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
);
183 * get_kernel_page() - pin a kernel page in memory
184 * @start: starting kernel address
185 * @write: pinning for read/write, currently ignored
186 * @pages: array that receives pointer to the page pinned.
187 * Must be at least nr_segs long.
189 * Returns 1 if page is pinned. If the page was not pinned, returns
190 * -errno. The page returned must be released with a put_page() call
191 * when it is finished with.
193 int get_kernel_page(unsigned long start
, int write
, struct page
**pages
)
195 const struct kvec kiov
= {
196 .iov_base
= (void *)start
,
200 return get_kernel_pages(&kiov
, 1, write
, pages
);
202 EXPORT_SYMBOL_GPL(get_kernel_page
);
204 static void pagevec_lru_move_fn(struct pagevec
*pvec
,
205 void (*move_fn
)(struct page
*page
, struct lruvec
*lruvec
))
208 struct lruvec
*lruvec
= NULL
;
209 unsigned long flags
= 0;
211 for (i
= 0; i
< pagevec_count(pvec
); i
++) {
212 struct page
*page
= pvec
->pages
[i
];
214 /* block memcg migration during page moving between lru */
215 if (!TestClearPageLRU(page
))
218 lruvec
= relock_page_lruvec_irqsave(page
, lruvec
, &flags
);
219 (*move_fn
)(page
, lruvec
);
224 unlock_page_lruvec_irqrestore(lruvec
, flags
);
225 release_pages(pvec
->pages
, pvec
->nr
);
226 pagevec_reinit(pvec
);
229 static void pagevec_move_tail_fn(struct page
*page
, struct lruvec
*lruvec
)
231 if (!PageUnevictable(page
)) {
232 del_page_from_lru_list(page
, lruvec
);
233 ClearPageActive(page
);
234 add_page_to_lru_list_tail(page
, lruvec
);
235 __count_vm_events(PGROTATED
, thp_nr_pages(page
));
239 /* return true if pagevec needs to drain */
240 static bool pagevec_add_and_need_flush(struct pagevec
*pvec
, struct page
*page
)
244 if (!pagevec_add(pvec
, page
) || PageCompound(page
) ||
245 lru_cache_disabled())
252 * Writeback is about to end against a page which has been marked for immediate
253 * reclaim. If it still appears to be reclaimable, move it to the tail of the
256 * rotate_reclaimable_page() must disable IRQs, to prevent nasty races.
258 void rotate_reclaimable_page(struct page
*page
)
260 if (!PageLocked(page
) && !PageDirty(page
) &&
261 !PageUnevictable(page
) && PageLRU(page
)) {
262 struct pagevec
*pvec
;
266 local_lock_irqsave(&lru_rotate
.lock
, flags
);
267 pvec
= this_cpu_ptr(&lru_rotate
.pvec
);
268 if (pagevec_add_and_need_flush(pvec
, page
))
269 pagevec_lru_move_fn(pvec
, pagevec_move_tail_fn
);
270 local_unlock_irqrestore(&lru_rotate
.lock
, flags
);
274 void lru_note_cost(struct lruvec
*lruvec
, bool file
, unsigned int nr_pages
)
277 unsigned long lrusize
;
280 * Hold lruvec->lru_lock is safe here, since
281 * 1) The pinned lruvec in reclaim, or
282 * 2) From a pre-LRU page during refault (which also holds the
283 * rcu lock, so would be safe even if the page was on the LRU
284 * and could move simultaneously to a new lruvec).
286 spin_lock_irq(&lruvec
->lru_lock
);
287 /* Record cost event */
289 lruvec
->file_cost
+= nr_pages
;
291 lruvec
->anon_cost
+= nr_pages
;
294 * Decay previous events
296 * Because workloads change over time (and to avoid
297 * overflow) we keep these statistics as a floating
298 * average, which ends up weighing recent refaults
299 * more than old ones.
301 lrusize
= lruvec_page_state(lruvec
, NR_INACTIVE_ANON
) +
302 lruvec_page_state(lruvec
, NR_ACTIVE_ANON
) +
303 lruvec_page_state(lruvec
, NR_INACTIVE_FILE
) +
304 lruvec_page_state(lruvec
, NR_ACTIVE_FILE
);
306 if (lruvec
->file_cost
+ lruvec
->anon_cost
> lrusize
/ 4) {
307 lruvec
->file_cost
/= 2;
308 lruvec
->anon_cost
/= 2;
310 spin_unlock_irq(&lruvec
->lru_lock
);
311 } while ((lruvec
= parent_lruvec(lruvec
)));
314 void lru_note_cost_page(struct page
*page
)
316 lru_note_cost(mem_cgroup_page_lruvec(page
),
317 page_is_file_lru(page
), thp_nr_pages(page
));
320 static void __activate_page(struct page
*page
, struct lruvec
*lruvec
)
322 if (!PageActive(page
) && !PageUnevictable(page
)) {
323 int nr_pages
= thp_nr_pages(page
);
325 del_page_from_lru_list(page
, lruvec
);
327 add_page_to_lru_list(page
, lruvec
);
328 trace_mm_lru_activate(page
);
330 __count_vm_events(PGACTIVATE
, nr_pages
);
331 __count_memcg_events(lruvec_memcg(lruvec
), PGACTIVATE
,
337 static void activate_page_drain(int cpu
)
339 struct pagevec
*pvec
= &per_cpu(lru_pvecs
.activate_page
, cpu
);
341 if (pagevec_count(pvec
))
342 pagevec_lru_move_fn(pvec
, __activate_page
);
345 static bool need_activate_page_drain(int cpu
)
347 return pagevec_count(&per_cpu(lru_pvecs
.activate_page
, cpu
)) != 0;
350 static void activate_page(struct page
*page
)
352 page
= compound_head(page
);
353 if (PageLRU(page
) && !PageActive(page
) && !PageUnevictable(page
)) {
354 struct pagevec
*pvec
;
356 local_lock(&lru_pvecs
.lock
);
357 pvec
= this_cpu_ptr(&lru_pvecs
.activate_page
);
359 if (pagevec_add_and_need_flush(pvec
, page
))
360 pagevec_lru_move_fn(pvec
, __activate_page
);
361 local_unlock(&lru_pvecs
.lock
);
366 static inline void activate_page_drain(int cpu
)
370 static void activate_page(struct page
*page
)
372 struct lruvec
*lruvec
;
374 page
= compound_head(page
);
375 if (TestClearPageLRU(page
)) {
376 lruvec
= lock_page_lruvec_irq(page
);
377 __activate_page(page
, lruvec
);
378 unlock_page_lruvec_irq(lruvec
);
384 static void __lru_cache_activate_page(struct page
*page
)
386 struct pagevec
*pvec
;
389 local_lock(&lru_pvecs
.lock
);
390 pvec
= this_cpu_ptr(&lru_pvecs
.lru_add
);
393 * Search backwards on the optimistic assumption that the page being
394 * activated has just been added to this pagevec. Note that only
395 * the local pagevec is examined as a !PageLRU page could be in the
396 * process of being released, reclaimed, migrated or on a remote
397 * pagevec that is currently being drained. Furthermore, marking
398 * a remote pagevec's page PageActive potentially hits a race where
399 * a page is marked PageActive just after it is added to the inactive
400 * list causing accounting errors and BUG_ON checks to trigger.
402 for (i
= pagevec_count(pvec
) - 1; i
>= 0; i
--) {
403 struct page
*pagevec_page
= pvec
->pages
[i
];
405 if (pagevec_page
== page
) {
411 local_unlock(&lru_pvecs
.lock
);
415 * Mark a page as having seen activity.
417 * inactive,unreferenced -> inactive,referenced
418 * inactive,referenced -> active,unreferenced
419 * active,unreferenced -> active,referenced
421 * When a newly allocated page is not yet visible, so safe for non-atomic ops,
422 * __SetPageReferenced(page) may be substituted for mark_page_accessed(page).
424 void mark_page_accessed(struct page
*page
)
426 page
= compound_head(page
);
428 if (!PageReferenced(page
)) {
429 SetPageReferenced(page
);
430 } else if (PageUnevictable(page
)) {
432 * Unevictable pages are on the "LRU_UNEVICTABLE" list. But,
433 * this list is never rotated or maintained, so marking an
434 * evictable page accessed has no effect.
436 } else if (!PageActive(page
)) {
438 * If the page is on the LRU, queue it for activation via
439 * lru_pvecs.activate_page. Otherwise, assume the page is on a
440 * pagevec, mark it active and it'll be moved to the active
441 * LRU on the next drain.
446 __lru_cache_activate_page(page
);
447 ClearPageReferenced(page
);
448 workingset_activation(page
);
450 if (page_is_idle(page
))
451 clear_page_idle(page
);
453 EXPORT_SYMBOL(mark_page_accessed
);
456 * lru_cache_add - add a page to a page list
457 * @page: the page to be added to the LRU.
459 * Queue the page for addition to the LRU via pagevec. The decision on whether
460 * to add the page to the [in]active [file|anon] list is deferred until the
461 * pagevec is drained. This gives a chance for the caller of lru_cache_add()
462 * have the page added to the active list using mark_page_accessed().
464 void lru_cache_add(struct page
*page
)
466 struct pagevec
*pvec
;
468 VM_BUG_ON_PAGE(PageActive(page
) && PageUnevictable(page
), page
);
469 VM_BUG_ON_PAGE(PageLRU(page
), page
);
472 local_lock(&lru_pvecs
.lock
);
473 pvec
= this_cpu_ptr(&lru_pvecs
.lru_add
);
474 if (pagevec_add_and_need_flush(pvec
, page
))
475 __pagevec_lru_add(pvec
);
476 local_unlock(&lru_pvecs
.lock
);
478 EXPORT_SYMBOL(lru_cache_add
);
481 * lru_cache_add_inactive_or_unevictable
482 * @page: the page to be added to LRU
483 * @vma: vma in which page is mapped for determining reclaimability
485 * Place @page on the inactive or unevictable LRU list, depending on its
488 void lru_cache_add_inactive_or_unevictable(struct page
*page
,
489 struct vm_area_struct
*vma
)
493 VM_BUG_ON_PAGE(PageLRU(page
), page
);
495 unevictable
= (vma
->vm_flags
& (VM_LOCKED
| VM_SPECIAL
)) == VM_LOCKED
;
496 if (unlikely(unevictable
) && !TestSetPageMlocked(page
)) {
497 int nr_pages
= thp_nr_pages(page
);
499 * We use the irq-unsafe __mod_zone_page_state because this
500 * counter is not modified from interrupt context, and the pte
501 * lock is held(spinlock), which implies preemption disabled.
503 __mod_zone_page_state(page_zone(page
), NR_MLOCK
, nr_pages
);
504 count_vm_events(UNEVICTABLE_PGMLOCKED
, nr_pages
);
510 * If the page can not be invalidated, it is moved to the
511 * inactive list to speed up its reclaim. It is moved to the
512 * head of the list, rather than the tail, to give the flusher
513 * threads some time to write it out, as this is much more
514 * effective than the single-page writeout from reclaim.
516 * If the page isn't page_mapped and dirty/writeback, the page
517 * could reclaim asap using PG_reclaim.
519 * 1. active, mapped page -> none
520 * 2. active, dirty/writeback page -> inactive, head, PG_reclaim
521 * 3. inactive, mapped page -> none
522 * 4. inactive, dirty/writeback page -> inactive, head, PG_reclaim
523 * 5. inactive, clean -> inactive, tail
526 * In 4, why it moves inactive's head, the VM expects the page would
527 * be write it out by flusher threads as this is much more effective
528 * than the single-page writeout from reclaim.
530 static void lru_deactivate_file_fn(struct page
*page
, struct lruvec
*lruvec
)
532 bool active
= PageActive(page
);
533 int nr_pages
= thp_nr_pages(page
);
535 if (PageUnevictable(page
))
538 /* Some processes are using the page */
539 if (page_mapped(page
))
542 del_page_from_lru_list(page
, lruvec
);
543 ClearPageActive(page
);
544 ClearPageReferenced(page
);
546 if (PageWriteback(page
) || PageDirty(page
)) {
548 * PG_reclaim could be raced with end_page_writeback
549 * It can make readahead confusing. But race window
550 * is _really_ small and it's non-critical problem.
552 add_page_to_lru_list(page
, lruvec
);
553 SetPageReclaim(page
);
556 * The page's writeback ends up during pagevec
557 * We move that page into tail of inactive.
559 add_page_to_lru_list_tail(page
, lruvec
);
560 __count_vm_events(PGROTATED
, nr_pages
);
564 __count_vm_events(PGDEACTIVATE
, nr_pages
);
565 __count_memcg_events(lruvec_memcg(lruvec
), PGDEACTIVATE
,
570 static void lru_deactivate_fn(struct page
*page
, struct lruvec
*lruvec
)
572 if (PageActive(page
) && !PageUnevictable(page
)) {
573 int nr_pages
= thp_nr_pages(page
);
575 del_page_from_lru_list(page
, lruvec
);
576 ClearPageActive(page
);
577 ClearPageReferenced(page
);
578 add_page_to_lru_list(page
, lruvec
);
580 __count_vm_events(PGDEACTIVATE
, nr_pages
);
581 __count_memcg_events(lruvec_memcg(lruvec
), PGDEACTIVATE
,
586 static void lru_lazyfree_fn(struct page
*page
, struct lruvec
*lruvec
)
588 if (PageAnon(page
) && PageSwapBacked(page
) &&
589 !PageSwapCache(page
) && !PageUnevictable(page
)) {
590 int nr_pages
= thp_nr_pages(page
);
592 del_page_from_lru_list(page
, lruvec
);
593 ClearPageActive(page
);
594 ClearPageReferenced(page
);
596 * Lazyfree pages are clean anonymous pages. They have
597 * PG_swapbacked flag cleared, to distinguish them from normal
600 ClearPageSwapBacked(page
);
601 add_page_to_lru_list(page
, lruvec
);
603 __count_vm_events(PGLAZYFREE
, nr_pages
);
604 __count_memcg_events(lruvec_memcg(lruvec
), PGLAZYFREE
,
610 * Drain pages out of the cpu's pagevecs.
611 * Either "cpu" is the current CPU, and preemption has already been
612 * disabled; or "cpu" is being hot-unplugged, and is already dead.
614 void lru_add_drain_cpu(int cpu
)
616 struct pagevec
*pvec
= &per_cpu(lru_pvecs
.lru_add
, cpu
);
618 if (pagevec_count(pvec
))
619 __pagevec_lru_add(pvec
);
621 pvec
= &per_cpu(lru_rotate
.pvec
, cpu
);
622 /* Disabling interrupts below acts as a compiler barrier. */
623 if (data_race(pagevec_count(pvec
))) {
626 /* No harm done if a racing interrupt already did this */
627 local_lock_irqsave(&lru_rotate
.lock
, flags
);
628 pagevec_lru_move_fn(pvec
, pagevec_move_tail_fn
);
629 local_unlock_irqrestore(&lru_rotate
.lock
, flags
);
632 pvec
= &per_cpu(lru_pvecs
.lru_deactivate_file
, cpu
);
633 if (pagevec_count(pvec
))
634 pagevec_lru_move_fn(pvec
, lru_deactivate_file_fn
);
636 pvec
= &per_cpu(lru_pvecs
.lru_deactivate
, cpu
);
637 if (pagevec_count(pvec
))
638 pagevec_lru_move_fn(pvec
, lru_deactivate_fn
);
640 pvec
= &per_cpu(lru_pvecs
.lru_lazyfree
, cpu
);
641 if (pagevec_count(pvec
))
642 pagevec_lru_move_fn(pvec
, lru_lazyfree_fn
);
644 activate_page_drain(cpu
);
645 invalidate_bh_lrus_cpu(cpu
);
649 * deactivate_file_page - forcefully deactivate a file page
650 * @page: page to deactivate
652 * This function hints the VM that @page is a good reclaim candidate,
653 * for example if its invalidation fails due to the page being dirty
654 * or under writeback.
656 void deactivate_file_page(struct page
*page
)
659 * In a workload with many unevictable page such as mprotect,
660 * unevictable page deactivation for accelerating reclaim is pointless.
662 if (PageUnevictable(page
))
665 if (likely(get_page_unless_zero(page
))) {
666 struct pagevec
*pvec
;
668 local_lock(&lru_pvecs
.lock
);
669 pvec
= this_cpu_ptr(&lru_pvecs
.lru_deactivate_file
);
671 if (pagevec_add_and_need_flush(pvec
, page
))
672 pagevec_lru_move_fn(pvec
, lru_deactivate_file_fn
);
673 local_unlock(&lru_pvecs
.lock
);
678 * deactivate_page - deactivate a page
679 * @page: page to deactivate
681 * deactivate_page() moves @page to the inactive list if @page was on the active
682 * list and was not an unevictable page. This is done to accelerate the reclaim
685 void deactivate_page(struct page
*page
)
687 if (PageLRU(page
) && PageActive(page
) && !PageUnevictable(page
)) {
688 struct pagevec
*pvec
;
690 local_lock(&lru_pvecs
.lock
);
691 pvec
= this_cpu_ptr(&lru_pvecs
.lru_deactivate
);
693 if (pagevec_add_and_need_flush(pvec
, page
))
694 pagevec_lru_move_fn(pvec
, lru_deactivate_fn
);
695 local_unlock(&lru_pvecs
.lock
);
700 * mark_page_lazyfree - make an anon page lazyfree
701 * @page: page to deactivate
703 * mark_page_lazyfree() moves @page to the inactive file list.
704 * This is done to accelerate the reclaim of @page.
706 void mark_page_lazyfree(struct page
*page
)
708 if (PageLRU(page
) && PageAnon(page
) && PageSwapBacked(page
) &&
709 !PageSwapCache(page
) && !PageUnevictable(page
)) {
710 struct pagevec
*pvec
;
712 local_lock(&lru_pvecs
.lock
);
713 pvec
= this_cpu_ptr(&lru_pvecs
.lru_lazyfree
);
715 if (pagevec_add_and_need_flush(pvec
, page
))
716 pagevec_lru_move_fn(pvec
, lru_lazyfree_fn
);
717 local_unlock(&lru_pvecs
.lock
);
721 void lru_add_drain(void)
723 local_lock(&lru_pvecs
.lock
);
724 lru_add_drain_cpu(smp_processor_id());
725 local_unlock(&lru_pvecs
.lock
);
728 void lru_add_drain_cpu_zone(struct zone
*zone
)
730 local_lock(&lru_pvecs
.lock
);
731 lru_add_drain_cpu(smp_processor_id());
732 drain_local_pages(zone
);
733 local_unlock(&lru_pvecs
.lock
);
738 static DEFINE_PER_CPU(struct work_struct
, lru_add_drain_work
);
740 static void lru_add_drain_per_cpu(struct work_struct
*dummy
)
746 * Doesn't need any cpu hotplug locking because we do rely on per-cpu
747 * kworkers being shut down before our page_alloc_cpu_dead callback is
748 * executed on the offlined cpu.
749 * Calling this function with cpu hotplug locks held can actually lead
750 * to obscure indirect dependencies via WQ context.
752 inline void __lru_add_drain_all(bool force_all_cpus
)
755 * lru_drain_gen - Global pages generation number
757 * (A) Definition: global lru_drain_gen = x implies that all generations
758 * 0 < n <= x are already *scheduled* for draining.
760 * This is an optimization for the highly-contended use case where a
761 * user space workload keeps constantly generating a flow of pages for
764 static unsigned int lru_drain_gen
;
765 static struct cpumask has_work
;
766 static DEFINE_MUTEX(lock
);
767 unsigned cpu
, this_gen
;
770 * Make sure nobody triggers this path before mm_percpu_wq is fully
773 if (WARN_ON(!mm_percpu_wq
))
777 * Guarantee pagevec counter stores visible by this CPU are visible to
778 * other CPUs before loading the current drain generation.
783 * (B) Locally cache global LRU draining generation number
785 * The read barrier ensures that the counter is loaded before the mutex
786 * is taken. It pairs with smp_mb() inside the mutex critical section
789 this_gen
= smp_load_acquire(&lru_drain_gen
);
794 * (C) Exit the draining operation if a newer generation, from another
795 * lru_add_drain_all(), was already scheduled for draining. Check (A).
797 if (unlikely(this_gen
!= lru_drain_gen
&& !force_all_cpus
))
801 * (D) Increment global generation number
803 * Pairs with smp_load_acquire() at (B), outside of the critical
804 * section. Use a full memory barrier to guarantee that the new global
805 * drain generation number is stored before loading pagevec counters.
807 * This pairing must be done here, before the for_each_online_cpu loop
808 * below which drains the page vectors.
810 * Let x, y, and z represent some system CPU numbers, where x < y < z.
811 * Assume CPU #z is in the middle of the for_each_online_cpu loop
812 * below and has already reached CPU #y's per-cpu data. CPU #x comes
813 * along, adds some pages to its per-cpu vectors, then calls
814 * lru_add_drain_all().
816 * If the paired barrier is done at any later step, e.g. after the
817 * loop, CPU #x will just exit at (C) and miss flushing out all of its
820 WRITE_ONCE(lru_drain_gen
, lru_drain_gen
+ 1);
823 cpumask_clear(&has_work
);
824 for_each_online_cpu(cpu
) {
825 struct work_struct
*work
= &per_cpu(lru_add_drain_work
, cpu
);
827 if (force_all_cpus
||
828 pagevec_count(&per_cpu(lru_pvecs
.lru_add
, cpu
)) ||
829 data_race(pagevec_count(&per_cpu(lru_rotate
.pvec
, cpu
))) ||
830 pagevec_count(&per_cpu(lru_pvecs
.lru_deactivate_file
, cpu
)) ||
831 pagevec_count(&per_cpu(lru_pvecs
.lru_deactivate
, cpu
)) ||
832 pagevec_count(&per_cpu(lru_pvecs
.lru_lazyfree
, cpu
)) ||
833 need_activate_page_drain(cpu
) ||
834 has_bh_in_lru(cpu
, NULL
)) {
835 INIT_WORK(work
, lru_add_drain_per_cpu
);
836 queue_work_on(cpu
, mm_percpu_wq
, work
);
837 __cpumask_set_cpu(cpu
, &has_work
);
841 for_each_cpu(cpu
, &has_work
)
842 flush_work(&per_cpu(lru_add_drain_work
, cpu
));
848 void lru_add_drain_all(void)
850 __lru_add_drain_all(false);
853 void lru_add_drain_all(void)
857 #endif /* CONFIG_SMP */
859 atomic_t lru_disable_count
= ATOMIC_INIT(0);
862 * lru_cache_disable() needs to be called before we start compiling
863 * a list of pages to be migrated using isolate_lru_page().
864 * It drains pages on LRU cache and then disable on all cpus until
865 * lru_cache_enable is called.
867 * Must be paired with a call to lru_cache_enable().
869 void lru_cache_disable(void)
871 atomic_inc(&lru_disable_count
);
874 * lru_add_drain_all in the force mode will schedule draining on
875 * all online CPUs so any calls of lru_cache_disabled wrapped by
876 * local_lock or preemption disabled would be ordered by that.
877 * The atomic operation doesn't need to have stronger ordering
878 * requirements because that is enforeced by the scheduling
881 __lru_add_drain_all(true);
888 * release_pages - batched put_page()
889 * @pages: array of pages to release
890 * @nr: number of pages
892 * Decrement the reference count on all the pages in @pages. If it
893 * fell to zero, remove the page from the LRU and free it.
895 void release_pages(struct page
**pages
, int nr
)
898 LIST_HEAD(pages_to_free
);
899 struct lruvec
*lruvec
= NULL
;
901 unsigned int lock_batch
;
903 for (i
= 0; i
< nr
; i
++) {
904 struct page
*page
= pages
[i
];
907 * Make sure the IRQ-safe lock-holding time does not get
908 * excessive with a continuous string of pages from the
909 * same lruvec. The lock is held only if lruvec != NULL.
911 if (lruvec
&& ++lock_batch
== SWAP_CLUSTER_MAX
) {
912 unlock_page_lruvec_irqrestore(lruvec
, flags
);
916 page
= compound_head(page
);
917 if (is_huge_zero_page(page
))
920 if (is_zone_device_page(page
)) {
922 unlock_page_lruvec_irqrestore(lruvec
, flags
);
926 * ZONE_DEVICE pages that return 'false' from
927 * page_is_devmap_managed() do not require special
928 * processing, and instead, expect a call to
929 * put_page_testzero().
931 if (page_is_devmap_managed(page
)) {
932 put_devmap_managed_page(page
);
935 if (put_page_testzero(page
))
936 put_dev_pagemap(page
->pgmap
);
940 if (!put_page_testzero(page
))
943 if (PageCompound(page
)) {
945 unlock_page_lruvec_irqrestore(lruvec
, flags
);
948 __put_compound_page(page
);
953 struct lruvec
*prev_lruvec
= lruvec
;
955 lruvec
= relock_page_lruvec_irqsave(page
, lruvec
,
957 if (prev_lruvec
!= lruvec
)
960 del_page_from_lru_list(page
, lruvec
);
961 __clear_page_lru_flags(page
);
964 __ClearPageWaiters(page
);
966 list_add(&page
->lru
, &pages_to_free
);
969 unlock_page_lruvec_irqrestore(lruvec
, flags
);
971 mem_cgroup_uncharge_list(&pages_to_free
);
972 free_unref_page_list(&pages_to_free
);
974 EXPORT_SYMBOL(release_pages
);
977 * The pages which we're about to release may be in the deferred lru-addition
978 * queues. That would prevent them from really being freed right now. That's
979 * OK from a correctness point of view but is inefficient - those pages may be
980 * cache-warm and we want to give them back to the page allocator ASAP.
982 * So __pagevec_release() will drain those queues here. __pagevec_lru_add()
983 * and __pagevec_lru_add_active() call release_pages() directly to avoid
986 void __pagevec_release(struct pagevec
*pvec
)
988 if (!pvec
->percpu_pvec_drained
) {
990 pvec
->percpu_pvec_drained
= true;
992 release_pages(pvec
->pages
, pagevec_count(pvec
));
993 pagevec_reinit(pvec
);
995 EXPORT_SYMBOL(__pagevec_release
);
997 static void __pagevec_lru_add_fn(struct page
*page
, struct lruvec
*lruvec
)
999 int was_unevictable
= TestClearPageUnevictable(page
);
1000 int nr_pages
= thp_nr_pages(page
);
1002 VM_BUG_ON_PAGE(PageLRU(page
), page
);
1005 * Page becomes evictable in two ways:
1006 * 1) Within LRU lock [munlock_vma_page() and __munlock_pagevec()].
1007 * 2) Before acquiring LRU lock to put the page to correct LRU and then
1008 * a) do PageLRU check with lock [check_move_unevictable_pages]
1009 * b) do PageLRU check before lock [clear_page_mlock]
1011 * (1) & (2a) are ok as LRU lock will serialize them. For (2b), we need
1012 * following strict ordering:
1014 * #0: __pagevec_lru_add_fn #1: clear_page_mlock
1016 * SetPageLRU() TestClearPageMlocked()
1017 * smp_mb() // explicit ordering // above provides strict
1019 * PageMlocked() PageLRU()
1022 * if '#1' does not observe setting of PG_lru by '#0' and fails
1023 * isolation, the explicit barrier will make sure that page_evictable
1024 * check will put the page in correct LRU. Without smp_mb(), SetPageLRU
1025 * can be reordered after PageMlocked check and can make '#1' to fail
1026 * the isolation of the page whose Mlocked bit is cleared (#0 is also
1027 * looking at the same page) and the evictable page will be stranded
1028 * in an unevictable LRU.
1031 smp_mb__after_atomic();
1033 if (page_evictable(page
)) {
1034 if (was_unevictable
)
1035 __count_vm_events(UNEVICTABLE_PGRESCUED
, nr_pages
);
1037 ClearPageActive(page
);
1038 SetPageUnevictable(page
);
1039 if (!was_unevictable
)
1040 __count_vm_events(UNEVICTABLE_PGCULLED
, nr_pages
);
1043 add_page_to_lru_list(page
, lruvec
);
1044 trace_mm_lru_insertion(page
);
1048 * Add the passed pages to the LRU, then drop the caller's refcount
1049 * on them. Reinitialises the caller's pagevec.
1051 void __pagevec_lru_add(struct pagevec
*pvec
)
1054 struct lruvec
*lruvec
= NULL
;
1055 unsigned long flags
= 0;
1057 for (i
= 0; i
< pagevec_count(pvec
); i
++) {
1058 struct page
*page
= pvec
->pages
[i
];
1060 lruvec
= relock_page_lruvec_irqsave(page
, lruvec
, &flags
);
1061 __pagevec_lru_add_fn(page
, lruvec
);
1064 unlock_page_lruvec_irqrestore(lruvec
, flags
);
1065 release_pages(pvec
->pages
, pvec
->nr
);
1066 pagevec_reinit(pvec
);
1070 * pagevec_remove_exceptionals - pagevec exceptionals pruning
1071 * @pvec: The pagevec to prune
1073 * find_get_entries() fills both pages and XArray value entries (aka
1074 * exceptional entries) into the pagevec. This function prunes all
1075 * exceptionals from @pvec without leaving holes, so that it can be
1076 * passed on to page-only pagevec operations.
1078 void pagevec_remove_exceptionals(struct pagevec
*pvec
)
1082 for (i
= 0, j
= 0; i
< pagevec_count(pvec
); i
++) {
1083 struct page
*page
= pvec
->pages
[i
];
1084 if (!xa_is_value(page
))
1085 pvec
->pages
[j
++] = page
;
1091 * pagevec_lookup_range - gang pagecache lookup
1092 * @pvec: Where the resulting pages are placed
1093 * @mapping: The address_space to search
1094 * @start: The starting page index
1095 * @end: The final page index
1097 * pagevec_lookup_range() will search for & return a group of up to PAGEVEC_SIZE
1098 * pages in the mapping starting from index @start and upto index @end
1099 * (inclusive). The pages are placed in @pvec. pagevec_lookup() takes a
1100 * reference against the pages in @pvec.
1102 * The search returns a group of mapping-contiguous pages with ascending
1103 * indexes. There may be holes in the indices due to not-present pages. We
1104 * also update @start to index the next page for the traversal.
1106 * pagevec_lookup_range() returns the number of pages which were found. If this
1107 * number is smaller than PAGEVEC_SIZE, the end of specified range has been
1110 unsigned pagevec_lookup_range(struct pagevec
*pvec
,
1111 struct address_space
*mapping
, pgoff_t
*start
, pgoff_t end
)
1113 pvec
->nr
= find_get_pages_range(mapping
, start
, end
, PAGEVEC_SIZE
,
1115 return pagevec_count(pvec
);
1117 EXPORT_SYMBOL(pagevec_lookup_range
);
1119 unsigned pagevec_lookup_range_tag(struct pagevec
*pvec
,
1120 struct address_space
*mapping
, pgoff_t
*index
, pgoff_t end
,
1123 pvec
->nr
= find_get_pages_range_tag(mapping
, index
, end
, tag
,
1124 PAGEVEC_SIZE
, pvec
->pages
);
1125 return pagevec_count(pvec
);
1127 EXPORT_SYMBOL(pagevec_lookup_range_tag
);
1130 * Perform any setup for the swap system
1132 void __init
swap_setup(void)
1134 unsigned long megs
= totalram_pages() >> (20 - PAGE_SHIFT
);
1136 /* Use a smaller cluster for small-memory machines */
1142 * Right now other parts of the system means that we
1143 * _really_ don't want to cluster much more
1147 #ifdef CONFIG_DEV_PAGEMAP_OPS
1148 void put_devmap_managed_page(struct page
*page
)
1152 if (WARN_ON_ONCE(!page_is_devmap_managed(page
)))
1155 count
= page_ref_dec_return(page
);
1158 * devmap page refcounts are 1-based, rather than 0-based: if
1159 * refcount is 1, then the page is free and the refcount is
1160 * stable because nobody holds a reference on the page.
1163 free_devmap_managed_page(page
);
1167 EXPORT_SYMBOL(put_devmap_managed_page
);