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.fbatch which requires disabling interrupts */
52 struct folio_batch fbatch;
54 static DEFINE_PER_CPU(struct lru_rotate, lru_rotate) = {
55 .lock = INIT_LOCAL_LOCK(lock),
59 * The following folio batches are grouped together because they are protected
60 * by disabling preemption (and interrupts remain enabled).
64 struct folio_batch lru_add;
65 struct folio_batch lru_deactivate_file;
66 struct folio_batch lru_deactivate;
67 struct folio_batch lru_lazyfree;
69 struct folio_batch activate;
72 static DEFINE_PER_CPU(struct cpu_fbatches, cpu_fbatches) = {
73 .lock = INIT_LOCAL_LOCK(lock),
77 * This path almost never happens for VM activity - pages are normally freed
78 * via pagevecs. But it gets used by networking - and for compound pages.
80 static void __page_cache_release(struct folio *folio)
82 if (folio_test_lru(folio)) {
83 struct lruvec *lruvec;
86 lruvec = folio_lruvec_lock_irqsave(folio, &flags);
87 lruvec_del_folio(lruvec, folio);
88 __folio_clear_lru_flags(folio);
89 unlock_page_lruvec_irqrestore(lruvec, flags);
91 /* See comment on folio_test_mlocked in release_pages() */
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);
101 static void __folio_put_small(struct folio *folio)
103 __page_cache_release(folio);
104 mem_cgroup_uncharge(folio);
105 free_unref_page(&folio->page, 0);
108 static void __folio_put_large(struct folio *folio)
111 * __page_cache_release() is supposed to be called for thp, not for
112 * hugetlb. This is because hugetlb page does never have PageLRU set
113 * (it's never listed to any LRU lists) and no memcg routines should
114 * be called for hugetlb (it has a separate hugetlb_cgroup.)
116 if (!folio_test_hugetlb(folio))
117 __page_cache_release(folio);
118 destroy_large_folio(folio);
121 void __folio_put(struct folio *folio)
123 if (unlikely(folio_is_zone_device(folio)))
124 free_zone_device_page(&folio->page);
125 else if (unlikely(folio_test_large(folio)))
126 __folio_put_large(folio);
128 __folio_put_small(folio);
130 EXPORT_SYMBOL(__folio_put);
133 * put_pages_list() - release a list of pages
134 * @pages: list of pages threaded on page->lru
136 * Release a list of pages which are strung together on page.lru.
138 void put_pages_list(struct list_head *pages)
140 struct folio *folio, *next;
142 list_for_each_entry_safe(folio, next, pages, lru) {
143 if (!folio_put_testzero(folio)) {
144 list_del(&folio->lru);
147 if (folio_test_large(folio)) {
148 list_del(&folio->lru);
149 __folio_put_large(folio);
152 /* LRU flag must be clear because it's passed using the lru */
155 free_unref_page_list(pages);
156 INIT_LIST_HEAD(pages);
158 EXPORT_SYMBOL(put_pages_list);
161 * get_kernel_pages() - pin kernel pages in memory
162 * @kiov: An array of struct kvec structures
163 * @nr_segs: number of segments to pin
164 * @write: pinning for read/write, currently ignored
165 * @pages: array that receives pointers to the pages pinned.
166 * Should be at least nr_segs long.
168 * Returns number of pages pinned. This may be fewer than the number requested.
169 * If nr_segs is 0 or negative, returns 0. If no pages were pinned, returns 0.
170 * Each page returned must be released with a put_page() call when it is
173 int get_kernel_pages(const struct kvec *kiov, int nr_segs, int write,
178 for (seg = 0; seg < nr_segs; seg++) {
179 if (WARN_ON(kiov[seg].iov_len != PAGE_SIZE))
182 pages[seg] = kmap_to_page(kiov[seg].iov_base);
183 get_page(pages[seg]);
188 EXPORT_SYMBOL_GPL(get_kernel_pages);
190 typedef void (*move_fn_t)(struct lruvec *lruvec, struct folio *folio);
192 static void lru_add_fn(struct lruvec *lruvec, struct folio *folio)
194 int was_unevictable = folio_test_clear_unevictable(folio);
195 long nr_pages = folio_nr_pages(folio);
197 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
200 * Is an smp_mb__after_atomic() still required here, before
201 * folio_evictable() tests the mlocked flag, to rule out the possibility
202 * of stranding an evictable folio on an unevictable LRU? I think
203 * not, because __munlock_page() only clears the mlocked flag
204 * while the LRU lock is held.
206 * (That is not true of __page_cache_release(), and not necessarily
207 * true of release_pages(): but those only clear the mlocked flag after
208 * folio_put_testzero() has excluded any other users of the folio.)
210 if (folio_evictable(folio)) {
212 __count_vm_events(UNEVICTABLE_PGRESCUED, nr_pages);
214 folio_clear_active(folio);
215 folio_set_unevictable(folio);
217 * folio->mlock_count = !!folio_test_mlocked(folio)?
218 * But that leaves __mlock_page() in doubt whether another
219 * actor has already counted the mlock or not. Err on the
220 * safe side, underestimate, let page reclaim fix it, rather
221 * than leaving a page on the unevictable LRU indefinitely.
223 folio->mlock_count = 0;
224 if (!was_unevictable)
225 __count_vm_events(UNEVICTABLE_PGCULLED, nr_pages);
228 lruvec_add_folio(lruvec, folio);
229 trace_mm_lru_insertion(folio);
232 static void folio_batch_move_lru(struct folio_batch *fbatch, move_fn_t move_fn)
235 struct lruvec *lruvec = NULL;
236 unsigned long flags = 0;
238 for (i = 0; i < folio_batch_count(fbatch); i++) {
239 struct folio *folio = fbatch->folios[i];
241 /* block memcg migration while the folio moves between lru */
242 if (move_fn != lru_add_fn && !folio_test_clear_lru(folio))
245 lruvec = folio_lruvec_relock_irqsave(folio, lruvec, &flags);
246 move_fn(lruvec, folio);
248 folio_set_lru(folio);
252 unlock_page_lruvec_irqrestore(lruvec, flags);
253 folios_put(fbatch->folios, folio_batch_count(fbatch));
254 folio_batch_init(fbatch);
257 static void folio_batch_add_and_move(struct folio_batch *fbatch,
258 struct folio *folio, move_fn_t move_fn)
260 if (folio_batch_add(fbatch, folio) && !folio_test_large(folio) &&
261 !lru_cache_disabled())
263 folio_batch_move_lru(fbatch, move_fn);
266 static void lru_move_tail_fn(struct lruvec *lruvec, struct folio *folio)
268 if (!folio_test_unevictable(folio)) {
269 lruvec_del_folio(lruvec, folio);
270 folio_clear_active(folio);
271 lruvec_add_folio_tail(lruvec, folio);
272 __count_vm_events(PGROTATED, folio_nr_pages(folio));
277 * Writeback is about to end against a folio which has been marked for
278 * immediate reclaim. If it still appears to be reclaimable, move it
279 * to the tail of the inactive list.
281 * folio_rotate_reclaimable() must disable IRQs, to prevent nasty races.
283 void folio_rotate_reclaimable(struct folio *folio)
285 if (!folio_test_locked(folio) && !folio_test_dirty(folio) &&
286 !folio_test_unevictable(folio) && folio_test_lru(folio)) {
287 struct folio_batch *fbatch;
291 local_lock_irqsave(&lru_rotate.lock, flags);
292 fbatch = this_cpu_ptr(&lru_rotate.fbatch);
293 folio_batch_add_and_move(fbatch, folio, lru_move_tail_fn);
294 local_unlock_irqrestore(&lru_rotate.lock, flags);
298 void lru_note_cost(struct lruvec *lruvec, bool file,
299 unsigned int nr_io, unsigned int nr_rotated)
304 * Reflect the relative cost of incurring IO and spending CPU
305 * time on rotations. This doesn't attempt to make a precise
306 * comparison, it just says: if reloads are about comparable
307 * between the LRU lists, or rotations are overwhelmingly
308 * different between them, adjust scan balance for CPU work.
310 cost = nr_io * SWAP_CLUSTER_MAX + nr_rotated;
313 unsigned long lrusize;
316 * Hold lruvec->lru_lock is safe here, since
317 * 1) The pinned lruvec in reclaim, or
318 * 2) From a pre-LRU page during refault (which also holds the
319 * rcu lock, so would be safe even if the page was on the LRU
320 * and could move simultaneously to a new lruvec).
322 spin_lock_irq(&lruvec->lru_lock);
323 /* Record cost event */
325 lruvec->file_cost += cost;
327 lruvec->anon_cost += cost;
330 * Decay previous events
332 * Because workloads change over time (and to avoid
333 * overflow) we keep these statistics as a floating
334 * average, which ends up weighing recent refaults
335 * more than old ones.
337 lrusize = lruvec_page_state(lruvec, NR_INACTIVE_ANON) +
338 lruvec_page_state(lruvec, NR_ACTIVE_ANON) +
339 lruvec_page_state(lruvec, NR_INACTIVE_FILE) +
340 lruvec_page_state(lruvec, NR_ACTIVE_FILE);
342 if (lruvec->file_cost + lruvec->anon_cost > lrusize / 4) {
343 lruvec->file_cost /= 2;
344 lruvec->anon_cost /= 2;
346 spin_unlock_irq(&lruvec->lru_lock);
347 } while ((lruvec = parent_lruvec(lruvec)));
350 void lru_note_cost_refault(struct folio *folio)
352 lru_note_cost(folio_lruvec(folio), folio_is_file_lru(folio),
353 folio_nr_pages(folio), 0);
356 static void folio_activate_fn(struct lruvec *lruvec, struct folio *folio)
358 if (!folio_test_active(folio) && !folio_test_unevictable(folio)) {
359 long nr_pages = folio_nr_pages(folio);
361 lruvec_del_folio(lruvec, folio);
362 folio_set_active(folio);
363 lruvec_add_folio(lruvec, folio);
364 trace_mm_lru_activate(folio);
366 __count_vm_events(PGACTIVATE, nr_pages);
367 __count_memcg_events(lruvec_memcg(lruvec), PGACTIVATE,
373 static void folio_activate_drain(int cpu)
375 struct folio_batch *fbatch = &per_cpu(cpu_fbatches.activate, cpu);
377 if (folio_batch_count(fbatch))
378 folio_batch_move_lru(fbatch, folio_activate_fn);
381 void folio_activate(struct folio *folio)
383 if (folio_test_lru(folio) && !folio_test_active(folio) &&
384 !folio_test_unevictable(folio)) {
385 struct folio_batch *fbatch;
388 local_lock(&cpu_fbatches.lock);
389 fbatch = this_cpu_ptr(&cpu_fbatches.activate);
390 folio_batch_add_and_move(fbatch, folio, folio_activate_fn);
391 local_unlock(&cpu_fbatches.lock);
396 static inline void folio_activate_drain(int cpu)
400 void folio_activate(struct folio *folio)
402 struct lruvec *lruvec;
404 if (folio_test_clear_lru(folio)) {
405 lruvec = folio_lruvec_lock_irq(folio);
406 folio_activate_fn(lruvec, folio);
407 unlock_page_lruvec_irq(lruvec);
408 folio_set_lru(folio);
413 static void __lru_cache_activate_folio(struct folio *folio)
415 struct folio_batch *fbatch;
418 local_lock(&cpu_fbatches.lock);
419 fbatch = this_cpu_ptr(&cpu_fbatches.lru_add);
422 * Search backwards on the optimistic assumption that the folio being
423 * activated has just been added to this batch. Note that only
424 * the local batch is examined as a !LRU folio could be in the
425 * process of being released, reclaimed, migrated or on a remote
426 * batch that is currently being drained. Furthermore, marking
427 * a remote batch's folio active potentially hits a race where
428 * a folio is marked active just after it is added to the inactive
429 * list causing accounting errors and BUG_ON checks to trigger.
431 for (i = folio_batch_count(fbatch) - 1; i >= 0; i--) {
432 struct folio *batch_folio = fbatch->folios[i];
434 if (batch_folio == folio) {
435 folio_set_active(folio);
440 local_unlock(&cpu_fbatches.lock);
443 #ifdef CONFIG_LRU_GEN
444 static void folio_inc_refs(struct folio *folio)
446 unsigned long new_flags, old_flags = READ_ONCE(folio->flags);
448 if (folio_test_unevictable(folio))
451 if (!folio_test_referenced(folio)) {
452 folio_set_referenced(folio);
456 if (!folio_test_workingset(folio)) {
457 folio_set_workingset(folio);
461 /* see the comment on MAX_NR_TIERS */
463 new_flags = old_flags & LRU_REFS_MASK;
464 if (new_flags == LRU_REFS_MASK)
467 new_flags += BIT(LRU_REFS_PGOFF);
468 new_flags |= old_flags & ~LRU_REFS_MASK;
469 } while (!try_cmpxchg(&folio->flags, &old_flags, new_flags));
472 static void folio_inc_refs(struct folio *folio)
475 #endif /* CONFIG_LRU_GEN */
478 * Mark a page as having seen activity.
480 * inactive,unreferenced -> inactive,referenced
481 * inactive,referenced -> active,unreferenced
482 * active,unreferenced -> active,referenced
484 * When a newly allocated page is not yet visible, so safe for non-atomic ops,
485 * __SetPageReferenced(page) may be substituted for mark_page_accessed(page).
487 void folio_mark_accessed(struct folio *folio)
489 if (lru_gen_enabled()) {
490 folio_inc_refs(folio);
494 if (!folio_test_referenced(folio)) {
495 folio_set_referenced(folio);
496 } else if (folio_test_unevictable(folio)) {
498 * Unevictable pages are on the "LRU_UNEVICTABLE" list. But,
499 * this list is never rotated or maintained, so marking an
500 * unevictable page accessed has no effect.
502 } else if (!folio_test_active(folio)) {
504 * If the folio is on the LRU, queue it for activation via
505 * cpu_fbatches.activate. Otherwise, assume the folio is in a
506 * folio_batch, mark it active and it'll be moved to the active
507 * LRU on the next drain.
509 if (folio_test_lru(folio))
510 folio_activate(folio);
512 __lru_cache_activate_folio(folio);
513 folio_clear_referenced(folio);
514 workingset_activation(folio);
516 if (folio_test_idle(folio))
517 folio_clear_idle(folio);
519 EXPORT_SYMBOL(folio_mark_accessed);
522 * folio_add_lru - Add a folio to an LRU list.
523 * @folio: The folio to be added to the LRU.
525 * Queue the folio for addition to the LRU. The decision on whether
526 * to add the page to the [in]active [file|anon] list is deferred until the
527 * folio_batch is drained. This gives a chance for the caller of folio_add_lru()
528 * have the folio added to the active list using folio_mark_accessed().
530 void folio_add_lru(struct folio *folio)
532 struct folio_batch *fbatch;
534 VM_BUG_ON_FOLIO(folio_test_active(folio) &&
535 folio_test_unevictable(folio), folio);
536 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
538 /* see the comment in lru_gen_add_folio() */
539 if (lru_gen_enabled() && !folio_test_unevictable(folio) &&
540 lru_gen_in_fault() && !(current->flags & PF_MEMALLOC))
541 folio_set_active(folio);
544 local_lock(&cpu_fbatches.lock);
545 fbatch = this_cpu_ptr(&cpu_fbatches.lru_add);
546 folio_batch_add_and_move(fbatch, folio, lru_add_fn);
547 local_unlock(&cpu_fbatches.lock);
549 EXPORT_SYMBOL(folio_add_lru);
552 * folio_add_lru_vma() - Add a folio to the appropate LRU list for this VMA.
553 * @folio: The folio to be added to the LRU.
554 * @vma: VMA in which the folio is mapped.
556 * If the VMA is mlocked, @folio is added to the unevictable list.
557 * Otherwise, it is treated the same way as folio_add_lru().
559 void folio_add_lru_vma(struct folio *folio, struct vm_area_struct *vma)
561 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
563 if (unlikely((vma->vm_flags & (VM_LOCKED | VM_SPECIAL)) == VM_LOCKED))
564 mlock_new_page(&folio->page);
566 folio_add_lru(folio);
570 * If the folio cannot be invalidated, it is moved to the
571 * inactive list to speed up its reclaim. It is moved to the
572 * head of the list, rather than the tail, to give the flusher
573 * threads some time to write it out, as this is much more
574 * effective than the single-page writeout from reclaim.
576 * If the folio isn't mapped and dirty/writeback, the folio
577 * could be reclaimed asap using the reclaim flag.
579 * 1. active, mapped folio -> none
580 * 2. active, dirty/writeback folio -> inactive, head, reclaim
581 * 3. inactive, mapped folio -> none
582 * 4. inactive, dirty/writeback folio -> inactive, head, reclaim
583 * 5. inactive, clean -> inactive, tail
586 * In 4, it moves to the head of the inactive list so the folio is
587 * written out by flusher threads as this is much more efficient
588 * than the single-page writeout from reclaim.
590 static void lru_deactivate_file_fn(struct lruvec *lruvec, struct folio *folio)
592 bool active = folio_test_active(folio);
593 long nr_pages = folio_nr_pages(folio);
595 if (folio_test_unevictable(folio))
598 /* Some processes are using the folio */
599 if (folio_mapped(folio))
602 lruvec_del_folio(lruvec, folio);
603 folio_clear_active(folio);
604 folio_clear_referenced(folio);
606 if (folio_test_writeback(folio) || folio_test_dirty(folio)) {
608 * Setting the reclaim flag could race with
609 * folio_end_writeback() and confuse readahead. But the
610 * race window is _really_ small and it's not a critical
613 lruvec_add_folio(lruvec, folio);
614 folio_set_reclaim(folio);
617 * The folio's writeback ended while it was in the batch.
618 * We move that folio to the tail of the inactive list.
620 lruvec_add_folio_tail(lruvec, folio);
621 __count_vm_events(PGROTATED, nr_pages);
625 __count_vm_events(PGDEACTIVATE, nr_pages);
626 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
631 static void lru_deactivate_fn(struct lruvec *lruvec, struct folio *folio)
633 if (!folio_test_unevictable(folio) && (folio_test_active(folio) || lru_gen_enabled())) {
634 long nr_pages = folio_nr_pages(folio);
636 lruvec_del_folio(lruvec, folio);
637 folio_clear_active(folio);
638 folio_clear_referenced(folio);
639 lruvec_add_folio(lruvec, folio);
641 __count_vm_events(PGDEACTIVATE, nr_pages);
642 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
647 static void lru_lazyfree_fn(struct lruvec *lruvec, struct folio *folio)
649 if (folio_test_anon(folio) && folio_test_swapbacked(folio) &&
650 !folio_test_swapcache(folio) && !folio_test_unevictable(folio)) {
651 long nr_pages = folio_nr_pages(folio);
653 lruvec_del_folio(lruvec, folio);
654 folio_clear_active(folio);
655 folio_clear_referenced(folio);
657 * Lazyfree folios are clean anonymous folios. They have
658 * the swapbacked flag cleared, to distinguish them from normal
661 folio_clear_swapbacked(folio);
662 lruvec_add_folio(lruvec, folio);
664 __count_vm_events(PGLAZYFREE, nr_pages);
665 __count_memcg_events(lruvec_memcg(lruvec), PGLAZYFREE,
671 * Drain pages out of the cpu's folio_batch.
672 * Either "cpu" is the current CPU, and preemption has already been
673 * disabled; or "cpu" is being hot-unplugged, and is already dead.
675 void lru_add_drain_cpu(int cpu)
677 struct cpu_fbatches *fbatches = &per_cpu(cpu_fbatches, cpu);
678 struct folio_batch *fbatch = &fbatches->lru_add;
680 if (folio_batch_count(fbatch))
681 folio_batch_move_lru(fbatch, lru_add_fn);
683 fbatch = &per_cpu(lru_rotate.fbatch, cpu);
684 /* Disabling interrupts below acts as a compiler barrier. */
685 if (data_race(folio_batch_count(fbatch))) {
688 /* No harm done if a racing interrupt already did this */
689 local_lock_irqsave(&lru_rotate.lock, flags);
690 folio_batch_move_lru(fbatch, lru_move_tail_fn);
691 local_unlock_irqrestore(&lru_rotate.lock, flags);
694 fbatch = &fbatches->lru_deactivate_file;
695 if (folio_batch_count(fbatch))
696 folio_batch_move_lru(fbatch, lru_deactivate_file_fn);
698 fbatch = &fbatches->lru_deactivate;
699 if (folio_batch_count(fbatch))
700 folio_batch_move_lru(fbatch, lru_deactivate_fn);
702 fbatch = &fbatches->lru_lazyfree;
703 if (folio_batch_count(fbatch))
704 folio_batch_move_lru(fbatch, lru_lazyfree_fn);
706 folio_activate_drain(cpu);
710 * deactivate_file_folio() - Deactivate a file folio.
711 * @folio: Folio to deactivate.
713 * This function hints to the VM that @folio is a good reclaim candidate,
714 * for example if its invalidation fails due to the folio being dirty
715 * or under writeback.
717 * Context: Caller holds a reference on the folio.
719 void deactivate_file_folio(struct folio *folio)
721 struct folio_batch *fbatch;
723 /* Deactivating an unevictable folio will not accelerate reclaim */
724 if (folio_test_unevictable(folio))
728 local_lock(&cpu_fbatches.lock);
729 fbatch = this_cpu_ptr(&cpu_fbatches.lru_deactivate_file);
730 folio_batch_add_and_move(fbatch, folio, lru_deactivate_file_fn);
731 local_unlock(&cpu_fbatches.lock);
735 * deactivate_page - deactivate a page
736 * @page: page to deactivate
738 * deactivate_page() moves @page to the inactive list if @page was on the active
739 * list and was not an unevictable page. This is done to accelerate the reclaim
742 void deactivate_page(struct page *page)
744 struct folio *folio = page_folio(page);
746 if (folio_test_lru(folio) && !folio_test_unevictable(folio) &&
747 (folio_test_active(folio) || lru_gen_enabled())) {
748 struct folio_batch *fbatch;
751 local_lock(&cpu_fbatches.lock);
752 fbatch = this_cpu_ptr(&cpu_fbatches.lru_deactivate);
753 folio_batch_add_and_move(fbatch, folio, lru_deactivate_fn);
754 local_unlock(&cpu_fbatches.lock);
759 * mark_page_lazyfree - make an anon page lazyfree
760 * @page: page to deactivate
762 * mark_page_lazyfree() moves @page to the inactive file list.
763 * This is done to accelerate the reclaim of @page.
765 void mark_page_lazyfree(struct page *page)
767 struct folio *folio = page_folio(page);
769 if (folio_test_lru(folio) && folio_test_anon(folio) &&
770 folio_test_swapbacked(folio) && !folio_test_swapcache(folio) &&
771 !folio_test_unevictable(folio)) {
772 struct folio_batch *fbatch;
775 local_lock(&cpu_fbatches.lock);
776 fbatch = this_cpu_ptr(&cpu_fbatches.lru_lazyfree);
777 folio_batch_add_and_move(fbatch, folio, lru_lazyfree_fn);
778 local_unlock(&cpu_fbatches.lock);
782 void lru_add_drain(void)
784 local_lock(&cpu_fbatches.lock);
785 lru_add_drain_cpu(smp_processor_id());
786 local_unlock(&cpu_fbatches.lock);
787 mlock_page_drain_local();
791 * It's called from per-cpu workqueue context in SMP case so
792 * lru_add_drain_cpu and invalidate_bh_lrus_cpu should run on
793 * the same cpu. It shouldn't be a problem in !SMP case since
794 * the core is only one and the locks will disable preemption.
796 static void lru_add_and_bh_lrus_drain(void)
798 local_lock(&cpu_fbatches.lock);
799 lru_add_drain_cpu(smp_processor_id());
800 local_unlock(&cpu_fbatches.lock);
801 invalidate_bh_lrus_cpu();
802 mlock_page_drain_local();
805 void lru_add_drain_cpu_zone(struct zone *zone)
807 local_lock(&cpu_fbatches.lock);
808 lru_add_drain_cpu(smp_processor_id());
809 drain_local_pages(zone);
810 local_unlock(&cpu_fbatches.lock);
811 mlock_page_drain_local();
816 static DEFINE_PER_CPU(struct work_struct, lru_add_drain_work);
818 static void lru_add_drain_per_cpu(struct work_struct *dummy)
820 lru_add_and_bh_lrus_drain();
823 static bool cpu_needs_drain(unsigned int cpu)
825 struct cpu_fbatches *fbatches = &per_cpu(cpu_fbatches, cpu);
827 /* Check these in order of likelihood that they're not zero */
828 return folio_batch_count(&fbatches->lru_add) ||
829 data_race(folio_batch_count(&per_cpu(lru_rotate.fbatch, cpu))) ||
830 folio_batch_count(&fbatches->lru_deactivate_file) ||
831 folio_batch_count(&fbatches->lru_deactivate) ||
832 folio_batch_count(&fbatches->lru_lazyfree) ||
833 folio_batch_count(&fbatches->activate) ||
834 need_mlock_page_drain(cpu) ||
835 has_bh_in_lru(cpu, NULL);
839 * Doesn't need any cpu hotplug locking because we do rely on per-cpu
840 * kworkers being shut down before our page_alloc_cpu_dead callback is
841 * executed on the offlined cpu.
842 * Calling this function with cpu hotplug locks held can actually lead
843 * to obscure indirect dependencies via WQ context.
845 static inline void __lru_add_drain_all(bool force_all_cpus)
848 * lru_drain_gen - Global pages generation number
850 * (A) Definition: global lru_drain_gen = x implies that all generations
851 * 0 < n <= x are already *scheduled* for draining.
853 * This is an optimization for the highly-contended use case where a
854 * user space workload keeps constantly generating a flow of pages for
857 static unsigned int lru_drain_gen;
858 static struct cpumask has_work;
859 static DEFINE_MUTEX(lock);
860 unsigned cpu, this_gen;
863 * Make sure nobody triggers this path before mm_percpu_wq is fully
866 if (WARN_ON(!mm_percpu_wq))
870 * Guarantee folio_batch counter stores visible by this CPU
871 * are visible to other CPUs before loading the current drain
877 * (B) Locally cache global LRU draining generation number
879 * The read barrier ensures that the counter is loaded before the mutex
880 * is taken. It pairs with smp_mb() inside the mutex critical section
883 this_gen = smp_load_acquire(&lru_drain_gen);
888 * (C) Exit the draining operation if a newer generation, from another
889 * lru_add_drain_all(), was already scheduled for draining. Check (A).
891 if (unlikely(this_gen != lru_drain_gen && !force_all_cpus))
895 * (D) Increment global generation number
897 * Pairs with smp_load_acquire() at (B), outside of the critical
898 * section. Use a full memory barrier to guarantee that the
899 * new global drain generation number is stored before loading
900 * folio_batch counters.
902 * This pairing must be done here, before the for_each_online_cpu loop
903 * below which drains the page vectors.
905 * Let x, y, and z represent some system CPU numbers, where x < y < z.
906 * Assume CPU #z is in the middle of the for_each_online_cpu loop
907 * below and has already reached CPU #y's per-cpu data. CPU #x comes
908 * along, adds some pages to its per-cpu vectors, then calls
909 * lru_add_drain_all().
911 * If the paired barrier is done at any later step, e.g. after the
912 * loop, CPU #x will just exit at (C) and miss flushing out all of its
915 WRITE_ONCE(lru_drain_gen, lru_drain_gen + 1);
918 cpumask_clear(&has_work);
919 for_each_online_cpu(cpu) {
920 struct work_struct *work = &per_cpu(lru_add_drain_work, cpu);
922 if (cpu_needs_drain(cpu)) {
923 INIT_WORK(work, lru_add_drain_per_cpu);
924 queue_work_on(cpu, mm_percpu_wq, work);
925 __cpumask_set_cpu(cpu, &has_work);
929 for_each_cpu(cpu, &has_work)
930 flush_work(&per_cpu(lru_add_drain_work, cpu));
936 void lru_add_drain_all(void)
938 __lru_add_drain_all(false);
941 void lru_add_drain_all(void)
945 #endif /* CONFIG_SMP */
947 atomic_t lru_disable_count = ATOMIC_INIT(0);
950 * lru_cache_disable() needs to be called before we start compiling
951 * a list of pages to be migrated using isolate_lru_page().
952 * It drains pages on LRU cache and then disable on all cpus until
953 * lru_cache_enable is called.
955 * Must be paired with a call to lru_cache_enable().
957 void lru_cache_disable(void)
959 atomic_inc(&lru_disable_count);
961 * Readers of lru_disable_count are protected by either disabling
962 * preemption or rcu_read_lock:
964 * preempt_disable, local_irq_disable [bh_lru_lock()]
965 * rcu_read_lock [rt_spin_lock CONFIG_PREEMPT_RT]
966 * preempt_disable [local_lock !CONFIG_PREEMPT_RT]
968 * Since v5.1 kernel, synchronize_rcu() is guaranteed to wait on
969 * preempt_disable() regions of code. So any CPU which sees
970 * lru_disable_count = 0 will have exited the critical
971 * section when synchronize_rcu() returns.
973 synchronize_rcu_expedited();
975 __lru_add_drain_all(true);
977 lru_add_and_bh_lrus_drain();
982 * release_pages - batched put_page()
983 * @pages: array of pages to release
984 * @nr: number of pages
986 * Decrement the reference count on all the pages in @pages. If it
987 * fell to zero, remove the page from the LRU and free it.
989 void release_pages(struct page **pages, int nr)
992 LIST_HEAD(pages_to_free);
993 struct lruvec *lruvec = NULL;
994 unsigned long flags = 0;
995 unsigned int lock_batch;
997 for (i = 0; i < nr; i++) {
998 struct folio *folio = page_folio(pages[i]);
1001 * Make sure the IRQ-safe lock-holding time does not get
1002 * excessive with a continuous string of pages from the
1003 * same lruvec. The lock is held only if lruvec != NULL.
1005 if (lruvec && ++lock_batch == SWAP_CLUSTER_MAX) {
1006 unlock_page_lruvec_irqrestore(lruvec, flags);
1010 if (is_huge_zero_page(&folio->page))
1013 if (folio_is_zone_device(folio)) {
1015 unlock_page_lruvec_irqrestore(lruvec, flags);
1018 if (put_devmap_managed_page(&folio->page))
1020 if (folio_put_testzero(folio))
1021 free_zone_device_page(&folio->page);
1025 if (!folio_put_testzero(folio))
1028 if (folio_test_large(folio)) {
1030 unlock_page_lruvec_irqrestore(lruvec, flags);
1033 __folio_put_large(folio);
1037 if (folio_test_lru(folio)) {
1038 struct lruvec *prev_lruvec = lruvec;
1040 lruvec = folio_lruvec_relock_irqsave(folio, lruvec,
1042 if (prev_lruvec != lruvec)
1045 lruvec_del_folio(lruvec, folio);
1046 __folio_clear_lru_flags(folio);
1050 * In rare cases, when truncation or holepunching raced with
1051 * munlock after VM_LOCKED was cleared, Mlocked may still be
1052 * found set here. This does not indicate a problem, unless
1053 * "unevictable_pgs_cleared" appears worryingly large.
1055 if (unlikely(folio_test_mlocked(folio))) {
1056 __folio_clear_mlocked(folio);
1057 zone_stat_sub_folio(folio, NR_MLOCK);
1058 count_vm_event(UNEVICTABLE_PGCLEARED);
1061 list_add(&folio->lru, &pages_to_free);
1064 unlock_page_lruvec_irqrestore(lruvec, flags);
1066 mem_cgroup_uncharge_list(&pages_to_free);
1067 free_unref_page_list(&pages_to_free);
1069 EXPORT_SYMBOL(release_pages);
1072 * The pages which we're about to release may be in the deferred lru-addition
1073 * queues. That would prevent them from really being freed right now. That's
1074 * OK from a correctness point of view but is inefficient - those pages may be
1075 * cache-warm and we want to give them back to the page allocator ASAP.
1077 * So __pagevec_release() will drain those queues here.
1078 * folio_batch_move_lru() calls folios_put() directly to avoid
1081 void __pagevec_release(struct pagevec *pvec)
1083 if (!pvec->percpu_pvec_drained) {
1085 pvec->percpu_pvec_drained = true;
1087 release_pages(pvec->pages, pagevec_count(pvec));
1088 pagevec_reinit(pvec);
1090 EXPORT_SYMBOL(__pagevec_release);
1093 * folio_batch_remove_exceptionals() - Prune non-folios from a batch.
1094 * @fbatch: The batch to prune
1096 * find_get_entries() fills a batch with both folios and shadow/swap/DAX
1097 * entries. This function prunes all the non-folio entries from @fbatch
1098 * without leaving holes, so that it can be passed on to folio-only batch
1101 void folio_batch_remove_exceptionals(struct folio_batch *fbatch)
1105 for (i = 0, j = 0; i < folio_batch_count(fbatch); i++) {
1106 struct folio *folio = fbatch->folios[i];
1107 if (!xa_is_value(folio))
1108 fbatch->folios[j++] = folio;
1113 unsigned pagevec_lookup_range_tag(struct pagevec *pvec,
1114 struct address_space *mapping, pgoff_t *index, pgoff_t end,
1117 pvec->nr = find_get_pages_range_tag(mapping, index, end, tag,
1118 PAGEVEC_SIZE, pvec->pages);
1119 return pagevec_count(pvec);
1121 EXPORT_SYMBOL(pagevec_lookup_range_tag);
1124 * Perform any setup for the swap system
1126 void __init swap_setup(void)
1128 unsigned long megs = totalram_pages() >> (20 - PAGE_SHIFT);
1130 /* Use a smaller cluster for small-memory machines */
1136 * Right now other parts of the system means that we
1137 * _really_ don't want to cluster much more