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_deactivate_pvecs);
50
51/*
52 * This path almost never happens for VM activity - pages are normally
53 * freed via pagevecs. But it gets used by networking.
54 */
55static void __page_cache_release(struct page *page)
56{
57 if (PageLRU(page)) {
58 struct zone *zone = page_zone(page);
59 struct lruvec *lruvec;
60 unsigned long flags;
61
62 spin_lock_irqsave(&zone->lru_lock, flags);
63 lruvec = mem_cgroup_page_lruvec(page, zone);
64 VM_BUG_ON_PAGE(!PageLRU(page), page);
65 __ClearPageLRU(page);
66 del_page_from_lru_list(page, lruvec, page_off_lru(page));
67 spin_unlock_irqrestore(&zone->lru_lock, flags);
68 }
69 mem_cgroup_uncharge(page);
70}
71
72static void __put_single_page(struct page *page)
73{
74 __page_cache_release(page);
75 free_hot_cold_page(page, false);
76}
77
78static void __put_compound_page(struct page *page)
79{
80 compound_page_dtor *dtor;
81
82 /*
83 * __page_cache_release() is supposed to be called for thp, not for
84 * hugetlb. This is because hugetlb page does never have PageLRU set
85 * (it's never listed to any LRU lists) and no memcg routines should
86 * be called for hugetlb (it has a separate hugetlb_cgroup.)
87 */
88 if (!PageHuge(page))
89 __page_cache_release(page);
90 dtor = get_compound_page_dtor(page);
91 (*dtor)(page);
92}
93
94void __put_page(struct page *page)
95{
96 if (unlikely(PageCompound(page)))
97 __put_compound_page(page);
98 else
99 __put_single_page(page);
100}
101EXPORT_SYMBOL(__put_page);
102
103/**
104 * put_pages_list() - release a list of pages
105 * @pages: list of pages threaded on page->lru
106 *
107 * Release a list of pages which are strung together on page.lru. Currently
108 * used by read_cache_pages() and related error recovery code.
109 */
110void put_pages_list(struct list_head *pages)
111{
112 while (!list_empty(pages)) {
113 struct page *victim;
114
115 victim = list_entry(pages->prev, struct page, lru);
116 list_del(&victim->lru);
117 put_page(victim);
118 }
119}
120EXPORT_SYMBOL(put_pages_list);
121
122/*
123 * get_kernel_pages() - pin kernel pages in memory
124 * @kiov: An array of struct kvec structures
125 * @nr_segs: number of segments to pin
126 * @write: pinning for read/write, currently ignored
127 * @pages: array that receives pointers to the pages pinned.
128 * Should be at least nr_segs long.
129 *
130 * Returns number of pages pinned. This may be fewer than the number
131 * requested. If nr_pages is 0 or negative, returns 0. If no pages
132 * were pinned, returns -errno. Each page returned must be released
133 * with a put_page() call when it is finished with.
134 */
135int get_kernel_pages(const struct kvec *kiov, int nr_segs, int write,
136 struct page **pages)
137{
138 int seg;
139
140 for (seg = 0; seg < nr_segs; seg++) {
141 if (WARN_ON(kiov[seg].iov_len != PAGE_SIZE))
142 return seg;
143
144 pages[seg] = kmap_to_page(kiov[seg].iov_base);
145 get_page(pages[seg]);
146 }
147
148 return seg;
149}
150EXPORT_SYMBOL_GPL(get_kernel_pages);
151
152/*
153 * get_kernel_page() - pin a kernel page in memory
154 * @start: starting kernel address
155 * @write: pinning for read/write, currently ignored
156 * @pages: array that receives pointer to the page pinned.
157 * Must be at least nr_segs long.
158 *
159 * Returns 1 if page is pinned. If the page was not pinned, returns
160 * -errno. The page returned must be released with a put_page() call
161 * when it is finished with.
162 */
163int get_kernel_page(unsigned long start, int write, struct page **pages)
164{
165 const struct kvec kiov = {
166 .iov_base = (void *)start,
167 .iov_len = PAGE_SIZE
168 };
169
170 return get_kernel_pages(&kiov, 1, write, pages);
171}
172EXPORT_SYMBOL_GPL(get_kernel_page);
173
174static void pagevec_lru_move_fn(struct pagevec *pvec,
175 void (*move_fn)(struct page *page, struct lruvec *lruvec, void *arg),
176 void *arg)
177{
178 int i;
179 struct zone *zone = NULL;
180 struct lruvec *lruvec;
181 unsigned long flags = 0;
182
183 for (i = 0; i < pagevec_count(pvec); i++) {
184 struct page *page = pvec->pages[i];
185 struct zone *pagezone = page_zone(page);
186
187 if (pagezone != zone) {
188 if (zone)
189 spin_unlock_irqrestore(&zone->lru_lock, flags);
190 zone = pagezone;
191 spin_lock_irqsave(&zone->lru_lock, flags);
192 }
193
194 lruvec = mem_cgroup_page_lruvec(page, zone);
195 (*move_fn)(page, lruvec, arg);
196 }
197 if (zone)
198 spin_unlock_irqrestore(&zone->lru_lock, flags);
199 release_pages(pvec->pages, pvec->nr, pvec->cold);
200 pagevec_reinit(pvec);
201}
202
203static void pagevec_move_tail_fn(struct page *page, struct lruvec *lruvec,
204 void *arg)
205{
206 int *pgmoved = arg;
207
208 if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
209 enum lru_list lru = page_lru_base_type(page);
210 list_move_tail(&page->lru, &lruvec->lists[lru]);
211 (*pgmoved)++;
212 }
213}
214
215/*
216 * pagevec_move_tail() must be called with IRQ disabled.
217 * Otherwise this may cause nasty races.
218 */
219static void pagevec_move_tail(struct pagevec *pvec)
220{
221 int pgmoved = 0;
222
223 pagevec_lru_move_fn(pvec, pagevec_move_tail_fn, &pgmoved);
224 __count_vm_events(PGROTATED, pgmoved);
225}
226
227/*
228 * Writeback is about to end against a page which has been marked for immediate
229 * reclaim. If it still appears to be reclaimable, move it to the tail of the
230 * inactive list.
231 */
232void rotate_reclaimable_page(struct page *page)
233{
234 if (!PageLocked(page) && !PageDirty(page) && !PageActive(page) &&
235 !PageUnevictable(page) && PageLRU(page)) {
236 struct pagevec *pvec;
237 unsigned long flags;
238
239 get_page(page);
240 local_irq_save(flags);
241 pvec = this_cpu_ptr(&lru_rotate_pvecs);
242 if (!pagevec_add(pvec, page))
243 pagevec_move_tail(pvec);
244 local_irq_restore(flags);
245 }
246}
247
248static void update_page_reclaim_stat(struct lruvec *lruvec,
249 int file, int rotated)
250{
251 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
252
253 reclaim_stat->recent_scanned[file]++;
254 if (rotated)
255 reclaim_stat->recent_rotated[file]++;
256}
257
258static void __activate_page(struct page *page, struct lruvec *lruvec,
259 void *arg)
260{
261 if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
262 int file = page_is_file_cache(page);
263 int lru = page_lru_base_type(page);
264
265 del_page_from_lru_list(page, lruvec, lru);
266 SetPageActive(page);
267 lru += LRU_ACTIVE;
268 add_page_to_lru_list(page, lruvec, lru);
269 trace_mm_lru_activate(page);
270
271 __count_vm_event(PGACTIVATE);
272 update_page_reclaim_stat(lruvec, file, 1);
273 }
274}
275
276#ifdef CONFIG_SMP
277static DEFINE_PER_CPU(struct pagevec, activate_page_pvecs);
278
279static void activate_page_drain(int cpu)
280{
281 struct pagevec *pvec = &per_cpu(activate_page_pvecs, cpu);
282
283 if (pagevec_count(pvec))
284 pagevec_lru_move_fn(pvec, __activate_page, NULL);
285}
286
287static bool need_activate_page_drain(int cpu)
288{
289 return pagevec_count(&per_cpu(activate_page_pvecs, cpu)) != 0;
290}
291
292void activate_page(struct page *page)
293{
294 if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
295 struct pagevec *pvec = &get_cpu_var(activate_page_pvecs);
296
297 get_page(page);
298 if (!pagevec_add(pvec, page))
299 pagevec_lru_move_fn(pvec, __activate_page, NULL);
300 put_cpu_var(activate_page_pvecs);
301 }
302}
303
304#else
305static inline void activate_page_drain(int cpu)
306{
307}
308
309static bool need_activate_page_drain(int cpu)
310{
311 return false;
312}
313
314void activate_page(struct page *page)
315{
316 struct zone *zone = page_zone(page);
317
318 spin_lock_irq(&zone->lru_lock);
319 __activate_page(page, mem_cgroup_page_lruvec(page, zone), NULL);
320 spin_unlock_irq(&zone->lru_lock);
321}
322#endif
323
324static void __lru_cache_activate_page(struct page *page)
325{
326 struct pagevec *pvec = &get_cpu_var(lru_add_pvec);
327 int i;
328
329 /*
330 * Search backwards on the optimistic assumption that the page being
331 * activated has just been added to this pagevec. Note that only
332 * the local pagevec is examined as a !PageLRU page could be in the
333 * process of being released, reclaimed, migrated or on a remote
334 * pagevec that is currently being drained. Furthermore, marking
335 * a remote pagevec's page PageActive potentially hits a race where
336 * a page is marked PageActive just after it is added to the inactive
337 * list causing accounting errors and BUG_ON checks to trigger.
338 */
339 for (i = pagevec_count(pvec) - 1; i >= 0; i--) {
340 struct page *pagevec_page = pvec->pages[i];
341
342 if (pagevec_page == page) {
343 SetPageActive(page);
344 break;
345 }
346 }
347
348 put_cpu_var(lru_add_pvec);
349}
350
351/*
352 * Mark a page as having seen activity.
353 *
354 * inactive,unreferenced -> inactive,referenced
355 * inactive,referenced -> active,unreferenced
356 * active,unreferenced -> active,referenced
357 *
358 * When a newly allocated page is not yet visible, so safe for non-atomic ops,
359 * __SetPageReferenced(page) may be substituted for mark_page_accessed(page).
360 */
361void mark_page_accessed(struct page *page)
362{
363 page = compound_head(page);
364 if (!PageActive(page) && !PageUnevictable(page) &&
365 PageReferenced(page)) {
366
367 /*
368 * If the page is on the LRU, queue it for activation via
369 * activate_page_pvecs. Otherwise, assume the page is on a
370 * pagevec, mark it active and it'll be moved to the active
371 * LRU on the next drain.
372 */
373 if (PageLRU(page))
374 activate_page(page);
375 else
376 __lru_cache_activate_page(page);
377 ClearPageReferenced(page);
378 if (page_is_file_cache(page))
379 workingset_activation(page);
380 } else if (!PageReferenced(page)) {
381 SetPageReferenced(page);
382 }
383 if (page_is_idle(page))
384 clear_page_idle(page);
385}
386EXPORT_SYMBOL(mark_page_accessed);
387
388static void __lru_cache_add(struct page *page)
389{
390 struct pagevec *pvec = &get_cpu_var(lru_add_pvec);
391
392 get_page(page);
393 if (!pagevec_space(pvec))
394 __pagevec_lru_add(pvec);
395 pagevec_add(pvec, page);
396 put_cpu_var(lru_add_pvec);
397}
398
399/**
400 * lru_cache_add: add a page to the page lists
401 * @page: the page to add
402 */
403void lru_cache_add_anon(struct page *page)
404{
405 if (PageActive(page))
406 ClearPageActive(page);
407 __lru_cache_add(page);
408}
409
410void lru_cache_add_file(struct page *page)
411{
412 if (PageActive(page))
413 ClearPageActive(page);
414 __lru_cache_add(page);
415}
416EXPORT_SYMBOL(lru_cache_add_file);
417
418/**
419 * lru_cache_add - add a page to a page list
420 * @page: the page to be added to the LRU.
421 *
422 * Queue the page for addition to the LRU via pagevec. The decision on whether
423 * to add the page to the [in]active [file|anon] list is deferred until the
424 * pagevec is drained. This gives a chance for the caller of lru_cache_add()
425 * have the page added to the active list using mark_page_accessed().
426 */
427void lru_cache_add(struct page *page)
428{
429 VM_BUG_ON_PAGE(PageActive(page) && PageUnevictable(page), page);
430 VM_BUG_ON_PAGE(PageLRU(page), page);
431 __lru_cache_add(page);
432}
433
434/**
435 * add_page_to_unevictable_list - add a page to the unevictable list
436 * @page: the page to be added to the unevictable list
437 *
438 * Add page directly to its zone's unevictable list. To avoid races with
439 * tasks that might be making the page evictable, through eg. munlock,
440 * munmap or exit, while it's not on the lru, we want to add the page
441 * while it's locked or otherwise "invisible" to other tasks. This is
442 * difficult to do when using the pagevec cache, so bypass that.
443 */
444void add_page_to_unevictable_list(struct page *page)
445{
446 struct zone *zone = page_zone(page);
447 struct lruvec *lruvec;
448
449 spin_lock_irq(&zone->lru_lock);
450 lruvec = mem_cgroup_page_lruvec(page, zone);
451 ClearPageActive(page);
452 SetPageUnevictable(page);
453 SetPageLRU(page);
454 add_page_to_lru_list(page, lruvec, LRU_UNEVICTABLE);
455 spin_unlock_irq(&zone->lru_lock);
456}
457
458/**
459 * lru_cache_add_active_or_unevictable
460 * @page: the page to be added to LRU
461 * @vma: vma in which page is mapped for determining reclaimability
462 *
463 * Place @page on the active or unevictable LRU list, depending on its
464 * evictability. Note that if the page is not evictable, it goes
465 * directly back onto it's zone's unevictable list, it does NOT use a
466 * per cpu pagevec.
467 */
468void lru_cache_add_active_or_unevictable(struct page *page,
469 struct vm_area_struct *vma)
470{
471 VM_BUG_ON_PAGE(PageLRU(page), page);
472
473 if (likely((vma->vm_flags & (VM_LOCKED | VM_SPECIAL)) != VM_LOCKED)) {
474 SetPageActive(page);
475 lru_cache_add(page);
476 return;
477 }
478
479 if (!TestSetPageMlocked(page)) {
480 /*
481 * We use the irq-unsafe __mod_zone_page_stat because this
482 * counter is not modified from interrupt context, and the pte
483 * lock is held(spinlock), which implies preemption disabled.
484 */
485 __mod_zone_page_state(page_zone(page), NR_MLOCK,
486 hpage_nr_pages(page));
487 count_vm_event(UNEVICTABLE_PGMLOCKED);
488 }
489 add_page_to_unevictable_list(page);
490}
491
492/*
493 * If the page can not be invalidated, it is moved to the
494 * inactive list to speed up its reclaim. It is moved to the
495 * head of the list, rather than the tail, to give the flusher
496 * threads some time to write it out, as this is much more
497 * effective than the single-page writeout from reclaim.
498 *
499 * If the page isn't page_mapped and dirty/writeback, the page
500 * could reclaim asap using PG_reclaim.
501 *
502 * 1. active, mapped page -> none
503 * 2. active, dirty/writeback page -> inactive, head, PG_reclaim
504 * 3. inactive, mapped page -> none
505 * 4. inactive, dirty/writeback page -> inactive, head, PG_reclaim
506 * 5. inactive, clean -> inactive, tail
507 * 6. Others -> none
508 *
509 * In 4, why it moves inactive's head, the VM expects the page would
510 * be write it out by flusher threads as this is much more effective
511 * than the single-page writeout from reclaim.
512 */
513static void lru_deactivate_file_fn(struct page *page, struct lruvec *lruvec,
514 void *arg)
515{
516 int lru, file;
517 bool active;
518
519 if (!PageLRU(page))
520 return;
521
522 if (PageUnevictable(page))
523 return;
524
525 /* Some processes are using the page */
526 if (page_mapped(page))
527 return;
528
529 active = PageActive(page);
530 file = page_is_file_cache(page);
531 lru = page_lru_base_type(page);
532
533 del_page_from_lru_list(page, lruvec, lru + active);
534 ClearPageActive(page);
535 ClearPageReferenced(page);
536 add_page_to_lru_list(page, lruvec, lru);
537
538 if (PageWriteback(page) || PageDirty(page)) {
539 /*
540 * PG_reclaim could be raced with end_page_writeback
541 * It can make readahead confusing. But race window
542 * is _really_ small and it's non-critical problem.
543 */
544 SetPageReclaim(page);
545 } else {
546 /*
547 * The page's writeback ends up during pagevec
548 * We moves tha page into tail of inactive.
549 */
550 list_move_tail(&page->lru, &lruvec->lists[lru]);
551 __count_vm_event(PGROTATED);
552 }
553
554 if (active)
555 __count_vm_event(PGDEACTIVATE);
556 update_page_reclaim_stat(lruvec, file, 0);
557}
558
559
560static void lru_deactivate_fn(struct page *page, struct lruvec *lruvec,
561 void *arg)
562{
563 if (PageLRU(page) && PageActive(page) && !PageUnevictable(page)) {
564 int file = page_is_file_cache(page);
565 int lru = page_lru_base_type(page);
566
567 del_page_from_lru_list(page, lruvec, lru + LRU_ACTIVE);
568 ClearPageActive(page);
569 ClearPageReferenced(page);
570 add_page_to_lru_list(page, lruvec, lru);
571
572 __count_vm_event(PGDEACTIVATE);
573 update_page_reclaim_stat(lruvec, file, 0);
574 }
575}
576
577/*
578 * Drain pages out of the cpu's pagevecs.
579 * Either "cpu" is the current CPU, and preemption has already been
580 * disabled; or "cpu" is being hot-unplugged, and is already dead.
581 */
582void lru_add_drain_cpu(int cpu)
583{
584 struct pagevec *pvec = &per_cpu(lru_add_pvec, cpu);
585
586 if (pagevec_count(pvec))
587 __pagevec_lru_add(pvec);
588
589 pvec = &per_cpu(lru_rotate_pvecs, cpu);
590 if (pagevec_count(pvec)) {
591 unsigned long flags;
592
593 /* No harm done if a racing interrupt already did this */
594 local_irq_save(flags);
595 pagevec_move_tail(pvec);
596 local_irq_restore(flags);
597 }
598
599 pvec = &per_cpu(lru_deactivate_file_pvecs, cpu);
600 if (pagevec_count(pvec))
601 pagevec_lru_move_fn(pvec, lru_deactivate_file_fn, NULL);
602
603 pvec = &per_cpu(lru_deactivate_pvecs, cpu);
604 if (pagevec_count(pvec))
605 pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
606
607 activate_page_drain(cpu);
608}
609
610/**
611 * deactivate_file_page - forcefully deactivate a file page
612 * @page: page to deactivate
613 *
614 * This function hints the VM that @page is a good reclaim candidate,
615 * for example if its invalidation fails due to the page being dirty
616 * or under writeback.
617 */
618void deactivate_file_page(struct page *page)
619{
620 /*
621 * In a workload with many unevictable page such as mprotect,
622 * unevictable page deactivation for accelerating reclaim is pointless.
623 */
624 if (PageUnevictable(page))
625 return;
626
627 if (likely(get_page_unless_zero(page))) {
628 struct pagevec *pvec = &get_cpu_var(lru_deactivate_file_pvecs);
629
630 if (!pagevec_add(pvec, page))
631 pagevec_lru_move_fn(pvec, lru_deactivate_file_fn, NULL);
632 put_cpu_var(lru_deactivate_file_pvecs);
633 }
634}
635
636/**
637 * deactivate_page - deactivate a page
638 * @page: page to deactivate
639 *
640 * deactivate_page() moves @page to the inactive list if @page was on the active
641 * list and was not an unevictable page. This is done to accelerate the reclaim
642 * of @page.
643 */
644void deactivate_page(struct page *page)
645{
646 if (PageLRU(page) && PageActive(page) && !PageUnevictable(page)) {
647 struct pagevec *pvec = &get_cpu_var(lru_deactivate_pvecs);
648
649 get_page(page);
650 if (!pagevec_add(pvec, page))
651 pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
652 put_cpu_var(lru_deactivate_pvecs);
653 }
654}
655
656void lru_add_drain(void)
657{
658 lru_add_drain_cpu(get_cpu());
659 put_cpu();
660}
661
662static void lru_add_drain_per_cpu(struct work_struct *dummy)
663{
664 lru_add_drain();
665}
666
667static DEFINE_PER_CPU(struct work_struct, lru_add_drain_work);
668
669void lru_add_drain_all(void)
670{
671 static DEFINE_MUTEX(lock);
672 static struct cpumask has_work;
673 int cpu;
674
675 mutex_lock(&lock);
676 get_online_cpus();
677 cpumask_clear(&has_work);
678
679 for_each_online_cpu(cpu) {
680 struct work_struct *work = &per_cpu(lru_add_drain_work, cpu);
681
682 if (pagevec_count(&per_cpu(lru_add_pvec, cpu)) ||
683 pagevec_count(&per_cpu(lru_rotate_pvecs, cpu)) ||
684 pagevec_count(&per_cpu(lru_deactivate_file_pvecs, cpu)) ||
685 pagevec_count(&per_cpu(lru_deactivate_pvecs, cpu)) ||
686 need_activate_page_drain(cpu)) {
687 INIT_WORK(work, lru_add_drain_per_cpu);
688 schedule_work_on(cpu, work);
689 cpumask_set_cpu(cpu, &has_work);
690 }
691 }
692
693 for_each_cpu(cpu, &has_work)
694 flush_work(&per_cpu(lru_add_drain_work, cpu));
695
696 put_online_cpus();
697 mutex_unlock(&lock);
698}
699
700/**
701 * release_pages - batched put_page()
702 * @pages: array of pages to release
703 * @nr: number of pages
704 * @cold: whether the pages are cache cold
705 *
706 * Decrement the reference count on all the pages in @pages. If it
707 * fell to zero, remove the page from the LRU and free it.
708 */
709void release_pages(struct page **pages, int nr, bool cold)
710{
711 int i;
712 LIST_HEAD(pages_to_free);
713 struct zone *zone = NULL;
714 struct lruvec *lruvec;
715 unsigned long uninitialized_var(flags);
716 unsigned int uninitialized_var(lock_batch);
717
718 for (i = 0; i < nr; i++) {
719 struct page *page = pages[i];
720
721 /*
722 * Make sure the IRQ-safe lock-holding time does not get
723 * excessive with a continuous string of pages from the
724 * same zone. The lock is held only if zone != NULL.
725 */
726 if (zone && ++lock_batch == SWAP_CLUSTER_MAX) {
727 spin_unlock_irqrestore(&zone->lru_lock, flags);
728 zone = NULL;
729 }
730
731 if (is_huge_zero_page(page)) {
732 put_huge_zero_page();
733 continue;
734 }
735
736 page = compound_head(page);
737 if (!put_page_testzero(page))
738 continue;
739
740 if (PageCompound(page)) {
741 if (zone) {
742 spin_unlock_irqrestore(&zone->lru_lock, flags);
743 zone = NULL;
744 }
745 __put_compound_page(page);
746 continue;
747 }
748
749 if (PageLRU(page)) {
750 struct zone *pagezone = page_zone(page);
751
752 if (pagezone != zone) {
753 if (zone)
754 spin_unlock_irqrestore(&zone->lru_lock,
755 flags);
756 lock_batch = 0;
757 zone = pagezone;
758 spin_lock_irqsave(&zone->lru_lock, flags);
759 }
760
761 lruvec = mem_cgroup_page_lruvec(page, zone);
762 VM_BUG_ON_PAGE(!PageLRU(page), page);
763 __ClearPageLRU(page);
764 del_page_from_lru_list(page, lruvec, page_off_lru(page));
765 }
766
767 /* Clear Active bit in case of parallel mark_page_accessed */
768 __ClearPageActive(page);
769
770 list_add(&page->lru, &pages_to_free);
771 }
772 if (zone)
773 spin_unlock_irqrestore(&zone->lru_lock, flags);
774
775 mem_cgroup_uncharge_list(&pages_to_free);
776 free_hot_cold_page_list(&pages_to_free, cold);
777}
778EXPORT_SYMBOL(release_pages);
779
780/*
781 * The pages which we're about to release may be in the deferred lru-addition
782 * queues. That would prevent them from really being freed right now. That's
783 * OK from a correctness point of view but is inefficient - those pages may be
784 * cache-warm and we want to give them back to the page allocator ASAP.
785 *
786 * So __pagevec_release() will drain those queues here. __pagevec_lru_add()
787 * and __pagevec_lru_add_active() call release_pages() directly to avoid
788 * mutual recursion.
789 */
790void __pagevec_release(struct pagevec *pvec)
791{
792 lru_add_drain();
793 release_pages(pvec->pages, pagevec_count(pvec), pvec->cold);
794 pagevec_reinit(pvec);
795}
796EXPORT_SYMBOL(__pagevec_release);
797
798#ifdef CONFIG_TRANSPARENT_HUGEPAGE
799/* used by __split_huge_page_refcount() */
800void lru_add_page_tail(struct page *page, struct page *page_tail,
801 struct lruvec *lruvec, struct list_head *list)
802{
803 const int file = 0;
804
805 VM_BUG_ON_PAGE(!PageHead(page), page);
806 VM_BUG_ON_PAGE(PageCompound(page_tail), page);
807 VM_BUG_ON_PAGE(PageLRU(page_tail), page);
808 VM_BUG_ON(NR_CPUS != 1 &&
809 !spin_is_locked(&lruvec_zone(lruvec)->lru_lock));
810
811 if (!list)
812 SetPageLRU(page_tail);
813
814 if (likely(PageLRU(page)))
815 list_add_tail(&page_tail->lru, &page->lru);
816 else if (list) {
817 /* page reclaim is reclaiming a huge page */
818 get_page(page_tail);
819 list_add_tail(&page_tail->lru, list);
820 } else {
821 struct list_head *list_head;
822 /*
823 * Head page has not yet been counted, as an hpage,
824 * so we must account for each subpage individually.
825 *
826 * Use the standard add function to put page_tail on the list,
827 * but then correct its position so they all end up in order.
828 */
829 add_page_to_lru_list(page_tail, lruvec, page_lru(page_tail));
830 list_head = page_tail->lru.prev;
831 list_move_tail(&page_tail->lru, list_head);
832 }
833
834 if (!PageUnevictable(page))
835 update_page_reclaim_stat(lruvec, file, PageActive(page_tail));
836}
837#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
838
839static void __pagevec_lru_add_fn(struct page *page, struct lruvec *lruvec,
840 void *arg)
841{
842 int file = page_is_file_cache(page);
843 int active = PageActive(page);
844 enum lru_list lru = page_lru(page);
845
846 VM_BUG_ON_PAGE(PageLRU(page), page);
847
848 SetPageLRU(page);
849 add_page_to_lru_list(page, lruvec, lru);
850 update_page_reclaim_stat(lruvec, file, active);
851 trace_mm_lru_insertion(page, lru);
852}
853
854/*
855 * Add the passed pages to the LRU, then drop the caller's refcount
856 * on them. Reinitialises the caller's pagevec.
857 */
858void __pagevec_lru_add(struct pagevec *pvec)
859{
860 pagevec_lru_move_fn(pvec, __pagevec_lru_add_fn, NULL);
861}
862EXPORT_SYMBOL(__pagevec_lru_add);
863
864/**
865 * pagevec_lookup_entries - gang pagecache lookup
866 * @pvec: Where the resulting entries are placed
867 * @mapping: The address_space to search
868 * @start: The starting entry index
869 * @nr_entries: The maximum number of entries
870 * @indices: The cache indices corresponding to the entries in @pvec
871 *
872 * pagevec_lookup_entries() will search for and return a group of up
873 * to @nr_entries pages and shadow entries in the mapping. All
874 * entries are placed in @pvec. pagevec_lookup_entries() takes a
875 * reference against actual pages in @pvec.
876 *
877 * The search returns a group of mapping-contiguous entries with
878 * ascending indexes. There may be holes in the indices due to
879 * not-present entries.
880 *
881 * pagevec_lookup_entries() returns the number of entries which were
882 * found.
883 */
884unsigned pagevec_lookup_entries(struct pagevec *pvec,
885 struct address_space *mapping,
886 pgoff_t start, unsigned nr_pages,
887 pgoff_t *indices)
888{
889 pvec->nr = find_get_entries(mapping, start, nr_pages,
890 pvec->pages, indices);
891 return pagevec_count(pvec);
892}
893
894/**
895 * pagevec_remove_exceptionals - pagevec exceptionals pruning
896 * @pvec: The pagevec to prune
897 *
898 * pagevec_lookup_entries() fills both pages and exceptional radix
899 * tree entries into the pagevec. This function prunes all
900 * exceptionals from @pvec without leaving holes, so that it can be
901 * passed on to page-only pagevec operations.
902 */
903void pagevec_remove_exceptionals(struct pagevec *pvec)
904{
905 int i, j;
906
907 for (i = 0, j = 0; i < pagevec_count(pvec); i++) {
908 struct page *page = pvec->pages[i];
909 if (!radix_tree_exceptional_entry(page))
910 pvec->pages[j++] = page;
911 }
912 pvec->nr = j;
913}
914
915/**
916 * pagevec_lookup - gang pagecache lookup
917 * @pvec: Where the resulting pages are placed
918 * @mapping: The address_space to search
919 * @start: The starting page index
920 * @nr_pages: The maximum number of pages
921 *
922 * pagevec_lookup() will search for and return a group of up to @nr_pages pages
923 * in the mapping. The pages are placed in @pvec. pagevec_lookup() takes a
924 * reference against the pages in @pvec.
925 *
926 * The search returns a group of mapping-contiguous pages with ascending
927 * indexes. There may be holes in the indices due to not-present pages.
928 *
929 * pagevec_lookup() returns the number of pages which were found.
930 */
931unsigned pagevec_lookup(struct pagevec *pvec, struct address_space *mapping,
932 pgoff_t start, unsigned nr_pages)
933{
934 pvec->nr = find_get_pages(mapping, start, nr_pages, pvec->pages);
935 return pagevec_count(pvec);
936}
937EXPORT_SYMBOL(pagevec_lookup);
938
939unsigned pagevec_lookup_tag(struct pagevec *pvec, struct address_space *mapping,
940 pgoff_t *index, int tag, unsigned nr_pages)
941{
942 pvec->nr = find_get_pages_tag(mapping, index, tag,
943 nr_pages, pvec->pages);
944 return pagevec_count(pvec);
945}
946EXPORT_SYMBOL(pagevec_lookup_tag);
947
948/*
949 * Perform any setup for the swap system
950 */
951void __init swap_setup(void)
952{
953 unsigned long megs = totalram_pages >> (20 - PAGE_SHIFT);
954#ifdef CONFIG_SWAP
955 int i;
956
957 for (i = 0; i < MAX_SWAPFILES; i++)
958 spin_lock_init(&swapper_spaces[i].tree_lock);
959#endif
960
961 /* Use a smaller cluster for small-memory machines */
962 if (megs < 16)
963 page_cluster = 2;
964 else
965 page_cluster = 3;
966 /*
967 * Right now other parts of the system means that we
968 * _really_ don't want to cluster much more
969 */
970}
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? */
47int page_cluster;
48
49/* Protecting only lru_rotate.pvec which requires disabling interrupts */
50struct lru_rotate {
51 local_lock_t lock;
52 struct pagevec pvec;
53};
54static DEFINE_PER_CPU(struct lru_rotate, lru_rotate) = {
55 .lock = INIT_LOCAL_LOCK(lock),
56};
57
58/*
59 * The following struct pagevec are grouped together because they are protected
60 * by disabling preemption (and interrupts remain enabled).
61 */
62struct lru_pvecs {
63 local_lock_t lock;
64 struct pagevec lru_add;
65 struct pagevec lru_deactivate_file;
66 struct pagevec lru_deactivate;
67 struct pagevec lru_lazyfree;
68#ifdef CONFIG_SMP
69 struct pagevec activate_page;
70#endif
71};
72static DEFINE_PER_CPU(struct lru_pvecs, lru_pvecs) = {
73 .lock = INIT_LOCAL_LOCK(lock),
74};
75
76/*
77 * This path almost never happens for VM activity - pages are normally
78 * freed via pagevecs. But it gets used by networking.
79 */
80static void __page_cache_release(struct page *page)
81{
82 if (PageLRU(page)) {
83 struct lruvec *lruvec;
84 unsigned long flags;
85
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);
90 }
91 __ClearPageWaiters(page);
92}
93
94static void __put_single_page(struct page *page)
95{
96 __page_cache_release(page);
97 mem_cgroup_uncharge(page);
98 free_unref_page(page, 0);
99}
100
101static void __put_compound_page(struct page *page)
102{
103 /*
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.)
108 */
109 if (!PageHuge(page))
110 __page_cache_release(page);
111 destroy_compound_page(page);
112}
113
114void __put_page(struct page *page)
115{
116 if (is_zone_device_page(page)) {
117 put_dev_pagemap(page->pgmap);
118
119 /*
120 * The page belongs to the device that created pgmap. Do
121 * not return it to page allocator.
122 */
123 return;
124 }
125
126 if (unlikely(PageCompound(page)))
127 __put_compound_page(page);
128 else
129 __put_single_page(page);
130}
131EXPORT_SYMBOL(__put_page);
132
133/**
134 * put_pages_list() - release a list of pages
135 * @pages: list of pages threaded on page->lru
136 *
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.
139 */
140void put_pages_list(struct list_head *pages)
141{
142 while (!list_empty(pages)) {
143 struct page *victim;
144
145 victim = lru_to_page(pages);
146 list_del(&victim->lru);
147 put_page(victim);
148 }
149}
150EXPORT_SYMBOL(put_pages_list);
151
152/*
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.
159 *
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.
164 */
165int get_kernel_pages(const struct kvec *kiov, int nr_segs, int write,
166 struct page **pages)
167{
168 int seg;
169
170 for (seg = 0; seg < nr_segs; seg++) {
171 if (WARN_ON(kiov[seg].iov_len != PAGE_SIZE))
172 return seg;
173
174 pages[seg] = kmap_to_page(kiov[seg].iov_base);
175 get_page(pages[seg]);
176 }
177
178 return seg;
179}
180EXPORT_SYMBOL_GPL(get_kernel_pages);
181
182/*
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.
188 *
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.
192 */
193int get_kernel_page(unsigned long start, int write, struct page **pages)
194{
195 const struct kvec kiov = {
196 .iov_base = (void *)start,
197 .iov_len = PAGE_SIZE
198 };
199
200 return get_kernel_pages(&kiov, 1, write, pages);
201}
202EXPORT_SYMBOL_GPL(get_kernel_page);
203
204static void pagevec_lru_move_fn(struct pagevec *pvec,
205 void (*move_fn)(struct page *page, struct lruvec *lruvec))
206{
207 int i;
208 struct lruvec *lruvec = NULL;
209 unsigned long flags = 0;
210
211 for (i = 0; i < pagevec_count(pvec); i++) {
212 struct page *page = pvec->pages[i];
213
214 /* block memcg migration during page moving between lru */
215 if (!TestClearPageLRU(page))
216 continue;
217
218 lruvec = relock_page_lruvec_irqsave(page, lruvec, &flags);
219 (*move_fn)(page, lruvec);
220
221 SetPageLRU(page);
222 }
223 if (lruvec)
224 unlock_page_lruvec_irqrestore(lruvec, flags);
225 release_pages(pvec->pages, pvec->nr);
226 pagevec_reinit(pvec);
227}
228
229static void pagevec_move_tail_fn(struct page *page, struct lruvec *lruvec)
230{
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));
236 }
237}
238
239/* return true if pagevec needs to drain */
240static bool pagevec_add_and_need_flush(struct pagevec *pvec, struct page *page)
241{
242 bool ret = false;
243
244 if (!pagevec_add(pvec, page) || PageCompound(page) ||
245 lru_cache_disabled())
246 ret = true;
247
248 return ret;
249}
250
251/*
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
254 * inactive list.
255 *
256 * rotate_reclaimable_page() must disable IRQs, to prevent nasty races.
257 */
258void rotate_reclaimable_page(struct page *page)
259{
260 if (!PageLocked(page) && !PageDirty(page) &&
261 !PageUnevictable(page) && PageLRU(page)) {
262 struct pagevec *pvec;
263 unsigned long flags;
264
265 get_page(page);
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);
271 }
272}
273
274void lru_note_cost(struct lruvec *lruvec, bool file, unsigned int nr_pages)
275{
276 do {
277 unsigned long lrusize;
278
279 /*
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).
285 */
286 spin_lock_irq(&lruvec->lru_lock);
287 /* Record cost event */
288 if (file)
289 lruvec->file_cost += nr_pages;
290 else
291 lruvec->anon_cost += nr_pages;
292
293 /*
294 * Decay previous events
295 *
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.
300 */
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);
305
306 if (lruvec->file_cost + lruvec->anon_cost > lrusize / 4) {
307 lruvec->file_cost /= 2;
308 lruvec->anon_cost /= 2;
309 }
310 spin_unlock_irq(&lruvec->lru_lock);
311 } while ((lruvec = parent_lruvec(lruvec)));
312}
313
314void lru_note_cost_page(struct page *page)
315{
316 lru_note_cost(mem_cgroup_page_lruvec(page),
317 page_is_file_lru(page), thp_nr_pages(page));
318}
319
320static void __activate_page(struct page *page, struct lruvec *lruvec)
321{
322 if (!PageActive(page) && !PageUnevictable(page)) {
323 int nr_pages = thp_nr_pages(page);
324
325 del_page_from_lru_list(page, lruvec);
326 SetPageActive(page);
327 add_page_to_lru_list(page, lruvec);
328 trace_mm_lru_activate(page);
329
330 __count_vm_events(PGACTIVATE, nr_pages);
331 __count_memcg_events(lruvec_memcg(lruvec), PGACTIVATE,
332 nr_pages);
333 }
334}
335
336#ifdef CONFIG_SMP
337static void activate_page_drain(int cpu)
338{
339 struct pagevec *pvec = &per_cpu(lru_pvecs.activate_page, cpu);
340
341 if (pagevec_count(pvec))
342 pagevec_lru_move_fn(pvec, __activate_page);
343}
344
345static bool need_activate_page_drain(int cpu)
346{
347 return pagevec_count(&per_cpu(lru_pvecs.activate_page, cpu)) != 0;
348}
349
350static void activate_page(struct page *page)
351{
352 page = compound_head(page);
353 if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
354 struct pagevec *pvec;
355
356 local_lock(&lru_pvecs.lock);
357 pvec = this_cpu_ptr(&lru_pvecs.activate_page);
358 get_page(page);
359 if (pagevec_add_and_need_flush(pvec, page))
360 pagevec_lru_move_fn(pvec, __activate_page);
361 local_unlock(&lru_pvecs.lock);
362 }
363}
364
365#else
366static inline void activate_page_drain(int cpu)
367{
368}
369
370static void activate_page(struct page *page)
371{
372 struct lruvec *lruvec;
373
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);
379 SetPageLRU(page);
380 }
381}
382#endif
383
384static void __lru_cache_activate_page(struct page *page)
385{
386 struct pagevec *pvec;
387 int i;
388
389 local_lock(&lru_pvecs.lock);
390 pvec = this_cpu_ptr(&lru_pvecs.lru_add);
391
392 /*
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.
401 */
402 for (i = pagevec_count(pvec) - 1; i >= 0; i--) {
403 struct page *pagevec_page = pvec->pages[i];
404
405 if (pagevec_page == page) {
406 SetPageActive(page);
407 break;
408 }
409 }
410
411 local_unlock(&lru_pvecs.lock);
412}
413
414/*
415 * Mark a page as having seen activity.
416 *
417 * inactive,unreferenced -> inactive,referenced
418 * inactive,referenced -> active,unreferenced
419 * active,unreferenced -> active,referenced
420 *
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).
423 */
424void mark_page_accessed(struct page *page)
425{
426 page = compound_head(page);
427
428 if (!PageReferenced(page)) {
429 SetPageReferenced(page);
430 } else if (PageUnevictable(page)) {
431 /*
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.
435 */
436 } else if (!PageActive(page)) {
437 /*
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.
442 */
443 if (PageLRU(page))
444 activate_page(page);
445 else
446 __lru_cache_activate_page(page);
447 ClearPageReferenced(page);
448 workingset_activation(page);
449 }
450 if (page_is_idle(page))
451 clear_page_idle(page);
452}
453EXPORT_SYMBOL(mark_page_accessed);
454
455/**
456 * lru_cache_add - add a page to a page list
457 * @page: the page to be added to the LRU.
458 *
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().
463 */
464void lru_cache_add(struct page *page)
465{
466 struct pagevec *pvec;
467
468 VM_BUG_ON_PAGE(PageActive(page) && PageUnevictable(page), page);
469 VM_BUG_ON_PAGE(PageLRU(page), page);
470
471 get_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);
477}
478EXPORT_SYMBOL(lru_cache_add);
479
480/**
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
484 *
485 * Place @page on the inactive or unevictable LRU list, depending on its
486 * evictability.
487 */
488void lru_cache_add_inactive_or_unevictable(struct page *page,
489 struct vm_area_struct *vma)
490{
491 bool unevictable;
492
493 VM_BUG_ON_PAGE(PageLRU(page), page);
494
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);
498 /*
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.
502 */
503 __mod_zone_page_state(page_zone(page), NR_MLOCK, nr_pages);
504 count_vm_events(UNEVICTABLE_PGMLOCKED, nr_pages);
505 }
506 lru_cache_add(page);
507}
508
509/*
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.
515 *
516 * If the page isn't page_mapped and dirty/writeback, the page
517 * could reclaim asap using PG_reclaim.
518 *
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
524 * 6. Others -> none
525 *
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.
529 */
530static void lru_deactivate_file_fn(struct page *page, struct lruvec *lruvec)
531{
532 bool active = PageActive(page);
533 int nr_pages = thp_nr_pages(page);
534
535 if (PageUnevictable(page))
536 return;
537
538 /* Some processes are using the page */
539 if (page_mapped(page))
540 return;
541
542 del_page_from_lru_list(page, lruvec);
543 ClearPageActive(page);
544 ClearPageReferenced(page);
545
546 if (PageWriteback(page) || PageDirty(page)) {
547 /*
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.
551 */
552 add_page_to_lru_list(page, lruvec);
553 SetPageReclaim(page);
554 } else {
555 /*
556 * The page's writeback ends up during pagevec
557 * We move that page into tail of inactive.
558 */
559 add_page_to_lru_list_tail(page, lruvec);
560 __count_vm_events(PGROTATED, nr_pages);
561 }
562
563 if (active) {
564 __count_vm_events(PGDEACTIVATE, nr_pages);
565 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
566 nr_pages);
567 }
568}
569
570static void lru_deactivate_fn(struct page *page, struct lruvec *lruvec)
571{
572 if (PageActive(page) && !PageUnevictable(page)) {
573 int nr_pages = thp_nr_pages(page);
574
575 del_page_from_lru_list(page, lruvec);
576 ClearPageActive(page);
577 ClearPageReferenced(page);
578 add_page_to_lru_list(page, lruvec);
579
580 __count_vm_events(PGDEACTIVATE, nr_pages);
581 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
582 nr_pages);
583 }
584}
585
586static void lru_lazyfree_fn(struct page *page, struct lruvec *lruvec)
587{
588 if (PageAnon(page) && PageSwapBacked(page) &&
589 !PageSwapCache(page) && !PageUnevictable(page)) {
590 int nr_pages = thp_nr_pages(page);
591
592 del_page_from_lru_list(page, lruvec);
593 ClearPageActive(page);
594 ClearPageReferenced(page);
595 /*
596 * Lazyfree pages are clean anonymous pages. They have
597 * PG_swapbacked flag cleared, to distinguish them from normal
598 * anonymous pages
599 */
600 ClearPageSwapBacked(page);
601 add_page_to_lru_list(page, lruvec);
602
603 __count_vm_events(PGLAZYFREE, nr_pages);
604 __count_memcg_events(lruvec_memcg(lruvec), PGLAZYFREE,
605 nr_pages);
606 }
607}
608
609/*
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.
613 */
614void lru_add_drain_cpu(int cpu)
615{
616 struct pagevec *pvec = &per_cpu(lru_pvecs.lru_add, cpu);
617
618 if (pagevec_count(pvec))
619 __pagevec_lru_add(pvec);
620
621 pvec = &per_cpu(lru_rotate.pvec, cpu);
622 /* Disabling interrupts below acts as a compiler barrier. */
623 if (data_race(pagevec_count(pvec))) {
624 unsigned long flags;
625
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);
630 }
631
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);
635
636 pvec = &per_cpu(lru_pvecs.lru_deactivate, cpu);
637 if (pagevec_count(pvec))
638 pagevec_lru_move_fn(pvec, lru_deactivate_fn);
639
640 pvec = &per_cpu(lru_pvecs.lru_lazyfree, cpu);
641 if (pagevec_count(pvec))
642 pagevec_lru_move_fn(pvec, lru_lazyfree_fn);
643
644 activate_page_drain(cpu);
645 invalidate_bh_lrus_cpu(cpu);
646}
647
648/**
649 * deactivate_file_page - forcefully deactivate a file page
650 * @page: page to deactivate
651 *
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.
655 */
656void deactivate_file_page(struct page *page)
657{
658 /*
659 * In a workload with many unevictable page such as mprotect,
660 * unevictable page deactivation for accelerating reclaim is pointless.
661 */
662 if (PageUnevictable(page))
663 return;
664
665 if (likely(get_page_unless_zero(page))) {
666 struct pagevec *pvec;
667
668 local_lock(&lru_pvecs.lock);
669 pvec = this_cpu_ptr(&lru_pvecs.lru_deactivate_file);
670
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);
674 }
675}
676
677/*
678 * deactivate_page - deactivate a page
679 * @page: page to deactivate
680 *
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
683 * of @page.
684 */
685void deactivate_page(struct page *page)
686{
687 if (PageLRU(page) && PageActive(page) && !PageUnevictable(page)) {
688 struct pagevec *pvec;
689
690 local_lock(&lru_pvecs.lock);
691 pvec = this_cpu_ptr(&lru_pvecs.lru_deactivate);
692 get_page(page);
693 if (pagevec_add_and_need_flush(pvec, page))
694 pagevec_lru_move_fn(pvec, lru_deactivate_fn);
695 local_unlock(&lru_pvecs.lock);
696 }
697}
698
699/**
700 * mark_page_lazyfree - make an anon page lazyfree
701 * @page: page to deactivate
702 *
703 * mark_page_lazyfree() moves @page to the inactive file list.
704 * This is done to accelerate the reclaim of @page.
705 */
706void mark_page_lazyfree(struct page *page)
707{
708 if (PageLRU(page) && PageAnon(page) && PageSwapBacked(page) &&
709 !PageSwapCache(page) && !PageUnevictable(page)) {
710 struct pagevec *pvec;
711
712 local_lock(&lru_pvecs.lock);
713 pvec = this_cpu_ptr(&lru_pvecs.lru_lazyfree);
714 get_page(page);
715 if (pagevec_add_and_need_flush(pvec, page))
716 pagevec_lru_move_fn(pvec, lru_lazyfree_fn);
717 local_unlock(&lru_pvecs.lock);
718 }
719}
720
721void lru_add_drain(void)
722{
723 local_lock(&lru_pvecs.lock);
724 lru_add_drain_cpu(smp_processor_id());
725 local_unlock(&lru_pvecs.lock);
726}
727
728void lru_add_drain_cpu_zone(struct zone *zone)
729{
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);
734}
735
736#ifdef CONFIG_SMP
737
738static DEFINE_PER_CPU(struct work_struct, lru_add_drain_work);
739
740static void lru_add_drain_per_cpu(struct work_struct *dummy)
741{
742 lru_add_drain();
743}
744
745/*
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.
751 */
752inline void __lru_add_drain_all(bool force_all_cpus)
753{
754 /*
755 * lru_drain_gen - Global pages generation number
756 *
757 * (A) Definition: global lru_drain_gen = x implies that all generations
758 * 0 < n <= x are already *scheduled* for draining.
759 *
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
762 * each CPU.
763 */
764 static unsigned int lru_drain_gen;
765 static struct cpumask has_work;
766 static DEFINE_MUTEX(lock);
767 unsigned cpu, this_gen;
768
769 /*
770 * Make sure nobody triggers this path before mm_percpu_wq is fully
771 * initialized.
772 */
773 if (WARN_ON(!mm_percpu_wq))
774 return;
775
776 /*
777 * Guarantee pagevec counter stores visible by this CPU are visible to
778 * other CPUs before loading the current drain generation.
779 */
780 smp_mb();
781
782 /*
783 * (B) Locally cache global LRU draining generation number
784 *
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
787 * at (D).
788 */
789 this_gen = smp_load_acquire(&lru_drain_gen);
790
791 mutex_lock(&lock);
792
793 /*
794 * (C) Exit the draining operation if a newer generation, from another
795 * lru_add_drain_all(), was already scheduled for draining. Check (A).
796 */
797 if (unlikely(this_gen != lru_drain_gen && !force_all_cpus))
798 goto done;
799
800 /*
801 * (D) Increment global generation number
802 *
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.
806 *
807 * This pairing must be done here, before the for_each_online_cpu loop
808 * below which drains the page vectors.
809 *
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().
815 *
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
818 * added pages.
819 */
820 WRITE_ONCE(lru_drain_gen, lru_drain_gen + 1);
821 smp_mb();
822
823 cpumask_clear(&has_work);
824 for_each_online_cpu(cpu) {
825 struct work_struct *work = &per_cpu(lru_add_drain_work, cpu);
826
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);
838 }
839 }
840
841 for_each_cpu(cpu, &has_work)
842 flush_work(&per_cpu(lru_add_drain_work, cpu));
843
844done:
845 mutex_unlock(&lock);
846}
847
848void lru_add_drain_all(void)
849{
850 __lru_add_drain_all(false);
851}
852#else
853void lru_add_drain_all(void)
854{
855 lru_add_drain();
856}
857#endif /* CONFIG_SMP */
858
859atomic_t lru_disable_count = ATOMIC_INIT(0);
860
861/*
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.
866 *
867 * Must be paired with a call to lru_cache_enable().
868 */
869void lru_cache_disable(void)
870{
871 atomic_inc(&lru_disable_count);
872#ifdef CONFIG_SMP
873 /*
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
879 * guarantees.
880 */
881 __lru_add_drain_all(true);
882#else
883 lru_add_drain();
884#endif
885}
886
887/**
888 * release_pages - batched put_page()
889 * @pages: array of pages to release
890 * @nr: number of pages
891 *
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.
894 */
895void release_pages(struct page **pages, int nr)
896{
897 int i;
898 LIST_HEAD(pages_to_free);
899 struct lruvec *lruvec = NULL;
900 unsigned long flags;
901 unsigned int lock_batch;
902
903 for (i = 0; i < nr; i++) {
904 struct page *page = pages[i];
905
906 /*
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.
910 */
911 if (lruvec && ++lock_batch == SWAP_CLUSTER_MAX) {
912 unlock_page_lruvec_irqrestore(lruvec, flags);
913 lruvec = NULL;
914 }
915
916 page = compound_head(page);
917 if (is_huge_zero_page(page))
918 continue;
919
920 if (is_zone_device_page(page)) {
921 if (lruvec) {
922 unlock_page_lruvec_irqrestore(lruvec, flags);
923 lruvec = NULL;
924 }
925 /*
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().
930 */
931 if (page_is_devmap_managed(page)) {
932 put_devmap_managed_page(page);
933 continue;
934 }
935 if (put_page_testzero(page))
936 put_dev_pagemap(page->pgmap);
937 continue;
938 }
939
940 if (!put_page_testzero(page))
941 continue;
942
943 if (PageCompound(page)) {
944 if (lruvec) {
945 unlock_page_lruvec_irqrestore(lruvec, flags);
946 lruvec = NULL;
947 }
948 __put_compound_page(page);
949 continue;
950 }
951
952 if (PageLRU(page)) {
953 struct lruvec *prev_lruvec = lruvec;
954
955 lruvec = relock_page_lruvec_irqsave(page, lruvec,
956 &flags);
957 if (prev_lruvec != lruvec)
958 lock_batch = 0;
959
960 del_page_from_lru_list(page, lruvec);
961 __clear_page_lru_flags(page);
962 }
963
964 __ClearPageWaiters(page);
965
966 list_add(&page->lru, &pages_to_free);
967 }
968 if (lruvec)
969 unlock_page_lruvec_irqrestore(lruvec, flags);
970
971 mem_cgroup_uncharge_list(&pages_to_free);
972 free_unref_page_list(&pages_to_free);
973}
974EXPORT_SYMBOL(release_pages);
975
976/*
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.
981 *
982 * So __pagevec_release() will drain those queues here. __pagevec_lru_add()
983 * and __pagevec_lru_add_active() call release_pages() directly to avoid
984 * mutual recursion.
985 */
986void __pagevec_release(struct pagevec *pvec)
987{
988 if (!pvec->percpu_pvec_drained) {
989 lru_add_drain();
990 pvec->percpu_pvec_drained = true;
991 }
992 release_pages(pvec->pages, pagevec_count(pvec));
993 pagevec_reinit(pvec);
994}
995EXPORT_SYMBOL(__pagevec_release);
996
997static void __pagevec_lru_add_fn(struct page *page, struct lruvec *lruvec)
998{
999 int was_unevictable = TestClearPageUnevictable(page);
1000 int nr_pages = thp_nr_pages(page);
1001
1002 VM_BUG_ON_PAGE(PageLRU(page), page);
1003
1004 /*
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]
1010 *
1011 * (1) & (2a) are ok as LRU lock will serialize them. For (2b), we need
1012 * following strict ordering:
1013 *
1014 * #0: __pagevec_lru_add_fn #1: clear_page_mlock
1015 *
1016 * SetPageLRU() TestClearPageMlocked()
1017 * smp_mb() // explicit ordering // above provides strict
1018 * // ordering
1019 * PageMlocked() PageLRU()
1020 *
1021 *
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.
1029 */
1030 SetPageLRU(page);
1031 smp_mb__after_atomic();
1032
1033 if (page_evictable(page)) {
1034 if (was_unevictable)
1035 __count_vm_events(UNEVICTABLE_PGRESCUED, nr_pages);
1036 } else {
1037 ClearPageActive(page);
1038 SetPageUnevictable(page);
1039 if (!was_unevictable)
1040 __count_vm_events(UNEVICTABLE_PGCULLED, nr_pages);
1041 }
1042
1043 add_page_to_lru_list(page, lruvec);
1044 trace_mm_lru_insertion(page);
1045}
1046
1047/*
1048 * Add the passed pages to the LRU, then drop the caller's refcount
1049 * on them. Reinitialises the caller's pagevec.
1050 */
1051void __pagevec_lru_add(struct pagevec *pvec)
1052{
1053 int i;
1054 struct lruvec *lruvec = NULL;
1055 unsigned long flags = 0;
1056
1057 for (i = 0; i < pagevec_count(pvec); i++) {
1058 struct page *page = pvec->pages[i];
1059
1060 lruvec = relock_page_lruvec_irqsave(page, lruvec, &flags);
1061 __pagevec_lru_add_fn(page, lruvec);
1062 }
1063 if (lruvec)
1064 unlock_page_lruvec_irqrestore(lruvec, flags);
1065 release_pages(pvec->pages, pvec->nr);
1066 pagevec_reinit(pvec);
1067}
1068
1069/**
1070 * pagevec_remove_exceptionals - pagevec exceptionals pruning
1071 * @pvec: The pagevec to prune
1072 *
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.
1077 */
1078void pagevec_remove_exceptionals(struct pagevec *pvec)
1079{
1080 int i, j;
1081
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;
1086 }
1087 pvec->nr = j;
1088}
1089
1090/**
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
1096 *
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.
1101 *
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.
1105 *
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
1108 * reached.
1109 */
1110unsigned pagevec_lookup_range(struct pagevec *pvec,
1111 struct address_space *mapping, pgoff_t *start, pgoff_t end)
1112{
1113 pvec->nr = find_get_pages_range(mapping, start, end, PAGEVEC_SIZE,
1114 pvec->pages);
1115 return pagevec_count(pvec);
1116}
1117EXPORT_SYMBOL(pagevec_lookup_range);
1118
1119unsigned pagevec_lookup_range_tag(struct pagevec *pvec,
1120 struct address_space *mapping, pgoff_t *index, pgoff_t end,
1121 xa_mark_t tag)
1122{
1123 pvec->nr = find_get_pages_range_tag(mapping, index, end, tag,
1124 PAGEVEC_SIZE, pvec->pages);
1125 return pagevec_count(pvec);
1126}
1127EXPORT_SYMBOL(pagevec_lookup_range_tag);
1128
1129/*
1130 * Perform any setup for the swap system
1131 */
1132void __init swap_setup(void)
1133{
1134 unsigned long megs = totalram_pages() >> (20 - PAGE_SHIFT);
1135
1136 /* Use a smaller cluster for small-memory machines */
1137 if (megs < 16)
1138 page_cluster = 2;
1139 else
1140 page_cluster = 3;
1141 /*
1142 * Right now other parts of the system means that we
1143 * _really_ don't want to cluster much more
1144 */
1145}
1146
1147#ifdef CONFIG_DEV_PAGEMAP_OPS
1148void put_devmap_managed_page(struct page *page)
1149{
1150 int count;
1151
1152 if (WARN_ON_ONCE(!page_is_devmap_managed(page)))
1153 return;
1154
1155 count = page_ref_dec_return(page);
1156
1157 /*
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.
1161 */
1162 if (count == 1)
1163 free_devmap_managed_page(page);
1164 else if (!count)
1165 __put_page(page);
1166}
1167EXPORT_SYMBOL(put_devmap_managed_page);
1168#endif