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/module.h>
25#include <linux/mm_inline.h>
26#include <linux/buffer_head.h> /* for try_to_release_page() */
27#include <linux/percpu_counter.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
35#include "internal.h"
36
37/* How many pages do we try to swap or page in/out together? */
38int page_cluster;
39
40static DEFINE_PER_CPU(struct pagevec[NR_LRU_LISTS], lru_add_pvecs);
41static DEFINE_PER_CPU(struct pagevec, lru_rotate_pvecs);
42static DEFINE_PER_CPU(struct pagevec, lru_deactivate_pvecs);
43
44/*
45 * This path almost never happens for VM activity - pages are normally
46 * freed via pagevecs. But it gets used by networking.
47 */
48static void __page_cache_release(struct page *page)
49{
50 if (PageLRU(page)) {
51 unsigned long flags;
52 struct zone *zone = page_zone(page);
53
54 spin_lock_irqsave(&zone->lru_lock, flags);
55 VM_BUG_ON(!PageLRU(page));
56 __ClearPageLRU(page);
57 del_page_from_lru(zone, page);
58 spin_unlock_irqrestore(&zone->lru_lock, flags);
59 }
60}
61
62static void __put_single_page(struct page *page)
63{
64 __page_cache_release(page);
65 free_hot_cold_page(page, 0);
66}
67
68static void __put_compound_page(struct page *page)
69{
70 compound_page_dtor *dtor;
71
72 __page_cache_release(page);
73 dtor = get_compound_page_dtor(page);
74 (*dtor)(page);
75}
76
77static void put_compound_page(struct page *page)
78{
79 if (unlikely(PageTail(page))) {
80 /* __split_huge_page_refcount can run under us */
81 struct page *page_head = page->first_page;
82 smp_rmb();
83 /*
84 * If PageTail is still set after smp_rmb() we can be sure
85 * that the page->first_page we read wasn't a dangling pointer.
86 * See __split_huge_page_refcount() smp_wmb().
87 */
88 if (likely(PageTail(page) && get_page_unless_zero(page_head))) {
89 unsigned long flags;
90 /*
91 * Verify that our page_head wasn't converted
92 * to a a regular page before we got a
93 * reference on it.
94 */
95 if (unlikely(!PageHead(page_head))) {
96 /* PageHead is cleared after PageTail */
97 smp_rmb();
98 VM_BUG_ON(PageTail(page));
99 goto out_put_head;
100 }
101 /*
102 * Only run compound_lock on a valid PageHead,
103 * after having it pinned with
104 * get_page_unless_zero() above.
105 */
106 smp_mb();
107 /* page_head wasn't a dangling pointer */
108 flags = compound_lock_irqsave(page_head);
109 if (unlikely(!PageTail(page))) {
110 /* __split_huge_page_refcount run before us */
111 compound_unlock_irqrestore(page_head, flags);
112 VM_BUG_ON(PageHead(page_head));
113 out_put_head:
114 if (put_page_testzero(page_head))
115 __put_single_page(page_head);
116 out_put_single:
117 if (put_page_testzero(page))
118 __put_single_page(page);
119 return;
120 }
121 VM_BUG_ON(page_head != page->first_page);
122 /*
123 * We can release the refcount taken by
124 * get_page_unless_zero now that
125 * split_huge_page_refcount is blocked on the
126 * compound_lock.
127 */
128 if (put_page_testzero(page_head))
129 VM_BUG_ON(1);
130 /* __split_huge_page_refcount will wait now */
131 VM_BUG_ON(atomic_read(&page->_count) <= 0);
132 atomic_dec(&page->_count);
133 VM_BUG_ON(atomic_read(&page_head->_count) <= 0);
134 compound_unlock_irqrestore(page_head, flags);
135 if (put_page_testzero(page_head)) {
136 if (PageHead(page_head))
137 __put_compound_page(page_head);
138 else
139 __put_single_page(page_head);
140 }
141 } else {
142 /* page_head is a dangling pointer */
143 VM_BUG_ON(PageTail(page));
144 goto out_put_single;
145 }
146 } else if (put_page_testzero(page)) {
147 if (PageHead(page))
148 __put_compound_page(page);
149 else
150 __put_single_page(page);
151 }
152}
153
154void put_page(struct page *page)
155{
156 if (unlikely(PageCompound(page)))
157 put_compound_page(page);
158 else if (put_page_testzero(page))
159 __put_single_page(page);
160}
161EXPORT_SYMBOL(put_page);
162
163/**
164 * put_pages_list() - release a list of pages
165 * @pages: list of pages threaded on page->lru
166 *
167 * Release a list of pages which are strung together on page.lru. Currently
168 * used by read_cache_pages() and related error recovery code.
169 */
170void put_pages_list(struct list_head *pages)
171{
172 while (!list_empty(pages)) {
173 struct page *victim;
174
175 victim = list_entry(pages->prev, struct page, lru);
176 list_del(&victim->lru);
177 page_cache_release(victim);
178 }
179}
180EXPORT_SYMBOL(put_pages_list);
181
182static void pagevec_lru_move_fn(struct pagevec *pvec,
183 void (*move_fn)(struct page *page, void *arg),
184 void *arg)
185{
186 int i;
187 struct zone *zone = NULL;
188 unsigned long flags = 0;
189
190 for (i = 0; i < pagevec_count(pvec); i++) {
191 struct page *page = pvec->pages[i];
192 struct zone *pagezone = page_zone(page);
193
194 if (pagezone != zone) {
195 if (zone)
196 spin_unlock_irqrestore(&zone->lru_lock, flags);
197 zone = pagezone;
198 spin_lock_irqsave(&zone->lru_lock, flags);
199 }
200
201 (*move_fn)(page, arg);
202 }
203 if (zone)
204 spin_unlock_irqrestore(&zone->lru_lock, flags);
205 release_pages(pvec->pages, pvec->nr, pvec->cold);
206 pagevec_reinit(pvec);
207}
208
209static void pagevec_move_tail_fn(struct page *page, void *arg)
210{
211 int *pgmoved = arg;
212 struct zone *zone = page_zone(page);
213
214 if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
215 enum lru_list lru = page_lru_base_type(page);
216 list_move_tail(&page->lru, &zone->lru[lru].list);
217 mem_cgroup_rotate_reclaimable_page(page);
218 (*pgmoved)++;
219 }
220}
221
222/*
223 * pagevec_move_tail() must be called with IRQ disabled.
224 * Otherwise this may cause nasty races.
225 */
226static void pagevec_move_tail(struct pagevec *pvec)
227{
228 int pgmoved = 0;
229
230 pagevec_lru_move_fn(pvec, pagevec_move_tail_fn, &pgmoved);
231 __count_vm_events(PGROTATED, pgmoved);
232}
233
234/*
235 * Writeback is about to end against a page which has been marked for immediate
236 * reclaim. If it still appears to be reclaimable, move it to the tail of the
237 * inactive list.
238 */
239void rotate_reclaimable_page(struct page *page)
240{
241 if (!PageLocked(page) && !PageDirty(page) && !PageActive(page) &&
242 !PageUnevictable(page) && PageLRU(page)) {
243 struct pagevec *pvec;
244 unsigned long flags;
245
246 page_cache_get(page);
247 local_irq_save(flags);
248 pvec = &__get_cpu_var(lru_rotate_pvecs);
249 if (!pagevec_add(pvec, page))
250 pagevec_move_tail(pvec);
251 local_irq_restore(flags);
252 }
253}
254
255static void update_page_reclaim_stat(struct zone *zone, struct page *page,
256 int file, int rotated)
257{
258 struct zone_reclaim_stat *reclaim_stat = &zone->reclaim_stat;
259 struct zone_reclaim_stat *memcg_reclaim_stat;
260
261 memcg_reclaim_stat = mem_cgroup_get_reclaim_stat_from_page(page);
262
263 reclaim_stat->recent_scanned[file]++;
264 if (rotated)
265 reclaim_stat->recent_rotated[file]++;
266
267 if (!memcg_reclaim_stat)
268 return;
269
270 memcg_reclaim_stat->recent_scanned[file]++;
271 if (rotated)
272 memcg_reclaim_stat->recent_rotated[file]++;
273}
274
275static void __activate_page(struct page *page, void *arg)
276{
277 struct zone *zone = page_zone(page);
278
279 if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
280 int file = page_is_file_cache(page);
281 int lru = page_lru_base_type(page);
282 del_page_from_lru_list(zone, page, lru);
283
284 SetPageActive(page);
285 lru += LRU_ACTIVE;
286 add_page_to_lru_list(zone, page, lru);
287 __count_vm_event(PGACTIVATE);
288
289 update_page_reclaim_stat(zone, page, file, 1);
290 }
291}
292
293#ifdef CONFIG_SMP
294static DEFINE_PER_CPU(struct pagevec, activate_page_pvecs);
295
296static void activate_page_drain(int cpu)
297{
298 struct pagevec *pvec = &per_cpu(activate_page_pvecs, cpu);
299
300 if (pagevec_count(pvec))
301 pagevec_lru_move_fn(pvec, __activate_page, NULL);
302}
303
304void activate_page(struct page *page)
305{
306 if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
307 struct pagevec *pvec = &get_cpu_var(activate_page_pvecs);
308
309 page_cache_get(page);
310 if (!pagevec_add(pvec, page))
311 pagevec_lru_move_fn(pvec, __activate_page, NULL);
312 put_cpu_var(activate_page_pvecs);
313 }
314}
315
316#else
317static inline void activate_page_drain(int cpu)
318{
319}
320
321void activate_page(struct page *page)
322{
323 struct zone *zone = page_zone(page);
324
325 spin_lock_irq(&zone->lru_lock);
326 __activate_page(page, NULL);
327 spin_unlock_irq(&zone->lru_lock);
328}
329#endif
330
331/*
332 * Mark a page as having seen activity.
333 *
334 * inactive,unreferenced -> inactive,referenced
335 * inactive,referenced -> active,unreferenced
336 * active,unreferenced -> active,referenced
337 */
338void mark_page_accessed(struct page *page)
339{
340 if (!PageActive(page) && !PageUnevictable(page) &&
341 PageReferenced(page) && PageLRU(page)) {
342 activate_page(page);
343 ClearPageReferenced(page);
344 } else if (!PageReferenced(page)) {
345 SetPageReferenced(page);
346 }
347}
348
349EXPORT_SYMBOL(mark_page_accessed);
350
351void __lru_cache_add(struct page *page, enum lru_list lru)
352{
353 struct pagevec *pvec = &get_cpu_var(lru_add_pvecs)[lru];
354
355 page_cache_get(page);
356 if (!pagevec_add(pvec, page))
357 ____pagevec_lru_add(pvec, lru);
358 put_cpu_var(lru_add_pvecs);
359}
360EXPORT_SYMBOL(__lru_cache_add);
361
362/**
363 * lru_cache_add_lru - add a page to a page list
364 * @page: the page to be added to the LRU.
365 * @lru: the LRU list to which the page is added.
366 */
367void lru_cache_add_lru(struct page *page, enum lru_list lru)
368{
369 if (PageActive(page)) {
370 VM_BUG_ON(PageUnevictable(page));
371 ClearPageActive(page);
372 } else if (PageUnevictable(page)) {
373 VM_BUG_ON(PageActive(page));
374 ClearPageUnevictable(page);
375 }
376
377 VM_BUG_ON(PageLRU(page) || PageActive(page) || PageUnevictable(page));
378 __lru_cache_add(page, lru);
379}
380
381/**
382 * add_page_to_unevictable_list - add a page to the unevictable list
383 * @page: the page to be added to the unevictable list
384 *
385 * Add page directly to its zone's unevictable list. To avoid races with
386 * tasks that might be making the page evictable, through eg. munlock,
387 * munmap or exit, while it's not on the lru, we want to add the page
388 * while it's locked or otherwise "invisible" to other tasks. This is
389 * difficult to do when using the pagevec cache, so bypass that.
390 */
391void add_page_to_unevictable_list(struct page *page)
392{
393 struct zone *zone = page_zone(page);
394
395 spin_lock_irq(&zone->lru_lock);
396 SetPageUnevictable(page);
397 SetPageLRU(page);
398 add_page_to_lru_list(zone, page, LRU_UNEVICTABLE);
399 spin_unlock_irq(&zone->lru_lock);
400}
401
402/*
403 * If the page can not be invalidated, it is moved to the
404 * inactive list to speed up its reclaim. It is moved to the
405 * head of the list, rather than the tail, to give the flusher
406 * threads some time to write it out, as this is much more
407 * effective than the single-page writeout from reclaim.
408 *
409 * If the page isn't page_mapped and dirty/writeback, the page
410 * could reclaim asap using PG_reclaim.
411 *
412 * 1. active, mapped page -> none
413 * 2. active, dirty/writeback page -> inactive, head, PG_reclaim
414 * 3. inactive, mapped page -> none
415 * 4. inactive, dirty/writeback page -> inactive, head, PG_reclaim
416 * 5. inactive, clean -> inactive, tail
417 * 6. Others -> none
418 *
419 * In 4, why it moves inactive's head, the VM expects the page would
420 * be write it out by flusher threads as this is much more effective
421 * than the single-page writeout from reclaim.
422 */
423static void lru_deactivate_fn(struct page *page, void *arg)
424{
425 int lru, file;
426 bool active;
427 struct zone *zone = page_zone(page);
428
429 if (!PageLRU(page))
430 return;
431
432 if (PageUnevictable(page))
433 return;
434
435 /* Some processes are using the page */
436 if (page_mapped(page))
437 return;
438
439 active = PageActive(page);
440
441 file = page_is_file_cache(page);
442 lru = page_lru_base_type(page);
443 del_page_from_lru_list(zone, page, lru + active);
444 ClearPageActive(page);
445 ClearPageReferenced(page);
446 add_page_to_lru_list(zone, page, lru);
447
448 if (PageWriteback(page) || PageDirty(page)) {
449 /*
450 * PG_reclaim could be raced with end_page_writeback
451 * It can make readahead confusing. But race window
452 * is _really_ small and it's non-critical problem.
453 */
454 SetPageReclaim(page);
455 } else {
456 /*
457 * The page's writeback ends up during pagevec
458 * We moves tha page into tail of inactive.
459 */
460 list_move_tail(&page->lru, &zone->lru[lru].list);
461 mem_cgroup_rotate_reclaimable_page(page);
462 __count_vm_event(PGROTATED);
463 }
464
465 if (active)
466 __count_vm_event(PGDEACTIVATE);
467 update_page_reclaim_stat(zone, page, file, 0);
468}
469
470/*
471 * Drain pages out of the cpu's pagevecs.
472 * Either "cpu" is the current CPU, and preemption has already been
473 * disabled; or "cpu" is being hot-unplugged, and is already dead.
474 */
475static void drain_cpu_pagevecs(int cpu)
476{
477 struct pagevec *pvecs = per_cpu(lru_add_pvecs, cpu);
478 struct pagevec *pvec;
479 int lru;
480
481 for_each_lru(lru) {
482 pvec = &pvecs[lru - LRU_BASE];
483 if (pagevec_count(pvec))
484 ____pagevec_lru_add(pvec, lru);
485 }
486
487 pvec = &per_cpu(lru_rotate_pvecs, cpu);
488 if (pagevec_count(pvec)) {
489 unsigned long flags;
490
491 /* No harm done if a racing interrupt already did this */
492 local_irq_save(flags);
493 pagevec_move_tail(pvec);
494 local_irq_restore(flags);
495 }
496
497 pvec = &per_cpu(lru_deactivate_pvecs, cpu);
498 if (pagevec_count(pvec))
499 pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
500
501 activate_page_drain(cpu);
502}
503
504/**
505 * deactivate_page - forcefully deactivate a page
506 * @page: page to deactivate
507 *
508 * This function hints the VM that @page is a good reclaim candidate,
509 * for example if its invalidation fails due to the page being dirty
510 * or under writeback.
511 */
512void deactivate_page(struct page *page)
513{
514 /*
515 * In a workload with many unevictable page such as mprotect, unevictable
516 * page deactivation for accelerating reclaim is pointless.
517 */
518 if (PageUnevictable(page))
519 return;
520
521 if (likely(get_page_unless_zero(page))) {
522 struct pagevec *pvec = &get_cpu_var(lru_deactivate_pvecs);
523
524 if (!pagevec_add(pvec, page))
525 pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
526 put_cpu_var(lru_deactivate_pvecs);
527 }
528}
529
530void lru_add_drain(void)
531{
532 drain_cpu_pagevecs(get_cpu());
533 put_cpu();
534}
535
536static void lru_add_drain_per_cpu(struct work_struct *dummy)
537{
538 lru_add_drain();
539}
540
541/*
542 * Returns 0 for success
543 */
544int lru_add_drain_all(void)
545{
546 return schedule_on_each_cpu(lru_add_drain_per_cpu);
547}
548
549/*
550 * Batched page_cache_release(). Decrement the reference count on all the
551 * passed pages. If it fell to zero then remove the page from the LRU and
552 * free it.
553 *
554 * Avoid taking zone->lru_lock if possible, but if it is taken, retain it
555 * for the remainder of the operation.
556 *
557 * The locking in this function is against shrink_inactive_list(): we recheck
558 * the page count inside the lock to see whether shrink_inactive_list()
559 * grabbed the page via the LRU. If it did, give up: shrink_inactive_list()
560 * will free it.
561 */
562void release_pages(struct page **pages, int nr, int cold)
563{
564 int i;
565 struct pagevec pages_to_free;
566 struct zone *zone = NULL;
567 unsigned long uninitialized_var(flags);
568
569 pagevec_init(&pages_to_free, cold);
570 for (i = 0; i < nr; i++) {
571 struct page *page = pages[i];
572
573 if (unlikely(PageCompound(page))) {
574 if (zone) {
575 spin_unlock_irqrestore(&zone->lru_lock, flags);
576 zone = NULL;
577 }
578 put_compound_page(page);
579 continue;
580 }
581
582 if (!put_page_testzero(page))
583 continue;
584
585 if (PageLRU(page)) {
586 struct zone *pagezone = page_zone(page);
587
588 if (pagezone != zone) {
589 if (zone)
590 spin_unlock_irqrestore(&zone->lru_lock,
591 flags);
592 zone = pagezone;
593 spin_lock_irqsave(&zone->lru_lock, flags);
594 }
595 VM_BUG_ON(!PageLRU(page));
596 __ClearPageLRU(page);
597 del_page_from_lru(zone, page);
598 }
599
600 if (!pagevec_add(&pages_to_free, page)) {
601 if (zone) {
602 spin_unlock_irqrestore(&zone->lru_lock, flags);
603 zone = NULL;
604 }
605 __pagevec_free(&pages_to_free);
606 pagevec_reinit(&pages_to_free);
607 }
608 }
609 if (zone)
610 spin_unlock_irqrestore(&zone->lru_lock, flags);
611
612 pagevec_free(&pages_to_free);
613}
614EXPORT_SYMBOL(release_pages);
615
616/*
617 * The pages which we're about to release may be in the deferred lru-addition
618 * queues. That would prevent them from really being freed right now. That's
619 * OK from a correctness point of view but is inefficient - those pages may be
620 * cache-warm and we want to give them back to the page allocator ASAP.
621 *
622 * So __pagevec_release() will drain those queues here. __pagevec_lru_add()
623 * and __pagevec_lru_add_active() call release_pages() directly to avoid
624 * mutual recursion.
625 */
626void __pagevec_release(struct pagevec *pvec)
627{
628 lru_add_drain();
629 release_pages(pvec->pages, pagevec_count(pvec), pvec->cold);
630 pagevec_reinit(pvec);
631}
632
633EXPORT_SYMBOL(__pagevec_release);
634
635/* used by __split_huge_page_refcount() */
636void lru_add_page_tail(struct zone* zone,
637 struct page *page, struct page *page_tail)
638{
639 int active;
640 enum lru_list lru;
641 const int file = 0;
642 struct list_head *head;
643
644 VM_BUG_ON(!PageHead(page));
645 VM_BUG_ON(PageCompound(page_tail));
646 VM_BUG_ON(PageLRU(page_tail));
647 VM_BUG_ON(!spin_is_locked(&zone->lru_lock));
648
649 SetPageLRU(page_tail);
650
651 if (page_evictable(page_tail, NULL)) {
652 if (PageActive(page)) {
653 SetPageActive(page_tail);
654 active = 1;
655 lru = LRU_ACTIVE_ANON;
656 } else {
657 active = 0;
658 lru = LRU_INACTIVE_ANON;
659 }
660 update_page_reclaim_stat(zone, page_tail, file, active);
661 if (likely(PageLRU(page)))
662 head = page->lru.prev;
663 else
664 head = &zone->lru[lru].list;
665 __add_page_to_lru_list(zone, page_tail, lru, head);
666 } else {
667 SetPageUnevictable(page_tail);
668 add_page_to_lru_list(zone, page_tail, LRU_UNEVICTABLE);
669 }
670}
671
672static void ____pagevec_lru_add_fn(struct page *page, void *arg)
673{
674 enum lru_list lru = (enum lru_list)arg;
675 struct zone *zone = page_zone(page);
676 int file = is_file_lru(lru);
677 int active = is_active_lru(lru);
678
679 VM_BUG_ON(PageActive(page));
680 VM_BUG_ON(PageUnevictable(page));
681 VM_BUG_ON(PageLRU(page));
682
683 SetPageLRU(page);
684 if (active)
685 SetPageActive(page);
686 update_page_reclaim_stat(zone, page, file, active);
687 add_page_to_lru_list(zone, page, lru);
688}
689
690/*
691 * Add the passed pages to the LRU, then drop the caller's refcount
692 * on them. Reinitialises the caller's pagevec.
693 */
694void ____pagevec_lru_add(struct pagevec *pvec, enum lru_list lru)
695{
696 VM_BUG_ON(is_unevictable_lru(lru));
697
698 pagevec_lru_move_fn(pvec, ____pagevec_lru_add_fn, (void *)lru);
699}
700
701EXPORT_SYMBOL(____pagevec_lru_add);
702
703/*
704 * Try to drop buffers from the pages in a pagevec
705 */
706void pagevec_strip(struct pagevec *pvec)
707{
708 int i;
709
710 for (i = 0; i < pagevec_count(pvec); i++) {
711 struct page *page = pvec->pages[i];
712
713 if (page_has_private(page) && trylock_page(page)) {
714 if (page_has_private(page))
715 try_to_release_page(page, 0);
716 unlock_page(page);
717 }
718 }
719}
720
721/**
722 * pagevec_lookup - gang pagecache lookup
723 * @pvec: Where the resulting pages are placed
724 * @mapping: The address_space to search
725 * @start: The starting page index
726 * @nr_pages: The maximum number of pages
727 *
728 * pagevec_lookup() will search for and return a group of up to @nr_pages pages
729 * in the mapping. The pages are placed in @pvec. pagevec_lookup() takes a
730 * reference against the pages in @pvec.
731 *
732 * The search returns a group of mapping-contiguous pages with ascending
733 * indexes. There may be holes in the indices due to not-present pages.
734 *
735 * pagevec_lookup() returns the number of pages which were found.
736 */
737unsigned pagevec_lookup(struct pagevec *pvec, struct address_space *mapping,
738 pgoff_t start, unsigned nr_pages)
739{
740 pvec->nr = find_get_pages(mapping, start, nr_pages, pvec->pages);
741 return pagevec_count(pvec);
742}
743
744EXPORT_SYMBOL(pagevec_lookup);
745
746unsigned pagevec_lookup_tag(struct pagevec *pvec, struct address_space *mapping,
747 pgoff_t *index, int tag, unsigned nr_pages)
748{
749 pvec->nr = find_get_pages_tag(mapping, index, tag,
750 nr_pages, pvec->pages);
751 return pagevec_count(pvec);
752}
753
754EXPORT_SYMBOL(pagevec_lookup_tag);
755
756/*
757 * Perform any setup for the swap system
758 */
759void __init swap_setup(void)
760{
761 unsigned long megs = totalram_pages >> (20 - PAGE_SHIFT);
762
763#ifdef CONFIG_SWAP
764 bdi_init(swapper_space.backing_dev_info);
765#endif
766
767 /* Use a smaller cluster for small-memory machines */
768 if (megs < 16)
769 page_cluster = 2;
770 else
771 page_cluster = 3;
772 /*
773 * Right now other parts of the system means that we
774 * _really_ don't want to cluster much more
775 */
776}
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
40#include "internal.h"
41
42#define CREATE_TRACE_POINTS
43#include <trace/events/pagemap.h>
44
45/* How many pages do we try to swap or page in/out together? */
46int page_cluster;
47
48/* Protecting only lru_rotate.pvec which requires disabling interrupts */
49struct lru_rotate {
50 local_lock_t lock;
51 struct pagevec pvec;
52};
53static DEFINE_PER_CPU(struct lru_rotate, lru_rotate) = {
54 .lock = INIT_LOCAL_LOCK(lock),
55};
56
57/*
58 * The following struct pagevec are grouped together because they are protected
59 * by disabling preemption (and interrupts remain enabled).
60 */
61struct lru_pvecs {
62 local_lock_t lock;
63 struct pagevec lru_add;
64 struct pagevec lru_deactivate_file;
65 struct pagevec lru_deactivate;
66 struct pagevec lru_lazyfree;
67#ifdef CONFIG_SMP
68 struct pagevec activate_page;
69#endif
70};
71static DEFINE_PER_CPU(struct lru_pvecs, lru_pvecs) = {
72 .lock = INIT_LOCAL_LOCK(lock),
73};
74
75/*
76 * This path almost never happens for VM activity - pages are normally
77 * freed via pagevecs. But it gets used by networking.
78 */
79static void __page_cache_release(struct page *page)
80{
81 if (PageLRU(page)) {
82 pg_data_t *pgdat = page_pgdat(page);
83 struct lruvec *lruvec;
84 unsigned long flags;
85
86 spin_lock_irqsave(&pgdat->lru_lock, flags);
87 lruvec = mem_cgroup_page_lruvec(page, pgdat);
88 VM_BUG_ON_PAGE(!PageLRU(page), page);
89 __ClearPageLRU(page);
90 del_page_from_lru_list(page, lruvec, page_off_lru(page));
91 spin_unlock_irqrestore(&pgdat->lru_lock, flags);
92 }
93 __ClearPageWaiters(page);
94}
95
96static void __put_single_page(struct page *page)
97{
98 __page_cache_release(page);
99 mem_cgroup_uncharge(page);
100 free_unref_page(page);
101}
102
103static void __put_compound_page(struct page *page)
104{
105 /*
106 * __page_cache_release() is supposed to be called for thp, not for
107 * hugetlb. This is because hugetlb page does never have PageLRU set
108 * (it's never listed to any LRU lists) and no memcg routines should
109 * be called for hugetlb (it has a separate hugetlb_cgroup.)
110 */
111 if (!PageHuge(page))
112 __page_cache_release(page);
113 destroy_compound_page(page);
114}
115
116void __put_page(struct page *page)
117{
118 if (is_zone_device_page(page)) {
119 put_dev_pagemap(page->pgmap);
120
121 /*
122 * The page belongs to the device that created pgmap. Do
123 * not return it to page allocator.
124 */
125 return;
126 }
127
128 if (unlikely(PageCompound(page)))
129 __put_compound_page(page);
130 else
131 __put_single_page(page);
132}
133EXPORT_SYMBOL(__put_page);
134
135/**
136 * put_pages_list() - release a list of pages
137 * @pages: list of pages threaded on page->lru
138 *
139 * Release a list of pages which are strung together on page.lru. Currently
140 * used by read_cache_pages() and related error recovery code.
141 */
142void put_pages_list(struct list_head *pages)
143{
144 while (!list_empty(pages)) {
145 struct page *victim;
146
147 victim = lru_to_page(pages);
148 list_del(&victim->lru);
149 put_page(victim);
150 }
151}
152EXPORT_SYMBOL(put_pages_list);
153
154/*
155 * get_kernel_pages() - pin kernel pages in memory
156 * @kiov: An array of struct kvec structures
157 * @nr_segs: number of segments to pin
158 * @write: pinning for read/write, currently ignored
159 * @pages: array that receives pointers to the pages pinned.
160 * Should be at least nr_segs long.
161 *
162 * Returns number of pages pinned. This may be fewer than the number
163 * requested. If nr_pages is 0 or negative, returns 0. If no pages
164 * were pinned, returns -errno. Each page returned must be released
165 * with a put_page() call when it is finished with.
166 */
167int get_kernel_pages(const struct kvec *kiov, int nr_segs, int write,
168 struct page **pages)
169{
170 int seg;
171
172 for (seg = 0; seg < nr_segs; seg++) {
173 if (WARN_ON(kiov[seg].iov_len != PAGE_SIZE))
174 return seg;
175
176 pages[seg] = kmap_to_page(kiov[seg].iov_base);
177 get_page(pages[seg]);
178 }
179
180 return seg;
181}
182EXPORT_SYMBOL_GPL(get_kernel_pages);
183
184/*
185 * get_kernel_page() - pin a kernel page in memory
186 * @start: starting kernel address
187 * @write: pinning for read/write, currently ignored
188 * @pages: array that receives pointer to the page pinned.
189 * Must be at least nr_segs long.
190 *
191 * Returns 1 if page is pinned. If the page was not pinned, returns
192 * -errno. The page returned must be released with a put_page() call
193 * when it is finished with.
194 */
195int get_kernel_page(unsigned long start, int write, struct page **pages)
196{
197 const struct kvec kiov = {
198 .iov_base = (void *)start,
199 .iov_len = PAGE_SIZE
200 };
201
202 return get_kernel_pages(&kiov, 1, write, pages);
203}
204EXPORT_SYMBOL_GPL(get_kernel_page);
205
206static void pagevec_lru_move_fn(struct pagevec *pvec,
207 void (*move_fn)(struct page *page, struct lruvec *lruvec, void *arg),
208 void *arg)
209{
210 int i;
211 struct pglist_data *pgdat = NULL;
212 struct lruvec *lruvec;
213 unsigned long flags = 0;
214
215 for (i = 0; i < pagevec_count(pvec); i++) {
216 struct page *page = pvec->pages[i];
217 struct pglist_data *pagepgdat = page_pgdat(page);
218
219 if (pagepgdat != pgdat) {
220 if (pgdat)
221 spin_unlock_irqrestore(&pgdat->lru_lock, flags);
222 pgdat = pagepgdat;
223 spin_lock_irqsave(&pgdat->lru_lock, flags);
224 }
225
226 lruvec = mem_cgroup_page_lruvec(page, pgdat);
227 (*move_fn)(page, lruvec, arg);
228 }
229 if (pgdat)
230 spin_unlock_irqrestore(&pgdat->lru_lock, flags);
231 release_pages(pvec->pages, pvec->nr);
232 pagevec_reinit(pvec);
233}
234
235static void pagevec_move_tail_fn(struct page *page, struct lruvec *lruvec,
236 void *arg)
237{
238 int *pgmoved = arg;
239
240 if (PageLRU(page) && !PageUnevictable(page)) {
241 del_page_from_lru_list(page, lruvec, page_lru(page));
242 ClearPageActive(page);
243 add_page_to_lru_list_tail(page, lruvec, page_lru(page));
244 (*pgmoved) += thp_nr_pages(page);
245 }
246}
247
248/*
249 * pagevec_move_tail() must be called with IRQ disabled.
250 * Otherwise this may cause nasty races.
251 */
252static void pagevec_move_tail(struct pagevec *pvec)
253{
254 int pgmoved = 0;
255
256 pagevec_lru_move_fn(pvec, pagevec_move_tail_fn, &pgmoved);
257 __count_vm_events(PGROTATED, pgmoved);
258}
259
260/*
261 * Writeback is about to end against a page which has been marked for immediate
262 * reclaim. If it still appears to be reclaimable, move it to the tail of the
263 * inactive list.
264 */
265void rotate_reclaimable_page(struct page *page)
266{
267 if (!PageLocked(page) && !PageDirty(page) &&
268 !PageUnevictable(page) && PageLRU(page)) {
269 struct pagevec *pvec;
270 unsigned long flags;
271
272 get_page(page);
273 local_lock_irqsave(&lru_rotate.lock, flags);
274 pvec = this_cpu_ptr(&lru_rotate.pvec);
275 if (!pagevec_add(pvec, page) || PageCompound(page))
276 pagevec_move_tail(pvec);
277 local_unlock_irqrestore(&lru_rotate.lock, flags);
278 }
279}
280
281void lru_note_cost(struct lruvec *lruvec, bool file, unsigned int nr_pages)
282{
283 do {
284 unsigned long lrusize;
285
286 /* Record cost event */
287 if (file)
288 lruvec->file_cost += nr_pages;
289 else
290 lruvec->anon_cost += nr_pages;
291
292 /*
293 * Decay previous events
294 *
295 * Because workloads change over time (and to avoid
296 * overflow) we keep these statistics as a floating
297 * average, which ends up weighing recent refaults
298 * more than old ones.
299 */
300 lrusize = lruvec_page_state(lruvec, NR_INACTIVE_ANON) +
301 lruvec_page_state(lruvec, NR_ACTIVE_ANON) +
302 lruvec_page_state(lruvec, NR_INACTIVE_FILE) +
303 lruvec_page_state(lruvec, NR_ACTIVE_FILE);
304
305 if (lruvec->file_cost + lruvec->anon_cost > lrusize / 4) {
306 lruvec->file_cost /= 2;
307 lruvec->anon_cost /= 2;
308 }
309 } while ((lruvec = parent_lruvec(lruvec)));
310}
311
312void lru_note_cost_page(struct page *page)
313{
314 lru_note_cost(mem_cgroup_page_lruvec(page, page_pgdat(page)),
315 page_is_file_lru(page), thp_nr_pages(page));
316}
317
318static void __activate_page(struct page *page, struct lruvec *lruvec,
319 void *arg)
320{
321 if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
322 int lru = page_lru_base_type(page);
323 int nr_pages = thp_nr_pages(page);
324
325 del_page_from_lru_list(page, lruvec, lru);
326 SetPageActive(page);
327 lru += LRU_ACTIVE;
328 add_page_to_lru_list(page, lruvec, lru);
329 trace_mm_lru_activate(page);
330
331 __count_vm_events(PGACTIVATE, nr_pages);
332 __count_memcg_events(lruvec_memcg(lruvec), PGACTIVATE,
333 nr_pages);
334 }
335}
336
337#ifdef CONFIG_SMP
338static void activate_page_drain(int cpu)
339{
340 struct pagevec *pvec = &per_cpu(lru_pvecs.activate_page, cpu);
341
342 if (pagevec_count(pvec))
343 pagevec_lru_move_fn(pvec, __activate_page, NULL);
344}
345
346static bool need_activate_page_drain(int cpu)
347{
348 return pagevec_count(&per_cpu(lru_pvecs.activate_page, cpu)) != 0;
349}
350
351void activate_page(struct page *page)
352{
353 page = compound_head(page);
354 if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
355 struct pagevec *pvec;
356
357 local_lock(&lru_pvecs.lock);
358 pvec = this_cpu_ptr(&lru_pvecs.activate_page);
359 get_page(page);
360 if (!pagevec_add(pvec, page) || PageCompound(page))
361 pagevec_lru_move_fn(pvec, __activate_page, NULL);
362 local_unlock(&lru_pvecs.lock);
363 }
364}
365
366#else
367static inline void activate_page_drain(int cpu)
368{
369}
370
371void activate_page(struct page *page)
372{
373 pg_data_t *pgdat = page_pgdat(page);
374
375 page = compound_head(page);
376 spin_lock_irq(&pgdat->lru_lock);
377 __activate_page(page, mem_cgroup_page_lruvec(page, pgdat), NULL);
378 spin_unlock_irq(&pgdat->lru_lock);
379}
380#endif
381
382static void __lru_cache_activate_page(struct page *page)
383{
384 struct pagevec *pvec;
385 int i;
386
387 local_lock(&lru_pvecs.lock);
388 pvec = this_cpu_ptr(&lru_pvecs.lru_add);
389
390 /*
391 * Search backwards on the optimistic assumption that the page being
392 * activated has just been added to this pagevec. Note that only
393 * the local pagevec is examined as a !PageLRU page could be in the
394 * process of being released, reclaimed, migrated or on a remote
395 * pagevec that is currently being drained. Furthermore, marking
396 * a remote pagevec's page PageActive potentially hits a race where
397 * a page is marked PageActive just after it is added to the inactive
398 * list causing accounting errors and BUG_ON checks to trigger.
399 */
400 for (i = pagevec_count(pvec) - 1; i >= 0; i--) {
401 struct page *pagevec_page = pvec->pages[i];
402
403 if (pagevec_page == page) {
404 SetPageActive(page);
405 break;
406 }
407 }
408
409 local_unlock(&lru_pvecs.lock);
410}
411
412/*
413 * Mark a page as having seen activity.
414 *
415 * inactive,unreferenced -> inactive,referenced
416 * inactive,referenced -> active,unreferenced
417 * active,unreferenced -> active,referenced
418 *
419 * When a newly allocated page is not yet visible, so safe for non-atomic ops,
420 * __SetPageReferenced(page) may be substituted for mark_page_accessed(page).
421 */
422void mark_page_accessed(struct page *page)
423{
424 page = compound_head(page);
425
426 if (!PageReferenced(page)) {
427 SetPageReferenced(page);
428 } else if (PageUnevictable(page)) {
429 /*
430 * Unevictable pages are on the "LRU_UNEVICTABLE" list. But,
431 * this list is never rotated or maintained, so marking an
432 * evictable page accessed has no effect.
433 */
434 } else if (!PageActive(page)) {
435 /*
436 * If the page is on the LRU, queue it for activation via
437 * lru_pvecs.activate_page. Otherwise, assume the page is on a
438 * pagevec, mark it active and it'll be moved to the active
439 * LRU on the next drain.
440 */
441 if (PageLRU(page))
442 activate_page(page);
443 else
444 __lru_cache_activate_page(page);
445 ClearPageReferenced(page);
446 workingset_activation(page);
447 }
448 if (page_is_idle(page))
449 clear_page_idle(page);
450}
451EXPORT_SYMBOL(mark_page_accessed);
452
453/**
454 * lru_cache_add - add a page to a page list
455 * @page: the page to be added to the LRU.
456 *
457 * Queue the page for addition to the LRU via pagevec. The decision on whether
458 * to add the page to the [in]active [file|anon] list is deferred until the
459 * pagevec is drained. This gives a chance for the caller of lru_cache_add()
460 * have the page added to the active list using mark_page_accessed().
461 */
462void lru_cache_add(struct page *page)
463{
464 struct pagevec *pvec;
465
466 VM_BUG_ON_PAGE(PageActive(page) && PageUnevictable(page), page);
467 VM_BUG_ON_PAGE(PageLRU(page), page);
468
469 get_page(page);
470 local_lock(&lru_pvecs.lock);
471 pvec = this_cpu_ptr(&lru_pvecs.lru_add);
472 if (!pagevec_add(pvec, page) || PageCompound(page))
473 __pagevec_lru_add(pvec);
474 local_unlock(&lru_pvecs.lock);
475}
476EXPORT_SYMBOL(lru_cache_add);
477
478/**
479 * lru_cache_add_inactive_or_unevictable
480 * @page: the page to be added to LRU
481 * @vma: vma in which page is mapped for determining reclaimability
482 *
483 * Place @page on the inactive or unevictable LRU list, depending on its
484 * evictability. Note that if the page is not evictable, it goes
485 * directly back onto it's zone's unevictable list, it does NOT use a
486 * per cpu pagevec.
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_stat 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 void *arg)
532{
533 int lru;
534 bool active;
535 int nr_pages = thp_nr_pages(page);
536
537 if (!PageLRU(page))
538 return;
539
540 if (PageUnevictable(page))
541 return;
542
543 /* Some processes are using the page */
544 if (page_mapped(page))
545 return;
546
547 active = PageActive(page);
548 lru = page_lru_base_type(page);
549
550 del_page_from_lru_list(page, lruvec, lru + active);
551 ClearPageActive(page);
552 ClearPageReferenced(page);
553
554 if (PageWriteback(page) || PageDirty(page)) {
555 /*
556 * PG_reclaim could be raced with end_page_writeback
557 * It can make readahead confusing. But race window
558 * is _really_ small and it's non-critical problem.
559 */
560 add_page_to_lru_list(page, lruvec, lru);
561 SetPageReclaim(page);
562 } else {
563 /*
564 * The page's writeback ends up during pagevec
565 * We moves tha page into tail of inactive.
566 */
567 add_page_to_lru_list_tail(page, lruvec, lru);
568 __count_vm_events(PGROTATED, nr_pages);
569 }
570
571 if (active) {
572 __count_vm_events(PGDEACTIVATE, nr_pages);
573 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
574 nr_pages);
575 }
576}
577
578static void lru_deactivate_fn(struct page *page, struct lruvec *lruvec,
579 void *arg)
580{
581 if (PageLRU(page) && PageActive(page) && !PageUnevictable(page)) {
582 int lru = page_lru_base_type(page);
583 int nr_pages = thp_nr_pages(page);
584
585 del_page_from_lru_list(page, lruvec, lru + LRU_ACTIVE);
586 ClearPageActive(page);
587 ClearPageReferenced(page);
588 add_page_to_lru_list(page, lruvec, lru);
589
590 __count_vm_events(PGDEACTIVATE, nr_pages);
591 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
592 nr_pages);
593 }
594}
595
596static void lru_lazyfree_fn(struct page *page, struct lruvec *lruvec,
597 void *arg)
598{
599 if (PageLRU(page) && PageAnon(page) && PageSwapBacked(page) &&
600 !PageSwapCache(page) && !PageUnevictable(page)) {
601 bool active = PageActive(page);
602 int nr_pages = thp_nr_pages(page);
603
604 del_page_from_lru_list(page, lruvec,
605 LRU_INACTIVE_ANON + active);
606 ClearPageActive(page);
607 ClearPageReferenced(page);
608 /*
609 * Lazyfree pages are clean anonymous pages. They have
610 * PG_swapbacked flag cleared, to distinguish them from normal
611 * anonymous pages
612 */
613 ClearPageSwapBacked(page);
614 add_page_to_lru_list(page, lruvec, LRU_INACTIVE_FILE);
615
616 __count_vm_events(PGLAZYFREE, nr_pages);
617 __count_memcg_events(lruvec_memcg(lruvec), PGLAZYFREE,
618 nr_pages);
619 }
620}
621
622/*
623 * Drain pages out of the cpu's pagevecs.
624 * Either "cpu" is the current CPU, and preemption has already been
625 * disabled; or "cpu" is being hot-unplugged, and is already dead.
626 */
627void lru_add_drain_cpu(int cpu)
628{
629 struct pagevec *pvec = &per_cpu(lru_pvecs.lru_add, cpu);
630
631 if (pagevec_count(pvec))
632 __pagevec_lru_add(pvec);
633
634 pvec = &per_cpu(lru_rotate.pvec, cpu);
635 /* Disabling interrupts below acts as a compiler barrier. */
636 if (data_race(pagevec_count(pvec))) {
637 unsigned long flags;
638
639 /* No harm done if a racing interrupt already did this */
640 local_lock_irqsave(&lru_rotate.lock, flags);
641 pagevec_move_tail(pvec);
642 local_unlock_irqrestore(&lru_rotate.lock, flags);
643 }
644
645 pvec = &per_cpu(lru_pvecs.lru_deactivate_file, cpu);
646 if (pagevec_count(pvec))
647 pagevec_lru_move_fn(pvec, lru_deactivate_file_fn, NULL);
648
649 pvec = &per_cpu(lru_pvecs.lru_deactivate, cpu);
650 if (pagevec_count(pvec))
651 pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
652
653 pvec = &per_cpu(lru_pvecs.lru_lazyfree, cpu);
654 if (pagevec_count(pvec))
655 pagevec_lru_move_fn(pvec, lru_lazyfree_fn, NULL);
656
657 activate_page_drain(cpu);
658}
659
660/**
661 * deactivate_file_page - forcefully deactivate a file page
662 * @page: page to deactivate
663 *
664 * This function hints the VM that @page is a good reclaim candidate,
665 * for example if its invalidation fails due to the page being dirty
666 * or under writeback.
667 */
668void deactivate_file_page(struct page *page)
669{
670 /*
671 * In a workload with many unevictable page such as mprotect,
672 * unevictable page deactivation for accelerating reclaim is pointless.
673 */
674 if (PageUnevictable(page))
675 return;
676
677 if (likely(get_page_unless_zero(page))) {
678 struct pagevec *pvec;
679
680 local_lock(&lru_pvecs.lock);
681 pvec = this_cpu_ptr(&lru_pvecs.lru_deactivate_file);
682
683 if (!pagevec_add(pvec, page) || PageCompound(page))
684 pagevec_lru_move_fn(pvec, lru_deactivate_file_fn, NULL);
685 local_unlock(&lru_pvecs.lock);
686 }
687}
688
689/*
690 * deactivate_page - deactivate a page
691 * @page: page to deactivate
692 *
693 * deactivate_page() moves @page to the inactive list if @page was on the active
694 * list and was not an unevictable page. This is done to accelerate the reclaim
695 * of @page.
696 */
697void deactivate_page(struct page *page)
698{
699 if (PageLRU(page) && PageActive(page) && !PageUnevictable(page)) {
700 struct pagevec *pvec;
701
702 local_lock(&lru_pvecs.lock);
703 pvec = this_cpu_ptr(&lru_pvecs.lru_deactivate);
704 get_page(page);
705 if (!pagevec_add(pvec, page) || PageCompound(page))
706 pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
707 local_unlock(&lru_pvecs.lock);
708 }
709}
710
711/**
712 * mark_page_lazyfree - make an anon page lazyfree
713 * @page: page to deactivate
714 *
715 * mark_page_lazyfree() moves @page to the inactive file list.
716 * This is done to accelerate the reclaim of @page.
717 */
718void mark_page_lazyfree(struct page *page)
719{
720 if (PageLRU(page) && PageAnon(page) && PageSwapBacked(page) &&
721 !PageSwapCache(page) && !PageUnevictable(page)) {
722 struct pagevec *pvec;
723
724 local_lock(&lru_pvecs.lock);
725 pvec = this_cpu_ptr(&lru_pvecs.lru_lazyfree);
726 get_page(page);
727 if (!pagevec_add(pvec, page) || PageCompound(page))
728 pagevec_lru_move_fn(pvec, lru_lazyfree_fn, NULL);
729 local_unlock(&lru_pvecs.lock);
730 }
731}
732
733void lru_add_drain(void)
734{
735 local_lock(&lru_pvecs.lock);
736 lru_add_drain_cpu(smp_processor_id());
737 local_unlock(&lru_pvecs.lock);
738}
739
740void lru_add_drain_cpu_zone(struct zone *zone)
741{
742 local_lock(&lru_pvecs.lock);
743 lru_add_drain_cpu(smp_processor_id());
744 drain_local_pages(zone);
745 local_unlock(&lru_pvecs.lock);
746}
747
748#ifdef CONFIG_SMP
749
750static DEFINE_PER_CPU(struct work_struct, lru_add_drain_work);
751
752static void lru_add_drain_per_cpu(struct work_struct *dummy)
753{
754 lru_add_drain();
755}
756
757/*
758 * Doesn't need any cpu hotplug locking because we do rely on per-cpu
759 * kworkers being shut down before our page_alloc_cpu_dead callback is
760 * executed on the offlined cpu.
761 * Calling this function with cpu hotplug locks held can actually lead
762 * to obscure indirect dependencies via WQ context.
763 */
764void lru_add_drain_all(void)
765{
766 static seqcount_t seqcount = SEQCNT_ZERO(seqcount);
767 static DEFINE_MUTEX(lock);
768 static struct cpumask has_work;
769 int cpu, seq;
770
771 /*
772 * Make sure nobody triggers this path before mm_percpu_wq is fully
773 * initialized.
774 */
775 if (WARN_ON(!mm_percpu_wq))
776 return;
777
778 seq = raw_read_seqcount_latch(&seqcount);
779
780 mutex_lock(&lock);
781
782 /*
783 * Piggyback on drain started and finished while we waited for lock:
784 * all pages pended at the time of our enter were drained from vectors.
785 */
786 if (__read_seqcount_retry(&seqcount, seq))
787 goto done;
788
789 raw_write_seqcount_latch(&seqcount);
790
791 cpumask_clear(&has_work);
792
793 for_each_online_cpu(cpu) {
794 struct work_struct *work = &per_cpu(lru_add_drain_work, cpu);
795
796 if (pagevec_count(&per_cpu(lru_pvecs.lru_add, cpu)) ||
797 data_race(pagevec_count(&per_cpu(lru_rotate.pvec, cpu))) ||
798 pagevec_count(&per_cpu(lru_pvecs.lru_deactivate_file, cpu)) ||
799 pagevec_count(&per_cpu(lru_pvecs.lru_deactivate, cpu)) ||
800 pagevec_count(&per_cpu(lru_pvecs.lru_lazyfree, cpu)) ||
801 need_activate_page_drain(cpu)) {
802 INIT_WORK(work, lru_add_drain_per_cpu);
803 queue_work_on(cpu, mm_percpu_wq, work);
804 cpumask_set_cpu(cpu, &has_work);
805 }
806 }
807
808 for_each_cpu(cpu, &has_work)
809 flush_work(&per_cpu(lru_add_drain_work, cpu));
810
811done:
812 mutex_unlock(&lock);
813}
814#else
815void lru_add_drain_all(void)
816{
817 lru_add_drain();
818}
819#endif
820
821/**
822 * release_pages - batched put_page()
823 * @pages: array of pages to release
824 * @nr: number of pages
825 *
826 * Decrement the reference count on all the pages in @pages. If it
827 * fell to zero, remove the page from the LRU and free it.
828 */
829void release_pages(struct page **pages, int nr)
830{
831 int i;
832 LIST_HEAD(pages_to_free);
833 struct pglist_data *locked_pgdat = NULL;
834 struct lruvec *lruvec;
835 unsigned long flags;
836 unsigned int lock_batch;
837
838 for (i = 0; i < nr; i++) {
839 struct page *page = pages[i];
840
841 /*
842 * Make sure the IRQ-safe lock-holding time does not get
843 * excessive with a continuous string of pages from the
844 * same pgdat. The lock is held only if pgdat != NULL.
845 */
846 if (locked_pgdat && ++lock_batch == SWAP_CLUSTER_MAX) {
847 spin_unlock_irqrestore(&locked_pgdat->lru_lock, flags);
848 locked_pgdat = NULL;
849 }
850
851 if (is_huge_zero_page(page))
852 continue;
853
854 if (is_zone_device_page(page)) {
855 if (locked_pgdat) {
856 spin_unlock_irqrestore(&locked_pgdat->lru_lock,
857 flags);
858 locked_pgdat = NULL;
859 }
860 /*
861 * ZONE_DEVICE pages that return 'false' from
862 * put_devmap_managed_page() do not require special
863 * processing, and instead, expect a call to
864 * put_page_testzero().
865 */
866 if (page_is_devmap_managed(page)) {
867 put_devmap_managed_page(page);
868 continue;
869 }
870 }
871
872 page = compound_head(page);
873 if (!put_page_testzero(page))
874 continue;
875
876 if (PageCompound(page)) {
877 if (locked_pgdat) {
878 spin_unlock_irqrestore(&locked_pgdat->lru_lock, flags);
879 locked_pgdat = NULL;
880 }
881 __put_compound_page(page);
882 continue;
883 }
884
885 if (PageLRU(page)) {
886 struct pglist_data *pgdat = page_pgdat(page);
887
888 if (pgdat != locked_pgdat) {
889 if (locked_pgdat)
890 spin_unlock_irqrestore(&locked_pgdat->lru_lock,
891 flags);
892 lock_batch = 0;
893 locked_pgdat = pgdat;
894 spin_lock_irqsave(&locked_pgdat->lru_lock, flags);
895 }
896
897 lruvec = mem_cgroup_page_lruvec(page, locked_pgdat);
898 VM_BUG_ON_PAGE(!PageLRU(page), page);
899 __ClearPageLRU(page);
900 del_page_from_lru_list(page, lruvec, page_off_lru(page));
901 }
902
903 /* Clear Active bit in case of parallel mark_page_accessed */
904 __ClearPageActive(page);
905 __ClearPageWaiters(page);
906
907 list_add(&page->lru, &pages_to_free);
908 }
909 if (locked_pgdat)
910 spin_unlock_irqrestore(&locked_pgdat->lru_lock, flags);
911
912 mem_cgroup_uncharge_list(&pages_to_free);
913 free_unref_page_list(&pages_to_free);
914}
915EXPORT_SYMBOL(release_pages);
916
917/*
918 * The pages which we're about to release may be in the deferred lru-addition
919 * queues. That would prevent them from really being freed right now. That's
920 * OK from a correctness point of view but is inefficient - those pages may be
921 * cache-warm and we want to give them back to the page allocator ASAP.
922 *
923 * So __pagevec_release() will drain those queues here. __pagevec_lru_add()
924 * and __pagevec_lru_add_active() call release_pages() directly to avoid
925 * mutual recursion.
926 */
927void __pagevec_release(struct pagevec *pvec)
928{
929 if (!pvec->percpu_pvec_drained) {
930 lru_add_drain();
931 pvec->percpu_pvec_drained = true;
932 }
933 release_pages(pvec->pages, pagevec_count(pvec));
934 pagevec_reinit(pvec);
935}
936EXPORT_SYMBOL(__pagevec_release);
937
938#ifdef CONFIG_TRANSPARENT_HUGEPAGE
939/* used by __split_huge_page_refcount() */
940void lru_add_page_tail(struct page *page, struct page *page_tail,
941 struct lruvec *lruvec, struct list_head *list)
942{
943 VM_BUG_ON_PAGE(!PageHead(page), page);
944 VM_BUG_ON_PAGE(PageCompound(page_tail), page);
945 VM_BUG_ON_PAGE(PageLRU(page_tail), page);
946 lockdep_assert_held(&lruvec_pgdat(lruvec)->lru_lock);
947
948 if (!list)
949 SetPageLRU(page_tail);
950
951 if (likely(PageLRU(page)))
952 list_add_tail(&page_tail->lru, &page->lru);
953 else if (list) {
954 /* page reclaim is reclaiming a huge page */
955 get_page(page_tail);
956 list_add_tail(&page_tail->lru, list);
957 } else {
958 /*
959 * Head page has not yet been counted, as an hpage,
960 * so we must account for each subpage individually.
961 *
962 * Put page_tail on the list at the correct position
963 * so they all end up in order.
964 */
965 add_page_to_lru_list_tail(page_tail, lruvec,
966 page_lru(page_tail));
967 }
968}
969#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
970
971static void __pagevec_lru_add_fn(struct page *page, struct lruvec *lruvec,
972 void *arg)
973{
974 enum lru_list lru;
975 int was_unevictable = TestClearPageUnevictable(page);
976 int nr_pages = thp_nr_pages(page);
977
978 VM_BUG_ON_PAGE(PageLRU(page), page);
979
980 /*
981 * Page becomes evictable in two ways:
982 * 1) Within LRU lock [munlock_vma_page() and __munlock_pagevec()].
983 * 2) Before acquiring LRU lock to put the page to correct LRU and then
984 * a) do PageLRU check with lock [check_move_unevictable_pages]
985 * b) do PageLRU check before lock [clear_page_mlock]
986 *
987 * (1) & (2a) are ok as LRU lock will serialize them. For (2b), we need
988 * following strict ordering:
989 *
990 * #0: __pagevec_lru_add_fn #1: clear_page_mlock
991 *
992 * SetPageLRU() TestClearPageMlocked()
993 * smp_mb() // explicit ordering // above provides strict
994 * // ordering
995 * PageMlocked() PageLRU()
996 *
997 *
998 * if '#1' does not observe setting of PG_lru by '#0' and fails
999 * isolation, the explicit barrier will make sure that page_evictable
1000 * check will put the page in correct LRU. Without smp_mb(), SetPageLRU
1001 * can be reordered after PageMlocked check and can make '#1' to fail
1002 * the isolation of the page whose Mlocked bit is cleared (#0 is also
1003 * looking at the same page) and the evictable page will be stranded
1004 * in an unevictable LRU.
1005 */
1006 SetPageLRU(page);
1007 smp_mb__after_atomic();
1008
1009 if (page_evictable(page)) {
1010 lru = page_lru(page);
1011 if (was_unevictable)
1012 __count_vm_events(UNEVICTABLE_PGRESCUED, nr_pages);
1013 } else {
1014 lru = LRU_UNEVICTABLE;
1015 ClearPageActive(page);
1016 SetPageUnevictable(page);
1017 if (!was_unevictable)
1018 __count_vm_events(UNEVICTABLE_PGCULLED, nr_pages);
1019 }
1020
1021 add_page_to_lru_list(page, lruvec, lru);
1022 trace_mm_lru_insertion(page, lru);
1023}
1024
1025/*
1026 * Add the passed pages to the LRU, then drop the caller's refcount
1027 * on them. Reinitialises the caller's pagevec.
1028 */
1029void __pagevec_lru_add(struct pagevec *pvec)
1030{
1031 pagevec_lru_move_fn(pvec, __pagevec_lru_add_fn, NULL);
1032}
1033
1034/**
1035 * pagevec_lookup_entries - gang pagecache lookup
1036 * @pvec: Where the resulting entries are placed
1037 * @mapping: The address_space to search
1038 * @start: The starting entry index
1039 * @nr_entries: The maximum number of pages
1040 * @indices: The cache indices corresponding to the entries in @pvec
1041 *
1042 * pagevec_lookup_entries() will search for and return a group of up
1043 * to @nr_pages pages and shadow entries in the mapping. All
1044 * entries are placed in @pvec. pagevec_lookup_entries() takes a
1045 * reference against actual pages in @pvec.
1046 *
1047 * The search returns a group of mapping-contiguous entries with
1048 * ascending indexes. There may be holes in the indices due to
1049 * not-present entries.
1050 *
1051 * Only one subpage of a Transparent Huge Page is returned in one call:
1052 * allowing truncate_inode_pages_range() to evict the whole THP without
1053 * cycling through a pagevec of extra references.
1054 *
1055 * pagevec_lookup_entries() returns the number of entries which were
1056 * found.
1057 */
1058unsigned pagevec_lookup_entries(struct pagevec *pvec,
1059 struct address_space *mapping,
1060 pgoff_t start, unsigned nr_entries,
1061 pgoff_t *indices)
1062{
1063 pvec->nr = find_get_entries(mapping, start, nr_entries,
1064 pvec->pages, indices);
1065 return pagevec_count(pvec);
1066}
1067
1068/**
1069 * pagevec_remove_exceptionals - pagevec exceptionals pruning
1070 * @pvec: The pagevec to prune
1071 *
1072 * pagevec_lookup_entries() fills both pages and exceptional radix
1073 * tree entries into the pagevec. This function prunes all
1074 * exceptionals from @pvec without leaving holes, so that it can be
1075 * passed on to page-only pagevec operations.
1076 */
1077void pagevec_remove_exceptionals(struct pagevec *pvec)
1078{
1079 int i, j;
1080
1081 for (i = 0, j = 0; i < pagevec_count(pvec); i++) {
1082 struct page *page = pvec->pages[i];
1083 if (!xa_is_value(page))
1084 pvec->pages[j++] = page;
1085 }
1086 pvec->nr = j;
1087}
1088
1089/**
1090 * pagevec_lookup_range - gang pagecache lookup
1091 * @pvec: Where the resulting pages are placed
1092 * @mapping: The address_space to search
1093 * @start: The starting page index
1094 * @end: The final page index
1095 *
1096 * pagevec_lookup_range() will search for & return a group of up to PAGEVEC_SIZE
1097 * pages in the mapping starting from index @start and upto index @end
1098 * (inclusive). The pages are placed in @pvec. pagevec_lookup() takes a
1099 * reference against the pages in @pvec.
1100 *
1101 * The search returns a group of mapping-contiguous pages with ascending
1102 * indexes. There may be holes in the indices due to not-present pages. We
1103 * also update @start to index the next page for the traversal.
1104 *
1105 * pagevec_lookup_range() returns the number of pages which were found. If this
1106 * number is smaller than PAGEVEC_SIZE, the end of specified range has been
1107 * reached.
1108 */
1109unsigned pagevec_lookup_range(struct pagevec *pvec,
1110 struct address_space *mapping, pgoff_t *start, pgoff_t end)
1111{
1112 pvec->nr = find_get_pages_range(mapping, start, end, PAGEVEC_SIZE,
1113 pvec->pages);
1114 return pagevec_count(pvec);
1115}
1116EXPORT_SYMBOL(pagevec_lookup_range);
1117
1118unsigned pagevec_lookup_range_tag(struct pagevec *pvec,
1119 struct address_space *mapping, pgoff_t *index, pgoff_t end,
1120 xa_mark_t tag)
1121{
1122 pvec->nr = find_get_pages_range_tag(mapping, index, end, tag,
1123 PAGEVEC_SIZE, pvec->pages);
1124 return pagevec_count(pvec);
1125}
1126EXPORT_SYMBOL(pagevec_lookup_range_tag);
1127
1128unsigned pagevec_lookup_range_nr_tag(struct pagevec *pvec,
1129 struct address_space *mapping, pgoff_t *index, pgoff_t end,
1130 xa_mark_t tag, unsigned max_pages)
1131{
1132 pvec->nr = find_get_pages_range_tag(mapping, index, end, tag,
1133 min_t(unsigned int, max_pages, PAGEVEC_SIZE), pvec->pages);
1134 return pagevec_count(pvec);
1135}
1136EXPORT_SYMBOL(pagevec_lookup_range_nr_tag);
1137/*
1138 * Perform any setup for the swap system
1139 */
1140void __init swap_setup(void)
1141{
1142 unsigned long megs = totalram_pages() >> (20 - PAGE_SHIFT);
1143
1144 /* Use a smaller cluster for small-memory machines */
1145 if (megs < 16)
1146 page_cluster = 2;
1147 else
1148 page_cluster = 3;
1149 /*
1150 * Right now other parts of the system means that we
1151 * _really_ don't want to cluster much more
1152 */
1153}
1154
1155#ifdef CONFIG_DEV_PAGEMAP_OPS
1156void put_devmap_managed_page(struct page *page)
1157{
1158 int count;
1159
1160 if (WARN_ON_ONCE(!page_is_devmap_managed(page)))
1161 return;
1162
1163 count = page_ref_dec_return(page);
1164
1165 /*
1166 * devmap page refcounts are 1-based, rather than 0-based: if
1167 * refcount is 1, then the page is free and the refcount is
1168 * stable because nobody holds a reference on the page.
1169 */
1170 if (count == 1)
1171 free_devmap_managed_page(page);
1172 else if (!count)
1173 __put_page(page);
1174}
1175EXPORT_SYMBOL(put_devmap_managed_page);
1176#endif