Loading...
1/*
2 * linux/mm/swap.c
3 *
4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 */
6
7/*
8 * This file contains the default values for the operation of the
9 * Linux VM subsystem. Fine-tuning documentation can be found in
10 * Documentation/sysctl/vm.txt.
11 * Started 18.12.91
12 * Swap aging added 23.2.95, Stephen Tweedie.
13 * Buffermem limits added 12.3.98, Rik van Riel.
14 */
15
16#include <linux/mm.h>
17#include <linux/sched.h>
18#include <linux/kernel_stat.h>
19#include <linux/swap.h>
20#include <linux/mman.h>
21#include <linux/pagemap.h>
22#include <linux/pagevec.h>
23#include <linux/init.h>
24#include <linux/export.h>
25#include <linux/mm_inline.h>
26#include <linux/percpu_counter.h>
27#include <linux/memremap.h>
28#include <linux/percpu.h>
29#include <linux/cpu.h>
30#include <linux/notifier.h>
31#include <linux/backing-dev.h>
32#include <linux/memcontrol.h>
33#include <linux/gfp.h>
34#include <linux/uio.h>
35#include <linux/hugetlb.h>
36#include <linux/page_idle.h>
37
38#include "internal.h"
39
40#define CREATE_TRACE_POINTS
41#include <trace/events/pagemap.h>
42
43/* How many pages do we try to swap or page in/out together? */
44int page_cluster;
45
46static DEFINE_PER_CPU(struct pagevec, lru_add_pvec);
47static DEFINE_PER_CPU(struct pagevec, lru_rotate_pvecs);
48static DEFINE_PER_CPU(struct pagevec, lru_deactivate_file_pvecs);
49static DEFINE_PER_CPU(struct pagevec, lru_lazyfree_pvecs);
50#ifdef CONFIG_SMP
51static DEFINE_PER_CPU(struct pagevec, activate_page_pvecs);
52#endif
53
54/*
55 * This path almost never happens for VM activity - pages are normally
56 * freed via pagevecs. But it gets used by networking.
57 */
58static void __page_cache_release(struct page *page)
59{
60 if (PageLRU(page)) {
61 struct zone *zone = page_zone(page);
62 struct lruvec *lruvec;
63 unsigned long flags;
64
65 spin_lock_irqsave(zone_lru_lock(zone), flags);
66 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
67 VM_BUG_ON_PAGE(!PageLRU(page), page);
68 __ClearPageLRU(page);
69 del_page_from_lru_list(page, lruvec, page_off_lru(page));
70 spin_unlock_irqrestore(zone_lru_lock(zone), flags);
71 }
72 __ClearPageWaiters(page);
73 mem_cgroup_uncharge(page);
74}
75
76static void __put_single_page(struct page *page)
77{
78 __page_cache_release(page);
79 free_unref_page(page);
80}
81
82static void __put_compound_page(struct page *page)
83{
84 compound_page_dtor *dtor;
85
86 /*
87 * __page_cache_release() is supposed to be called for thp, not for
88 * hugetlb. This is because hugetlb page does never have PageLRU set
89 * (it's never listed to any LRU lists) and no memcg routines should
90 * be called for hugetlb (it has a separate hugetlb_cgroup.)
91 */
92 if (!PageHuge(page))
93 __page_cache_release(page);
94 dtor = get_compound_page_dtor(page);
95 (*dtor)(page);
96}
97
98void __put_page(struct page *page)
99{
100 if (is_zone_device_page(page)) {
101 put_dev_pagemap(page->pgmap);
102
103 /*
104 * The page belongs to the device that created pgmap. Do
105 * not return it to page allocator.
106 */
107 return;
108 }
109
110 if (unlikely(PageCompound(page)))
111 __put_compound_page(page);
112 else
113 __put_single_page(page);
114}
115EXPORT_SYMBOL(__put_page);
116
117/**
118 * put_pages_list() - release a list of pages
119 * @pages: list of pages threaded on page->lru
120 *
121 * Release a list of pages which are strung together on page.lru. Currently
122 * used by read_cache_pages() and related error recovery code.
123 */
124void put_pages_list(struct list_head *pages)
125{
126 while (!list_empty(pages)) {
127 struct page *victim;
128
129 victim = list_entry(pages->prev, struct page, lru);
130 list_del(&victim->lru);
131 put_page(victim);
132 }
133}
134EXPORT_SYMBOL(put_pages_list);
135
136/*
137 * get_kernel_pages() - pin kernel pages in memory
138 * @kiov: An array of struct kvec structures
139 * @nr_segs: number of segments to pin
140 * @write: pinning for read/write, currently ignored
141 * @pages: array that receives pointers to the pages pinned.
142 * Should be at least nr_segs long.
143 *
144 * Returns number of pages pinned. This may be fewer than the number
145 * requested. If nr_pages is 0 or negative, returns 0. If no pages
146 * were pinned, returns -errno. Each page returned must be released
147 * with a put_page() call when it is finished with.
148 */
149int get_kernel_pages(const struct kvec *kiov, int nr_segs, int write,
150 struct page **pages)
151{
152 int seg;
153
154 for (seg = 0; seg < nr_segs; seg++) {
155 if (WARN_ON(kiov[seg].iov_len != PAGE_SIZE))
156 return seg;
157
158 pages[seg] = kmap_to_page(kiov[seg].iov_base);
159 get_page(pages[seg]);
160 }
161
162 return seg;
163}
164EXPORT_SYMBOL_GPL(get_kernel_pages);
165
166/*
167 * get_kernel_page() - pin a kernel page in memory
168 * @start: starting kernel address
169 * @write: pinning for read/write, currently ignored
170 * @pages: array that receives pointer to the page pinned.
171 * Must be at least nr_segs long.
172 *
173 * Returns 1 if page is pinned. If the page was not pinned, returns
174 * -errno. The page returned must be released with a put_page() call
175 * when it is finished with.
176 */
177int get_kernel_page(unsigned long start, int write, struct page **pages)
178{
179 const struct kvec kiov = {
180 .iov_base = (void *)start,
181 .iov_len = PAGE_SIZE
182 };
183
184 return get_kernel_pages(&kiov, 1, write, pages);
185}
186EXPORT_SYMBOL_GPL(get_kernel_page);
187
188static void pagevec_lru_move_fn(struct pagevec *pvec,
189 void (*move_fn)(struct page *page, struct lruvec *lruvec, void *arg),
190 void *arg)
191{
192 int i;
193 struct pglist_data *pgdat = NULL;
194 struct lruvec *lruvec;
195 unsigned long flags = 0;
196
197 for (i = 0; i < pagevec_count(pvec); i++) {
198 struct page *page = pvec->pages[i];
199 struct pglist_data *pagepgdat = page_pgdat(page);
200
201 if (pagepgdat != pgdat) {
202 if (pgdat)
203 spin_unlock_irqrestore(&pgdat->lru_lock, flags);
204 pgdat = pagepgdat;
205 spin_lock_irqsave(&pgdat->lru_lock, flags);
206 }
207
208 lruvec = mem_cgroup_page_lruvec(page, pgdat);
209 (*move_fn)(page, lruvec, arg);
210 }
211 if (pgdat)
212 spin_unlock_irqrestore(&pgdat->lru_lock, flags);
213 release_pages(pvec->pages, pvec->nr);
214 pagevec_reinit(pvec);
215}
216
217static void pagevec_move_tail_fn(struct page *page, struct lruvec *lruvec,
218 void *arg)
219{
220 int *pgmoved = arg;
221
222 if (PageLRU(page) && !PageUnevictable(page)) {
223 del_page_from_lru_list(page, lruvec, page_lru(page));
224 ClearPageActive(page);
225 add_page_to_lru_list_tail(page, lruvec, page_lru(page));
226 (*pgmoved)++;
227 }
228}
229
230/*
231 * pagevec_move_tail() must be called with IRQ disabled.
232 * Otherwise this may cause nasty races.
233 */
234static void pagevec_move_tail(struct pagevec *pvec)
235{
236 int pgmoved = 0;
237
238 pagevec_lru_move_fn(pvec, pagevec_move_tail_fn, &pgmoved);
239 __count_vm_events(PGROTATED, pgmoved);
240}
241
242/*
243 * Writeback is about to end against a page which has been marked for immediate
244 * reclaim. If it still appears to be reclaimable, move it to the tail of the
245 * inactive list.
246 */
247void rotate_reclaimable_page(struct page *page)
248{
249 if (!PageLocked(page) && !PageDirty(page) &&
250 !PageUnevictable(page) && PageLRU(page)) {
251 struct pagevec *pvec;
252 unsigned long flags;
253
254 get_page(page);
255 local_irq_save(flags);
256 pvec = this_cpu_ptr(&lru_rotate_pvecs);
257 if (!pagevec_add(pvec, page) || PageCompound(page))
258 pagevec_move_tail(pvec);
259 local_irq_restore(flags);
260 }
261}
262
263static void update_page_reclaim_stat(struct lruvec *lruvec,
264 int file, int rotated)
265{
266 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
267
268 reclaim_stat->recent_scanned[file]++;
269 if (rotated)
270 reclaim_stat->recent_rotated[file]++;
271}
272
273static void __activate_page(struct page *page, struct lruvec *lruvec,
274 void *arg)
275{
276 if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
277 int file = page_is_file_cache(page);
278 int lru = page_lru_base_type(page);
279
280 del_page_from_lru_list(page, lruvec, lru);
281 SetPageActive(page);
282 lru += LRU_ACTIVE;
283 add_page_to_lru_list(page, lruvec, lru);
284 trace_mm_lru_activate(page);
285
286 __count_vm_event(PGACTIVATE);
287 update_page_reclaim_stat(lruvec, file, 1);
288 }
289}
290
291#ifdef CONFIG_SMP
292static void activate_page_drain(int cpu)
293{
294 struct pagevec *pvec = &per_cpu(activate_page_pvecs, cpu);
295
296 if (pagevec_count(pvec))
297 pagevec_lru_move_fn(pvec, __activate_page, NULL);
298}
299
300static bool need_activate_page_drain(int cpu)
301{
302 return pagevec_count(&per_cpu(activate_page_pvecs, cpu)) != 0;
303}
304
305void activate_page(struct page *page)
306{
307 page = compound_head(page);
308 if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
309 struct pagevec *pvec = &get_cpu_var(activate_page_pvecs);
310
311 get_page(page);
312 if (!pagevec_add(pvec, page) || PageCompound(page))
313 pagevec_lru_move_fn(pvec, __activate_page, NULL);
314 put_cpu_var(activate_page_pvecs);
315 }
316}
317
318#else
319static inline void activate_page_drain(int cpu)
320{
321}
322
323static bool need_activate_page_drain(int cpu)
324{
325 return false;
326}
327
328void activate_page(struct page *page)
329{
330 struct zone *zone = page_zone(page);
331
332 page = compound_head(page);
333 spin_lock_irq(zone_lru_lock(zone));
334 __activate_page(page, mem_cgroup_page_lruvec(page, zone->zone_pgdat), NULL);
335 spin_unlock_irq(zone_lru_lock(zone));
336}
337#endif
338
339static void __lru_cache_activate_page(struct page *page)
340{
341 struct pagevec *pvec = &get_cpu_var(lru_add_pvec);
342 int i;
343
344 /*
345 * Search backwards on the optimistic assumption that the page being
346 * activated has just been added to this pagevec. Note that only
347 * the local pagevec is examined as a !PageLRU page could be in the
348 * process of being released, reclaimed, migrated or on a remote
349 * pagevec that is currently being drained. Furthermore, marking
350 * a remote pagevec's page PageActive potentially hits a race where
351 * a page is marked PageActive just after it is added to the inactive
352 * list causing accounting errors and BUG_ON checks to trigger.
353 */
354 for (i = pagevec_count(pvec) - 1; i >= 0; i--) {
355 struct page *pagevec_page = pvec->pages[i];
356
357 if (pagevec_page == page) {
358 SetPageActive(page);
359 break;
360 }
361 }
362
363 put_cpu_var(lru_add_pvec);
364}
365
366/*
367 * Mark a page as having seen activity.
368 *
369 * inactive,unreferenced -> inactive,referenced
370 * inactive,referenced -> active,unreferenced
371 * active,unreferenced -> active,referenced
372 *
373 * When a newly allocated page is not yet visible, so safe for non-atomic ops,
374 * __SetPageReferenced(page) may be substituted for mark_page_accessed(page).
375 */
376void mark_page_accessed(struct page *page)
377{
378 page = compound_head(page);
379 if (!PageActive(page) && !PageUnevictable(page) &&
380 PageReferenced(page)) {
381
382 /*
383 * If the page is on the LRU, queue it for activation via
384 * activate_page_pvecs. Otherwise, assume the page is on a
385 * pagevec, mark it active and it'll be moved to the active
386 * LRU on the next drain.
387 */
388 if (PageLRU(page))
389 activate_page(page);
390 else
391 __lru_cache_activate_page(page);
392 ClearPageReferenced(page);
393 if (page_is_file_cache(page))
394 workingset_activation(page);
395 } else if (!PageReferenced(page)) {
396 SetPageReferenced(page);
397 }
398 if (page_is_idle(page))
399 clear_page_idle(page);
400}
401EXPORT_SYMBOL(mark_page_accessed);
402
403static void __lru_cache_add(struct page *page)
404{
405 struct pagevec *pvec = &get_cpu_var(lru_add_pvec);
406
407 get_page(page);
408 if (!pagevec_add(pvec, page) || PageCompound(page))
409 __pagevec_lru_add(pvec);
410 put_cpu_var(lru_add_pvec);
411}
412
413/**
414 * lru_cache_add_anon - add a page to the page lists
415 * @page: the page to add
416 */
417void lru_cache_add_anon(struct page *page)
418{
419 if (PageActive(page))
420 ClearPageActive(page);
421 __lru_cache_add(page);
422}
423
424void lru_cache_add_file(struct page *page)
425{
426 if (PageActive(page))
427 ClearPageActive(page);
428 __lru_cache_add(page);
429}
430EXPORT_SYMBOL(lru_cache_add_file);
431
432/**
433 * lru_cache_add - add a page to a page list
434 * @page: the page to be added to the LRU.
435 *
436 * Queue the page for addition to the LRU via pagevec. The decision on whether
437 * to add the page to the [in]active [file|anon] list is deferred until the
438 * pagevec is drained. This gives a chance for the caller of lru_cache_add()
439 * have the page added to the active list using mark_page_accessed().
440 */
441void lru_cache_add(struct page *page)
442{
443 VM_BUG_ON_PAGE(PageActive(page) && PageUnevictable(page), page);
444 VM_BUG_ON_PAGE(PageLRU(page), page);
445 __lru_cache_add(page);
446}
447
448/**
449 * lru_cache_add_active_or_unevictable
450 * @page: the page to be added to LRU
451 * @vma: vma in which page is mapped for determining reclaimability
452 *
453 * Place @page on the active or unevictable LRU list, depending on its
454 * evictability. Note that if the page is not evictable, it goes
455 * directly back onto it's zone's unevictable list, it does NOT use a
456 * per cpu pagevec.
457 */
458void lru_cache_add_active_or_unevictable(struct page *page,
459 struct vm_area_struct *vma)
460{
461 VM_BUG_ON_PAGE(PageLRU(page), page);
462
463 if (likely((vma->vm_flags & (VM_LOCKED | VM_SPECIAL)) != VM_LOCKED))
464 SetPageActive(page);
465 else if (!TestSetPageMlocked(page)) {
466 /*
467 * We use the irq-unsafe __mod_zone_page_stat because this
468 * counter is not modified from interrupt context, and the pte
469 * lock is held(spinlock), which implies preemption disabled.
470 */
471 __mod_zone_page_state(page_zone(page), NR_MLOCK,
472 hpage_nr_pages(page));
473 count_vm_event(UNEVICTABLE_PGMLOCKED);
474 }
475 lru_cache_add(page);
476}
477
478/*
479 * If the page can not be invalidated, it is moved to the
480 * inactive list to speed up its reclaim. It is moved to the
481 * head of the list, rather than the tail, to give the flusher
482 * threads some time to write it out, as this is much more
483 * effective than the single-page writeout from reclaim.
484 *
485 * If the page isn't page_mapped and dirty/writeback, the page
486 * could reclaim asap using PG_reclaim.
487 *
488 * 1. active, mapped page -> none
489 * 2. active, dirty/writeback page -> inactive, head, PG_reclaim
490 * 3. inactive, mapped page -> none
491 * 4. inactive, dirty/writeback page -> inactive, head, PG_reclaim
492 * 5. inactive, clean -> inactive, tail
493 * 6. Others -> none
494 *
495 * In 4, why it moves inactive's head, the VM expects the page would
496 * be write it out by flusher threads as this is much more effective
497 * than the single-page writeout from reclaim.
498 */
499static void lru_deactivate_file_fn(struct page *page, struct lruvec *lruvec,
500 void *arg)
501{
502 int lru, file;
503 bool active;
504
505 if (!PageLRU(page))
506 return;
507
508 if (PageUnevictable(page))
509 return;
510
511 /* Some processes are using the page */
512 if (page_mapped(page))
513 return;
514
515 active = PageActive(page);
516 file = page_is_file_cache(page);
517 lru = page_lru_base_type(page);
518
519 del_page_from_lru_list(page, lruvec, lru + active);
520 ClearPageActive(page);
521 ClearPageReferenced(page);
522 add_page_to_lru_list(page, lruvec, lru);
523
524 if (PageWriteback(page) || PageDirty(page)) {
525 /*
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.
529 */
530 SetPageReclaim(page);
531 } else {
532 /*
533 * The page's writeback ends up during pagevec
534 * We moves tha page into tail of inactive.
535 */
536 list_move_tail(&page->lru, &lruvec->lists[lru]);
537 __count_vm_event(PGROTATED);
538 }
539
540 if (active)
541 __count_vm_event(PGDEACTIVATE);
542 update_page_reclaim_stat(lruvec, file, 0);
543}
544
545
546static void lru_lazyfree_fn(struct page *page, struct lruvec *lruvec,
547 void *arg)
548{
549 if (PageLRU(page) && PageAnon(page) && PageSwapBacked(page) &&
550 !PageSwapCache(page) && !PageUnevictable(page)) {
551 bool active = PageActive(page);
552
553 del_page_from_lru_list(page, lruvec,
554 LRU_INACTIVE_ANON + active);
555 ClearPageActive(page);
556 ClearPageReferenced(page);
557 /*
558 * lazyfree pages are clean anonymous pages. They have
559 * SwapBacked flag cleared to distinguish normal anonymous
560 * pages
561 */
562 ClearPageSwapBacked(page);
563 add_page_to_lru_list(page, lruvec, LRU_INACTIVE_FILE);
564
565 __count_vm_events(PGLAZYFREE, hpage_nr_pages(page));
566 count_memcg_page_event(page, PGLAZYFREE);
567 update_page_reclaim_stat(lruvec, 1, 0);
568 }
569}
570
571/*
572 * Drain pages out of the cpu's pagevecs.
573 * Either "cpu" is the current CPU, and preemption has already been
574 * disabled; or "cpu" is being hot-unplugged, and is already dead.
575 */
576void lru_add_drain_cpu(int cpu)
577{
578 struct pagevec *pvec = &per_cpu(lru_add_pvec, cpu);
579
580 if (pagevec_count(pvec))
581 __pagevec_lru_add(pvec);
582
583 pvec = &per_cpu(lru_rotate_pvecs, cpu);
584 if (pagevec_count(pvec)) {
585 unsigned long flags;
586
587 /* No harm done if a racing interrupt already did this */
588 local_irq_save(flags);
589 pagevec_move_tail(pvec);
590 local_irq_restore(flags);
591 }
592
593 pvec = &per_cpu(lru_deactivate_file_pvecs, cpu);
594 if (pagevec_count(pvec))
595 pagevec_lru_move_fn(pvec, lru_deactivate_file_fn, NULL);
596
597 pvec = &per_cpu(lru_lazyfree_pvecs, cpu);
598 if (pagevec_count(pvec))
599 pagevec_lru_move_fn(pvec, lru_lazyfree_fn, NULL);
600
601 activate_page_drain(cpu);
602}
603
604/**
605 * deactivate_file_page - forcefully deactivate a file page
606 * @page: page to deactivate
607 *
608 * This function hints the VM that @page is a good reclaim candidate,
609 * for example if its invalidation fails due to the page being dirty
610 * or under writeback.
611 */
612void deactivate_file_page(struct page *page)
613{
614 /*
615 * In a workload with many unevictable page such as mprotect,
616 * unevictable page deactivation for accelerating reclaim is pointless.
617 */
618 if (PageUnevictable(page))
619 return;
620
621 if (likely(get_page_unless_zero(page))) {
622 struct pagevec *pvec = &get_cpu_var(lru_deactivate_file_pvecs);
623
624 if (!pagevec_add(pvec, page) || PageCompound(page))
625 pagevec_lru_move_fn(pvec, lru_deactivate_file_fn, NULL);
626 put_cpu_var(lru_deactivate_file_pvecs);
627 }
628}
629
630/**
631 * mark_page_lazyfree - make an anon page lazyfree
632 * @page: page to deactivate
633 *
634 * mark_page_lazyfree() moves @page to the inactive file list.
635 * This is done to accelerate the reclaim of @page.
636 */
637void mark_page_lazyfree(struct page *page)
638{
639 if (PageLRU(page) && PageAnon(page) && PageSwapBacked(page) &&
640 !PageSwapCache(page) && !PageUnevictable(page)) {
641 struct pagevec *pvec = &get_cpu_var(lru_lazyfree_pvecs);
642
643 get_page(page);
644 if (!pagevec_add(pvec, page) || PageCompound(page))
645 pagevec_lru_move_fn(pvec, lru_lazyfree_fn, NULL);
646 put_cpu_var(lru_lazyfree_pvecs);
647 }
648}
649
650void lru_add_drain(void)
651{
652 lru_add_drain_cpu(get_cpu());
653 put_cpu();
654}
655
656static void lru_add_drain_per_cpu(struct work_struct *dummy)
657{
658 lru_add_drain();
659}
660
661static DEFINE_PER_CPU(struct work_struct, lru_add_drain_work);
662
663/*
664 * Doesn't need any cpu hotplug locking because we do rely on per-cpu
665 * kworkers being shut down before our page_alloc_cpu_dead callback is
666 * executed on the offlined cpu.
667 * Calling this function with cpu hotplug locks held can actually lead
668 * to obscure indirect dependencies via WQ context.
669 */
670void lru_add_drain_all(void)
671{
672 static DEFINE_MUTEX(lock);
673 static struct cpumask has_work;
674 int cpu;
675
676 /*
677 * Make sure nobody triggers this path before mm_percpu_wq is fully
678 * initialized.
679 */
680 if (WARN_ON(!mm_percpu_wq))
681 return;
682
683 mutex_lock(&lock);
684 cpumask_clear(&has_work);
685
686 for_each_online_cpu(cpu) {
687 struct work_struct *work = &per_cpu(lru_add_drain_work, cpu);
688
689 if (pagevec_count(&per_cpu(lru_add_pvec, cpu)) ||
690 pagevec_count(&per_cpu(lru_rotate_pvecs, cpu)) ||
691 pagevec_count(&per_cpu(lru_deactivate_file_pvecs, cpu)) ||
692 pagevec_count(&per_cpu(lru_lazyfree_pvecs, cpu)) ||
693 need_activate_page_drain(cpu)) {
694 INIT_WORK(work, lru_add_drain_per_cpu);
695 queue_work_on(cpu, mm_percpu_wq, work);
696 cpumask_set_cpu(cpu, &has_work);
697 }
698 }
699
700 for_each_cpu(cpu, &has_work)
701 flush_work(&per_cpu(lru_add_drain_work, cpu));
702
703 mutex_unlock(&lock);
704}
705
706/**
707 * release_pages - batched put_page()
708 * @pages: array of pages to release
709 * @nr: number of pages
710 *
711 * Decrement the reference count on all the pages in @pages. If it
712 * fell to zero, remove the page from the LRU and free it.
713 */
714void release_pages(struct page **pages, int nr)
715{
716 int i;
717 LIST_HEAD(pages_to_free);
718 struct pglist_data *locked_pgdat = NULL;
719 struct lruvec *lruvec;
720 unsigned long uninitialized_var(flags);
721 unsigned int uninitialized_var(lock_batch);
722
723 for (i = 0; i < nr; i++) {
724 struct page *page = pages[i];
725
726 /*
727 * Make sure the IRQ-safe lock-holding time does not get
728 * excessive with a continuous string of pages from the
729 * same pgdat. The lock is held only if pgdat != NULL.
730 */
731 if (locked_pgdat && ++lock_batch == SWAP_CLUSTER_MAX) {
732 spin_unlock_irqrestore(&locked_pgdat->lru_lock, flags);
733 locked_pgdat = NULL;
734 }
735
736 if (is_huge_zero_page(page))
737 continue;
738
739 /* Device public page can not be huge page */
740 if (is_device_public_page(page)) {
741 if (locked_pgdat) {
742 spin_unlock_irqrestore(&locked_pgdat->lru_lock,
743 flags);
744 locked_pgdat = NULL;
745 }
746 put_zone_device_private_or_public_page(page);
747 continue;
748 }
749
750 page = compound_head(page);
751 if (!put_page_testzero(page))
752 continue;
753
754 if (PageCompound(page)) {
755 if (locked_pgdat) {
756 spin_unlock_irqrestore(&locked_pgdat->lru_lock, flags);
757 locked_pgdat = NULL;
758 }
759 __put_compound_page(page);
760 continue;
761 }
762
763 if (PageLRU(page)) {
764 struct pglist_data *pgdat = page_pgdat(page);
765
766 if (pgdat != locked_pgdat) {
767 if (locked_pgdat)
768 spin_unlock_irqrestore(&locked_pgdat->lru_lock,
769 flags);
770 lock_batch = 0;
771 locked_pgdat = pgdat;
772 spin_lock_irqsave(&locked_pgdat->lru_lock, flags);
773 }
774
775 lruvec = mem_cgroup_page_lruvec(page, locked_pgdat);
776 VM_BUG_ON_PAGE(!PageLRU(page), page);
777 __ClearPageLRU(page);
778 del_page_from_lru_list(page, lruvec, page_off_lru(page));
779 }
780
781 /* Clear Active bit in case of parallel mark_page_accessed */
782 __ClearPageActive(page);
783 __ClearPageWaiters(page);
784
785 list_add(&page->lru, &pages_to_free);
786 }
787 if (locked_pgdat)
788 spin_unlock_irqrestore(&locked_pgdat->lru_lock, flags);
789
790 mem_cgroup_uncharge_list(&pages_to_free);
791 free_unref_page_list(&pages_to_free);
792}
793EXPORT_SYMBOL(release_pages);
794
795/*
796 * The pages which we're about to release may be in the deferred lru-addition
797 * queues. That would prevent them from really being freed right now. That's
798 * OK from a correctness point of view but is inefficient - those pages may be
799 * cache-warm and we want to give them back to the page allocator ASAP.
800 *
801 * So __pagevec_release() will drain those queues here. __pagevec_lru_add()
802 * and __pagevec_lru_add_active() call release_pages() directly to avoid
803 * mutual recursion.
804 */
805void __pagevec_release(struct pagevec *pvec)
806{
807 if (!pvec->percpu_pvec_drained) {
808 lru_add_drain();
809 pvec->percpu_pvec_drained = true;
810 }
811 release_pages(pvec->pages, pagevec_count(pvec));
812 pagevec_reinit(pvec);
813}
814EXPORT_SYMBOL(__pagevec_release);
815
816#ifdef CONFIG_TRANSPARENT_HUGEPAGE
817/* used by __split_huge_page_refcount() */
818void lru_add_page_tail(struct page *page, struct page *page_tail,
819 struct lruvec *lruvec, struct list_head *list)
820{
821 const int file = 0;
822
823 VM_BUG_ON_PAGE(!PageHead(page), page);
824 VM_BUG_ON_PAGE(PageCompound(page_tail), page);
825 VM_BUG_ON_PAGE(PageLRU(page_tail), page);
826 VM_BUG_ON(NR_CPUS != 1 &&
827 !spin_is_locked(&lruvec_pgdat(lruvec)->lru_lock));
828
829 if (!list)
830 SetPageLRU(page_tail);
831
832 if (likely(PageLRU(page)))
833 list_add_tail(&page_tail->lru, &page->lru);
834 else if (list) {
835 /* page reclaim is reclaiming a huge page */
836 get_page(page_tail);
837 list_add_tail(&page_tail->lru, list);
838 } else {
839 struct list_head *list_head;
840 /*
841 * Head page has not yet been counted, as an hpage,
842 * so we must account for each subpage individually.
843 *
844 * Use the standard add function to put page_tail on the list,
845 * but then correct its position so they all end up in order.
846 */
847 add_page_to_lru_list(page_tail, lruvec, page_lru(page_tail));
848 list_head = page_tail->lru.prev;
849 list_move_tail(&page_tail->lru, list_head);
850 }
851
852 if (!PageUnevictable(page))
853 update_page_reclaim_stat(lruvec, file, PageActive(page_tail));
854}
855#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
856
857static void __pagevec_lru_add_fn(struct page *page, struct lruvec *lruvec,
858 void *arg)
859{
860 enum lru_list lru;
861 int was_unevictable = TestClearPageUnevictable(page);
862
863 VM_BUG_ON_PAGE(PageLRU(page), page);
864
865 SetPageLRU(page);
866 /*
867 * Page becomes evictable in two ways:
868 * 1) Within LRU lock [munlock_vma_pages() and __munlock_pagevec()].
869 * 2) Before acquiring LRU lock to put the page to correct LRU and then
870 * a) do PageLRU check with lock [check_move_unevictable_pages]
871 * b) do PageLRU check before lock [clear_page_mlock]
872 *
873 * (1) & (2a) are ok as LRU lock will serialize them. For (2b), we need
874 * following strict ordering:
875 *
876 * #0: __pagevec_lru_add_fn #1: clear_page_mlock
877 *
878 * SetPageLRU() TestClearPageMlocked()
879 * smp_mb() // explicit ordering // above provides strict
880 * // ordering
881 * PageMlocked() PageLRU()
882 *
883 *
884 * if '#1' does not observe setting of PG_lru by '#0' and fails
885 * isolation, the explicit barrier will make sure that page_evictable
886 * check will put the page in correct LRU. Without smp_mb(), SetPageLRU
887 * can be reordered after PageMlocked check and can make '#1' to fail
888 * the isolation of the page whose Mlocked bit is cleared (#0 is also
889 * looking at the same page) and the evictable page will be stranded
890 * in an unevictable LRU.
891 */
892 smp_mb();
893
894 if (page_evictable(page)) {
895 lru = page_lru(page);
896 update_page_reclaim_stat(lruvec, page_is_file_cache(page),
897 PageActive(page));
898 if (was_unevictable)
899 count_vm_event(UNEVICTABLE_PGRESCUED);
900 } else {
901 lru = LRU_UNEVICTABLE;
902 ClearPageActive(page);
903 SetPageUnevictable(page);
904 if (!was_unevictable)
905 count_vm_event(UNEVICTABLE_PGCULLED);
906 }
907
908 add_page_to_lru_list(page, lruvec, lru);
909 trace_mm_lru_insertion(page, lru);
910}
911
912/*
913 * Add the passed pages to the LRU, then drop the caller's refcount
914 * on them. Reinitialises the caller's pagevec.
915 */
916void __pagevec_lru_add(struct pagevec *pvec)
917{
918 pagevec_lru_move_fn(pvec, __pagevec_lru_add_fn, NULL);
919}
920EXPORT_SYMBOL(__pagevec_lru_add);
921
922/**
923 * pagevec_lookup_entries - gang pagecache lookup
924 * @pvec: Where the resulting entries are placed
925 * @mapping: The address_space to search
926 * @start: The starting entry index
927 * @nr_entries: The maximum number of pages
928 * @indices: The cache indices corresponding to the entries in @pvec
929 *
930 * pagevec_lookup_entries() will search for and return a group of up
931 * to @nr_pages pages and shadow entries in the mapping. All
932 * entries are placed in @pvec. pagevec_lookup_entries() takes a
933 * reference against actual pages in @pvec.
934 *
935 * The search returns a group of mapping-contiguous entries with
936 * ascending indexes. There may be holes in the indices due to
937 * not-present entries.
938 *
939 * pagevec_lookup_entries() returns the number of entries which were
940 * found.
941 */
942unsigned pagevec_lookup_entries(struct pagevec *pvec,
943 struct address_space *mapping,
944 pgoff_t start, unsigned nr_entries,
945 pgoff_t *indices)
946{
947 pvec->nr = find_get_entries(mapping, start, nr_entries,
948 pvec->pages, indices);
949 return pagevec_count(pvec);
950}
951
952/**
953 * pagevec_remove_exceptionals - pagevec exceptionals pruning
954 * @pvec: The pagevec to prune
955 *
956 * pagevec_lookup_entries() fills both pages and exceptional radix
957 * tree entries into the pagevec. This function prunes all
958 * exceptionals from @pvec without leaving holes, so that it can be
959 * passed on to page-only pagevec operations.
960 */
961void pagevec_remove_exceptionals(struct pagevec *pvec)
962{
963 int i, j;
964
965 for (i = 0, j = 0; i < pagevec_count(pvec); i++) {
966 struct page *page = pvec->pages[i];
967 if (!radix_tree_exceptional_entry(page))
968 pvec->pages[j++] = page;
969 }
970 pvec->nr = j;
971}
972
973/**
974 * pagevec_lookup_range - gang pagecache lookup
975 * @pvec: Where the resulting pages are placed
976 * @mapping: The address_space to search
977 * @start: The starting page index
978 * @end: The final page index
979 *
980 * pagevec_lookup_range() will search for & return a group of up to PAGEVEC_SIZE
981 * pages in the mapping starting from index @start and upto index @end
982 * (inclusive). The pages are placed in @pvec. pagevec_lookup() takes a
983 * reference against the pages in @pvec.
984 *
985 * The search returns a group of mapping-contiguous pages with ascending
986 * indexes. There may be holes in the indices due to not-present pages. We
987 * also update @start to index the next page for the traversal.
988 *
989 * pagevec_lookup_range() returns the number of pages which were found. If this
990 * number is smaller than PAGEVEC_SIZE, the end of specified range has been
991 * reached.
992 */
993unsigned pagevec_lookup_range(struct pagevec *pvec,
994 struct address_space *mapping, pgoff_t *start, pgoff_t end)
995{
996 pvec->nr = find_get_pages_range(mapping, start, end, PAGEVEC_SIZE,
997 pvec->pages);
998 return pagevec_count(pvec);
999}
1000EXPORT_SYMBOL(pagevec_lookup_range);
1001
1002unsigned pagevec_lookup_range_tag(struct pagevec *pvec,
1003 struct address_space *mapping, pgoff_t *index, pgoff_t end,
1004 int tag)
1005{
1006 pvec->nr = find_get_pages_range_tag(mapping, index, end, tag,
1007 PAGEVEC_SIZE, pvec->pages);
1008 return pagevec_count(pvec);
1009}
1010EXPORT_SYMBOL(pagevec_lookup_range_tag);
1011
1012unsigned pagevec_lookup_range_nr_tag(struct pagevec *pvec,
1013 struct address_space *mapping, pgoff_t *index, pgoff_t end,
1014 int tag, unsigned max_pages)
1015{
1016 pvec->nr = find_get_pages_range_tag(mapping, index, end, tag,
1017 min_t(unsigned int, max_pages, PAGEVEC_SIZE), pvec->pages);
1018 return pagevec_count(pvec);
1019}
1020EXPORT_SYMBOL(pagevec_lookup_range_nr_tag);
1021/*
1022 * Perform any setup for the swap system
1023 */
1024void __init swap_setup(void)
1025{
1026 unsigned long megs = totalram_pages >> (20 - PAGE_SHIFT);
1027
1028 /* Use a smaller cluster for small-memory machines */
1029 if (megs < 16)
1030 page_cluster = 2;
1031 else
1032 page_cluster = 3;
1033 /*
1034 * Right now other parts of the system means that we
1035 * _really_ don't want to cluster much more
1036 */
1037}
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/mm/swap.c
4 *
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 */
7
8/*
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.
12 * Started 18.12.91
13 * Swap aging added 23.2.95, Stephen Tweedie.
14 * Buffermem limits added 12.3.98, Rik van Riel.
15 */
16
17#include <linux/mm.h>
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>
40
41#include "internal.h"
42
43#define CREATE_TRACE_POINTS
44#include <trace/events/pagemap.h>
45
46/* How many pages do we try to swap or page in/out together? As a power of 2 */
47int page_cluster;
48const int page_cluster_max = 31;
49
50struct cpu_fbatches {
51 /*
52 * The following folio batches are grouped together because they are protected
53 * by disabling preemption (and interrupts remain enabled).
54 */
55 local_lock_t lock;
56 struct folio_batch lru_add;
57 struct folio_batch lru_deactivate_file;
58 struct folio_batch lru_deactivate;
59 struct folio_batch lru_lazyfree;
60#ifdef CONFIG_SMP
61 struct folio_batch lru_activate;
62#endif
63 /* Protecting the following batches which require disabling interrupts */
64 local_lock_t lock_irq;
65 struct folio_batch lru_move_tail;
66};
67
68static DEFINE_PER_CPU(struct cpu_fbatches, cpu_fbatches) = {
69 .lock = INIT_LOCAL_LOCK(lock),
70 .lock_irq = INIT_LOCAL_LOCK(lock_irq),
71};
72
73static void __page_cache_release(struct folio *folio, struct lruvec **lruvecp,
74 unsigned long *flagsp)
75{
76 if (folio_test_lru(folio)) {
77 folio_lruvec_relock_irqsave(folio, lruvecp, flagsp);
78 lruvec_del_folio(*lruvecp, folio);
79 __folio_clear_lru_flags(folio);
80 }
81}
82
83/*
84 * This path almost never happens for VM activity - pages are normally freed
85 * in batches. But it gets used by networking - and for compound pages.
86 */
87static void page_cache_release(struct folio *folio)
88{
89 struct lruvec *lruvec = NULL;
90 unsigned long flags;
91
92 __page_cache_release(folio, &lruvec, &flags);
93 if (lruvec)
94 unlock_page_lruvec_irqrestore(lruvec, flags);
95}
96
97void __folio_put(struct folio *folio)
98{
99 if (unlikely(folio_is_zone_device(folio))) {
100 free_zone_device_folio(folio);
101 return;
102 }
103
104 if (folio_test_hugetlb(folio)) {
105 free_huge_folio(folio);
106 return;
107 }
108
109 page_cache_release(folio);
110 folio_unqueue_deferred_split(folio);
111 mem_cgroup_uncharge(folio);
112 free_unref_page(&folio->page, folio_order(folio));
113}
114EXPORT_SYMBOL(__folio_put);
115
116typedef void (*move_fn_t)(struct lruvec *lruvec, struct folio *folio);
117
118static void lru_add(struct lruvec *lruvec, struct folio *folio)
119{
120 int was_unevictable = folio_test_clear_unevictable(folio);
121 long nr_pages = folio_nr_pages(folio);
122
123 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
124
125 /*
126 * Is an smp_mb__after_atomic() still required here, before
127 * folio_evictable() tests the mlocked flag, to rule out the possibility
128 * of stranding an evictable folio on an unevictable LRU? I think
129 * not, because __munlock_folio() only clears the mlocked flag
130 * while the LRU lock is held.
131 *
132 * (That is not true of __page_cache_release(), and not necessarily
133 * true of folios_put(): but those only clear the mlocked flag after
134 * folio_put_testzero() has excluded any other users of the folio.)
135 */
136 if (folio_evictable(folio)) {
137 if (was_unevictable)
138 __count_vm_events(UNEVICTABLE_PGRESCUED, nr_pages);
139 } else {
140 folio_clear_active(folio);
141 folio_set_unevictable(folio);
142 /*
143 * folio->mlock_count = !!folio_test_mlocked(folio)?
144 * But that leaves __mlock_folio() in doubt whether another
145 * actor has already counted the mlock or not. Err on the
146 * safe side, underestimate, let page reclaim fix it, rather
147 * than leaving a page on the unevictable LRU indefinitely.
148 */
149 folio->mlock_count = 0;
150 if (!was_unevictable)
151 __count_vm_events(UNEVICTABLE_PGCULLED, nr_pages);
152 }
153
154 lruvec_add_folio(lruvec, folio);
155 trace_mm_lru_insertion(folio);
156}
157
158static void folio_batch_move_lru(struct folio_batch *fbatch, move_fn_t move_fn)
159{
160 int i;
161 struct lruvec *lruvec = NULL;
162 unsigned long flags = 0;
163
164 for (i = 0; i < folio_batch_count(fbatch); i++) {
165 struct folio *folio = fbatch->folios[i];
166
167 folio_lruvec_relock_irqsave(folio, &lruvec, &flags);
168 move_fn(lruvec, folio);
169
170 folio_set_lru(folio);
171 }
172
173 if (lruvec)
174 unlock_page_lruvec_irqrestore(lruvec, flags);
175 folios_put(fbatch);
176}
177
178static void __folio_batch_add_and_move(struct folio_batch __percpu *fbatch,
179 struct folio *folio, move_fn_t move_fn,
180 bool on_lru, bool disable_irq)
181{
182 unsigned long flags;
183
184 if (on_lru && !folio_test_clear_lru(folio))
185 return;
186
187 folio_get(folio);
188
189 if (disable_irq)
190 local_lock_irqsave(&cpu_fbatches.lock_irq, flags);
191 else
192 local_lock(&cpu_fbatches.lock);
193
194 if (!folio_batch_add(this_cpu_ptr(fbatch), folio) || folio_test_large(folio) ||
195 lru_cache_disabled())
196 folio_batch_move_lru(this_cpu_ptr(fbatch), move_fn);
197
198 if (disable_irq)
199 local_unlock_irqrestore(&cpu_fbatches.lock_irq, flags);
200 else
201 local_unlock(&cpu_fbatches.lock);
202}
203
204#define folio_batch_add_and_move(folio, op, on_lru) \
205 __folio_batch_add_and_move( \
206 &cpu_fbatches.op, \
207 folio, \
208 op, \
209 on_lru, \
210 offsetof(struct cpu_fbatches, op) >= offsetof(struct cpu_fbatches, lock_irq) \
211 )
212
213static void lru_move_tail(struct lruvec *lruvec, struct folio *folio)
214{
215 if (folio_test_unevictable(folio))
216 return;
217
218 lruvec_del_folio(lruvec, folio);
219 folio_clear_active(folio);
220 lruvec_add_folio_tail(lruvec, folio);
221 __count_vm_events(PGROTATED, folio_nr_pages(folio));
222}
223
224/*
225 * Writeback is about to end against a folio which has been marked for
226 * immediate reclaim. If it still appears to be reclaimable, move it
227 * to the tail of the inactive list.
228 *
229 * folio_rotate_reclaimable() must disable IRQs, to prevent nasty races.
230 */
231void folio_rotate_reclaimable(struct folio *folio)
232{
233 if (folio_test_locked(folio) || folio_test_dirty(folio) ||
234 folio_test_unevictable(folio))
235 return;
236
237 folio_batch_add_and_move(folio, lru_move_tail, true);
238}
239
240void lru_note_cost(struct lruvec *lruvec, bool file,
241 unsigned int nr_io, unsigned int nr_rotated)
242{
243 unsigned long cost;
244
245 /*
246 * Reflect the relative cost of incurring IO and spending CPU
247 * time on rotations. This doesn't attempt to make a precise
248 * comparison, it just says: if reloads are about comparable
249 * between the LRU lists, or rotations are overwhelmingly
250 * different between them, adjust scan balance for CPU work.
251 */
252 cost = nr_io * SWAP_CLUSTER_MAX + nr_rotated;
253
254 do {
255 unsigned long lrusize;
256
257 /*
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).
263 */
264 spin_lock_irq(&lruvec->lru_lock);
265 /* Record cost event */
266 if (file)
267 lruvec->file_cost += cost;
268 else
269 lruvec->anon_cost += cost;
270
271 /*
272 * Decay previous events
273 *
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.
278 */
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);
283
284 if (lruvec->file_cost + lruvec->anon_cost > lrusize / 4) {
285 lruvec->file_cost /= 2;
286 lruvec->anon_cost /= 2;
287 }
288 spin_unlock_irq(&lruvec->lru_lock);
289 } while ((lruvec = parent_lruvec(lruvec)));
290}
291
292void lru_note_cost_refault(struct folio *folio)
293{
294 lru_note_cost(folio_lruvec(folio), folio_is_file_lru(folio),
295 folio_nr_pages(folio), 0);
296}
297
298static void lru_activate(struct lruvec *lruvec, struct folio *folio)
299{
300 long nr_pages = folio_nr_pages(folio);
301
302 if (folio_test_active(folio) || folio_test_unevictable(folio))
303 return;
304
305
306 lruvec_del_folio(lruvec, folio);
307 folio_set_active(folio);
308 lruvec_add_folio(lruvec, folio);
309 trace_mm_lru_activate(folio);
310
311 __count_vm_events(PGACTIVATE, nr_pages);
312 __count_memcg_events(lruvec_memcg(lruvec), PGACTIVATE, nr_pages);
313}
314
315#ifdef CONFIG_SMP
316static void folio_activate_drain(int cpu)
317{
318 struct folio_batch *fbatch = &per_cpu(cpu_fbatches.lru_activate, cpu);
319
320 if (folio_batch_count(fbatch))
321 folio_batch_move_lru(fbatch, lru_activate);
322}
323
324void folio_activate(struct folio *folio)
325{
326 if (folio_test_active(folio) || folio_test_unevictable(folio))
327 return;
328
329 folio_batch_add_and_move(folio, lru_activate, true);
330}
331
332#else
333static inline void folio_activate_drain(int cpu)
334{
335}
336
337void folio_activate(struct folio *folio)
338{
339 struct lruvec *lruvec;
340
341 if (!folio_test_clear_lru(folio))
342 return;
343
344 lruvec = folio_lruvec_lock_irq(folio);
345 lru_activate(lruvec, folio);
346 unlock_page_lruvec_irq(lruvec);
347 folio_set_lru(folio);
348}
349#endif
350
351static void __lru_cache_activate_folio(struct folio *folio)
352{
353 struct folio_batch *fbatch;
354 int i;
355
356 local_lock(&cpu_fbatches.lock);
357 fbatch = this_cpu_ptr(&cpu_fbatches.lru_add);
358
359 /*
360 * Search backwards on the optimistic assumption that the folio being
361 * activated has just been added to this batch. Note that only
362 * the local batch is examined as a !LRU folio could be in the
363 * process of being released, reclaimed, migrated or on a remote
364 * batch that is currently being drained. Furthermore, marking
365 * a remote batch's folio active potentially hits a race where
366 * a folio is marked active just after it is added to the inactive
367 * list causing accounting errors and BUG_ON checks to trigger.
368 */
369 for (i = folio_batch_count(fbatch) - 1; i >= 0; i--) {
370 struct folio *batch_folio = fbatch->folios[i];
371
372 if (batch_folio == folio) {
373 folio_set_active(folio);
374 break;
375 }
376 }
377
378 local_unlock(&cpu_fbatches.lock);
379}
380
381#ifdef CONFIG_LRU_GEN
382static void folio_inc_refs(struct folio *folio)
383{
384 unsigned long new_flags, old_flags = READ_ONCE(folio->flags);
385
386 if (folio_test_unevictable(folio))
387 return;
388
389 if (!folio_test_referenced(folio)) {
390 folio_set_referenced(folio);
391 return;
392 }
393
394 if (!folio_test_workingset(folio)) {
395 folio_set_workingset(folio);
396 return;
397 }
398
399 /* see the comment on MAX_NR_TIERS */
400 do {
401 new_flags = old_flags & LRU_REFS_MASK;
402 if (new_flags == LRU_REFS_MASK)
403 break;
404
405 new_flags += BIT(LRU_REFS_PGOFF);
406 new_flags |= old_flags & ~LRU_REFS_MASK;
407 } while (!try_cmpxchg(&folio->flags, &old_flags, new_flags));
408}
409#else
410static void folio_inc_refs(struct folio *folio)
411{
412}
413#endif /* CONFIG_LRU_GEN */
414
415/**
416 * folio_mark_accessed - Mark a folio as having seen activity.
417 * @folio: The folio to mark.
418 *
419 * This function will perform one of the following transitions:
420 *
421 * * inactive,unreferenced -> inactive,referenced
422 * * inactive,referenced -> active,unreferenced
423 * * active,unreferenced -> active,referenced
424 *
425 * When a newly allocated folio is not yet visible, so safe for non-atomic ops,
426 * __folio_set_referenced() may be substituted for folio_mark_accessed().
427 */
428void folio_mark_accessed(struct folio *folio)
429{
430 if (lru_gen_enabled()) {
431 folio_inc_refs(folio);
432 return;
433 }
434
435 if (!folio_test_referenced(folio)) {
436 folio_set_referenced(folio);
437 } else if (folio_test_unevictable(folio)) {
438 /*
439 * Unevictable pages are on the "LRU_UNEVICTABLE" list. But,
440 * this list is never rotated or maintained, so marking an
441 * unevictable page accessed has no effect.
442 */
443 } else if (!folio_test_active(folio)) {
444 /*
445 * If the folio is on the LRU, queue it for activation via
446 * cpu_fbatches.lru_activate. Otherwise, assume the folio is in a
447 * folio_batch, mark it active and it'll be moved to the active
448 * LRU on the next drain.
449 */
450 if (folio_test_lru(folio))
451 folio_activate(folio);
452 else
453 __lru_cache_activate_folio(folio);
454 folio_clear_referenced(folio);
455 workingset_activation(folio);
456 }
457 if (folio_test_idle(folio))
458 folio_clear_idle(folio);
459}
460EXPORT_SYMBOL(folio_mark_accessed);
461
462/**
463 * folio_add_lru - Add a folio to an LRU list.
464 * @folio: The folio to be added to the LRU.
465 *
466 * Queue the folio for addition to the LRU. The decision on whether
467 * to add the page to the [in]active [file|anon] list is deferred until the
468 * folio_batch is drained. This gives a chance for the caller of folio_add_lru()
469 * have the folio added to the active list using folio_mark_accessed().
470 */
471void folio_add_lru(struct folio *folio)
472{
473 VM_BUG_ON_FOLIO(folio_test_active(folio) &&
474 folio_test_unevictable(folio), folio);
475 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
476
477 /* see the comment in lru_gen_add_folio() */
478 if (lru_gen_enabled() && !folio_test_unevictable(folio) &&
479 lru_gen_in_fault() && !(current->flags & PF_MEMALLOC))
480 folio_set_active(folio);
481
482 folio_batch_add_and_move(folio, lru_add, false);
483}
484EXPORT_SYMBOL(folio_add_lru);
485
486/**
487 * folio_add_lru_vma() - Add a folio to the appropate LRU list for this VMA.
488 * @folio: The folio to be added to the LRU.
489 * @vma: VMA in which the folio is mapped.
490 *
491 * If the VMA is mlocked, @folio is added to the unevictable list.
492 * Otherwise, it is treated the same way as folio_add_lru().
493 */
494void folio_add_lru_vma(struct folio *folio, struct vm_area_struct *vma)
495{
496 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
497
498 if (unlikely((vma->vm_flags & (VM_LOCKED | VM_SPECIAL)) == VM_LOCKED))
499 mlock_new_folio(folio);
500 else
501 folio_add_lru(folio);
502}
503
504/*
505 * If the folio cannot be invalidated, it is moved to the
506 * inactive list to speed up its reclaim. It is moved to the
507 * head of the list, rather than the tail, to give the flusher
508 * threads some time to write it out, as this is much more
509 * effective than the single-page writeout from reclaim.
510 *
511 * If the folio isn't mapped and dirty/writeback, the folio
512 * could be reclaimed asap using the reclaim flag.
513 *
514 * 1. active, mapped folio -> none
515 * 2. active, dirty/writeback folio -> inactive, head, reclaim
516 * 3. inactive, mapped folio -> none
517 * 4. inactive, dirty/writeback folio -> inactive, head, reclaim
518 * 5. inactive, clean -> inactive, tail
519 * 6. Others -> none
520 *
521 * In 4, it moves to the head of the inactive list so the folio is
522 * written out by flusher threads as this is much more efficient
523 * than the single-page writeout from reclaim.
524 */
525static void lru_deactivate_file(struct lruvec *lruvec, struct folio *folio)
526{
527 bool active = folio_test_active(folio);
528 long nr_pages = folio_nr_pages(folio);
529
530 if (folio_test_unevictable(folio))
531 return;
532
533 /* Some processes are using the folio */
534 if (folio_mapped(folio))
535 return;
536
537 lruvec_del_folio(lruvec, folio);
538 folio_clear_active(folio);
539 folio_clear_referenced(folio);
540
541 if (folio_test_writeback(folio) || folio_test_dirty(folio)) {
542 /*
543 * Setting the reclaim flag could race with
544 * folio_end_writeback() and confuse readahead. But the
545 * race window is _really_ small and it's not a critical
546 * problem.
547 */
548 lruvec_add_folio(lruvec, folio);
549 folio_set_reclaim(folio);
550 } else {
551 /*
552 * The folio's writeback ended while it was in the batch.
553 * We move that folio to the tail of the inactive list.
554 */
555 lruvec_add_folio_tail(lruvec, folio);
556 __count_vm_events(PGROTATED, nr_pages);
557 }
558
559 if (active) {
560 __count_vm_events(PGDEACTIVATE, nr_pages);
561 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
562 nr_pages);
563 }
564}
565
566static void lru_deactivate(struct lruvec *lruvec, struct folio *folio)
567{
568 long nr_pages = folio_nr_pages(folio);
569
570 if (folio_test_unevictable(folio) || !(folio_test_active(folio) || lru_gen_enabled()))
571 return;
572
573 lruvec_del_folio(lruvec, folio);
574 folio_clear_active(folio);
575 folio_clear_referenced(folio);
576 lruvec_add_folio(lruvec, folio);
577
578 __count_vm_events(PGDEACTIVATE, nr_pages);
579 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_pages);
580}
581
582static void lru_lazyfree(struct lruvec *lruvec, struct folio *folio)
583{
584 long nr_pages = folio_nr_pages(folio);
585
586 if (!folio_test_anon(folio) || !folio_test_swapbacked(folio) ||
587 folio_test_swapcache(folio) || folio_test_unevictable(folio))
588 return;
589
590 lruvec_del_folio(lruvec, folio);
591 folio_clear_active(folio);
592 folio_clear_referenced(folio);
593 /*
594 * Lazyfree folios are clean anonymous folios. They have
595 * the swapbacked flag cleared, to distinguish them from normal
596 * anonymous folios
597 */
598 folio_clear_swapbacked(folio);
599 lruvec_add_folio(lruvec, folio);
600
601 __count_vm_events(PGLAZYFREE, nr_pages);
602 __count_memcg_events(lruvec_memcg(lruvec), PGLAZYFREE, nr_pages);
603}
604
605/*
606 * Drain pages out of the cpu's folio_batch.
607 * Either "cpu" is the current CPU, and preemption has already been
608 * disabled; or "cpu" is being hot-unplugged, and is already dead.
609 */
610void lru_add_drain_cpu(int cpu)
611{
612 struct cpu_fbatches *fbatches = &per_cpu(cpu_fbatches, cpu);
613 struct folio_batch *fbatch = &fbatches->lru_add;
614
615 if (folio_batch_count(fbatch))
616 folio_batch_move_lru(fbatch, lru_add);
617
618 fbatch = &fbatches->lru_move_tail;
619 /* Disabling interrupts below acts as a compiler barrier. */
620 if (data_race(folio_batch_count(fbatch))) {
621 unsigned long flags;
622
623 /* No harm done if a racing interrupt already did this */
624 local_lock_irqsave(&cpu_fbatches.lock_irq, flags);
625 folio_batch_move_lru(fbatch, lru_move_tail);
626 local_unlock_irqrestore(&cpu_fbatches.lock_irq, flags);
627 }
628
629 fbatch = &fbatches->lru_deactivate_file;
630 if (folio_batch_count(fbatch))
631 folio_batch_move_lru(fbatch, lru_deactivate_file);
632
633 fbatch = &fbatches->lru_deactivate;
634 if (folio_batch_count(fbatch))
635 folio_batch_move_lru(fbatch, lru_deactivate);
636
637 fbatch = &fbatches->lru_lazyfree;
638 if (folio_batch_count(fbatch))
639 folio_batch_move_lru(fbatch, lru_lazyfree);
640
641 folio_activate_drain(cpu);
642}
643
644/**
645 * deactivate_file_folio() - Deactivate a file folio.
646 * @folio: Folio to deactivate.
647 *
648 * This function hints to the VM that @folio is a good reclaim candidate,
649 * for example if its invalidation fails due to the folio being dirty
650 * or under writeback.
651 *
652 * Context: Caller holds a reference on the folio.
653 */
654void deactivate_file_folio(struct folio *folio)
655{
656 /* Deactivating an unevictable folio will not accelerate reclaim */
657 if (folio_test_unevictable(folio))
658 return;
659
660 folio_batch_add_and_move(folio, lru_deactivate_file, true);
661}
662
663/*
664 * folio_deactivate - deactivate a folio
665 * @folio: folio to deactivate
666 *
667 * folio_deactivate() moves @folio to the inactive list if @folio was on the
668 * active list and was not unevictable. This is done to accelerate the
669 * reclaim of @folio.
670 */
671void folio_deactivate(struct folio *folio)
672{
673 if (folio_test_unevictable(folio) || !(folio_test_active(folio) || lru_gen_enabled()))
674 return;
675
676 folio_batch_add_and_move(folio, lru_deactivate, true);
677}
678
679/**
680 * folio_mark_lazyfree - make an anon folio lazyfree
681 * @folio: folio to deactivate
682 *
683 * folio_mark_lazyfree() moves @folio to the inactive file list.
684 * This is done to accelerate the reclaim of @folio.
685 */
686void folio_mark_lazyfree(struct folio *folio)
687{
688 if (!folio_test_anon(folio) || !folio_test_swapbacked(folio) ||
689 folio_test_swapcache(folio) || folio_test_unevictable(folio))
690 return;
691
692 folio_batch_add_and_move(folio, lru_lazyfree, true);
693}
694
695void lru_add_drain(void)
696{
697 local_lock(&cpu_fbatches.lock);
698 lru_add_drain_cpu(smp_processor_id());
699 local_unlock(&cpu_fbatches.lock);
700 mlock_drain_local();
701}
702
703/*
704 * It's called from per-cpu workqueue context in SMP case so
705 * lru_add_drain_cpu and invalidate_bh_lrus_cpu should run on
706 * the same cpu. It shouldn't be a problem in !SMP case since
707 * the core is only one and the locks will disable preemption.
708 */
709static void lru_add_and_bh_lrus_drain(void)
710{
711 local_lock(&cpu_fbatches.lock);
712 lru_add_drain_cpu(smp_processor_id());
713 local_unlock(&cpu_fbatches.lock);
714 invalidate_bh_lrus_cpu();
715 mlock_drain_local();
716}
717
718void lru_add_drain_cpu_zone(struct zone *zone)
719{
720 local_lock(&cpu_fbatches.lock);
721 lru_add_drain_cpu(smp_processor_id());
722 drain_local_pages(zone);
723 local_unlock(&cpu_fbatches.lock);
724 mlock_drain_local();
725}
726
727#ifdef CONFIG_SMP
728
729static DEFINE_PER_CPU(struct work_struct, lru_add_drain_work);
730
731static void lru_add_drain_per_cpu(struct work_struct *dummy)
732{
733 lru_add_and_bh_lrus_drain();
734}
735
736static bool cpu_needs_drain(unsigned int cpu)
737{
738 struct cpu_fbatches *fbatches = &per_cpu(cpu_fbatches, cpu);
739
740 /* Check these in order of likelihood that they're not zero */
741 return folio_batch_count(&fbatches->lru_add) ||
742 folio_batch_count(&fbatches->lru_move_tail) ||
743 folio_batch_count(&fbatches->lru_deactivate_file) ||
744 folio_batch_count(&fbatches->lru_deactivate) ||
745 folio_batch_count(&fbatches->lru_lazyfree) ||
746 folio_batch_count(&fbatches->lru_activate) ||
747 need_mlock_drain(cpu) ||
748 has_bh_in_lru(cpu, NULL);
749}
750
751/*
752 * Doesn't need any cpu hotplug locking because we do rely on per-cpu
753 * kworkers being shut down before our page_alloc_cpu_dead callback is
754 * executed on the offlined cpu.
755 * Calling this function with cpu hotplug locks held can actually lead
756 * to obscure indirect dependencies via WQ context.
757 */
758static inline void __lru_add_drain_all(bool force_all_cpus)
759{
760 /*
761 * lru_drain_gen - Global pages generation number
762 *
763 * (A) Definition: global lru_drain_gen = x implies that all generations
764 * 0 < n <= x are already *scheduled* for draining.
765 *
766 * This is an optimization for the highly-contended use case where a
767 * user space workload keeps constantly generating a flow of pages for
768 * each CPU.
769 */
770 static unsigned int lru_drain_gen;
771 static struct cpumask has_work;
772 static DEFINE_MUTEX(lock);
773 unsigned cpu, this_gen;
774
775 /*
776 * Make sure nobody triggers this path before mm_percpu_wq is fully
777 * initialized.
778 */
779 if (WARN_ON(!mm_percpu_wq))
780 return;
781
782 /*
783 * Guarantee folio_batch counter stores visible by this CPU
784 * are visible to other CPUs before loading the current drain
785 * generation.
786 */
787 smp_mb();
788
789 /*
790 * (B) Locally cache global LRU draining generation number
791 *
792 * The read barrier ensures that the counter is loaded before the mutex
793 * is taken. It pairs with smp_mb() inside the mutex critical section
794 * at (D).
795 */
796 this_gen = smp_load_acquire(&lru_drain_gen);
797
798 mutex_lock(&lock);
799
800 /*
801 * (C) Exit the draining operation if a newer generation, from another
802 * lru_add_drain_all(), was already scheduled for draining. Check (A).
803 */
804 if (unlikely(this_gen != lru_drain_gen && !force_all_cpus))
805 goto done;
806
807 /*
808 * (D) Increment global generation number
809 *
810 * Pairs with smp_load_acquire() at (B), outside of the critical
811 * section. Use a full memory barrier to guarantee that the
812 * new global drain generation number is stored before loading
813 * folio_batch counters.
814 *
815 * This pairing must be done here, before the for_each_online_cpu loop
816 * below which drains the page vectors.
817 *
818 * Let x, y, and z represent some system CPU numbers, where x < y < z.
819 * Assume CPU #z is in the middle of the for_each_online_cpu loop
820 * below and has already reached CPU #y's per-cpu data. CPU #x comes
821 * along, adds some pages to its per-cpu vectors, then calls
822 * lru_add_drain_all().
823 *
824 * If the paired barrier is done at any later step, e.g. after the
825 * loop, CPU #x will just exit at (C) and miss flushing out all of its
826 * added pages.
827 */
828 WRITE_ONCE(lru_drain_gen, lru_drain_gen + 1);
829 smp_mb();
830
831 cpumask_clear(&has_work);
832 for_each_online_cpu(cpu) {
833 struct work_struct *work = &per_cpu(lru_add_drain_work, cpu);
834
835 if (cpu_needs_drain(cpu)) {
836 INIT_WORK(work, lru_add_drain_per_cpu);
837 queue_work_on(cpu, mm_percpu_wq, work);
838 __cpumask_set_cpu(cpu, &has_work);
839 }
840 }
841
842 for_each_cpu(cpu, &has_work)
843 flush_work(&per_cpu(lru_add_drain_work, cpu));
844
845done:
846 mutex_unlock(&lock);
847}
848
849void lru_add_drain_all(void)
850{
851 __lru_add_drain_all(false);
852}
853#else
854void lru_add_drain_all(void)
855{
856 lru_add_drain();
857}
858#endif /* CONFIG_SMP */
859
860atomic_t lru_disable_count = ATOMIC_INIT(0);
861
862/*
863 * lru_cache_disable() needs to be called before we start compiling
864 * a list of folios to be migrated using folio_isolate_lru().
865 * It drains folios on LRU cache and then disable on all cpus until
866 * lru_cache_enable is called.
867 *
868 * Must be paired with a call to lru_cache_enable().
869 */
870void lru_cache_disable(void)
871{
872 atomic_inc(&lru_disable_count);
873 /*
874 * Readers of lru_disable_count are protected by either disabling
875 * preemption or rcu_read_lock:
876 *
877 * preempt_disable, local_irq_disable [bh_lru_lock()]
878 * rcu_read_lock [rt_spin_lock CONFIG_PREEMPT_RT]
879 * preempt_disable [local_lock !CONFIG_PREEMPT_RT]
880 *
881 * Since v5.1 kernel, synchronize_rcu() is guaranteed to wait on
882 * preempt_disable() regions of code. So any CPU which sees
883 * lru_disable_count = 0 will have exited the critical
884 * section when synchronize_rcu() returns.
885 */
886 synchronize_rcu_expedited();
887#ifdef CONFIG_SMP
888 __lru_add_drain_all(true);
889#else
890 lru_add_and_bh_lrus_drain();
891#endif
892}
893
894/**
895 * folios_put_refs - Reduce the reference count on a batch of folios.
896 * @folios: The folios.
897 * @refs: The number of refs to subtract from each folio.
898 *
899 * Like folio_put(), but for a batch of folios. This is more efficient
900 * than writing the loop yourself as it will optimise the locks which need
901 * to be taken if the folios are freed. The folios batch is returned
902 * empty and ready to be reused for another batch; there is no need
903 * to reinitialise it. If @refs is NULL, we subtract one from each
904 * folio refcount.
905 *
906 * Context: May be called in process or interrupt context, but not in NMI
907 * context. May be called while holding a spinlock.
908 */
909void folios_put_refs(struct folio_batch *folios, unsigned int *refs)
910{
911 int i, j;
912 struct lruvec *lruvec = NULL;
913 unsigned long flags = 0;
914
915 for (i = 0, j = 0; i < folios->nr; i++) {
916 struct folio *folio = folios->folios[i];
917 unsigned int nr_refs = refs ? refs[i] : 1;
918
919 if (is_huge_zero_folio(folio))
920 continue;
921
922 if (folio_is_zone_device(folio)) {
923 if (lruvec) {
924 unlock_page_lruvec_irqrestore(lruvec, flags);
925 lruvec = NULL;
926 }
927 if (put_devmap_managed_folio_refs(folio, nr_refs))
928 continue;
929 if (folio_ref_sub_and_test(folio, nr_refs))
930 free_zone_device_folio(folio);
931 continue;
932 }
933
934 if (!folio_ref_sub_and_test(folio, nr_refs))
935 continue;
936
937 /* hugetlb has its own memcg */
938 if (folio_test_hugetlb(folio)) {
939 if (lruvec) {
940 unlock_page_lruvec_irqrestore(lruvec, flags);
941 lruvec = NULL;
942 }
943 free_huge_folio(folio);
944 continue;
945 }
946 folio_unqueue_deferred_split(folio);
947 __page_cache_release(folio, &lruvec, &flags);
948
949 if (j != i)
950 folios->folios[j] = folio;
951 j++;
952 }
953 if (lruvec)
954 unlock_page_lruvec_irqrestore(lruvec, flags);
955 if (!j) {
956 folio_batch_reinit(folios);
957 return;
958 }
959
960 folios->nr = j;
961 mem_cgroup_uncharge_folios(folios);
962 free_unref_folios(folios);
963}
964EXPORT_SYMBOL(folios_put_refs);
965
966/**
967 * release_pages - batched put_page()
968 * @arg: array of pages to release
969 * @nr: number of pages
970 *
971 * Decrement the reference count on all the pages in @arg. If it
972 * fell to zero, remove the page from the LRU and free it.
973 *
974 * Note that the argument can be an array of pages, encoded pages,
975 * or folio pointers. We ignore any encoded bits, and turn any of
976 * them into just a folio that gets free'd.
977 */
978void release_pages(release_pages_arg arg, int nr)
979{
980 struct folio_batch fbatch;
981 int refs[PAGEVEC_SIZE];
982 struct encoded_page **encoded = arg.encoded_pages;
983 int i;
984
985 folio_batch_init(&fbatch);
986 for (i = 0; i < nr; i++) {
987 /* Turn any of the argument types into a folio */
988 struct folio *folio = page_folio(encoded_page_ptr(encoded[i]));
989
990 /* Is our next entry actually "nr_pages" -> "nr_refs" ? */
991 refs[fbatch.nr] = 1;
992 if (unlikely(encoded_page_flags(encoded[i]) &
993 ENCODED_PAGE_BIT_NR_PAGES_NEXT))
994 refs[fbatch.nr] = encoded_nr_pages(encoded[++i]);
995
996 if (folio_batch_add(&fbatch, folio) > 0)
997 continue;
998 folios_put_refs(&fbatch, refs);
999 }
1000
1001 if (fbatch.nr)
1002 folios_put_refs(&fbatch, refs);
1003}
1004EXPORT_SYMBOL(release_pages);
1005
1006/*
1007 * The folios which we're about to release may be in the deferred lru-addition
1008 * queues. That would prevent them from really being freed right now. That's
1009 * OK from a correctness point of view but is inefficient - those folios may be
1010 * cache-warm and we want to give them back to the page allocator ASAP.
1011 *
1012 * So __folio_batch_release() will drain those queues here.
1013 * folio_batch_move_lru() calls folios_put() directly to avoid
1014 * mutual recursion.
1015 */
1016void __folio_batch_release(struct folio_batch *fbatch)
1017{
1018 if (!fbatch->percpu_pvec_drained) {
1019 lru_add_drain();
1020 fbatch->percpu_pvec_drained = true;
1021 }
1022 folios_put(fbatch);
1023}
1024EXPORT_SYMBOL(__folio_batch_release);
1025
1026/**
1027 * folio_batch_remove_exceptionals() - Prune non-folios from a batch.
1028 * @fbatch: The batch to prune
1029 *
1030 * find_get_entries() fills a batch with both folios and shadow/swap/DAX
1031 * entries. This function prunes all the non-folio entries from @fbatch
1032 * without leaving holes, so that it can be passed on to folio-only batch
1033 * operations.
1034 */
1035void folio_batch_remove_exceptionals(struct folio_batch *fbatch)
1036{
1037 unsigned int i, j;
1038
1039 for (i = 0, j = 0; i < folio_batch_count(fbatch); i++) {
1040 struct folio *folio = fbatch->folios[i];
1041 if (!xa_is_value(folio))
1042 fbatch->folios[j++] = folio;
1043 }
1044 fbatch->nr = j;
1045}
1046
1047/*
1048 * Perform any setup for the swap system
1049 */
1050void __init swap_setup(void)
1051{
1052 unsigned long megs = totalram_pages() >> (20 - PAGE_SHIFT);
1053
1054 /* Use a smaller cluster for small-memory machines */
1055 if (megs < 16)
1056 page_cluster = 2;
1057 else
1058 page_cluster = 3;
1059 /*
1060 * Right now other parts of the system means that we
1061 * _really_ don't want to cluster much more
1062 */
1063}