(lruvec, folio);
617 __count_vm_events(PGDEACTIVATE, nr_pages);
618 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
623 static void lru_lazyfree_fn(struct lruvec *lruvec, struct folio *folio)
625 if (folio_test_anon(folio) && folio_test_swapbacked(folio) &&
626 !folio_test_swapcache(folio) && !folio_test_unevictable(folio)) {
627 long nr_pages = folio_nr_pages(folio);
629 lruvec_del_folio(lruvec, folio);
630 folio_clear_active(folio);
631 folio_clear_referenced(folio);
633 * Lazyfree folios are clean anonymous folios. They have
634 * the swapbacked flag cleared, to distinguish them from normal
637 folio_clear_swapbacked(folio);
638 lruvec_add_folio(lruvec, folio);
640 __count_vm_events(PGLAZYFREE, nr_pages);
641 __count_memcg_events(lruvec_memcg(lruvec), PGLAZYFREE,
647 * Drain pages out of the cpu's folio_batch.
648 * Either "cpu" is the current CPU, and preemption has already been
649 * disabled; or "cpu" is being hot-unplugged, and is already dead.
651 void lru_add_drain_cpu(int cpu)
653 struct cpu_fbatches *fbatches = &per_cpu(cpu_fbatches, cpu);
654 struct folio_batch *fbatch = &fbatches->lru_add;
656 if (folio_batch_count(fbatch))
657 folio_batch_move_lru(fbatch, lru_add_fn);
659 fbatch = &per_cpu(lru_rotate.fbatch, cpu);
660 /* Disabling interrupts below acts as a compiler barrier. */
661 if (data_race(folio_batch_count(fbatch))) {
664 /* No harm done if a racing interrupt already did this */
665 local_lock_irqsave(&lru_rotate.lock, flags);
666 folio_batch_move_lru(fbatch, lru_move_tail_fn);
667 local_unlock_irqrestore(&lru_rotate.lock, flags);
670 fbatch = &fbatches->lru_deactivate_file;
671 if (folio_batch_count(fbatch))
672 folio_batch_move_lru(fbatch, lru_deactivate_file_fn);
674 fbatch = &fbatches->lru_deactivate;
675 if (folio_batch_count(fbatch))
676 folio_batch_move_lru(fbatch, lru_deactivate_fn);
678 fbatch = &fbatches->lru_lazyfree;
679 if (folio_batch_count(fbatch))
680 folio_batch_move_lru(fbatch, lru_lazyfree_fn);
682 folio_activate_drain(cpu);
686 * deactivate_file_folio() - Deactivate a file folio.
687 * @folio: Folio to deactivate.
689 * This function hints to the VM that @folio is a good reclaim candidate,
690 * for example if its invalidation fails due to the folio being dirty
691 * or under writeback.
693 * Context: Caller holds a reference on the folio.
695 void deactivate_file_folio(struct folio *folio)
697 struct folio_batch *fbatch;
699 /* Deactivating an unevictable folio will not accelerate reclaim */
700 if (folio_test_unevictable(folio))
704 local_lock(&cpu_fbatches.lock);
705 fbatch = this_cpu_ptr(&cpu_fbatches.lru_deactivate_file);
706 folio_batch_add_and_move(fbatch, folio, lru_deactivate_file_fn);
707 local_unlock(&cpu_fbatches.lock);
711 * folio_deactivate - deactivate a folio
712 * @folio: folio to deactivate
714 * folio_deactivate() moves @folio to the inactive list if @folio was on the
715 * active list and was not unevictable. This is done to accelerate the
718 void folio_deactivate(struct folio *folio)
720 if (folio_test_lru(folio) && !folio_test_unevictable(folio) &&
721 (folio_test_active(folio) || lru_gen_enabled())) {
722 struct folio_batch *fbatch;
725 local_lock(&cpu_fbatches.lock);
726 fbatch = this_cpu_ptr(&cpu_fbatches.lru_deactivate);
727 folio_batch_add_and_move(fbatch, folio, lru_deactivate_fn);
728 local_unlock(&cpu_fbatches.lock);
733 * folio_mark_lazyfree - make an anon folio lazyfree
734 * @folio: folio to deactivate
736 * folio_mark_lazyfree() moves @folio to the inactive file list.
737 * This is done to accelerate the reclaim of @folio.
739 void folio_mark_lazyfree(struct folio *folio)
741 if (folio_test_lru(folio) && folio_test_anon(folio) &&
742 folio_test_swapbacked(folio) && !folio_test_swapcache(folio) &&
743 !folio_test_unevictable(folio)) {
744 struct folio_batch *fbatch;
747 local_lock(&cpu_fbatches.lock);
748 fbatch = this_cpu_ptr(&cpu_fbatches.lru_lazyfree);
749 folio_batch_add_and_move(fbatch, folio, lru_lazyfree_fn);
750 local_unlock(&cpu_fbatches.lock);
754 void lru_add_drain(void)
756 local_lock(&cpu_fbatches.lock);
757 lru_add_drain_cpu(smp_processor_id());
758 local_unlock(&cpu_fbatches.lock);
763 * It's called from per-cpu workqueue context in SMP case so
764 * lru_add_drain_cpu and invalidate_bh_lrus_cpu should run on
765 * the same cpu. It shouldn't be a problem in !SMP case since
766 * the core is only one and the locks will disable preemption.
768 static void lru_add_and_bh_lrus_drain(void)
770 local_lock(&cpu_fbatches.lock);
771 lru_add_drain_cpu(smp_processor_id());
772 local_unlock(&cpu_fbatches.lock);
773 invalidate_bh_lrus_cpu();
777 void lru_add_drain_cpu_zone(struct zone *zone)
779 local_lock(&cpu_fbatches.lock);
780 lru_add_drain_cpu(smp_processor_id());
781 drain_local_pages(zone);
782 local_unlock(&cpu_fbatches.lock);
788 static DEFINE_PER_CPU(struct work_struct, lru_add_drain_work);
790 static void lru_add_drain_per_cpu(struct work_struct *dummy)
792 lru_add_and_bh_lrus_drain();
795 static bool cpu_needs_drain(unsigned int cpu)
797 struct cpu_fbatches *fbatches = &per_cpu(cpu_fbatches, cpu);
799 /* Check these in order of likelihood that they're not zero */
800 return folio_batch_count(&fbatches->lru_add) ||
801 data_race(folio_batch_count(&per_cpu(lru_rotate.fbatch, cpu))) ||
802 folio_batch_count(&fbatches->lru_deactivate_file) ||
803 folio_batch_count(&fbatches->lru_deactivate) ||
804 folio_batch_count(&fbatches->lru_lazyfree) ||
805 folio_batch_count(&fbatches->activate) ||
806 need_mlock_drain(cpu) ||
807 has_bh_in_lru(cpu, NULL);
811 * Doesn't need any cpu hotplug locking because we do rely on per-cpu
812 * kworkers being shut down before our page_alloc_cpu_dead callback is
813 * executed on the offlined cpu.
814 * Calling this function with cpu hotplug locks held can actually lead
815 * to obscure indirect dependencies via WQ context.
817 static inline void __lru_add_drain_all(bool force_all_cpus)
820 * lru_drain_gen - Global pages generation number
822 * (A) Definition: global lru_drain_gen = x implies that all generations
823 * 0 < n <= x are already *scheduled* for draining.
825 * This is an optimization for the highly-contended use case where a
826 * user space workload keeps constantly generating a flow of pages for
829 static unsigned int lru_drain_gen;
830 static struct cpumask has_work;
831 static DEFINE_MUTEX(lock);
832 unsigned cpu, this_gen;
835 * Make sure nobody triggers this path before mm_percpu_wq is fully
838 if (WARN_ON(!mm_percpu_wq))
842 * Guarantee folio_batch counter stores visible by this CPU
843 * are visible to other CPUs before loading the current drain
849 * (B) Locally cache global LRU draining generation number
851 * The read barrier ensures that the counter is loaded before the mutex
852 * is taken. It pairs with smp_mb() inside the mutex critical section
855 this_gen = smp_load_acquire(&lru_drain_gen);
860 * (C) Exit the draining operation if a newer generation, from another
861 * lru_add_drain_all(), was already scheduled for draining. Check (A).
863 if (unlikely(this_gen != lru_drain_gen && !force_all_cpus))
867 * (D) Increment global generation number
869 * Pairs with smp_load_acquire() at (B), outside of the critical
870 * section. Use a full memory barrier to guarantee that the
871 * new global drain generation number is stored before loading
872 * folio_batch counters.
874 * This pairing must be done here, before the for_each_online_cpu loop
875 * below which drains the page vectors.
877 * Let x, y, and z represent some system CPU numbers, where x < y < z.
878 * Assume CPU #z is in the middle of the for_each_online_cpu loop
879 * below and has already reached CPU #y's per-cpu data. CPU #x comes
880 * along, adds some pages to its per-cpu vectors, then calls
881 * lru_add_drain_all().
883 * If the paired barrier is done at any later step, e.g. after the
884 * loop, CPU #x will just exit at (C) and miss flushing out all of its
887 WRITE_ONCE(lru_drain_gen, lru_drain_gen + 1);
890 cpumask_clear(&has_work);
891 for_each_online_cpu(cpu) {
892 struct work_struct *work = &per_cpu(lru_add_drain_work, cpu);
894 if (cpu_needs_drain(cpu)) {
895 INIT_WORK(work, lru_add_drain_per_cpu);
896 queue_work_on(cpu, mm_percpu_wq, work);
897 __cpumask_set_cpu(cpu, &has_work);
901 for_each_cpu(cpu, &has_work)
902 flush_work(&per_cpu(lru_add_drain_work, cpu));
908 void lru_add_drain_all(void)
910 __lru_add_drain_all(false);
913 void lru_add_drain_all(void)
917 #endif /* CONFIG_SMP */
919 atomic_t lru_disable_count = ATOMIC_INIT(0);
922 * lru_cache_disable() needs to be called before we start compiling
923 * a list of pages to be migrated using isolate_lru_page().
924 * It drains pages on LRU cache and then disable on all cpus until
925 * lru_cache_enable is called.
927 * Must be paired with a call to lru_cache_enable().
929 void lru_cache_disable(void)
931 atomic_inc(&lru_disable_count);
933 * Readers of lru_disable_count are protected by either disabling
934 * preemption or rcu_read_lock:
936 * preempt_disable, local_irq_disable [bh_lru_lock()]
937 * rcu_read_lock [rt_spin_lock CONFIG_PREEMPT_RT]
938 * preempt_disable [local_lock !CONFIG_PREEMPT_RT]
940 * Since v5.1 kernel, synchronize_rcu() is guaranteed to wait on
941 * preempt_disable() regions of code. So any CPU which sees
942 * lru_disable_count = 0 will have exited the critical
943 * section when synchronize_rcu() returns.
945 synchronize_rcu_expedited();
947 __lru_add_drain_all(true);
949 lru_add_and_bh_lrus_drain();
954 * folios_put_refs - Reduce the reference count on a batch of folios.
955 * @folios: The folios.
956 * @refs: The number of refs to subtract from each folio.
958 * Like folio_put(), but for a batch of folios. This is more efficient
959 * than writing the loop yourself as it will optimise the locks which need
960 * to be taken if the folios are freed. The folios batch is returned
961 * empty and ready to be reused for another batch; there is no need
962 * to reinitialise it. If @refs is NULL, we subtract one from each
965 * Context: May be called in process or interrupt context, but not in NMI
966 * context. May be called while holding a spinlock.
968 void folios_put_refs(struct folio_batch *folios, unsigned int *refs)
971 struct lruvec *lruvec = NULL;
972 unsigned long flags = 0;
974 for (i = 0, j = 0; i < folios->nr; i++) {
975 struct folio *folio = folios->folios[i];
976 unsigned int nr_refs = refs ? refs[i] : 1;
978 if (is_huge_zero_folio(folio))
981 if (folio_is_zone_device(folio)) {
983 unlock_page_lruvec_irqrestore(lruvec, flags);
986 if (put_devmap_managed_folio_refs(folio, nr_refs))
988 if (folio_ref_sub_and_test(folio, nr_refs))
989 free_zone_device_folio(folio);
993 if (!folio_ref_sub_and_test(folio, nr_refs))
996 /* hugetlb has its own memcg */
997 if (folio_test_hugetlb(folio)) {
999 unlock_page_lruvec_irqrestore(lruvec, flags);
1002 free_huge_folio(folio);
1005 if (folio_test_large(folio) &&
1006 folio_test_large_rmappable(folio))
1007 folio_undo_large_rmappable(folio);
1009 __page_cache_release(folio, &lruvec, &flags);
1012 folios->folios[j] = folio;
1016 unlock_page_lruvec_irqrestore(lruvec, flags);
1018 folio_batch_reinit(folios);
1023 mem_cgroup_uncharge_folios(folios);
1024 free_unref_folios(folios);
1026 EXPORT_SYMBOL(folios_put_refs);
1029 * release_pages - batched put_page()
1030 * @arg: array of pages to release
1031 * @nr: number of pages
1033 * Decrement the reference count on all the pages in @arg. If it
1034 * fell to zero, remove the page from the LRU and free it.
1036 * Note that the argument can be an array of pages, encoded pages,
1037 * or folio pointers. We ignore any encoded bits, and turn any of
1038 * them into just a folio that gets free'd.
1040 void release_pages(release_pages_arg arg, int nr)
1042 struct folio_batch fbatch;
1043 int refs[PAGEVEC_SIZE];
1044 struct encoded_page **encoded = arg.encoded_pages;
1047 folio_batch_init(&fbatch);
1048 for (i = 0; i < nr; i++) {
1049 /* Turn any of the argument types into a folio */
1050 struct folio *folio = page_folio(encoded_page_ptr(encoded[i]));
1052 /* Is our next entry actually "nr_pages" -> "nr_refs" ? */
1053 refs[fbatch.nr] = 1;
1054 if (unlikely(encoded_page_flags(encoded[i]) &
1055 ENCODED_PAGE_BIT_NR_PAGES_NEXT))
1056 refs[fbatch.nr] = encoded_nr_pages(encoded[++i]);
1058 if (folio_batch_add(&fbatch, folio) > 0)
1060 folios_put_refs(&fbatch, refs);
1064 folios_put_refs(&fbatch, refs);
1066 EXPORT_SYMBOL(release_pages);
1069 * The folios which we're about to release may be in the deferred lru-addition
1070 * queues. That would prevent them from really being freed right now. That's
1071 * OK from a correctness point of view but is inefficient - those folios may be
1072 * cache-warm and we want to give them back to the page allocator ASAP.
1074 * So __folio_batch_release() will drain those queues here.
1075 * folio_batch_move_lru() calls folios_put() directly to avoid
1078 void __folio_batch_release(struct folio_batch *fbatch)
1080 if (!fbatch->percpu_pvec_drained) {
1082 fbatch->percpu_pvec_drained = true;
1086 EXPORT_SYMBOL(__folio_batch_release);
1089 * folio_batch_remove_exceptionals() - Prune non-folios from a batch.
1090 * @fbatch: The batch to prune
1092 * find_get_entries() fills a batch with both folios and shadow/swap/DAX
1093 * entries. This function prunes all the non-folio entries from @fbatch
1094 * without leaving holes, so that it can be passed on to folio-only batch
1097 void folio_batch_remove_exceptionals(struct folio_batch *fbatch)
1101 for (i = 0, j = 0; i < folio_batch_count(fbatch); i++) {
1102 struct folio *folio = fbatch->folios[i];
1103 if (!xa_is_value(folio))
1104 fbatch->folios[j++] = folio;
1110 * Perform any setup for the swap system
1112 void __init swap_setup(void)
1114 unsigned long megs = totalram_pages() >> (20 - PAGE_SHIFT);
1116 /* Use a smaller cluster for small-memory machines */
1122 * Right now other parts of the system means that we
1123 * _really_ don't want to cluster much more