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
);
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 if (!PageUnevictable(page
)) {
210 del_page_from_lru_list(page
, lruvec
);
211 ClearPageActive(page
);
212 add_page_to_lru_list_tail(page
, lruvec
);
213 __count_vm_events(PGROTATED
, thp_nr_pages(page
));
217 /* return true if pagevec needs to drain */
218 static bool pagevec_add_and_need_flush(struct pagevec
*pvec
, struct page
*page
)
222 if (!pagevec_add(pvec
, page
) || PageCompound(page
) ||
223 lru_cache_disabled())
230 * Writeback is about to end against a page which has been marked for immediate
231 * reclaim. If it still appears to be reclaimable, move it to the tail of the
234 * rotate_reclaimable_page() must disable IRQs, to prevent nasty races.
236 void rotate_reclaimable_page(struct page
*page
)
238 if (!PageLocked(page
) && !PageDirty(page
) &&
239 !PageUnevictable(page
) && PageLRU(page
)) {
240 struct pagevec
*pvec
;
244 local_lock_irqsave(&lru_rotate
.lock
, flags
);
245 pvec
= this_cpu_ptr(&lru_rotate
.pvec
);
246 if (pagevec_add_and_need_flush(pvec
, page
))
247 pagevec_lru_move_fn(pvec
, pagevec_move_tail_fn
);
248 local_unlock_irqrestore(&lru_rotate
.lock
, flags
);
252 void lru_note_cost(struct lruvec
*lruvec
, bool file
, unsigned int nr_pages
)
255 unsigned long lrusize
;
258 * Hold lruvec->lru_lock is safe here, since
259 * 1) The pinned lruvec in reclaim, or
260 * 2) From a pre-LRU page during refault (which also holds the
261 * rcu lock, so would be safe even if the page was on the LRU
262 * and could move simultaneously to a new lruvec).
264 spin_lock_irq(&lruvec
->lru_lock
);
265 /* Record cost event */
267 lruvec
->file_cost
+= nr_pages
;
269 lruvec
->anon_cost
+= nr_pages
;
272 * Decay previous events
274 * Because workloads change over time (and to avoid
275 * overflow) we keep these statistics as a floating
276 * average, which ends up weighing recent refaults
277 * more than old ones.
279 lrusize
= lruvec_page_state(lruvec
, NR_INACTIVE_ANON
) +
280 lruvec_page_state(lruvec
, NR_ACTIVE_ANON
) +
281 lruvec_page_state(lruvec
, NR_INACTIVE_FILE
) +
282 lruvec_page_state(lruvec
, NR_ACTIVE_FILE
);
284 if (lruvec
->file_cost
+ lruvec
->anon_cost
> lrusize
/ 4) {
285 lruvec
->file_cost
/= 2;
286 lruvec
->anon_cost
/= 2;
288 spin_unlock_irq(&lruvec
->lru_lock
);
289 } while ((lruvec
= parent_lruvec(lruvec
)));
292 void lru_note_cost_page(struct page
*page
)
294 lru_note_cost(mem_cgroup_page_lruvec(page
),
295 page_is_file_lru(page
), thp_nr_pages(page
));
298 static void __activate_page(struct page
*page
, struct lruvec
*lruvec
)
300 if (!PageActive(page
) && !PageUnevictable(page
)) {
301 int nr_pages
= thp_nr_pages(page
);
303 del_page_from_lru_list(page
, lruvec
);
305 add_page_to_lru_list(page
, lruvec
);
306 trace_mm_lru_activate(page
);
308 __count_vm_events(PGACTIVATE
, nr_pages
);
309 __count_memcg_events(lruvec_memcg(lruvec
), PGACTIVATE
,
315 static void activate_page_drain(int cpu
)
317 struct pagevec
*pvec
= &per_cpu(lru_pvecs
.activate_page
, cpu
);
319 if (pagevec_count(pvec
))
320 pagevec_lru_move_fn(pvec
, __activate_page
);
323 static bool need_activate_page_drain(int cpu
)
325 return pagevec_count(&per_cpu(lru_pvecs
.activate_page
, cpu
)) != 0;
328 static void activate_page(struct page
*page
)
330 page
= compound_head(page
);
331 if (PageLRU(page
) && !PageActive(page
) && !PageUnevictable(page
)) {
332 struct pagevec
*pvec
;
334 local_lock(&lru_pvecs
.lock
);
335 pvec
= this_cpu_ptr(&lru_pvecs
.activate_page
);
337 if (pagevec_add_and_need_flush(pvec
, page
))
338 pagevec_lru_move_fn(pvec
, __activate_page
);
339 local_unlock(&lru_pvecs
.lock
);
344 static inline void activate_page_drain(int cpu
)
348 static void activate_page(struct page
*page
)
350 struct lruvec
*lruvec
;
352 page
= compound_head(page
);
353 if (TestClearPageLRU(page
)) {
354 lruvec
= lock_page_lruvec_irq(page
);
355 __activate_page(page
, lruvec
);
356 unlock_page_lruvec_irq(lruvec
);
362 static void __lru_cache_activate_page(struct page
*page
)
364 struct pagevec
*pvec
;
367 local_lock(&lru_pvecs
.lock
);
368 pvec
= this_cpu_ptr(&lru_pvecs
.lru_add
);
371 * Search backwards on the optimistic assumption that the page being
372 * activated has just been added to this pagevec. Note that only
373 * the local pagevec is examined as a !PageLRU page could be in the
374 * process of being released, reclaimed, migrated or on a remote
375 * pagevec that is currently being drained. Furthermore, marking
376 * a remote pagevec's page PageActive potentially hits a race where
377 * a page is marked PageActive just after it is added to the inactive
378 * list causing accounting errors and BUG_ON checks to trigger.
380 for (i
= pagevec_count(pvec
) - 1; i
>= 0; i
--) {
381 struct page
*pagevec_page
= pvec
->pages
[i
];
383 if (pagevec_page
== page
) {
389 local_unlock(&lru_pvecs
.lock
);
393 * Mark a page as having seen activity.
395 * inactive,unreferenced -> inactive,referenced
396 * inactive,referenced -> active,unreferenced
397 * active,unreferenced -> active,referenced
399 * When a newly allocated page is not yet visible, so safe for non-atomic ops,
400 * __SetPageReferenced(page) may be substituted for mark_page_accessed(page).
402 void mark_page_accessed(struct page
*page
)
404 page
= compound_head(page
);
406 if (!PageReferenced(page
)) {
407 SetPageReferenced(page
);
408 } else if (PageUnevictable(page
)) {
410 * Unevictable pages are on the "LRU_UNEVICTABLE" list. But,
411 * this list is never rotated or maintained, so marking an
412 * evictable page accessed has no effect.
414 } else if (!PageActive(page
)) {
416 * If the page is on the LRU, queue it for activation via
417 * lru_pvecs.activate_page. Otherwise, assume the page is on a
418 * pagevec, mark it active and it'll be moved to the active
419 * LRU on the next drain.
424 __lru_cache_activate_page(page
);
425 ClearPageReferenced(page
);
426 workingset_activation(page
);
428 if (page_is_idle(page
))
429 clear_page_idle(page
);
431 EXPORT_SYMBOL(mark_page_accessed
);
434 * lru_cache_add - add a page to a page list
435 * @page: the page to be added to the LRU.
437 * Queue the page for addition to the LRU via pagevec. The decision on whether
438 * to add the page to the [in]active [file|anon] list is deferred until the
439 * pagevec is drained. This gives a chance for the caller of lru_cache_add()
440 * have the page added to the active list using mark_page_accessed().
442 void lru_cache_add(struct page
*page
)
444 struct pagevec
*pvec
;
446 VM_BUG_ON_PAGE(PageActive(page
) && PageUnevictable(page
), page
);
447 VM_BUG_ON_PAGE(PageLRU(page
), page
);
450 local_lock(&lru_pvecs
.lock
);
451 pvec
= this_cpu_ptr(&lru_pvecs
.lru_add
);
452 if (pagevec_add_and_need_flush(pvec
, page
))
453 __pagevec_lru_add(pvec
);
454 local_unlock(&lru_pvecs
.lock
);
456 EXPORT_SYMBOL(lru_cache_add
);
459 * lru_cache_add_inactive_or_unevictable
460 * @page: the page to be added to LRU
461 * @vma: vma in which page is mapped for determining reclaimability
463 * Place @page on the inactive or unevictable LRU list, depending on its
466 void lru_cache_add_inactive_or_unevictable(struct page
*page
,
467 struct vm_area_struct
*vma
)
471 VM_BUG_ON_PAGE(PageLRU(page
), page
);
473 unevictable
= (vma
->vm_flags
& (VM_LOCKED
| VM_SPECIAL
)) == VM_LOCKED
;
474 if (unlikely(unevictable
) && !TestSetPageMlocked(page
)) {
475 int nr_pages
= thp_nr_pages(page
);
477 * We use the irq-unsafe __mod_zone_page_state because this
478 * counter is not modified from interrupt context, and the pte
479 * lock is held(spinlock), which implies preemption disabled.
481 __mod_zone_page_state(page_zone(page
), NR_MLOCK
, nr_pages
);
482 count_vm_events(UNEVICTABLE_PGMLOCKED
, nr_pages
);
488 * If the page can not be invalidated, it is moved to the
489 * inactive list to speed up its reclaim. It is moved to the
490 * head of the list, rather than the tail, to give the flusher
491 * threads some time to write it out, as this is much more
492 * effective than the single-page writeout from reclaim.
494 * If the page isn't page_mapped and dirty/writeback, the page
495 * could reclaim asap using PG_reclaim.
497 * 1. active, mapped page -> none
498 * 2. active, dirty/writeback page -> inactive, head, PG_reclaim
499 * 3. inactive, mapped page -> none
500 * 4. inactive, dirty/writeback page -> inactive, head, PG_reclaim
501 * 5. inactive, clean -> inactive, tail
504 * In 4, why it moves inactive's head, the VM expects the page would
505 * be write it out by flusher threads as this is much more effective
506 * than the single-page writeout from reclaim.
508 static void lru_deactivate_file_fn(struct page
*page
, struct lruvec
*lruvec
)
510 bool active
= PageActive(page
);
511 int nr_pages
= thp_nr_pages(page
);
513 if (PageUnevictable(page
))
516 /* Some processes are using the page */
517 if (page_mapped(page
))
520 del_page_from_lru_list(page
, lruvec
);
521 ClearPageActive(page
);
522 ClearPageReferenced(page
);
524 if (PageWriteback(page
) || PageDirty(page
)) {
526 * PG_reclaim could be raced with end_page_writeback
527 * It can make readahead confusing. But race window
528 * is _really_ small and it's non-critical problem.
530 add_page_to_lru_list(page
, lruvec
);
531 SetPageReclaim(page
);
534 * The page's writeback ends up during pagevec
535 * We move that page into tail of inactive.
537 add_page_to_lru_list_tail(page
, lruvec
);
538 __count_vm_events(PGROTATED
, nr_pages
);
542 __count_vm_events(PGDEACTIVATE
, nr_pages
);
543 __count_memcg_events(lruvec_memcg(lruvec
), PGDEACTIVATE
,
548 static void lru_deactivate_fn(struct page
*page
, struct lruvec
*lruvec
)
550 if (PageActive(page
) && !PageUnevictable(page
)) {
551 int nr_pages
= thp_nr_pages(page
);
553 del_page_from_lru_list(page
, lruvec
);
554 ClearPageActive(page
);
555 ClearPageReferenced(page
);
556 add_page_to_lru_list(page
, lruvec
);
558 __count_vm_events(PGDEACTIVATE
, nr_pages
);
559 __count_memcg_events(lruvec_memcg(lruvec
), PGDEACTIVATE
,
564 static void lru_lazyfree_fn(struct page
*page
, struct lruvec
*lruvec
)
566 if (PageAnon(page
) && PageSwapBacked(page
) &&
567 !PageSwapCache(page
) && !PageUnevictable(page
)) {
568 int nr_pages
= thp_nr_pages(page
);
570 del_page_from_lru_list(page
, lruvec
);
571 ClearPageActive(page
);
572 ClearPageReferenced(page
);
574 * Lazyfree pages are clean anonymous pages. They have
575 * PG_swapbacked flag cleared, to distinguish them from normal
578 ClearPageSwapBacked(page
);
579 add_page_to_lru_list(page
, lruvec
);
581 __count_vm_events(PGLAZYFREE
, nr_pages
);
582 __count_memcg_events(lruvec_memcg(lruvec
), PGLAZYFREE
,
588 * Drain pages out of the cpu's pagevecs.
589 * Either "cpu" is the current CPU, and preemption has already been
590 * disabled; or "cpu" is being hot-unplugged, and is already dead.
592 void lru_add_drain_cpu(int cpu
)
594 struct pagevec
*pvec
= &per_cpu(lru_pvecs
.lru_add
, cpu
);
596 if (pagevec_count(pvec
))
597 __pagevec_lru_add(pvec
);
599 pvec
= &per_cpu(lru_rotate
.pvec
, cpu
);
600 /* Disabling interrupts below acts as a compiler barrier. */
601 if (data_race(pagevec_count(pvec
))) {
604 /* No harm done if a racing interrupt already did this */
605 local_lock_irqsave(&lru_rotate
.lock
, flags
);
606 pagevec_lru_move_fn(pvec
, pagevec_move_tail_fn
);
607 local_unlock_irqrestore(&lru_rotate
.lock
, flags
);
610 pvec
= &per_cpu(lru_pvecs
.lru_deactivate_file
, cpu
);
611 if (pagevec_count(pvec
))
612 pagevec_lru_move_fn(pvec
, lru_deactivate_file_fn
);
614 pvec
= &per_cpu(lru_pvecs
.lru_deactivate
, cpu
);
615 if (pagevec_count(pvec
))
616 pagevec_lru_move_fn(pvec
, lru_deactivate_fn
);
618 pvec
= &per_cpu(lru_pvecs
.lru_lazyfree
, cpu
);
619 if (pagevec_count(pvec
))
620 pagevec_lru_move_fn(pvec
, lru_lazyfree_fn
);
622 activate_page_drain(cpu
);
623 invalidate_bh_lrus_cpu(cpu
);
627 * deactivate_file_page - forcefully deactivate a file page
628 * @page: page to deactivate
630 * This function hints the VM that @page is a good reclaim candidate,
631 * for example if its invalidation fails due to the page being dirty
632 * or under writeback.
634 void deactivate_file_page(struct page
*page
)
637 * In a workload with many unevictable page such as mprotect,
638 * unevictable page deactivation for accelerating reclaim is pointless.
640 if (PageUnevictable(page
))
643 if (likely(get_page_unless_zero(page
))) {
644 struct pagevec
*pvec
;
646 local_lock(&lru_pvecs
.lock
);
647 pvec
= this_cpu_ptr(&lru_pvecs
.lru_deactivate_file
);
649 if (pagevec_add_and_need_flush(pvec
, page
))
650 pagevec_lru_move_fn(pvec
, lru_deactivate_file_fn
);
651 local_unlock(&lru_pvecs
.lock
);
656 * deactivate_page - deactivate a page
657 * @page: page to deactivate
659 * deactivate_page() moves @page to the inactive list if @page was on the active
660 * list and was not an unevictable page. This is done to accelerate the reclaim
663 void deactivate_page(struct page
*page
)
665 if (PageLRU(page
) && PageActive(page
) && !PageUnevictable(page
)) {
666 struct pagevec
*pvec
;
668 local_lock(&lru_pvecs
.lock
);
669 pvec
= this_cpu_ptr(&lru_pvecs
.lru_deactivate
);
671 if (pagevec_add_and_need_flush(pvec
, page
))
672 pagevec_lru_move_fn(pvec
, lru_deactivate_fn
);
673 local_unlock(&lru_pvecs
.lock
);
678 * mark_page_lazyfree - make an anon page lazyfree
679 * @page: page to deactivate
681 * mark_page_lazyfree() moves @page to the inactive file list.
682 * This is done to accelerate the reclaim of @page.
684 void mark_page_lazyfree(struct page
*page
)
686 if (PageLRU(page
) && PageAnon(page
) && PageSwapBacked(page
) &&
687 !PageSwapCache(page
) && !PageUnevictable(page
)) {
688 struct pagevec
*pvec
;
690 local_lock(&lru_pvecs
.lock
);
691 pvec
= this_cpu_ptr(&lru_pvecs
.lru_lazyfree
);
693 if (pagevec_add_and_need_flush(pvec
, page
))
694 pagevec_lru_move_fn(pvec
, lru_lazyfree_fn
);
695 local_unlock(&lru_pvecs
.lock
);
699 void lru_add_drain(void)
701 local_lock(&lru_pvecs
.lock
);
702 lru_add_drain_cpu(smp_processor_id());
703 local_unlock(&lru_pvecs
.lock
);
706 void lru_add_drain_cpu_zone(struct zone
*zone
)
708 local_lock(&lru_pvecs
.lock
);
709 lru_add_drain_cpu(smp_processor_id());
710 drain_local_pages(zone
);
711 local_unlock(&lru_pvecs
.lock
);
716 static DEFINE_PER_CPU(struct work_struct
, lru_add_drain_work
);
718 static void lru_add_drain_per_cpu(struct work_struct
*dummy
)
724 * Doesn't need any cpu hotplug locking because we do rely on per-cpu
725 * kworkers being shut down before our page_alloc_cpu_dead callback is
726 * executed on the offlined cpu.
727 * Calling this function with cpu hotplug locks held can actually lead
728 * to obscure indirect dependencies via WQ context.
730 inline void __lru_add_drain_all(bool force_all_cpus
)
733 * lru_drain_gen - Global pages generation number
735 * (A) Definition: global lru_drain_gen = x implies that all generations
736 * 0 < n <= x are already *scheduled* for draining.
738 * This is an optimization for the highly-contended use case where a
739 * user space workload keeps constantly generating a flow of pages for
742 static unsigned int lru_drain_gen
;
743 static struct cpumask has_work
;
744 static DEFINE_MUTEX(lock
);
745 unsigned cpu
, this_gen
;
748 * Make sure nobody triggers this path before mm_percpu_wq is fully
751 if (WARN_ON(!mm_percpu_wq
))
755 * Guarantee pagevec counter stores visible by this CPU are visible to
756 * other CPUs before loading the current drain generation.
761 * (B) Locally cache global LRU draining generation number
763 * The read barrier ensures that the counter is loaded before the mutex
764 * is taken. It pairs with smp_mb() inside the mutex critical section
767 this_gen
= smp_load_acquire(&lru_drain_gen
);
772 * (C) Exit the draining operation if a newer generation, from another
773 * lru_add_drain_all(), was already scheduled for draining. Check (A).
775 if (unlikely(this_gen
!= lru_drain_gen
&& !force_all_cpus
))
779 * (D) Increment global generation number
781 * Pairs with smp_load_acquire() at (B), outside of the critical
782 * section. Use a full memory barrier to guarantee that the new global
783 * drain generation number is stored before loading pagevec counters.
785 * This pairing must be done here, before the for_each_online_cpu loop
786 * below which drains the page vectors.
788 * Let x, y, and z represent some system CPU numbers, where x < y < z.
789 * Assume CPU #z is in the middle of the for_each_online_cpu loop
790 * below and has already reached CPU #y's per-cpu data. CPU #x comes
791 * along, adds some pages to its per-cpu vectors, then calls
792 * lru_add_drain_all().
794 * If the paired barrier is done at any later step, e.g. after the
795 * loop, CPU #x will just exit at (C) and miss flushing out all of its
798 WRITE_ONCE(lru_drain_gen
, lru_drain_gen
+ 1);
801 cpumask_clear(&has_work
);
802 for_each_online_cpu(cpu
) {
803 struct work_struct
*work
= &per_cpu(lru_add_drain_work
, cpu
);
805 if (force_all_cpus
||
806 pagevec_count(&per_cpu(lru_pvecs
.lru_add
, cpu
)) ||
807 data_race(pagevec_count(&per_cpu(lru_rotate
.pvec
, cpu
))) ||
808 pagevec_count(&per_cpu(lru_pvecs
.lru_deactivate_file
, cpu
)) ||
809 pagevec_count(&per_cpu(lru_pvecs
.lru_deactivate
, cpu
)) ||
810 pagevec_count(&per_cpu(lru_pvecs
.lru_lazyfree
, cpu
)) ||
811 need_activate_page_drain(cpu
) ||
812 has_bh_in_lru(cpu
, NULL
)) {
813 INIT_WORK(work
, lru_add_drain_per_cpu
);
814 queue_work_on(cpu
, mm_percpu_wq
, work
);
815 __cpumask_set_cpu(cpu
, &has_work
);
819 for_each_cpu(cpu
, &has_work
)
820 flush_work(&per_cpu(lru_add_drain_work
, cpu
));
826 void lru_add_drain_all(void)
828 __lru_add_drain_all(false);
831 void lru_add_drain_all(void)
835 #endif /* CONFIG_SMP */
837 atomic_t lru_disable_count
= ATOMIC_INIT(0);
840 * lru_cache_disable() needs to be called before we start compiling
841 * a list of pages to be migrated using isolate_lru_page().
842 * It drains pages on LRU cache and then disable on all cpus until
843 * lru_cache_enable is called.
845 * Must be paired with a call to lru_cache_enable().
847 void lru_cache_disable(void)
849 atomic_inc(&lru_disable_count
);
852 * lru_add_drain_all in the force mode will schedule draining on
853 * all online CPUs so any calls of lru_cache_disabled wrapped by
854 * local_lock or preemption disabled would be ordered by that.
855 * The atomic operation doesn't need to have stronger ordering
856 * requirements because that is enforeced by the scheduling
859 __lru_add_drain_all(true);
866 * release_pages - batched put_page()
867 * @pages: array of pages to release
868 * @nr: number of pages
870 * Decrement the reference count on all the pages in @pages. If it
871 * fell to zero, remove the page from the LRU and free it.
873 void release_pages(struct page
**pages
, int nr
)
876 LIST_HEAD(pages_to_free
);
877 struct lruvec
*lruvec
= NULL
;
879 unsigned int lock_batch
;
881 for (i
= 0; i
< nr
; i
++) {
882 struct page
*page
= pages
[i
];
885 * Make sure the IRQ-safe lock-holding time does not get
886 * excessive with a continuous string of pages from the
887 * same lruvec. The lock is held only if lruvec != NULL.
889 if (lruvec
&& ++lock_batch
== SWAP_CLUSTER_MAX
) {
890 unlock_page_lruvec_irqrestore(lruvec
, flags
);
894 page
= compound_head(page
);
895 if (is_huge_zero_page(page
))
898 if (is_zone_device_page(page
)) {
900 unlock_page_lruvec_irqrestore(lruvec
, flags
);
904 * ZONE_DEVICE pages that return 'false' from
905 * page_is_devmap_managed() do not require special
906 * processing, and instead, expect a call to
907 * put_page_testzero().
909 if (page_is_devmap_managed(page
)) {
910 put_devmap_managed_page(page
);
913 if (put_page_testzero(page
))
914 put_dev_pagemap(page
->pgmap
);
918 if (!put_page_testzero(page
))
921 if (PageCompound(page
)) {
923 unlock_page_lruvec_irqrestore(lruvec
, flags
);
926 __put_compound_page(page
);
931 struct lruvec
*prev_lruvec
= lruvec
;
933 lruvec
= relock_page_lruvec_irqsave(page
, lruvec
,
935 if (prev_lruvec
!= lruvec
)
938 del_page_from_lru_list(page
, lruvec
);
939 __clear_page_lru_flags(page
);
942 __ClearPageWaiters(page
);
944 list_add(&page
->lru
, &pages_to_free
);
947 unlock_page_lruvec_irqrestore(lruvec
, flags
);
949 mem_cgroup_uncharge_list(&pages_to_free
);
950 free_unref_page_list(&pages_to_free
);
952 EXPORT_SYMBOL(release_pages
);
955 * The pages which we're about to release may be in the deferred lru-addition
956 * queues. That would prevent them from really being freed right now. That's
957 * OK from a correctness point of view but is inefficient - those pages may be
958 * cache-warm and we want to give them back to the page allocator ASAP.
960 * So __pagevec_release() will drain those queues here. __pagevec_lru_add()
961 * and __pagevec_lru_add_active() call release_pages() directly to avoid
964 void __pagevec_release(struct pagevec
*pvec
)
966 if (!pvec
->percpu_pvec_drained
) {
968 pvec
->percpu_pvec_drained
= true;
970 release_pages(pvec
->pages
, pagevec_count(pvec
));
971 pagevec_reinit(pvec
);
973 EXPORT_SYMBOL(__pagevec_release
);
975 static void __pagevec_lru_add_fn(struct page
*page
, struct lruvec
*lruvec
)
977 int was_unevictable
= TestClearPageUnevictable(page
);
978 int nr_pages
= thp_nr_pages(page
);
980 VM_BUG_ON_PAGE(PageLRU(page
), page
);
983 * Page becomes evictable in two ways:
984 * 1) Within LRU lock [munlock_vma_page() and __munlock_pagevec()].
985 * 2) Before acquiring LRU lock to put the page to correct LRU and then
986 * a) do PageLRU check with lock [check_move_unevictable_pages]
987 * b) do PageLRU check before lock [clear_page_mlock]
989 * (1) & (2a) are ok as LRU lock will serialize them. For (2b), we need
990 * following strict ordering:
992 * #0: __pagevec_lru_add_fn #1: clear_page_mlock
994 * SetPageLRU() TestClearPageMlocked()
995 * smp_mb() // explicit ordering // above provides strict
997 * PageMlocked() PageLRU()
1000 * if '#1' does not observe setting of PG_lru by '#0' and fails
1001 * isolation, the explicit barrier will make sure that page_evictable
1002 * check will put the page in correct LRU. Without smp_mb(), SetPageLRU
1003 * can be reordered after PageMlocked check and can make '#1' to fail
1004 * the isolation of the page whose Mlocked bit is cleared (#0 is also
1005 * looking at the same page) and the evictable page will be stranded
1006 * in an unevictable LRU.
1009 smp_mb__after_atomic();
1011 if (page_evictable(page
)) {
1012 if (was_unevictable
)
1013 __count_vm_events(UNEVICTABLE_PGRESCUED
, nr_pages
);
1015 ClearPageActive(page
);
1016 SetPageUnevictable(page
);
1017 if (!was_unevictable
)
1018 __count_vm_events(UNEVICTABLE_PGCULLED
, nr_pages
);
1021 add_page_to_lru_list(page
, lruvec
);
1022 trace_mm_lru_insertion(page
);
1026 * Add the passed pages to the LRU, then drop the caller's refcount
1027 * on them. Reinitialises the caller's pagevec.
1029 void __pagevec_lru_add(struct pagevec
*pvec
)
1032 struct lruvec
*lruvec
= NULL
;
1033 unsigned long flags
= 0;
1035 for (i
= 0; i
< pagevec_count(pvec
); i
++) {
1036 struct page
*page
= pvec
->pages
[i
];
1038 lruvec
= relock_page_lruvec_irqsave(page
, lruvec
, &flags
);
1039 __pagevec_lru_add_fn(page
, lruvec
);
1042 unlock_page_lruvec_irqrestore(lruvec
, flags
);
1043 release_pages(pvec
->pages
, pvec
->nr
);
1044 pagevec_reinit(pvec
);
1048 * pagevec_remove_exceptionals - pagevec exceptionals pruning
1049 * @pvec: The pagevec to prune
1051 * find_get_entries() fills both pages and XArray value entries (aka
1052 * exceptional entries) into the pagevec. This function prunes all
1053 * exceptionals from @pvec without leaving holes, so that it can be
1054 * passed on to page-only pagevec operations.
1056 void pagevec_remove_exceptionals(struct pagevec
*pvec
)
1060 for (i
= 0, j
= 0; i
< pagevec_count(pvec
); i
++) {
1061 struct page
*page
= pvec
->pages
[i
];
1062 if (!xa_is_value(page
))
1063 pvec
->pages
[j
++] = page
;
1069 * pagevec_lookup_range - gang pagecache lookup
1070 * @pvec: Where the resulting pages are placed
1071 * @mapping: The address_space to search
1072 * @start: The starting page index
1073 * @end: The final page index
1075 * pagevec_lookup_range() will search for & return a group of up to PAGEVEC_SIZE
1076 * pages in the mapping starting from index @start and upto index @end
1077 * (inclusive). The pages are placed in @pvec. pagevec_lookup() takes a
1078 * reference against the pages in @pvec.
1080 * The search returns a group of mapping-contiguous pages with ascending
1081 * indexes. There may be holes in the indices due to not-present pages. We
1082 * also update @start to index the next page for the traversal.
1084 * pagevec_lookup_range() returns the number of pages which were found. If this
1085 * number is smaller than PAGEVEC_SIZE, the end of specified range has been
1088 unsigned pagevec_lookup_range(struct pagevec
*pvec
,
1089 struct address_space
*mapping
, pgoff_t
*start
, pgoff_t end
)
1091 pvec
->nr
= find_get_pages_range(mapping
, start
, end
, PAGEVEC_SIZE
,
1093 return pagevec_count(pvec
);
1095 EXPORT_SYMBOL(pagevec_lookup_range
);
1097 unsigned pagevec_lookup_range_tag(struct pagevec
*pvec
,
1098 struct address_space
*mapping
, pgoff_t
*index
, pgoff_t end
,
1101 pvec
->nr
= find_get_pages_range_tag(mapping
, index
, end
, tag
,
1102 PAGEVEC_SIZE
, pvec
->pages
);
1103 return pagevec_count(pvec
);
1105 EXPORT_SYMBOL(pagevec_lookup_range_tag
);
1108 * Perform any setup for the swap system
1110 void __init
swap_setup(void)
1112 unsigned long megs
= totalram_pages() >> (20 - PAGE_SHIFT
);
1114 /* Use a smaller cluster for small-memory machines */
1120 * Right now other parts of the system means that we
1121 * _really_ don't want to cluster much more
1125 #ifdef CONFIG_DEV_PAGEMAP_OPS
1126 void put_devmap_managed_page(struct page
*page
)
1130 if (WARN_ON_ONCE(!page_is_devmap_managed(page
)))
1133 count
= page_ref_dec_return(page
);
1136 * devmap page refcounts are 1-based, rather than 0-based: if
1137 * refcount is 1, then the page is free and the refcount is
1138 * stable because nobody holds a reference on the page.
1141 free_devmap_managed_page(page
);
1145 EXPORT_SYMBOL(put_devmap_managed_page
);