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, unsigned int nr_pages)
301 unsigned long lrusize;
304 * Hold lruvec->lru_lock is safe here, since
305 * 1) The pinned lruvec in reclaim, or
306 * 2) From a pre-LRU page during refault (which also holds the
307 * rcu lock, so would be safe even if the page was on the LRU
308 * and could move simultaneously to a new lruvec).
310 spin_lock_irq(&lruvec->lru_lock);
311 /* Record cost event */
313 lruvec->file_cost += nr_pages;
315 lruvec->anon_cost += nr_pages;
318 * Decay previous events
320 * Because workloads change over time (and to avoid
321 * overflow) we keep these statistics as a floating
322 * average, which ends up weighing recent refaults
323 * more than old ones.
325 lrusize = lruvec_page_state(lruvec, NR_INACTIVE_ANON) +
326 lruvec_page_state(lruvec, NR_ACTIVE_ANON) +
327 lruvec_page_state(lruvec, NR_INACTIVE_FILE) +
328 lruvec_page_state(lruvec, NR_ACTIVE_FILE);
330 if (lruvec->file_cost + lruvec->anon_cost > lrusize / 4) {
331 lruvec->file_cost /= 2;
332 lruvec->anon_cost /= 2;
334 spin_unlock_irq(&lruvec->lru_lock);
335 } while ((lruvec = parent_lruvec(lruvec)));
338 void lru_note_cost_folio(struct folio *folio)
340 lru_note_cost(folio_lruvec(folio), folio_is_file_lru(folio),
341 folio_nr_pages(folio));
344 static void folio_activate_fn(struct lruvec *lruvec, struct folio *folio)
346 if (!folio_test_active(folio) && !folio_test_unevictable(folio)) {
347 long nr_pages = folio_nr_pages(folio);
349 lruvec_del_folio(lruvec, folio);
350 folio_set_active(folio);
351 lruvec_add_folio(lruvec, folio);
352 trace_mm_lru_activate(folio);
354 __count_vm_events(PGACTIVATE, nr_pages);
355 __count_memcg_events(lruvec_memcg(lruvec), PGACTIVATE,
361 static void folio_activate_drain(int cpu)
363 struct folio_batch *fbatch = &per_cpu(cpu_fbatches.activate, cpu);
365 if (folio_batch_count(fbatch))
366 folio_batch_move_lru(fbatch, folio_activate_fn);
369 void folio_activate(struct folio *folio)
371 if (folio_test_lru(folio) && !folio_test_active(folio) &&
372 !folio_test_unevictable(folio)) {
373 struct folio_batch *fbatch;
376 local_lock(&cpu_fbatches.lock);
377 fbatch = this_cpu_ptr(&cpu_fbatches.activate);
378 folio_batch_add_and_move(fbatch, folio, folio_activate_fn);
379 local_unlock(&cpu_fbatches.lock);
384 static inline void folio_activate_drain(int cpu)
388 void folio_activate(struct folio *folio)
390 struct lruvec *lruvec;
392 if (folio_test_clear_lru(folio)) {
393 lruvec = folio_lruvec_lock_irq(folio);
394 folio_activate_fn(lruvec, folio);
395 unlock_page_lruvec_irq(lruvec);
396 folio_set_lru(folio);
401 static void __lru_cache_activate_folio(struct folio *folio)
403 struct folio_batch *fbatch;
406 local_lock(&cpu_fbatches.lock);
407 fbatch = this_cpu_ptr(&cpu_fbatches.lru_add);
410 * Search backwards on the optimistic assumption that the folio being
411 * activated has just been added to this batch. Note that only
412 * the local batch is examined as a !LRU folio could be in the
413 * process of being released, reclaimed, migrated or on a remote
414 * batch that is currently being drained. Furthermore, marking
415 * a remote batch's folio active potentially hits a race where
416 * a folio is marked active just after it is added to the inactive
417 * list causing accounting errors and BUG_ON checks to trigger.
419 for (i = folio_batch_count(fbatch) - 1; i >= 0; i--) {
420 struct folio *batch_folio = fbatch->folios[i];
422 if (batch_folio == folio) {
423 folio_set_active(folio);
428 local_unlock(&cpu_fbatches.lock);
431 #ifdef CONFIG_LRU_GEN
432 static void folio_inc_refs(struct folio *folio)
434 unsigned long new_flags, old_flags = READ_ONCE(folio->flags);
436 if (folio_test_unevictable(folio))
439 if (!folio_test_referenced(folio)) {
440 folio_set_referenced(folio);
444 if (!folio_test_workingset(folio)) {
445 folio_set_workingset(folio);
449 /* see the comment on MAX_NR_TIERS */
451 new_flags = old_flags & LRU_REFS_MASK;
452 if (new_flags == LRU_REFS_MASK)
455 new_flags += BIT(LRU_REFS_PGOFF);
456 new_flags |= old_flags & ~LRU_REFS_MASK;
457 } while (!try_cmpxchg(&folio->flags, &old_flags, new_flags));
460 static void folio_inc_refs(struct folio *folio)
463 #endif /* CONFIG_LRU_GEN */
466 * Mark a page as having seen activity.
468 * inactive,unreferenced -> inactive,referenced
469 * inactive,referenced -> active,unreferenced
470 * active,unreferenced -> active,referenced
472 * When a newly allocated page is not yet visible, so safe for non-atomic ops,
473 * __SetPageReferenced(page) may be substituted for mark_page_accessed(page).
475 void folio_mark_accessed(struct folio *folio)
477 if (lru_gen_enabled()) {
478 folio_inc_refs(folio);
482 if (!folio_test_referenced(folio)) {
483 folio_set_referenced(folio);
484 } else if (folio_test_unevictable(folio)) {
486 * Unevictable pages are on the "LRU_UNEVICTABLE" list. But,
487 * this list is never rotated or maintained, so marking an
488 * unevictable page accessed has no effect.
490 } else if (!folio_test_active(folio)) {
492 * If the folio is on the LRU, queue it for activation via
493 * cpu_fbatches.activate. Otherwise, assume the folio is in a
494 * folio_batch, mark it active and it'll be moved to the active
495 * LRU on the next drain.
497 if (folio_test_lru(folio))
498 folio_activate(folio);
500 __lru_cache_activate_folio(folio);
501 folio_clear_referenced(folio);
502 workingset_activation(folio);
504 if (folio_test_idle(folio))
505 folio_clear_idle(folio);
507 EXPORT_SYMBOL(folio_mark_accessed);
510 * folio_add_lru - Add a folio to an LRU list.
511 * @folio: The folio to be added to the LRU.
513 * Queue the folio for addition to the LRU. The decision on whether
514 * to add the page to the [in]active [file|anon] list is deferred until the
515 * folio_batch is drained. This gives a chance for the caller of folio_add_lru()
516 * have the folio added to the active list using folio_mark_accessed().
518 void folio_add_lru(struct folio *folio)
520 struct folio_batch *fbatch;
522 VM_BUG_ON_FOLIO(folio_test_active(folio) &&
523 folio_test_unevictable(folio), folio);
524 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
526 /* see the comment in lru_gen_add_folio() */
527 if (lru_gen_enabled() && !folio_test_unevictable(folio) &&
528 lru_gen_in_fault() && !(current->flags & PF_MEMALLOC))
529 folio_set_active(folio);
532 local_lock(&cpu_fbatches.lock);
533 fbatch = this_cpu_ptr(&cpu_fbatches.lru_add);
534 folio_batch_add_and_move(fbatch, folio, lru_add_fn);
535 local_unlock(&cpu_fbatches.lock);
537 EXPORT_SYMBOL(folio_add_lru);
540 * folio_add_lru_vma() - Add a folio to the appropate LRU list for this VMA.
541 * @folio: The folio to be added to the LRU.
542 * @vma: VMA in which the folio is mapped.
544 * If the VMA is mlocked, @folio is added to the unevictable list.
545 * Otherwise, it is treated the same way as folio_add_lru().
547 void folio_add_lru_vma(struct folio *folio, struct vm_area_struct *vma)
549 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
551 if (unlikely((vma->vm_flags & (VM_LOCKED | VM_SPECIAL)) == VM_LOCKED))
552 mlock_new_page(&folio->page);
554 folio_add_lru(folio);
558 * If the folio cannot be invalidated, it is moved to the
559 * inactive list to speed up its reclaim. It is moved to the
560 * head of the list, rather than the tail, to give the flusher
561 * threads some time to write it out, as this is much more
562 * effective than the single-page writeout from reclaim.
564 * If the folio isn't mapped and dirty/writeback, the folio
565 * could be reclaimed asap using the reclaim flag.
567 * 1. active, mapped folio -> none
568 * 2. active, dirty/writeback folio -> inactive, head, reclaim
569 * 3. inactive, mapped folio -> none
570 * 4. inactive, dirty/writeback folio -> inactive, head, reclaim
571 * 5. inactive, clean -> inactive, tail
574 * In 4, it moves to the head of the inactive list so the folio is
575 * written out by flusher threads as this is much more efficient
576 * than the single-page writeout from reclaim.
578 static void lru_deactivate_file_fn(struct lruvec *lruvec, struct folio *folio)
580 bool active = folio_test_active(folio);
581 long nr_pages = folio_nr_pages(folio);
583 if (folio_test_unevictable(folio))
586 /* Some processes are using the folio */
587 if (folio_mapped(folio))
590 lruvec_del_folio(lruvec, folio);
591 folio_clear_active(folio);
592 folio_clear_referenced(folio);
594 if (folio_test_writeback(folio) || folio_test_dirty(folio)) {
596 * Setting the reclaim flag could race with
597 * folio_end_writeback() and confuse readahead. But the
598 * race window is _really_ small and it's not a critical
601 lruvec_add_folio(lruvec, folio);
602 folio_set_reclaim(folio);
605 * The folio's writeback ended while it was in the batch.
606 * We move that folio to the tail of the inactive list.
608 lruvec_add_folio_tail(lruvec, folio);
609 __count_vm_events(PGROTATED, nr_pages);
613 __count_vm_events(PGDEACTIVATE, nr_pages);
614 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
619 static void lru_deactivate_fn(struct lruvec *lruvec, struct folio *folio)
621 if (!folio_test_unevictable(folio) && (folio_test_active(folio) || lru_gen_enabled())) {
622 long nr_pages = folio_nr_pages(folio);
624 lruvec_del_folio(lruvec, folio);
625 folio_clear_active(folio);
626 folio_clear_referenced(folio);
627 lruvec_add_folio(lruvec, folio);
629 __count_vm_events(PGDEACTIVATE, nr_pages);
630 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
635 static void lru_lazyfree_fn(struct lruvec *lruvec, struct folio *folio)
637 if (folio_test_anon(folio) && folio_test_swapbacked(folio) &&
638 !folio_test_swapcache(folio) && !folio_test_unevictable(folio)) {
639 long nr_pages = folio_nr_pages(folio);
641 lruvec_del_folio(lruvec, folio);
642 folio_clear_active(folio);
643 folio_clear_referenced(folio);
645 * Lazyfree folios are clean anonymous folios. They have
646 * the swapbacked flag cleared, to distinguish them from normal
649 folio_clear_swapbacked(folio);
650 lruvec_add_folio(lruvec, folio);
652 __count_vm_events(PGLAZYFREE, nr_pages);
653 __count_memcg_events(lruvec_memcg(lruvec), PGLAZYFREE,
659 * Drain pages out of the cpu's folio_batch.
660 * Either "cpu" is the current CPU, and preemption has already been
661 * disabled; or "cpu" is being hot-unplugged, and is already dead.
663 void lru_add_drain_cpu(int cpu)
665 struct cpu_fbatches *fbatches = &per_cpu(cpu_fbatches, cpu);
666 struct folio_batch *fbatch = &fbatches->lru_add;
668 if (folio_batch_count(fbatch))
669 folio_batch_move_lru(fbatch, lru_add_fn);
671 fbatch = &per_cpu(lru_rotate.fbatch, cpu);
672 /* Disabling interrupts below acts as a compiler barrier. */
673 if (data_race(folio_batch_count(fbatch))) {
676 /* No harm done if a racing interrupt already did this */
677 local_lock_irqsave(&lru_rotate.lock, flags);
678 folio_batch_move_lru(fbatch, lru_move_tail_fn);
679 local_unlock_irqrestore(&lru_rotate.lock, flags);
682 fbatch = &fbatches->lru_deactivate_file;
683 if (folio_batch_count(fbatch))
684 folio_batch_move_lru(fbatch, lru_deactivate_file_fn);
686 fbatch = &fbatches->lru_deactivate;
687 if (folio_batch_count(fbatch))
688 folio_batch_move_lru(fbatch, lru_deactivate_fn);
690 fbatch = &fbatches->lru_lazyfree;
691 if (folio_batch_count(fbatch))
692 folio_batch_move_lru(fbatch, lru_lazyfree_fn);
694 folio_activate_drain(cpu);
698 * deactivate_file_folio() - Deactivate a file folio.
699 * @folio: Folio to deactivate.
701 * This function hints to the VM that @folio is a good reclaim candidate,
702 * for example if its invalidation fails due to the folio being dirty
703 * or under writeback.
705 * Context: Caller holds a reference on the folio.
707 void deactivate_file_folio(struct folio *folio)
709 struct folio_batch *fbatch;
711 /* Deactivating an unevictable folio will not accelerate reclaim */
712 if (folio_test_unevictable(folio))
716 local_lock(&cpu_fbatches.lock);
717 fbatch = this_cpu_ptr(&cpu_fbatches.lru_deactivate_file);
718 folio_batch_add_and_move(fbatch, folio, lru_deactivate_file_fn);
719 local_unlock(&cpu_fbatches.lock);
723 * deactivate_page - deactivate a page
724 * @page: page to deactivate
726 * deactivate_page() moves @page to the inactive list if @page was on the active
727 * list and was not an unevictable page. This is done to accelerate the reclaim
730 void deactivate_page(struct page *page)
732 struct folio *folio = page_folio(page);
734 if (folio_test_lru(folio) && !folio_test_unevictable(folio) &&
735 (folio_test_active(folio) || lru_gen_enabled())) {
736 struct folio_batch *fbatch;
739 local_lock(&cpu_fbatches.lock);
740 fbatch = this_cpu_ptr(&cpu_fbatches.lru_deactivate);
741 folio_batch_add_and_move(fbatch, folio, lru_deactivate_fn);
742 local_unlock(&cpu_fbatches.lock);
747 * mark_page_lazyfree - make an anon page lazyfree
748 * @page: page to deactivate
750 * mark_page_lazyfree() moves @page to the inactive file list.
751 * This is done to accelerate the reclaim of @page.
753 void mark_page_lazyfree(struct page *page)
755 struct folio *folio = page_folio(page);
757 if (folio_test_lru(folio) && folio_test_anon(folio) &&
758 folio_test_swapbacked(folio) && !folio_test_swapcache(folio) &&
759 !folio_test_unevictable(folio)) {
760 struct folio_batch *fbatch;
763 local_lock(&cpu_fbatches.lock);
764 fbatch = this_cpu_ptr(&cpu_fbatches.lru_lazyfree);
765 folio_batch_add_and_move(fbatch, folio, lru_lazyfree_fn);
766 local_unlock(&cpu_fbatches.lock);
770 void lru_add_drain(void)
772 local_lock(&cpu_fbatches.lock);
773 lru_add_drain_cpu(smp_processor_id());
774 local_unlock(&cpu_fbatches.lock);
775 mlock_page_drain_local();
779 * It's called from per-cpu workqueue context in SMP case so
780 * lru_add_drain_cpu and invalidate_bh_lrus_cpu should run on
781 * the same cpu. It shouldn't be a problem in !SMP case since
782 * the core is only one and the locks will disable preemption.
784 static void lru_add_and_bh_lrus_drain(void)
786 local_lock(&cpu_fbatches.lock);
787 lru_add_drain_cpu(smp_processor_id());
788 local_unlock(&cpu_fbatches.lock);
789 invalidate_bh_lrus_cpu();
790 mlock_page_drain_local();
793 void lru_add_drain_cpu_zone(struct zone *zone)
795 local_lock(&cpu_fbatches.lock);
796 lru_add_drain_cpu(smp_processor_id());
797 drain_local_pages(zone);
798 local_unlock(&cpu_fbatches.lock);
799 mlock_page_drain_local();
804 static DEFINE_PER_CPU(struct work_struct, lru_add_drain_work);
806 static void lru_add_drain_per_cpu(struct work_struct *dummy)
808 lru_add_and_bh_lrus_drain();
811 static bool cpu_needs_drain(unsigned int cpu)
813 struct cpu_fbatches *fbatches = &per_cpu(cpu_fbatches, cpu);
815 /* Check these in order of likelihood that they're not zero */
816 return folio_batch_count(&fbatches->lru_add) ||
817 data_race(folio_batch_count(&per_cpu(lru_rotate.fbatch, cpu))) ||
818 folio_batch_count(&fbatches->lru_deactivate_file) ||
819 folio_batch_count(&fbatches->lru_deactivate) ||
820 folio_batch_count(&fbatches->lru_lazyfree) ||
821 folio_batch_count(&fbatches->activate) ||
822 need_mlock_page_drain(cpu) ||
823 has_bh_in_lru(cpu, NULL);
827 * Doesn't need any cpu hotplug locking because we do rely on per-cpu
828 * kworkers being shut down before our page_alloc_cpu_dead callback is
829 * executed on the offlined cpu.
830 * Calling this function with cpu hotplug locks held can actually lead
831 * to obscure indirect dependencies via WQ context.
833 static inline void __lru_add_drain_all(bool force_all_cpus)
836 * lru_drain_gen - Global pages generation number
838 * (A) Definition: global lru_drain_gen = x implies that all generations
839 * 0 < n <= x are already *scheduled* for draining.
841 * This is an optimization for the highly-contended use case where a
842 * user space workload keeps constantly generating a flow of pages for
845 static unsigned int lru_drain_gen;
846 static struct cpumask has_work;
847 static DEFINE_MUTEX(lock);
848 unsigned cpu, this_gen;
851 * Make sure nobody triggers this path before mm_percpu_wq is fully
854 if (WARN_ON(!mm_percpu_wq))
858 * Guarantee folio_batch counter stores visible by this CPU
859 * are visible to other CPUs before loading the current drain
865 * (B) Locally cache global LRU draining generation number
867 * The read barrier ensures that the counter is loaded before the mutex
868 * is taken. It pairs with smp_mb() inside the mutex critical section
871 this_gen = smp_load_acquire(&lru_drain_gen);
876 * (C) Exit the draining operation if a newer generation, from another
877 * lru_add_drain_all(), was already scheduled for draining. Check (A).
879 if (unlikely(this_gen != lru_drain_gen && !force_all_cpus))
883 * (D) Increment global generation number
885 * Pairs with smp_load_acquire() at (B), outside of the critical
886 * section. Use a full memory barrier to guarantee that the
887 * new global drain generation number is stored before loading
888 * folio_batch counters.
890 * This pairing must be done here, before the for_each_online_cpu loop
891 * below which drains the page vectors.
893 * Let x, y, and z represent some system CPU numbers, where x < y < z.
894 * Assume CPU #z is in the middle of the for_each_online_cpu loop
895 * below and has already reached CPU #y's per-cpu data. CPU #x comes
896 * along, adds some pages to its per-cpu vectors, then calls
897 * lru_add_drain_all().
899 * If the paired barrier is done at any later step, e.g. after the
900 * loop, CPU #x will just exit at (C) and miss flushing out all of its
903 WRITE_ONCE(lru_drain_gen, lru_drain_gen + 1);
906 cpumask_clear(&has_work);
907 for_each_online_cpu(cpu) {
908 struct work_struct *work = &per_cpu(lru_add_drain_work, cpu);
910 if (cpu_needs_drain(cpu)) {
911 INIT_WORK(work, lru_add_drain_per_cpu);
912 queue_work_on(cpu, mm_percpu_wq, work);
913 __cpumask_set_cpu(cpu, &has_work);
917 for_each_cpu(cpu, &has_work)
918 flush_work(&per_cpu(lru_add_drain_work, cpu));
924 void lru_add_drain_all(void)
926 __lru_add_drain_all(false);
929 void lru_add_drain_all(void)
933 #endif /* CONFIG_SMP */
935 atomic_t lru_disable_count = ATOMIC_INIT(0);
938 * lru_cache_disable() needs to be called before we start compiling
939 * a list of pages to be migrated using isolate_lru_page().
940 * It drains pages on LRU cache and then disable on all cpus until
941 * lru_cache_enable is called.
943 * Must be paired with a call to lru_cache_enable().
945 void lru_cache_disable(void)
947 atomic_inc(&lru_disable_count);
949 * Readers of lru_disable_count are protected by either disabling
950 * preemption or rcu_read_lock:
952 * preempt_disable, local_irq_disable [bh_lru_lock()]
953 * rcu_read_lock [rt_spin_lock CONFIG_PREEMPT_RT]
954 * preempt_disable [local_lock !CONFIG_PREEMPT_RT]
956 * Since v5.1 kernel, synchronize_rcu() is guaranteed to wait on
957 * preempt_disable() regions of code. So any CPU which sees
958 * lru_disable_count = 0 will have exited the critical
959 * section when synchronize_rcu() returns.
961 synchronize_rcu_expedited();
963 __lru_add_drain_all(true);
965 lru_add_and_bh_lrus_drain();
970 * release_pages - batched put_page()
971 * @pages: array of pages to release
972 * @nr: number of pages
974 * Decrement the reference count on all the pages in @pages. If it
975 * fell to zero, remove the page from the LRU and free it.
977 void release_pages(struct page **pages, int nr)
980 LIST_HEAD(pages_to_free);
981 struct lruvec *lruvec = NULL;
982 unsigned long flags = 0;
983 unsigned int lock_batch;
985 for (i = 0; i < nr; i++) {
986 struct folio *folio = page_folio(pages[i]);
989 * Make sure the IRQ-safe lock-holding time does not get
990 * excessive with a continuous string of pages from the
991 * same lruvec. The lock is held only if lruvec != NULL.
993 if (lruvec && ++lock_batch == SWAP_CLUSTER_MAX) {
994 unlock_page_lruvec_irqrestore(lruvec, flags);
998 if (is_huge_zero_page(&folio->page))
1001 if (folio_is_zone_device(folio)) {
1003 unlock_page_lruvec_irqrestore(lruvec, flags);
1006 if (put_devmap_managed_page(&folio->page))
1008 if (folio_put_testzero(folio))
1009 free_zone_device_page(&folio->page);
1013 if (!folio_put_testzero(folio))
1016 if (folio_test_large(folio)) {
1018 unlock_page_lruvec_irqrestore(lruvec, flags);
1021 __folio_put_large(folio);
1025 if (folio_test_lru(folio)) {
1026 struct lruvec *prev_lruvec = lruvec;
1028 lruvec = folio_lruvec_relock_irqsave(folio, lruvec,
1030 if (prev_lruvec != lruvec)
1033 lruvec_del_folio(lruvec, folio);
1034 __folio_clear_lru_flags(folio);
1038 * In rare cases, when truncation or holepunching raced with
1039 * munlock after VM_LOCKED was cleared, Mlocked may still be
1040 * found set here. This does not indicate a problem, unless
1041 * "unevictable_pgs_cleared" appears worryingly large.
1043 if (unlikely(folio_test_mlocked(folio))) {
1044 __folio_clear_mlocked(folio);
1045 zone_stat_sub_folio(folio, NR_MLOCK);
1046 count_vm_event(UNEVICTABLE_PGCLEARED);
1049 list_add(&folio->lru, &pages_to_free);
1052 unlock_page_lruvec_irqrestore(lruvec, flags);
1054 mem_cgroup_uncharge_list(&pages_to_free);
1055 free_unref_page_list(&pages_to_free);
1057 EXPORT_SYMBOL(release_pages);
1060 * The pages which we're about to release may be in the deferred lru-addition
1061 * queues. That would prevent them from really being freed right now. That's
1062 * OK from a correctness point of view but is inefficient - those pages may be
1063 * cache-warm and we want to give them back to the page allocator ASAP.
1065 * So __pagevec_release() will drain those queues here.
1066 * folio_batch_move_lru() calls folios_put() directly to avoid
1069 void __pagevec_release(struct pagevec *pvec)
1071 if (!pvec->percpu_pvec_drained) {
1073 pvec->percpu_pvec_drained = true;
1075 release_pages(pvec->pages, pagevec_count(pvec));
1076 pagevec_reinit(pvec);
1078 EXPORT_SYMBOL(__pagevec_release);
1081 * folio_batch_remove_exceptionals() - Prune non-folios from a batch.
1082 * @fbatch: The batch to prune
1084 * find_get_entries() fills a batch with both folios and shadow/swap/DAX
1085 * entries. This function prunes all the non-folio entries from @fbatch
1086 * without leaving holes, so that it can be passed on to folio-only batch
1089 void folio_batch_remove_exceptionals(struct folio_batch *fbatch)
1093 for (i = 0, j = 0; i < folio_batch_count(fbatch); i++) {
1094 struct folio *folio = fbatch->folios[i];
1095 if (!xa_is_value(folio))
1096 fbatch->folios[j++] = folio;
1101 unsigned pagevec_lookup_range_tag(struct pagevec *pvec,
1102 struct address_space *mapping, pgoff_t *index, pgoff_t end,
1105 pvec->nr = find_get_pages_range_tag(mapping, index, end, tag,
1106 PAGEVEC_SIZE, pvec->pages);
1107 return pagevec_count(pvec);
1109 EXPORT_SYMBOL(pagevec_lookup_range_tag);
1112 * Perform any setup for the swap system
1114 void __init swap_setup(void)
1116 unsigned long megs = totalram_pages() >> (20 - PAGE_SHIFT);
1118 /* Use a smaller cluster for small-memory machines */
1124 * Right now other parts of the system means that we
1125 * _really_ don't want to cluster much more