1 // SPDX-License-Identifier: GPL-2.0-only
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
9 * This file contains the default values for the operation of the
10 * Linux VM subsystem. Fine-tuning documentation can be found in
11 * Documentation/admin-guide/sysctl/vm.rst.
13 * Swap aging added 23.2.95, Stephen Tweedie.
14 * Buffermem limits added 12.3.98, Rik van Riel.
18 #include <linux/sched.h>
19 #include <linux/kernel_stat.h>
20 #include <linux/swap.h>
21 #include <linux/mman.h>
22 #include <linux/pagemap.h>
23 #include <linux/pagevec.h>
24 #include <linux/init.h>
25 #include <linux/export.h>
26 #include <linux/mm_inline.h>
27 #include <linux/percpu_counter.h>
28 #include <linux/memremap.h>
29 #include <linux/percpu.h>
30 #include <linux/cpu.h>
31 #include <linux/notifier.h>
32 #include <linux/backing-dev.h>
33 #include <linux/memcontrol.h>
34 #include <linux/gfp.h>
35 #include <linux/uio.h>
36 #include <linux/hugetlb.h>
37 #include <linux/page_idle.h>
38 #include <linux/local_lock.h>
39 #include <linux/buffer_head.h>
43 #define CREATE_TRACE_POINTS
44 #include <trace/events/pagemap.h>
46 /* How many pages do we try to swap or page in/out together? */
49 /* Protecting only lru_rotate.pvec which requires disabling interrupts */
54 static DEFINE_PER_CPU(struct lru_rotate
, lru_rotate
) = {
55 .lock
= INIT_LOCAL_LOCK(lock
),
59 * The following struct pagevec are grouped together because they are protected
60 * by disabling preemption (and interrupts remain enabled).
64 struct pagevec lru_add
;
65 struct pagevec lru_deactivate_file
;
66 struct pagevec lru_deactivate
;
67 struct pagevec lru_lazyfree
;
69 struct pagevec activate_page
;
72 static DEFINE_PER_CPU(struct lru_pvecs
, lru_pvecs
) = {
73 .lock
= INIT_LOCAL_LOCK(lock
),
77 * This path almost never happens for VM activity - pages are normally
78 * freed via pagevecs. But it gets used by networking.
80 static void __page_cache_release(struct page
*page
)
83 struct lruvec
*lruvec
;
86 lruvec
= lock_page_lruvec_irqsave(page
, &flags
);
87 del_page_from_lru_list(page
, lruvec
);
88 __clear_page_lru_flags(page
);
89 unlock_page_lruvec_irqrestore(lruvec
, flags
);
91 __ClearPageWaiters(page
);
94 static void __put_single_page(struct page
*page
)
96 __page_cache_release(page
);
97 mem_cgroup_uncharge(page_folio(page
));
98 free_unref_page(page
, 0);
101 static void __put_compound_page(struct page
*page
)
104 * __page_cache_release() is supposed to be called for thp, not for
105 * hugetlb. This is because hugetlb page does never have PageLRU set
106 * (it's never listed to any LRU lists) and no memcg routines should
107 * be called for hugetlb (it has a separate hugetlb_cgroup.)
110 __page_cache_release(page
);
111 destroy_compound_page(page
);
114 void __put_page(struct page
*page
)
116 if (is_zone_device_page(page
)) {
117 put_dev_pagemap(page
->pgmap
);
120 * The page belongs to the device that created pgmap. Do
121 * not return it to page allocator.
126 if (unlikely(PageCompound(page
)))
127 __put_compound_page(page
);
129 __put_single_page(page
);
131 EXPORT_SYMBOL(__put_page
);
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. Currently
138 * used by read_cache_pages() and related error recovery code.
140 void put_pages_list(struct list_head
*pages
)
142 while (!list_empty(pages
)) {
145 victim
= lru_to_page(pages
);
146 list_del(&victim
->lru
);
150 EXPORT_SYMBOL(put_pages_list
);
153 * get_kernel_pages() - pin kernel pages in memory
154 * @kiov: An array of struct kvec structures
155 * @nr_segs: number of segments to pin
156 * @write: pinning for read/write, currently ignored
157 * @pages: array that receives pointers to the pages pinned.
158 * Should be at least nr_segs long.
160 * Returns number of pages pinned. This may be fewer than the number
161 * requested. If nr_pages is 0 or negative, returns 0. If no pages
162 * were pinned, returns -errno. Each page returned must be released
163 * with a put_page() call when it is finished with.
165 int get_kernel_pages(const struct kvec
*kiov
, int nr_segs
, int write
,
170 for (seg
= 0; seg
< nr_segs
; seg
++) {
171 if (WARN_ON(kiov
[seg
].iov_len
!= PAGE_SIZE
))
174 pages
[seg
] = kmap_to_page(kiov
[seg
].iov_base
);
175 get_page(pages
[seg
]);
180 EXPORT_SYMBOL_GPL(get_kernel_pages
);
182 static void pagevec_lru_move_fn(struct pagevec
*pvec
,
183 void (*move_fn
)(struct page
*page
, struct lruvec
*lruvec
))
186 struct lruvec
*lruvec
= NULL
;
187 unsigned long flags
= 0;
189 for (i
= 0; i
< pagevec_count(pvec
); i
++) {
190 struct page
*page
= pvec
->pages
[i
];
192 /* block memcg migration during page moving between lru */
193 if (!TestClearPageLRU(page
))
196 lruvec
= relock_page_lruvec_irqsave(page
, lruvec
, &flags
);
197 (*move_fn
)(page
, lruvec
);
202 unlock_page_lruvec_irqrestore(lruvec
, flags
);
203 release_pages(pvec
->pages
, pvec
->nr
);
204 pagevec_reinit(pvec
);
207 static void pagevec_move_tail_fn(struct page
*page
, struct lruvec
*lruvec
)
209 struct folio
*folio
= page_folio(page
);
211 if (!folio_test_unevictable(folio
)) {
212 lruvec_del_folio(lruvec
, folio
);
213 folio_clear_active(folio
);
214 lruvec_add_folio_tail(lruvec
, folio
);
215 __count_vm_events(PGROTATED
, folio_nr_pages(folio
));
219 /* return true if pagevec needs to drain */
220 static bool pagevec_add_and_need_flush(struct pagevec
*pvec
, struct page
*page
)
224 if (!pagevec_add(pvec
, page
) || PageCompound(page
) ||
225 lru_cache_disabled())
232 * Writeback is about to end against a folio which has been marked for
233 * immediate reclaim. If it still appears to be reclaimable, move it
234 * to the tail of the inactive list.
236 * folio_rotate_reclaimable() must disable IRQs, to prevent nasty races.
238 void folio_rotate_reclaimable(struct folio
*folio
)
240 if (!folio_test_locked(folio
) && !folio_test_dirty(folio
) &&
241 !folio_test_unevictable(folio
) && folio_test_lru(folio
)) {
242 struct pagevec
*pvec
;
246 local_lock_irqsave(&lru_rotate
.lock
, flags
);
247 pvec
= this_cpu_ptr(&lru_rotate
.pvec
);
248 if (pagevec_add_and_need_flush(pvec
, &folio
->page
))
249 pagevec_lru_move_fn(pvec
, pagevec_move_tail_fn
);
250 local_unlock_irqrestore(&lru_rotate
.lock
, flags
);
254 void lru_note_cost(struct lruvec
*lruvec
, bool file
, unsigned int nr_pages
)
257 unsigned long lrusize
;
260 * Hold lruvec->lru_lock is safe here, since
261 * 1) The pinned lruvec in reclaim, or
262 * 2) From a pre-LRU page during refault (which also holds the
263 * rcu lock, so would be safe even if the page was on the LRU
264 * and could move simultaneously to a new lruvec).
266 spin_lock_irq(&lruvec
->lru_lock
);
267 /* Record cost event */
269 lruvec
->file_cost
+= nr_pages
;
271 lruvec
->anon_cost
+= nr_pages
;
274 * Decay previous events
276 * Because workloads change over time (and to avoid
277 * overflow) we keep these statistics as a floating
278 * average, which ends up weighing recent refaults
279 * more than old ones.
281 lrusize
= lruvec_page_state(lruvec
, NR_INACTIVE_ANON
) +
282 lruvec_page_state(lruvec
, NR_ACTIVE_ANON
) +
283 lruvec_page_state(lruvec
, NR_INACTIVE_FILE
) +
284 lruvec_page_state(lruvec
, NR_ACTIVE_FILE
);
286 if (lruvec
->file_cost
+ lruvec
->anon_cost
> lrusize
/ 4) {
287 lruvec
->file_cost
/= 2;
288 lruvec
->anon_cost
/= 2;
290 spin_unlock_irq(&lruvec
->lru_lock
);
291 } while ((lruvec
= parent_lruvec(lruvec
)));
294 void lru_note_cost_page(struct page
*page
)
296 struct folio
*folio
= page_folio(page
);
297 lru_note_cost(folio_lruvec(folio
),
298 page_is_file_lru(page
), thp_nr_pages(page
));
301 static void __activate_page(struct page
*page
, struct lruvec
*lruvec
)
303 if (!PageActive(page
) && !PageUnevictable(page
)) {
304 int nr_pages
= thp_nr_pages(page
);
306 del_page_from_lru_list(page
, lruvec
);
308 add_page_to_lru_list(page
, lruvec
);
309 trace_mm_lru_activate(page
);
311 __count_vm_events(PGACTIVATE
, nr_pages
);
312 __count_memcg_events(lruvec_memcg(lruvec
), PGACTIVATE
,
318 static void activate_page_drain(int cpu
)
320 struct pagevec
*pvec
= &per_cpu(lru_pvecs
.activate_page
, cpu
);
322 if (pagevec_count(pvec
))
323 pagevec_lru_move_fn(pvec
, __activate_page
);
326 static bool need_activate_page_drain(int cpu
)
328 return pagevec_count(&per_cpu(lru_pvecs
.activate_page
, cpu
)) != 0;
331 static void activate_page(struct page
*page
)
333 page
= compound_head(page
);
334 if (PageLRU(page
) && !PageActive(page
) && !PageUnevictable(page
)) {
335 struct pagevec
*pvec
;
337 local_lock(&lru_pvecs
.lock
);
338 pvec
= this_cpu_ptr(&lru_pvecs
.activate_page
);
340 if (pagevec_add_and_need_flush(pvec
, page
))
341 pagevec_lru_move_fn(pvec
, __activate_page
);
342 local_unlock(&lru_pvecs
.lock
);
347 static inline void activate_page_drain(int cpu
)
351 static void activate_page(struct page
*page
)
353 struct lruvec
*lruvec
;
355 page
= compound_head(page
);
356 if (TestClearPageLRU(page
)) {
357 lruvec
= lock_page_lruvec_irq(page
);
358 __activate_page(page
, lruvec
);
359 unlock_page_lruvec_irq(lruvec
);
365 static void __lru_cache_activate_page(struct page
*page
)
367 struct pagevec
*pvec
;
370 local_lock(&lru_pvecs
.lock
);
371 pvec
= this_cpu_ptr(&lru_pvecs
.lru_add
);
374 * Search backwards on the optimistic assumption that the page being
375 * activated has just been added to this pagevec. Note that only
376 * the local pagevec is examined as a !PageLRU page could be in the
377 * process of being released, reclaimed, migrated or on a remote
378 * pagevec that is currently being drained. Furthermore, marking
379 * a remote pagevec's page PageActive potentially hits a race where
380 * a page is marked PageActive just after it is added to the inactive
381 * list causing accounting errors and BUG_ON checks to trigger.
383 for (i
= pagevec_count(pvec
) - 1; i
>= 0; i
--) {
384 struct page
*pagevec_page
= pvec
->pages
[i
];
386 if (pagevec_page
== page
) {
392 local_unlock(&lru_pvecs
.lock
);
396 * Mark a page as having seen activity.
398 * inactive,unreferenced -> inactive,referenced
399 * inactive,referenced -> active,unreferenced
400 * active,unreferenced -> active,referenced
402 * When a newly allocated page is not yet visible, so safe for non-atomic ops,
403 * __SetPageReferenced(page) may be substituted for mark_page_accessed(page).
405 void mark_page_accessed(struct page
*page
)
407 page
= compound_head(page
);
409 if (!PageReferenced(page
)) {
410 SetPageReferenced(page
);
411 } else if (PageUnevictable(page
)) {
413 * Unevictable pages are on the "LRU_UNEVICTABLE" list. But,
414 * this list is never rotated or maintained, so marking an
415 * evictable page accessed has no effect.
417 } else if (!PageActive(page
)) {
419 * If the page is on the LRU, queue it for activation via
420 * lru_pvecs.activate_page. Otherwise, assume the page is on a
421 * pagevec, mark it active and it'll be moved to the active
422 * LRU on the next drain.
427 __lru_cache_activate_page(page
);
428 ClearPageReferenced(page
);
429 workingset_activation(page
);
431 if (page_is_idle(page
))
432 clear_page_idle(page
);
434 EXPORT_SYMBOL(mark_page_accessed
);
437 * lru_cache_add - add a page to a page list
438 * @page: the page to be added to the LRU.
440 * Queue the page for addition to the LRU via pagevec. The decision on whether
441 * to add the page to the [in]active [file|anon] list is deferred until the
442 * pagevec is drained. This gives a chance for the caller of lru_cache_add()
443 * have the page added to the active list using mark_page_accessed().
445 void lru_cache_add(struct page
*page
)
447 struct pagevec
*pvec
;
449 VM_BUG_ON_PAGE(PageActive(page
) && PageUnevictable(page
), page
);
450 VM_BUG_ON_PAGE(PageLRU(page
), page
);
453 local_lock(&lru_pvecs
.lock
);
454 pvec
= this_cpu_ptr(&lru_pvecs
.lru_add
);
455 if (pagevec_add_and_need_flush(pvec
, page
))
456 __pagevec_lru_add(pvec
);
457 local_unlock(&lru_pvecs
.lock
);
459 EXPORT_SYMBOL(lru_cache_add
);
462 * lru_cache_add_inactive_or_unevictable
463 * @page: the page to be added to LRU
464 * @vma: vma in which page is mapped for determining reclaimability
466 * Place @page on the inactive or unevictable LRU list, depending on its
469 void lru_cache_add_inactive_or_unevictable(struct page
*page
,
470 struct vm_area_struct
*vma
)
474 VM_BUG_ON_PAGE(PageLRU(page
), page
);
476 unevictable
= (vma
->vm_flags
& (VM_LOCKED
| VM_SPECIAL
)) == VM_LOCKED
;
477 if (unlikely(unevictable
) && !TestSetPageMlocked(page
)) {
478 int nr_pages
= thp_nr_pages(page
);
480 * We use the irq-unsafe __mod_zone_page_state because this
481 * counter is not modified from interrupt context, and the pte
482 * lock is held(spinlock), which implies preemption disabled.
484 __mod_zone_page_state(page_zone(page
), NR_MLOCK
, nr_pages
);
485 count_vm_events(UNEVICTABLE_PGMLOCKED
, nr_pages
);
491 * If the page can not be invalidated, it is moved to the
492 * inactive list to speed up its reclaim. It is moved to the
493 * head of the list, rather than the tail, to give the flusher
494 * threads some time to write it out, as this is much more
495 * effective than the single-page writeout from reclaim.
497 * If the page isn't page_mapped and dirty/writeback, the page
498 * could reclaim asap using PG_reclaim.
500 * 1. active, mapped page -> none
501 * 2. active, dirty/writeback page -> inactive, head, PG_reclaim
502 * 3. inactive, mapped page -> none
503 * 4. inactive, dirty/writeback page -> inactive, head, PG_reclaim
504 * 5. inactive, clean -> inactive, tail
507 * In 4, why it moves inactive's head, the VM expects the page would
508 * be write it out by flusher threads as this is much more effective
509 * than the single-page writeout from reclaim.
511 static void lru_deactivate_file_fn(struct page
*page
, struct lruvec
*lruvec
)
513 bool active
= PageActive(page
);
514 int nr_pages
= thp_nr_pages(page
);
516 if (PageUnevictable(page
))
519 /* Some processes are using the page */
520 if (page_mapped(page
))
523 del_page_from_lru_list(page
, lruvec
);
524 ClearPageActive(page
);
525 ClearPageReferenced(page
);
527 if (PageWriteback(page
) || PageDirty(page
)) {
529 * PG_reclaim could be raced with end_page_writeback
530 * It can make readahead confusing. But race window
531 * is _really_ small and it's non-critical problem.
533 add_page_to_lru_list(page
, lruvec
);
534 SetPageReclaim(page
);
537 * The page's writeback ends up during pagevec
538 * We move that page into tail of inactive.
540 add_page_to_lru_list_tail(page
, lruvec
);
541 __count_vm_events(PGROTATED
, nr_pages
);
545 __count_vm_events(PGDEACTIVATE
, nr_pages
);
546 __count_memcg_events(lruvec_memcg(lruvec
), PGDEACTIVATE
,
551 static void lru_deactivate_fn(struct page
*page
, struct lruvec
*lruvec
)
553 if (PageActive(page
) && !PageUnevictable(page
)) {
554 int nr_pages
= thp_nr_pages(page
);
556 del_page_from_lru_list(page
, lruvec
);
557 ClearPageActive(page
);
558 ClearPageReferenced(page
);
559 add_page_to_lru_list(page
, lruvec
);
561 __count_vm_events(PGDEACTIVATE
, nr_pages
);
562 __count_memcg_events(lruvec_memcg(lruvec
), PGDEACTIVATE
,
567 static void lru_lazyfree_fn(struct page
*page
, struct lruvec
*lruvec
)
569 if (PageAnon(page
) && PageSwapBacked(page
) &&
570 !PageSwapCache(page
) && !PageUnevictable(page
)) {
571 int nr_pages
= thp_nr_pages(page
);
573 del_page_from_lru_list(page
, lruvec
);
574 ClearPageActive(page
);
575 ClearPageReferenced(page
);
577 * Lazyfree pages are clean anonymous pages. They have
578 * PG_swapbacked flag cleared, to distinguish them from normal
581 ClearPageSwapBacked(page
);
582 add_page_to_lru_list(page
, lruvec
);
584 __count_vm_events(PGLAZYFREE
, nr_pages
);
585 __count_memcg_events(lruvec_memcg(lruvec
), PGLAZYFREE
,
591 * Drain pages out of the cpu's pagevecs.
592 * Either "cpu" is the current CPU, and preemption has already been
593 * disabled; or "cpu" is being hot-unplugged, and is already dead.
595 void lru_add_drain_cpu(int cpu
)
597 struct pagevec
*pvec
= &per_cpu(lru_pvecs
.lru_add
, cpu
);
599 if (pagevec_count(pvec
))
600 __pagevec_lru_add(pvec
);
602 pvec
= &per_cpu(lru_rotate
.pvec
, cpu
);
603 /* Disabling interrupts below acts as a compiler barrier. */
604 if (data_race(pagevec_count(pvec
))) {
607 /* No harm done if a racing interrupt already did this */
608 local_lock_irqsave(&lru_rotate
.lock
, flags
);
609 pagevec_lru_move_fn(pvec
, pagevec_move_tail_fn
);
610 local_unlock_irqrestore(&lru_rotate
.lock
, flags
);
613 pvec
= &per_cpu(lru_pvecs
.lru_deactivate_file
, cpu
);
614 if (pagevec_count(pvec
))
615 pagevec_lru_move_fn(pvec
, lru_deactivate_file_fn
);
617 pvec
= &per_cpu(lru_pvecs
.lru_deactivate
, cpu
);
618 if (pagevec_count(pvec
))
619 pagevec_lru_move_fn(pvec
, lru_deactivate_fn
);
621 pvec
= &per_cpu(lru_pvecs
.lru_lazyfree
, cpu
);
622 if (pagevec_count(pvec
))
623 pagevec_lru_move_fn(pvec
, lru_lazyfree_fn
);
625 activate_page_drain(cpu
);
629 * deactivate_file_page - forcefully deactivate a file page
630 * @page: page to deactivate
632 * This function hints the VM that @page is a good reclaim candidate,
633 * for example if its invalidation fails due to the page being dirty
634 * or under writeback.
636 void deactivate_file_page(struct page
*page
)
639 * In a workload with many unevictable page such as mprotect,
640 * unevictable page deactivation for accelerating reclaim is pointless.
642 if (PageUnevictable(page
))
645 if (likely(get_page_unless_zero(page
))) {
646 struct pagevec
*pvec
;
648 local_lock(&lru_pvecs
.lock
);
649 pvec
= this_cpu_ptr(&lru_pvecs
.lru_deactivate_file
);
651 if (pagevec_add_and_need_flush(pvec
, page
))
652 pagevec_lru_move_fn(pvec
, lru_deactivate_file_fn
);
653 local_unlock(&lru_pvecs
.lock
);
658 * deactivate_page - deactivate a page
659 * @page: page to deactivate
661 * deactivate_page() moves @page to the inactive list if @page was on the active
662 * list and was not an unevictable page. This is done to accelerate the reclaim
665 void deactivate_page(struct page
*page
)
667 if (PageLRU(page
) && PageActive(page
) && !PageUnevictable(page
)) {
668 struct pagevec
*pvec
;
670 local_lock(&lru_pvecs
.lock
);
671 pvec
= this_cpu_ptr(&lru_pvecs
.lru_deactivate
);
673 if (pagevec_add_and_need_flush(pvec
, page
))
674 pagevec_lru_move_fn(pvec
, lru_deactivate_fn
);
675 local_unlock(&lru_pvecs
.lock
);
680 * mark_page_lazyfree - make an anon page lazyfree
681 * @page: page to deactivate
683 * mark_page_lazyfree() moves @page to the inactive file list.
684 * This is done to accelerate the reclaim of @page.
686 void mark_page_lazyfree(struct page
*page
)
688 if (PageLRU(page
) && PageAnon(page
) && PageSwapBacked(page
) &&
689 !PageSwapCache(page
) && !PageUnevictable(page
)) {
690 struct pagevec
*pvec
;
692 local_lock(&lru_pvecs
.lock
);
693 pvec
= this_cpu_ptr(&lru_pvecs
.lru_lazyfree
);
695 if (pagevec_add_and_need_flush(pvec
, page
))
696 pagevec_lru_move_fn(pvec
, lru_lazyfree_fn
);
697 local_unlock(&lru_pvecs
.lock
);
701 void lru_add_drain(void)
703 local_lock(&lru_pvecs
.lock
);
704 lru_add_drain_cpu(smp_processor_id());
705 local_unlock(&lru_pvecs
.lock
);
709 * It's called from per-cpu workqueue context in SMP case so
710 * lru_add_drain_cpu and invalidate_bh_lrus_cpu should run on
711 * the same cpu. It shouldn't be a problem in !SMP case since
712 * the core is only one and the locks will disable preemption.
714 static void lru_add_and_bh_lrus_drain(void)
716 local_lock(&lru_pvecs
.lock
);
717 lru_add_drain_cpu(smp_processor_id());
718 local_unlock(&lru_pvecs
.lock
);
719 invalidate_bh_lrus_cpu();
722 void lru_add_drain_cpu_zone(struct zone
*zone
)
724 local_lock(&lru_pvecs
.lock
);
725 lru_add_drain_cpu(smp_processor_id());
726 drain_local_pages(zone
);
727 local_unlock(&lru_pvecs
.lock
);
732 static DEFINE_PER_CPU(struct work_struct
, lru_add_drain_work
);
734 static void lru_add_drain_per_cpu(struct work_struct
*dummy
)
736 lru_add_and_bh_lrus_drain();
740 * Doesn't need any cpu hotplug locking because we do rely on per-cpu
741 * kworkers being shut down before our page_alloc_cpu_dead callback is
742 * executed on the offlined cpu.
743 * Calling this function with cpu hotplug locks held can actually lead
744 * to obscure indirect dependencies via WQ context.
746 inline void __lru_add_drain_all(bool force_all_cpus
)
749 * lru_drain_gen - Global pages generation number
751 * (A) Definition: global lru_drain_gen = x implies that all generations
752 * 0 < n <= x are already *scheduled* for draining.
754 * This is an optimization for the highly-contended use case where a
755 * user space workload keeps constantly generating a flow of pages for
758 static unsigned int lru_drain_gen
;
759 static struct cpumask has_work
;
760 static DEFINE_MUTEX(lock
);
761 unsigned cpu
, this_gen
;
764 * Make sure nobody triggers this path before mm_percpu_wq is fully
767 if (WARN_ON(!mm_percpu_wq
))
771 * Guarantee pagevec counter stores visible by this CPU are visible to
772 * other CPUs before loading the current drain generation.
777 * (B) Locally cache global LRU draining generation number
779 * The read barrier ensures that the counter is loaded before the mutex
780 * is taken. It pairs with smp_mb() inside the mutex critical section
783 this_gen
= smp_load_acquire(&lru_drain_gen
);
788 * (C) Exit the draining operation if a newer generation, from another
789 * lru_add_drain_all(), was already scheduled for draining. Check (A).
791 if (unlikely(this_gen
!= lru_drain_gen
&& !force_all_cpus
))
795 * (D) Increment global generation number
797 * Pairs with smp_load_acquire() at (B), outside of the critical
798 * section. Use a full memory barrier to guarantee that the new global
799 * drain generation number is stored before loading pagevec counters.
801 * This pairing must be done here, before the for_each_online_cpu loop
802 * below which drains the page vectors.
804 * Let x, y, and z represent some system CPU numbers, where x < y < z.
805 * Assume CPU #z is in the middle of the for_each_online_cpu loop
806 * below and has already reached CPU #y's per-cpu data. CPU #x comes
807 * along, adds some pages to its per-cpu vectors, then calls
808 * lru_add_drain_all().
810 * If the paired barrier is done at any later step, e.g. after the
811 * loop, CPU #x will just exit at (C) and miss flushing out all of its
814 WRITE_ONCE(lru_drain_gen
, lru_drain_gen
+ 1);
817 cpumask_clear(&has_work
);
818 for_each_online_cpu(cpu
) {
819 struct work_struct
*work
= &per_cpu(lru_add_drain_work
, cpu
);
821 if (force_all_cpus
||
822 pagevec_count(&per_cpu(lru_pvecs
.lru_add
, cpu
)) ||
823 data_race(pagevec_count(&per_cpu(lru_rotate
.pvec
, cpu
))) ||
824 pagevec_count(&per_cpu(lru_pvecs
.lru_deactivate_file
, cpu
)) ||
825 pagevec_count(&per_cpu(lru_pvecs
.lru_deactivate
, cpu
)) ||
826 pagevec_count(&per_cpu(lru_pvecs
.lru_lazyfree
, cpu
)) ||
827 need_activate_page_drain(cpu
) ||
828 has_bh_in_lru(cpu
, NULL
)) {
829 INIT_WORK(work
, lru_add_drain_per_cpu
);
830 queue_work_on(cpu
, mm_percpu_wq
, work
);
831 __cpumask_set_cpu(cpu
, &has_work
);
835 for_each_cpu(cpu
, &has_work
)
836 flush_work(&per_cpu(lru_add_drain_work
, cpu
));
842 void lru_add_drain_all(void)
844 __lru_add_drain_all(false);
847 void lru_add_drain_all(void)
851 #endif /* CONFIG_SMP */
853 atomic_t lru_disable_count
= ATOMIC_INIT(0);
856 * lru_cache_disable() needs to be called before we start compiling
857 * a list of pages to be migrated using isolate_lru_page().
858 * It drains pages on LRU cache and then disable on all cpus until
859 * lru_cache_enable is called.
861 * Must be paired with a call to lru_cache_enable().
863 void lru_cache_disable(void)
865 atomic_inc(&lru_disable_count
);
868 * lru_add_drain_all in the force mode will schedule draining on
869 * all online CPUs so any calls of lru_cache_disabled wrapped by
870 * local_lock or preemption disabled would be ordered by that.
871 * The atomic operation doesn't need to have stronger ordering
872 * requirements because that is enforeced by the scheduling
875 __lru_add_drain_all(true);
877 lru_add_and_bh_lrus_drain();
882 * release_pages - batched put_page()
883 * @pages: array of pages to release
884 * @nr: number of pages
886 * Decrement the reference count on all the pages in @pages. If it
887 * fell to zero, remove the page from the LRU and free it.
889 void release_pages(struct page
**pages
, int nr
)
892 LIST_HEAD(pages_to_free
);
893 struct lruvec
*lruvec
= NULL
;
895 unsigned int lock_batch
;
897 for (i
= 0; i
< nr
; i
++) {
898 struct page
*page
= pages
[i
];
901 * Make sure the IRQ-safe lock-holding time does not get
902 * excessive with a continuous string of pages from the
903 * same lruvec. The lock is held only if lruvec != NULL.
905 if (lruvec
&& ++lock_batch
== SWAP_CLUSTER_MAX
) {
906 unlock_page_lruvec_irqrestore(lruvec
, flags
);
910 page
= compound_head(page
);
911 if (is_huge_zero_page(page
))
914 if (is_zone_device_page(page
)) {
916 unlock_page_lruvec_irqrestore(lruvec
, flags
);
920 * ZONE_DEVICE pages that return 'false' from
921 * page_is_devmap_managed() do not require special
922 * processing, and instead, expect a call to
923 * put_page_testzero().
925 if (page_is_devmap_managed(page
)) {
926 put_devmap_managed_page(page
);
929 if (put_page_testzero(page
))
930 put_dev_pagemap(page
->pgmap
);
934 if (!put_page_testzero(page
))
937 if (PageCompound(page
)) {
939 unlock_page_lruvec_irqrestore(lruvec
, flags
);
942 __put_compound_page(page
);
947 struct lruvec
*prev_lruvec
= lruvec
;
949 lruvec
= relock_page_lruvec_irqsave(page
, lruvec
,
951 if (prev_lruvec
!= lruvec
)
954 del_page_from_lru_list(page
, lruvec
);
955 __clear_page_lru_flags(page
);
958 __ClearPageWaiters(page
);
960 list_add(&page
->lru
, &pages_to_free
);
963 unlock_page_lruvec_irqrestore(lruvec
, flags
);
965 mem_cgroup_uncharge_list(&pages_to_free
);
966 free_unref_page_list(&pages_to_free
);
968 EXPORT_SYMBOL(release_pages
);
971 * The pages which we're about to release may be in the deferred lru-addition
972 * queues. That would prevent them from really being freed right now. That's
973 * OK from a correctness point of view but is inefficient - those pages may be
974 * cache-warm and we want to give them back to the page allocator ASAP.
976 * So __pagevec_release() will drain those queues here. __pagevec_lru_add()
977 * and __pagevec_lru_add_active() call release_pages() directly to avoid
980 void __pagevec_release(struct pagevec
*pvec
)
982 if (!pvec
->percpu_pvec_drained
) {
984 pvec
->percpu_pvec_drained
= true;
986 release_pages(pvec
->pages
, pagevec_count(pvec
));
987 pagevec_reinit(pvec
);
989 EXPORT_SYMBOL(__pagevec_release
);
991 static void __pagevec_lru_add_fn(struct page
*page
, struct lruvec
*lruvec
)
993 int was_unevictable
= TestClearPageUnevictable(page
);
994 int nr_pages
= thp_nr_pages(page
);
996 VM_BUG_ON_PAGE(PageLRU(page
), page
);
999 * Page becomes evictable in two ways:
1000 * 1) Within LRU lock [munlock_vma_page() and __munlock_pagevec()].
1001 * 2) Before acquiring LRU lock to put the page to correct LRU and then
1002 * a) do PageLRU check with lock [check_move_unevictable_pages]
1003 * b) do PageLRU check before lock [clear_page_mlock]
1005 * (1) & (2a) are ok as LRU lock will serialize them. For (2b), we need
1006 * following strict ordering:
1008 * #0: __pagevec_lru_add_fn #1: clear_page_mlock
1010 * SetPageLRU() TestClearPageMlocked()
1011 * smp_mb() // explicit ordering // above provides strict
1013 * PageMlocked() PageLRU()
1016 * if '#1' does not observe setting of PG_lru by '#0' and fails
1017 * isolation, the explicit barrier will make sure that page_evictable
1018 * check will put the page in correct LRU. Without smp_mb(), SetPageLRU
1019 * can be reordered after PageMlocked check and can make '#1' to fail
1020 * the isolation of the page whose Mlocked bit is cleared (#0 is also
1021 * looking at the same page) and the evictable page will be stranded
1022 * in an unevictable LRU.
1025 smp_mb__after_atomic();
1027 if (page_evictable(page
)) {
1028 if (was_unevictable
)
1029 __count_vm_events(UNEVICTABLE_PGRESCUED
, nr_pages
);
1031 ClearPageActive(page
);
1032 SetPageUnevictable(page
);
1033 if (!was_unevictable
)
1034 __count_vm_events(UNEVICTABLE_PGCULLED
, nr_pages
);
1037 add_page_to_lru_list(page
, lruvec
);
1038 trace_mm_lru_insertion(page
);
1042 * Add the passed pages to the LRU, then drop the caller's refcount
1043 * on them. Reinitialises the caller's pagevec.
1045 void __pagevec_lru_add(struct pagevec
*pvec
)
1048 struct lruvec
*lruvec
= NULL
;
1049 unsigned long flags
= 0;
1051 for (i
= 0; i
< pagevec_count(pvec
); i
++) {
1052 struct page
*page
= pvec
->pages
[i
];
1054 lruvec
= relock_page_lruvec_irqsave(page
, lruvec
, &flags
);
1055 __pagevec_lru_add_fn(page
, lruvec
);
1058 unlock_page_lruvec_irqrestore(lruvec
, flags
);
1059 release_pages(pvec
->pages
, pvec
->nr
);
1060 pagevec_reinit(pvec
);
1064 * pagevec_remove_exceptionals - pagevec exceptionals pruning
1065 * @pvec: The pagevec to prune
1067 * find_get_entries() fills both pages and XArray value entries (aka
1068 * exceptional entries) into the pagevec. This function prunes all
1069 * exceptionals from @pvec without leaving holes, so that it can be
1070 * passed on to page-only pagevec operations.
1072 void pagevec_remove_exceptionals(struct pagevec
*pvec
)
1076 for (i
= 0, j
= 0; i
< pagevec_count(pvec
); i
++) {
1077 struct page
*page
= pvec
->pages
[i
];
1078 if (!xa_is_value(page
))
1079 pvec
->pages
[j
++] = page
;
1085 * pagevec_lookup_range - gang pagecache lookup
1086 * @pvec: Where the resulting pages are placed
1087 * @mapping: The address_space to search
1088 * @start: The starting page index
1089 * @end: The final page index
1091 * pagevec_lookup_range() will search for & return a group of up to PAGEVEC_SIZE
1092 * pages in the mapping starting from index @start and upto index @end
1093 * (inclusive). The pages are placed in @pvec. pagevec_lookup() takes a
1094 * reference against the pages in @pvec.
1096 * The search returns a group of mapping-contiguous pages with ascending
1097 * indexes. There may be holes in the indices due to not-present pages. We
1098 * also update @start to index the next page for the traversal.
1100 * pagevec_lookup_range() returns the number of pages which were found. If this
1101 * number is smaller than PAGEVEC_SIZE, the end of specified range has been
1104 unsigned pagevec_lookup_range(struct pagevec
*pvec
,
1105 struct address_space
*mapping
, pgoff_t
*start
, pgoff_t end
)
1107 pvec
->nr
= find_get_pages_range(mapping
, start
, end
, PAGEVEC_SIZE
,
1109 return pagevec_count(pvec
);
1111 EXPORT_SYMBOL(pagevec_lookup_range
);
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
1141 #ifdef CONFIG_DEV_PAGEMAP_OPS
1142 void put_devmap_managed_page(struct page
*page
)
1146 if (WARN_ON_ONCE(!page_is_devmap_managed(page
)))
1149 count
= page_ref_dec_return(page
);
1152 * devmap page refcounts are 1-based, rather than 0-based: if
1153 * refcount is 1, then the page is free and the refcount is
1154 * stable because nobody holds a reference on the page.
1157 free_devmap_managed_page(page
);
1161 EXPORT_SYMBOL(put_devmap_managed_page
);