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>
42 #define CREATE_TRACE_POINTS
43 #include <trace/events/pagemap.h>
45 /* How many pages do we try to swap or page in/out together? */
48 /* Protecting only lru_rotate.pvec which requires disabling interrupts */
53 static DEFINE_PER_CPU(struct lru_rotate
, lru_rotate
) = {
54 .lock
= INIT_LOCAL_LOCK(lock
),
58 * The following struct pagevec are grouped together because they are protected
59 * by disabling preemption (and interrupts remain enabled).
63 struct pagevec lru_add
;
64 struct pagevec lru_deactivate_file
;
65 struct pagevec lru_deactivate
;
66 struct pagevec lru_lazyfree
;
68 struct pagevec activate_page
;
71 static DEFINE_PER_CPU(struct lru_pvecs
, lru_pvecs
) = {
72 .lock
= INIT_LOCAL_LOCK(lock
),
76 * This path almost never happens for VM activity - pages are normally
77 * freed via pagevecs. But it gets used by networking.
79 static void __page_cache_release(struct page
*page
)
82 pg_data_t
*pgdat
= page_pgdat(page
);
83 struct lruvec
*lruvec
;
86 spin_lock_irqsave(&pgdat
->lru_lock
, flags
);
87 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
88 VM_BUG_ON_PAGE(!PageLRU(page
), page
);
90 del_page_from_lru_list(page
, lruvec
, page_off_lru(page
));
91 spin_unlock_irqrestore(&pgdat
->lru_lock
, flags
);
93 __ClearPageWaiters(page
);
96 static void __put_single_page(struct page
*page
)
98 __page_cache_release(page
);
99 mem_cgroup_uncharge(page
);
100 free_unref_page(page
);
103 static void __put_compound_page(struct page
*page
)
106 * __page_cache_release() is supposed to be called for thp, not for
107 * hugetlb. This is because hugetlb page does never have PageLRU set
108 * (it's never listed to any LRU lists) and no memcg routines should
109 * be called for hugetlb (it has a separate hugetlb_cgroup.)
112 __page_cache_release(page
);
113 destroy_compound_page(page
);
116 void __put_page(struct page
*page
)
118 if (is_zone_device_page(page
)) {
119 put_dev_pagemap(page
->pgmap
);
122 * The page belongs to the device that created pgmap. Do
123 * not return it to page allocator.
128 if (unlikely(PageCompound(page
)))
129 __put_compound_page(page
);
131 __put_single_page(page
);
133 EXPORT_SYMBOL(__put_page
);
136 * put_pages_list() - release a list of pages
137 * @pages: list of pages threaded on page->lru
139 * Release a list of pages which are strung together on page.lru. Currently
140 * used by read_cache_pages() and related error recovery code.
142 void put_pages_list(struct list_head
*pages
)
144 while (!list_empty(pages
)) {
147 victim
= lru_to_page(pages
);
148 list_del(&victim
->lru
);
152 EXPORT_SYMBOL(put_pages_list
);
155 * get_kernel_pages() - pin kernel pages in memory
156 * @kiov: An array of struct kvec structures
157 * @nr_segs: number of segments to pin
158 * @write: pinning for read/write, currently ignored
159 * @pages: array that receives pointers to the pages pinned.
160 * Should be at least nr_segs long.
162 * Returns number of pages pinned. This may be fewer than the number
163 * requested. If nr_pages is 0 or negative, returns 0. If no pages
164 * were pinned, returns -errno. Each page returned must be released
165 * with a put_page() call when it is finished with.
167 int get_kernel_pages(const struct kvec
*kiov
, int nr_segs
, int write
,
172 for (seg
= 0; seg
< nr_segs
; seg
++) {
173 if (WARN_ON(kiov
[seg
].iov_len
!= PAGE_SIZE
))
176 pages
[seg
] = kmap_to_page(kiov
[seg
].iov_base
);
177 get_page(pages
[seg
]);
182 EXPORT_SYMBOL_GPL(get_kernel_pages
);
185 * get_kernel_page() - pin a kernel page in memory
186 * @start: starting kernel address
187 * @write: pinning for read/write, currently ignored
188 * @pages: array that receives pointer to the page pinned.
189 * Must be at least nr_segs long.
191 * Returns 1 if page is pinned. If the page was not pinned, returns
192 * -errno. The page returned must be released with a put_page() call
193 * when it is finished with.
195 int get_kernel_page(unsigned long start
, int write
, struct page
**pages
)
197 const struct kvec kiov
= {
198 .iov_base
= (void *)start
,
202 return get_kernel_pages(&kiov
, 1, write
, pages
);
204 EXPORT_SYMBOL_GPL(get_kernel_page
);
206 static void pagevec_lru_move_fn(struct pagevec
*pvec
,
207 void (*move_fn
)(struct page
*page
, struct lruvec
*lruvec
, void *arg
),
211 struct pglist_data
*pgdat
= NULL
;
212 struct lruvec
*lruvec
;
213 unsigned long flags
= 0;
215 for (i
= 0; i
< pagevec_count(pvec
); i
++) {
216 struct page
*page
= pvec
->pages
[i
];
217 struct pglist_data
*pagepgdat
= page_pgdat(page
);
219 if (pagepgdat
!= pgdat
) {
221 spin_unlock_irqrestore(&pgdat
->lru_lock
, flags
);
223 spin_lock_irqsave(&pgdat
->lru_lock
, flags
);
226 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
227 (*move_fn
)(page
, lruvec
, arg
);
230 spin_unlock_irqrestore(&pgdat
->lru_lock
, flags
);
231 release_pages(pvec
->pages
, pvec
->nr
);
232 pagevec_reinit(pvec
);
235 static void pagevec_move_tail_fn(struct page
*page
, struct lruvec
*lruvec
,
240 if (PageLRU(page
) && !PageUnevictable(page
)) {
241 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
242 ClearPageActive(page
);
243 add_page_to_lru_list_tail(page
, lruvec
, page_lru(page
));
244 (*pgmoved
) += thp_nr_pages(page
);
249 * pagevec_move_tail() must be called with IRQ disabled.
250 * Otherwise this may cause nasty races.
252 static void pagevec_move_tail(struct pagevec
*pvec
)
256 pagevec_lru_move_fn(pvec
, pagevec_move_tail_fn
, &pgmoved
);
257 __count_vm_events(PGROTATED
, pgmoved
);
261 * Writeback is about to end against a page which has been marked for immediate
262 * reclaim. If it still appears to be reclaimable, move it to the tail of the
265 void rotate_reclaimable_page(struct page
*page
)
267 if (!PageLocked(page
) && !PageDirty(page
) &&
268 !PageUnevictable(page
) && PageLRU(page
)) {
269 struct pagevec
*pvec
;
273 local_lock_irqsave(&lru_rotate
.lock
, flags
);
274 pvec
= this_cpu_ptr(&lru_rotate
.pvec
);
275 if (!pagevec_add(pvec
, page
) || PageCompound(page
))
276 pagevec_move_tail(pvec
);
277 local_unlock_irqrestore(&lru_rotate
.lock
, flags
);
281 void lru_note_cost(struct lruvec
*lruvec
, bool file
, unsigned int nr_pages
)
284 unsigned long lrusize
;
286 /* Record cost event */
288 lruvec
->file_cost
+= nr_pages
;
290 lruvec
->anon_cost
+= nr_pages
;
293 * Decay previous events
295 * Because workloads change over time (and to avoid
296 * overflow) we keep these statistics as a floating
297 * average, which ends up weighing recent refaults
298 * more than old ones.
300 lrusize
= lruvec_page_state(lruvec
, NR_INACTIVE_ANON
) +
301 lruvec_page_state(lruvec
, NR_ACTIVE_ANON
) +
302 lruvec_page_state(lruvec
, NR_INACTIVE_FILE
) +
303 lruvec_page_state(lruvec
, NR_ACTIVE_FILE
);
305 if (lruvec
->file_cost
+ lruvec
->anon_cost
> lrusize
/ 4) {
306 lruvec
->file_cost
/= 2;
307 lruvec
->anon_cost
/= 2;
309 } while ((lruvec
= parent_lruvec(lruvec
)));
312 void lru_note_cost_page(struct page
*page
)
314 lru_note_cost(mem_cgroup_page_lruvec(page
, page_pgdat(page
)),
315 page_is_file_lru(page
), thp_nr_pages(page
));
318 static void __activate_page(struct page
*page
, struct lruvec
*lruvec
,
321 if (PageLRU(page
) && !PageActive(page
) && !PageUnevictable(page
)) {
322 int lru
= page_lru_base_type(page
);
323 int nr_pages
= thp_nr_pages(page
);
325 del_page_from_lru_list(page
, lruvec
, lru
);
328 add_page_to_lru_list(page
, lruvec
, lru
);
329 trace_mm_lru_activate(page
);
331 __count_vm_events(PGACTIVATE
, nr_pages
);
332 __count_memcg_events(lruvec_memcg(lruvec
), PGACTIVATE
,
338 static void activate_page_drain(int cpu
)
340 struct pagevec
*pvec
= &per_cpu(lru_pvecs
.activate_page
, cpu
);
342 if (pagevec_count(pvec
))
343 pagevec_lru_move_fn(pvec
, __activate_page
, NULL
);
346 static bool need_activate_page_drain(int cpu
)
348 return pagevec_count(&per_cpu(lru_pvecs
.activate_page
, cpu
)) != 0;
351 static void activate_page(struct page
*page
)
353 page
= compound_head(page
);
354 if (PageLRU(page
) && !PageActive(page
) && !PageUnevictable(page
)) {
355 struct pagevec
*pvec
;
357 local_lock(&lru_pvecs
.lock
);
358 pvec
= this_cpu_ptr(&lru_pvecs
.activate_page
);
360 if (!pagevec_add(pvec
, page
) || PageCompound(page
))
361 pagevec_lru_move_fn(pvec
, __activate_page
, NULL
);
362 local_unlock(&lru_pvecs
.lock
);
367 static inline void activate_page_drain(int cpu
)
371 static void activate_page(struct page
*page
)
373 pg_data_t
*pgdat
= page_pgdat(page
);
375 page
= compound_head(page
);
376 spin_lock_irq(&pgdat
->lru_lock
);
377 __activate_page(page
, mem_cgroup_page_lruvec(page
, pgdat
), NULL
);
378 spin_unlock_irq(&pgdat
->lru_lock
);
382 static void __lru_cache_activate_page(struct page
*page
)
384 struct pagevec
*pvec
;
387 local_lock(&lru_pvecs
.lock
);
388 pvec
= this_cpu_ptr(&lru_pvecs
.lru_add
);
391 * Search backwards on the optimistic assumption that the page being
392 * activated has just been added to this pagevec. Note that only
393 * the local pagevec is examined as a !PageLRU page could be in the
394 * process of being released, reclaimed, migrated or on a remote
395 * pagevec that is currently being drained. Furthermore, marking
396 * a remote pagevec's page PageActive potentially hits a race where
397 * a page is marked PageActive just after it is added to the inactive
398 * list causing accounting errors and BUG_ON checks to trigger.
400 for (i
= pagevec_count(pvec
) - 1; i
>= 0; i
--) {
401 struct page
*pagevec_page
= pvec
->pages
[i
];
403 if (pagevec_page
== page
) {
409 local_unlock(&lru_pvecs
.lock
);
413 * Mark a page as having seen activity.
415 * inactive,unreferenced -> inactive,referenced
416 * inactive,referenced -> active,unreferenced
417 * active,unreferenced -> active,referenced
419 * When a newly allocated page is not yet visible, so safe for non-atomic ops,
420 * __SetPageReferenced(page) may be substituted for mark_page_accessed(page).
422 void mark_page_accessed(struct page
*page
)
424 page
= compound_head(page
);
426 if (!PageReferenced(page
)) {
427 SetPageReferenced(page
);
428 } else if (PageUnevictable(page
)) {
430 * Unevictable pages are on the "LRU_UNEVICTABLE" list. But,
431 * this list is never rotated or maintained, so marking an
432 * evictable page accessed has no effect.
434 } else if (!PageActive(page
)) {
436 * If the page is on the LRU, queue it for activation via
437 * lru_pvecs.activate_page. Otherwise, assume the page is on a
438 * pagevec, mark it active and it'll be moved to the active
439 * LRU on the next drain.
444 __lru_cache_activate_page(page
);
445 ClearPageReferenced(page
);
446 workingset_activation(page
);
448 if (page_is_idle(page
))
449 clear_page_idle(page
);
451 EXPORT_SYMBOL(mark_page_accessed
);
454 * lru_cache_add - add a page to a page list
455 * @page: the page to be added to the LRU.
457 * Queue the page for addition to the LRU via pagevec. The decision on whether
458 * to add the page to the [in]active [file|anon] list is deferred until the
459 * pagevec is drained. This gives a chance for the caller of lru_cache_add()
460 * have the page added to the active list using mark_page_accessed().
462 void lru_cache_add(struct page
*page
)
464 struct pagevec
*pvec
;
466 VM_BUG_ON_PAGE(PageActive(page
) && PageUnevictable(page
), page
);
467 VM_BUG_ON_PAGE(PageLRU(page
), page
);
470 local_lock(&lru_pvecs
.lock
);
471 pvec
= this_cpu_ptr(&lru_pvecs
.lru_add
);
472 if (!pagevec_add(pvec
, page
) || PageCompound(page
))
473 __pagevec_lru_add(pvec
);
474 local_unlock(&lru_pvecs
.lock
);
476 EXPORT_SYMBOL(lru_cache_add
);
479 * lru_cache_add_inactive_or_unevictable
480 * @page: the page to be added to LRU
481 * @vma: vma in which page is mapped for determining reclaimability
483 * Place @page on the inactive or unevictable LRU list, depending on its
486 void lru_cache_add_inactive_or_unevictable(struct page
*page
,
487 struct vm_area_struct
*vma
)
491 VM_BUG_ON_PAGE(PageLRU(page
), page
);
493 unevictable
= (vma
->vm_flags
& (VM_LOCKED
| VM_SPECIAL
)) == VM_LOCKED
;
494 if (unlikely(unevictable
) && !TestSetPageMlocked(page
)) {
495 int nr_pages
= thp_nr_pages(page
);
497 * We use the irq-unsafe __mod_zone_page_stat because this
498 * counter is not modified from interrupt context, and the pte
499 * lock is held(spinlock), which implies preemption disabled.
501 __mod_zone_page_state(page_zone(page
), NR_MLOCK
, nr_pages
);
502 count_vm_events(UNEVICTABLE_PGMLOCKED
, nr_pages
);
508 * If the page can not be invalidated, it is moved to the
509 * inactive list to speed up its reclaim. It is moved to the
510 * head of the list, rather than the tail, to give the flusher
511 * threads some time to write it out, as this is much more
512 * effective than the single-page writeout from reclaim.
514 * If the page isn't page_mapped and dirty/writeback, the page
515 * could reclaim asap using PG_reclaim.
517 * 1. active, mapped page -> none
518 * 2. active, dirty/writeback page -> inactive, head, PG_reclaim
519 * 3. inactive, mapped page -> none
520 * 4. inactive, dirty/writeback page -> inactive, head, PG_reclaim
521 * 5. inactive, clean -> inactive, tail
524 * In 4, why it moves inactive's head, the VM expects the page would
525 * be write it out by flusher threads as this is much more effective
526 * than the single-page writeout from reclaim.
528 static void lru_deactivate_file_fn(struct page
*page
, struct lruvec
*lruvec
,
533 int nr_pages
= thp_nr_pages(page
);
538 if (PageUnevictable(page
))
541 /* Some processes are using the page */
542 if (page_mapped(page
))
545 active
= PageActive(page
);
546 lru
= page_lru_base_type(page
);
548 del_page_from_lru_list(page
, lruvec
, lru
+ active
);
549 ClearPageActive(page
);
550 ClearPageReferenced(page
);
552 if (PageWriteback(page
) || PageDirty(page
)) {
554 * PG_reclaim could be raced with end_page_writeback
555 * It can make readahead confusing. But race window
556 * is _really_ small and it's non-critical problem.
558 add_page_to_lru_list(page
, lruvec
, lru
);
559 SetPageReclaim(page
);
562 * The page's writeback ends up during pagevec
563 * We moves tha page into tail of inactive.
565 add_page_to_lru_list_tail(page
, lruvec
, lru
);
566 __count_vm_events(PGROTATED
, nr_pages
);
570 __count_vm_events(PGDEACTIVATE
, nr_pages
);
571 __count_memcg_events(lruvec_memcg(lruvec
), PGDEACTIVATE
,
576 static void lru_deactivate_fn(struct page
*page
, struct lruvec
*lruvec
,
579 if (PageLRU(page
) && PageActive(page
) && !PageUnevictable(page
)) {
580 int lru
= page_lru_base_type(page
);
581 int nr_pages
= thp_nr_pages(page
);
583 del_page_from_lru_list(page
, lruvec
, lru
+ LRU_ACTIVE
);
584 ClearPageActive(page
);
585 ClearPageReferenced(page
);
586 add_page_to_lru_list(page
, lruvec
, lru
);
588 __count_vm_events(PGDEACTIVATE
, nr_pages
);
589 __count_memcg_events(lruvec_memcg(lruvec
), PGDEACTIVATE
,
594 static void lru_lazyfree_fn(struct page
*page
, struct lruvec
*lruvec
,
597 if (PageLRU(page
) && PageAnon(page
) && PageSwapBacked(page
) &&
598 !PageSwapCache(page
) && !PageUnevictable(page
)) {
599 bool active
= PageActive(page
);
600 int nr_pages
= thp_nr_pages(page
);
602 del_page_from_lru_list(page
, lruvec
,
603 LRU_INACTIVE_ANON
+ active
);
604 ClearPageActive(page
);
605 ClearPageReferenced(page
);
607 * Lazyfree pages are clean anonymous pages. They have
608 * PG_swapbacked flag cleared, to distinguish them from normal
611 ClearPageSwapBacked(page
);
612 add_page_to_lru_list(page
, lruvec
, LRU_INACTIVE_FILE
);
614 __count_vm_events(PGLAZYFREE
, nr_pages
);
615 __count_memcg_events(lruvec_memcg(lruvec
), PGLAZYFREE
,
621 * Drain pages out of the cpu's pagevecs.
622 * Either "cpu" is the current CPU, and preemption has already been
623 * disabled; or "cpu" is being hot-unplugged, and is already dead.
625 void lru_add_drain_cpu(int cpu
)
627 struct pagevec
*pvec
= &per_cpu(lru_pvecs
.lru_add
, cpu
);
629 if (pagevec_count(pvec
))
630 __pagevec_lru_add(pvec
);
632 pvec
= &per_cpu(lru_rotate
.pvec
, cpu
);
633 /* Disabling interrupts below acts as a compiler barrier. */
634 if (data_race(pagevec_count(pvec
))) {
637 /* No harm done if a racing interrupt already did this */
638 local_lock_irqsave(&lru_rotate
.lock
, flags
);
639 pagevec_move_tail(pvec
);
640 local_unlock_irqrestore(&lru_rotate
.lock
, flags
);
643 pvec
= &per_cpu(lru_pvecs
.lru_deactivate_file
, cpu
);
644 if (pagevec_count(pvec
))
645 pagevec_lru_move_fn(pvec
, lru_deactivate_file_fn
, NULL
);
647 pvec
= &per_cpu(lru_pvecs
.lru_deactivate
, cpu
);
648 if (pagevec_count(pvec
))
649 pagevec_lru_move_fn(pvec
, lru_deactivate_fn
, NULL
);
651 pvec
= &per_cpu(lru_pvecs
.lru_lazyfree
, cpu
);
652 if (pagevec_count(pvec
))
653 pagevec_lru_move_fn(pvec
, lru_lazyfree_fn
, NULL
);
655 activate_page_drain(cpu
);
659 * deactivate_file_page - forcefully deactivate a file page
660 * @page: page to deactivate
662 * This function hints the VM that @page is a good reclaim candidate,
663 * for example if its invalidation fails due to the page being dirty
664 * or under writeback.
666 void deactivate_file_page(struct page
*page
)
669 * In a workload with many unevictable page such as mprotect,
670 * unevictable page deactivation for accelerating reclaim is pointless.
672 if (PageUnevictable(page
))
675 if (likely(get_page_unless_zero(page
))) {
676 struct pagevec
*pvec
;
678 local_lock(&lru_pvecs
.lock
);
679 pvec
= this_cpu_ptr(&lru_pvecs
.lru_deactivate_file
);
681 if (!pagevec_add(pvec
, page
) || PageCompound(page
))
682 pagevec_lru_move_fn(pvec
, lru_deactivate_file_fn
, NULL
);
683 local_unlock(&lru_pvecs
.lock
);
688 * deactivate_page - deactivate a page
689 * @page: page to deactivate
691 * deactivate_page() moves @page to the inactive list if @page was on the active
692 * list and was not an unevictable page. This is done to accelerate the reclaim
695 void deactivate_page(struct page
*page
)
697 if (PageLRU(page
) && PageActive(page
) && !PageUnevictable(page
)) {
698 struct pagevec
*pvec
;
700 local_lock(&lru_pvecs
.lock
);
701 pvec
= this_cpu_ptr(&lru_pvecs
.lru_deactivate
);
703 if (!pagevec_add(pvec
, page
) || PageCompound(page
))
704 pagevec_lru_move_fn(pvec
, lru_deactivate_fn
, NULL
);
705 local_unlock(&lru_pvecs
.lock
);
710 * mark_page_lazyfree - make an anon page lazyfree
711 * @page: page to deactivate
713 * mark_page_lazyfree() moves @page to the inactive file list.
714 * This is done to accelerate the reclaim of @page.
716 void mark_page_lazyfree(struct page
*page
)
718 if (PageLRU(page
) && PageAnon(page
) && PageSwapBacked(page
) &&
719 !PageSwapCache(page
) && !PageUnevictable(page
)) {
720 struct pagevec
*pvec
;
722 local_lock(&lru_pvecs
.lock
);
723 pvec
= this_cpu_ptr(&lru_pvecs
.lru_lazyfree
);
725 if (!pagevec_add(pvec
, page
) || PageCompound(page
))
726 pagevec_lru_move_fn(pvec
, lru_lazyfree_fn
, NULL
);
727 local_unlock(&lru_pvecs
.lock
);
731 void lru_add_drain(void)
733 local_lock(&lru_pvecs
.lock
);
734 lru_add_drain_cpu(smp_processor_id());
735 local_unlock(&lru_pvecs
.lock
);
738 void lru_add_drain_cpu_zone(struct zone
*zone
)
740 local_lock(&lru_pvecs
.lock
);
741 lru_add_drain_cpu(smp_processor_id());
742 drain_local_pages(zone
);
743 local_unlock(&lru_pvecs
.lock
);
748 static DEFINE_PER_CPU(struct work_struct
, lru_add_drain_work
);
750 static void lru_add_drain_per_cpu(struct work_struct
*dummy
)
756 * Doesn't need any cpu hotplug locking because we do rely on per-cpu
757 * kworkers being shut down before our page_alloc_cpu_dead callback is
758 * executed on the offlined cpu.
759 * Calling this function with cpu hotplug locks held can actually lead
760 * to obscure indirect dependencies via WQ context.
762 void lru_add_drain_all(void)
765 * lru_drain_gen - Global pages generation number
767 * (A) Definition: global lru_drain_gen = x implies that all generations
768 * 0 < n <= x are already *scheduled* for draining.
770 * This is an optimization for the highly-contended use case where a
771 * user space workload keeps constantly generating a flow of pages for
774 static unsigned int lru_drain_gen
;
775 static struct cpumask has_work
;
776 static DEFINE_MUTEX(lock
);
777 unsigned cpu
, this_gen
;
780 * Make sure nobody triggers this path before mm_percpu_wq is fully
783 if (WARN_ON(!mm_percpu_wq
))
787 * Guarantee pagevec counter stores visible by this CPU are visible to
788 * other CPUs before loading the current drain generation.
793 * (B) Locally cache global LRU draining generation number
795 * The read barrier ensures that the counter is loaded before the mutex
796 * is taken. It pairs with smp_mb() inside the mutex critical section
799 this_gen
= smp_load_acquire(&lru_drain_gen
);
804 * (C) Exit the draining operation if a newer generation, from another
805 * lru_add_drain_all(), was already scheduled for draining. Check (A).
807 if (unlikely(this_gen
!= lru_drain_gen
))
811 * (D) Increment global generation number
813 * Pairs with smp_load_acquire() at (B), outside of the critical
814 * section. Use a full memory barrier to guarantee that the new global
815 * drain generation number is stored before loading pagevec counters.
817 * This pairing must be done here, before the for_each_online_cpu loop
818 * below which drains the page vectors.
820 * Let x, y, and z represent some system CPU numbers, where x < y < z.
821 * Assume CPU #z is is in the middle of the for_each_online_cpu loop
822 * below and has already reached CPU #y's per-cpu data. CPU #x comes
823 * along, adds some pages to its per-cpu vectors, then calls
824 * lru_add_drain_all().
826 * If the paired barrier is done at any later step, e.g. after the
827 * loop, CPU #x will just exit at (C) and miss flushing out all of its
830 WRITE_ONCE(lru_drain_gen
, lru_drain_gen
+ 1);
833 cpumask_clear(&has_work
);
834 for_each_online_cpu(cpu
) {
835 struct work_struct
*work
= &per_cpu(lru_add_drain_work
, cpu
);
837 if (pagevec_count(&per_cpu(lru_pvecs
.lru_add
, cpu
)) ||
838 data_race(pagevec_count(&per_cpu(lru_rotate
.pvec
, cpu
))) ||
839 pagevec_count(&per_cpu(lru_pvecs
.lru_deactivate_file
, cpu
)) ||
840 pagevec_count(&per_cpu(lru_pvecs
.lru_deactivate
, cpu
)) ||
841 pagevec_count(&per_cpu(lru_pvecs
.lru_lazyfree
, cpu
)) ||
842 need_activate_page_drain(cpu
)) {
843 INIT_WORK(work
, lru_add_drain_per_cpu
);
844 queue_work_on(cpu
, mm_percpu_wq
, work
);
845 __cpumask_set_cpu(cpu
, &has_work
);
849 for_each_cpu(cpu
, &has_work
)
850 flush_work(&per_cpu(lru_add_drain_work
, cpu
));
856 void lru_add_drain_all(void)
860 #endif /* CONFIG_SMP */
863 * release_pages - batched put_page()
864 * @pages: array of pages to release
865 * @nr: number of pages
867 * Decrement the reference count on all the pages in @pages. If it
868 * fell to zero, remove the page from the LRU and free it.
870 void release_pages(struct page
**pages
, int nr
)
873 LIST_HEAD(pages_to_free
);
874 struct pglist_data
*locked_pgdat
= NULL
;
875 struct lruvec
*lruvec
;
877 unsigned int lock_batch
;
879 for (i
= 0; i
< nr
; i
++) {
880 struct page
*page
= pages
[i
];
883 * Make sure the IRQ-safe lock-holding time does not get
884 * excessive with a continuous string of pages from the
885 * same pgdat. The lock is held only if pgdat != NULL.
887 if (locked_pgdat
&& ++lock_batch
== SWAP_CLUSTER_MAX
) {
888 spin_unlock_irqrestore(&locked_pgdat
->lru_lock
, flags
);
892 page
= compound_head(page
);
893 if (is_huge_zero_page(page
))
896 if (is_zone_device_page(page
)) {
898 spin_unlock_irqrestore(&locked_pgdat
->lru_lock
,
903 * ZONE_DEVICE pages that return 'false' from
904 * page_is_devmap_managed() do not require special
905 * processing, and instead, expect a call to
906 * put_page_testzero().
908 if (page_is_devmap_managed(page
)) {
909 put_devmap_managed_page(page
);
914 if (!put_page_testzero(page
))
917 if (PageCompound(page
)) {
919 spin_unlock_irqrestore(&locked_pgdat
->lru_lock
, flags
);
922 __put_compound_page(page
);
927 struct pglist_data
*pgdat
= page_pgdat(page
);
929 if (pgdat
!= locked_pgdat
) {
931 spin_unlock_irqrestore(&locked_pgdat
->lru_lock
,
934 locked_pgdat
= pgdat
;
935 spin_lock_irqsave(&locked_pgdat
->lru_lock
, flags
);
938 lruvec
= mem_cgroup_page_lruvec(page
, locked_pgdat
);
939 VM_BUG_ON_PAGE(!PageLRU(page
), page
);
940 __ClearPageLRU(page
);
941 del_page_from_lru_list(page
, lruvec
, page_off_lru(page
));
944 __ClearPageWaiters(page
);
946 list_add(&page
->lru
, &pages_to_free
);
949 spin_unlock_irqrestore(&locked_pgdat
->lru_lock
, flags
);
951 mem_cgroup_uncharge_list(&pages_to_free
);
952 free_unref_page_list(&pages_to_free
);
954 EXPORT_SYMBOL(release_pages
);
957 * The pages which we're about to release may be in the deferred lru-addition
958 * queues. That would prevent them from really being freed right now. That's
959 * OK from a correctness point of view but is inefficient - those pages may be
960 * cache-warm and we want to give them back to the page allocator ASAP.
962 * So __pagevec_release() will drain those queues here. __pagevec_lru_add()
963 * and __pagevec_lru_add_active() call release_pages() directly to avoid
966 void __pagevec_release(struct pagevec
*pvec
)
968 if (!pvec
->percpu_pvec_drained
) {
970 pvec
->percpu_pvec_drained
= true;
972 release_pages(pvec
->pages
, pagevec_count(pvec
));
973 pagevec_reinit(pvec
);
975 EXPORT_SYMBOL(__pagevec_release
);
977 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
978 /* used by __split_huge_page_refcount() */
979 void lru_add_page_tail(struct page
*page
, struct page
*page_tail
,
980 struct lruvec
*lruvec
, struct list_head
*list
)
982 VM_BUG_ON_PAGE(!PageHead(page
), page
);
983 VM_BUG_ON_PAGE(PageCompound(page_tail
), page
);
984 VM_BUG_ON_PAGE(PageLRU(page_tail
), page
);
985 lockdep_assert_held(&lruvec_pgdat(lruvec
)->lru_lock
);
988 SetPageLRU(page_tail
);
990 if (likely(PageLRU(page
)))
991 list_add_tail(&page_tail
->lru
, &page
->lru
);
993 /* page reclaim is reclaiming a huge page */
995 list_add_tail(&page_tail
->lru
, list
);
998 * Head page has not yet been counted, as an hpage,
999 * so we must account for each subpage individually.
1001 * Put page_tail on the list at the correct position
1002 * so they all end up in order.
1004 add_page_to_lru_list_tail(page_tail
, lruvec
,
1005 page_lru(page_tail
));
1008 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1010 static void __pagevec_lru_add_fn(struct page
*page
, struct lruvec
*lruvec
,
1014 int was_unevictable
= TestClearPageUnevictable(page
);
1015 int nr_pages
= thp_nr_pages(page
);
1017 VM_BUG_ON_PAGE(PageLRU(page
), page
);
1020 * Page becomes evictable in two ways:
1021 * 1) Within LRU lock [munlock_vma_page() and __munlock_pagevec()].
1022 * 2) Before acquiring LRU lock to put the page to correct LRU and then
1023 * a) do PageLRU check with lock [check_move_unevictable_pages]
1024 * b) do PageLRU check before lock [clear_page_mlock]
1026 * (1) & (2a) are ok as LRU lock will serialize them. For (2b), we need
1027 * following strict ordering:
1029 * #0: __pagevec_lru_add_fn #1: clear_page_mlock
1031 * SetPageLRU() TestClearPageMlocked()
1032 * smp_mb() // explicit ordering // above provides strict
1034 * PageMlocked() PageLRU()
1037 * if '#1' does not observe setting of PG_lru by '#0' and fails
1038 * isolation, the explicit barrier will make sure that page_evictable
1039 * check will put the page in correct LRU. Without smp_mb(), SetPageLRU
1040 * can be reordered after PageMlocked check and can make '#1' to fail
1041 * the isolation of the page whose Mlocked bit is cleared (#0 is also
1042 * looking at the same page) and the evictable page will be stranded
1043 * in an unevictable LRU.
1046 smp_mb__after_atomic();
1048 if (page_evictable(page
)) {
1049 lru
= page_lru(page
);
1050 if (was_unevictable
)
1051 __count_vm_events(UNEVICTABLE_PGRESCUED
, nr_pages
);
1053 lru
= LRU_UNEVICTABLE
;
1054 ClearPageActive(page
);
1055 SetPageUnevictable(page
);
1056 if (!was_unevictable
)
1057 __count_vm_events(UNEVICTABLE_PGCULLED
, nr_pages
);
1060 add_page_to_lru_list(page
, lruvec
, lru
);
1061 trace_mm_lru_insertion(page
, lru
);
1065 * Add the passed pages to the LRU, then drop the caller's refcount
1066 * on them. Reinitialises the caller's pagevec.
1068 void __pagevec_lru_add(struct pagevec
*pvec
)
1070 pagevec_lru_move_fn(pvec
, __pagevec_lru_add_fn
, NULL
);
1074 * pagevec_lookup_entries - gang pagecache lookup
1075 * @pvec: Where the resulting entries are placed
1076 * @mapping: The address_space to search
1077 * @start: The starting entry index
1078 * @nr_entries: The maximum number of pages
1079 * @indices: The cache indices corresponding to the entries in @pvec
1081 * pagevec_lookup_entries() will search for and return a group of up
1082 * to @nr_pages pages and shadow entries in the mapping. All
1083 * entries are placed in @pvec. pagevec_lookup_entries() takes a
1084 * reference against actual pages in @pvec.
1086 * The search returns a group of mapping-contiguous entries with
1087 * ascending indexes. There may be holes in the indices due to
1088 * not-present entries.
1090 * Only one subpage of a Transparent Huge Page is returned in one call:
1091 * allowing truncate_inode_pages_range() to evict the whole THP without
1092 * cycling through a pagevec of extra references.
1094 * pagevec_lookup_entries() returns the number of entries which were
1097 unsigned pagevec_lookup_entries(struct pagevec
*pvec
,
1098 struct address_space
*mapping
,
1099 pgoff_t start
, unsigned nr_entries
,
1102 pvec
->nr
= find_get_entries(mapping
, start
, nr_entries
,
1103 pvec
->pages
, indices
);
1104 return pagevec_count(pvec
);
1108 * pagevec_remove_exceptionals - pagevec exceptionals pruning
1109 * @pvec: The pagevec to prune
1111 * pagevec_lookup_entries() fills both pages and exceptional radix
1112 * tree entries into the pagevec. This function prunes all
1113 * exceptionals from @pvec without leaving holes, so that it can be
1114 * passed on to page-only pagevec operations.
1116 void pagevec_remove_exceptionals(struct pagevec
*pvec
)
1120 for (i
= 0, j
= 0; i
< pagevec_count(pvec
); i
++) {
1121 struct page
*page
= pvec
->pages
[i
];
1122 if (!xa_is_value(page
))
1123 pvec
->pages
[j
++] = page
;
1129 * pagevec_lookup_range - gang pagecache lookup
1130 * @pvec: Where the resulting pages are placed
1131 * @mapping: The address_space to search
1132 * @start: The starting page index
1133 * @end: The final page index
1135 * pagevec_lookup_range() will search for & return a group of up to PAGEVEC_SIZE
1136 * pages in the mapping starting from index @start and upto index @end
1137 * (inclusive). The pages are placed in @pvec. pagevec_lookup() takes a
1138 * reference against the pages in @pvec.
1140 * The search returns a group of mapping-contiguous pages with ascending
1141 * indexes. There may be holes in the indices due to not-present pages. We
1142 * also update @start to index the next page for the traversal.
1144 * pagevec_lookup_range() returns the number of pages which were found. If this
1145 * number is smaller than PAGEVEC_SIZE, the end of specified range has been
1148 unsigned pagevec_lookup_range(struct pagevec
*pvec
,
1149 struct address_space
*mapping
, pgoff_t
*start
, pgoff_t end
)
1151 pvec
->nr
= find_get_pages_range(mapping
, start
, end
, PAGEVEC_SIZE
,
1153 return pagevec_count(pvec
);
1155 EXPORT_SYMBOL(pagevec_lookup_range
);
1157 unsigned pagevec_lookup_range_tag(struct pagevec
*pvec
,
1158 struct address_space
*mapping
, pgoff_t
*index
, pgoff_t end
,
1161 pvec
->nr
= find_get_pages_range_tag(mapping
, index
, end
, tag
,
1162 PAGEVEC_SIZE
, pvec
->pages
);
1163 return pagevec_count(pvec
);
1165 EXPORT_SYMBOL(pagevec_lookup_range_tag
);
1167 unsigned pagevec_lookup_range_nr_tag(struct pagevec
*pvec
,
1168 struct address_space
*mapping
, pgoff_t
*index
, pgoff_t end
,
1169 xa_mark_t tag
, unsigned max_pages
)
1171 pvec
->nr
= find_get_pages_range_tag(mapping
, index
, end
, tag
,
1172 min_t(unsigned int, max_pages
, PAGEVEC_SIZE
), pvec
->pages
);
1173 return pagevec_count(pvec
);
1175 EXPORT_SYMBOL(pagevec_lookup_range_nr_tag
);
1177 * Perform any setup for the swap system
1179 void __init
swap_setup(void)
1181 unsigned long megs
= totalram_pages() >> (20 - PAGE_SHIFT
);
1183 /* Use a smaller cluster for small-memory machines */
1189 * Right now other parts of the system means that we
1190 * _really_ don't want to cluster much more
1194 #ifdef CONFIG_DEV_PAGEMAP_OPS
1195 void put_devmap_managed_page(struct page
*page
)
1199 if (WARN_ON_ONCE(!page_is_devmap_managed(page
)))
1202 count
= page_ref_dec_return(page
);
1205 * devmap page refcounts are 1-based, rather than 0-based: if
1206 * refcount is 1, then the page is free and the refcount is
1207 * stable because nobody holds a reference on the page.
1210 free_devmap_managed_page(page
);
1214 EXPORT_SYMBOL(put_devmap_managed_page
);