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#include <linux/buffer_head.h>
40
41#include "internal.h"
42
43#define CREATE_TRACE_POINTS
44#include <trace/events/pagemap.h>
45
46/* How many pages do we try to swap or page in/out together? As a power of 2 */
47int page_cluster;
48const int page_cluster_max = 31;
49
50/* Protecting only lru_rotate.fbatch which requires disabling interrupts */
51struct lru_rotate {
52 local_lock_t lock;
53 struct folio_batch fbatch;
54};
55static DEFINE_PER_CPU(struct lru_rotate, lru_rotate) = {
56 .lock = INIT_LOCAL_LOCK(lock),
57};
58
59/*
60 * The following folio batches are grouped together because they are protected
61 * by disabling preemption (and interrupts remain enabled).
62 */
63struct cpu_fbatches {
64 local_lock_t lock;
65 struct folio_batch lru_add;
66 struct folio_batch lru_deactivate_file;
67 struct folio_batch lru_deactivate;
68 struct folio_batch lru_lazyfree;
69#ifdef CONFIG_SMP
70 struct folio_batch activate;
71#endif
72};
73static DEFINE_PER_CPU(struct cpu_fbatches, cpu_fbatches) = {
74 .lock = INIT_LOCAL_LOCK(lock),
75};
76
77/*
78 * This path almost never happens for VM activity - pages are normally freed
79 * via pagevecs. But it gets used by networking - and for compound pages.
80 */
81static void __page_cache_release(struct folio *folio)
82{
83 if (folio_test_lru(folio)) {
84 struct lruvec *lruvec;
85 unsigned long flags;
86
87 lruvec = folio_lruvec_lock_irqsave(folio, &flags);
88 lruvec_del_folio(lruvec, folio);
89 __folio_clear_lru_flags(folio);
90 unlock_page_lruvec_irqrestore(lruvec, flags);
91 }
92 /* See comment on folio_test_mlocked in release_pages() */
93 if (unlikely(folio_test_mlocked(folio))) {
94 long nr_pages = folio_nr_pages(folio);
95
96 __folio_clear_mlocked(folio);
97 zone_stat_mod_folio(folio, NR_MLOCK, -nr_pages);
98 count_vm_events(UNEVICTABLE_PGCLEARED, nr_pages);
99 }
100}
101
102static void __folio_put_small(struct folio *folio)
103{
104 __page_cache_release(folio);
105 mem_cgroup_uncharge(folio);
106 free_unref_page(&folio->page, 0);
107}
108
109static void __folio_put_large(struct folio *folio)
110{
111 /*
112 * __page_cache_release() is supposed to be called for thp, not for
113 * hugetlb. This is because hugetlb page does never have PageLRU set
114 * (it's never listed to any LRU lists) and no memcg routines should
115 * be called for hugetlb (it has a separate hugetlb_cgroup.)
116 */
117 if (!folio_test_hugetlb(folio))
118 __page_cache_release(folio);
119 destroy_large_folio(folio);
120}
121
122void __folio_put(struct folio *folio)
123{
124 if (unlikely(folio_is_zone_device(folio)))
125 free_zone_device_page(&folio->page);
126 else if (unlikely(folio_test_large(folio)))
127 __folio_put_large(folio);
128 else
129 __folio_put_small(folio);
130}
131EXPORT_SYMBOL(__folio_put);
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.
138 */
139void put_pages_list(struct list_head *pages)
140{
141 struct folio *folio, *next;
142
143 list_for_each_entry_safe(folio, next, pages, lru) {
144 if (!folio_put_testzero(folio)) {
145 list_del(&folio->lru);
146 continue;
147 }
148 if (folio_test_large(folio)) {
149 list_del(&folio->lru);
150 __folio_put_large(folio);
151 continue;
152 }
153 /* LRU flag must be clear because it's passed using the lru */
154 }
155
156 free_unref_page_list(pages);
157 INIT_LIST_HEAD(pages);
158}
159EXPORT_SYMBOL(put_pages_list);
160
161/*
162 * get_kernel_pages() - pin kernel pages in memory
163 * @kiov: An array of struct kvec structures
164 * @nr_segs: number of segments to pin
165 * @write: pinning for read/write, currently ignored
166 * @pages: array that receives pointers to the pages pinned.
167 * Should be at least nr_segs long.
168 *
169 * Returns number of pages pinned. This may be fewer than the number requested.
170 * If nr_segs is 0 or negative, returns 0. If no pages were pinned, returns 0.
171 * Each page returned must be released with a put_page() call when it is
172 * finished with.
173 */
174int get_kernel_pages(const struct kvec *kiov, int nr_segs, int write,
175 struct page **pages)
176{
177 int seg;
178
179 for (seg = 0; seg < nr_segs; seg++) {
180 if (WARN_ON(kiov[seg].iov_len != PAGE_SIZE))
181 return seg;
182
183 pages[seg] = kmap_to_page(kiov[seg].iov_base);
184 get_page(pages[seg]);
185 }
186
187 return seg;
188}
189EXPORT_SYMBOL_GPL(get_kernel_pages);
190
191typedef void (*move_fn_t)(struct lruvec *lruvec, struct folio *folio);
192
193static void lru_add_fn(struct lruvec *lruvec, struct folio *folio)
194{
195 int was_unevictable = folio_test_clear_unevictable(folio);
196 long nr_pages = folio_nr_pages(folio);
197
198 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
199
200 /*
201 * Is an smp_mb__after_atomic() still required here, before
202 * folio_evictable() tests the mlocked flag, to rule out the possibility
203 * of stranding an evictable folio on an unevictable LRU? I think
204 * not, because __munlock_page() only clears the mlocked flag
205 * while the LRU lock is held.
206 *
207 * (That is not true of __page_cache_release(), and not necessarily
208 * true of release_pages(): but those only clear the mlocked flag after
209 * folio_put_testzero() has excluded any other users of the folio.)
210 */
211 if (folio_evictable(folio)) {
212 if (was_unevictable)
213 __count_vm_events(UNEVICTABLE_PGRESCUED, nr_pages);
214 } else {
215 folio_clear_active(folio);
216 folio_set_unevictable(folio);
217 /*
218 * folio->mlock_count = !!folio_test_mlocked(folio)?
219 * But that leaves __mlock_page() in doubt whether another
220 * actor has already counted the mlock or not. Err on the
221 * safe side, underestimate, let page reclaim fix it, rather
222 * than leaving a page on the unevictable LRU indefinitely.
223 */
224 folio->mlock_count = 0;
225 if (!was_unevictable)
226 __count_vm_events(UNEVICTABLE_PGCULLED, nr_pages);
227 }
228
229 lruvec_add_folio(lruvec, folio);
230 trace_mm_lru_insertion(folio);
231}
232
233static void folio_batch_move_lru(struct folio_batch *fbatch, move_fn_t move_fn)
234{
235 int i;
236 struct lruvec *lruvec = NULL;
237 unsigned long flags = 0;
238
239 for (i = 0; i < folio_batch_count(fbatch); i++) {
240 struct folio *folio = fbatch->folios[i];
241
242 /* block memcg migration while the folio moves between lru */
243 if (move_fn != lru_add_fn && !folio_test_clear_lru(folio))
244 continue;
245
246 lruvec = folio_lruvec_relock_irqsave(folio, lruvec, &flags);
247 move_fn(lruvec, folio);
248
249 folio_set_lru(folio);
250 }
251
252 if (lruvec)
253 unlock_page_lruvec_irqrestore(lruvec, flags);
254 folios_put(fbatch->folios, folio_batch_count(fbatch));
255 folio_batch_init(fbatch);
256}
257
258static void folio_batch_add_and_move(struct folio_batch *fbatch,
259 struct folio *folio, move_fn_t move_fn)
260{
261 if (folio_batch_add(fbatch, folio) && !folio_test_large(folio) &&
262 !lru_cache_disabled())
263 return;
264 folio_batch_move_lru(fbatch, move_fn);
265}
266
267static void lru_move_tail_fn(struct lruvec *lruvec, struct folio *folio)
268{
269 if (!folio_test_unevictable(folio)) {
270 lruvec_del_folio(lruvec, folio);
271 folio_clear_active(folio);
272 lruvec_add_folio_tail(lruvec, folio);
273 __count_vm_events(PGROTATED, folio_nr_pages(folio));
274 }
275}
276
277/*
278 * Writeback is about to end against a folio which has been marked for
279 * immediate reclaim. If it still appears to be reclaimable, move it
280 * to the tail of the inactive list.
281 *
282 * folio_rotate_reclaimable() must disable IRQs, to prevent nasty races.
283 */
284void folio_rotate_reclaimable(struct folio *folio)
285{
286 if (!folio_test_locked(folio) && !folio_test_dirty(folio) &&
287 !folio_test_unevictable(folio) && folio_test_lru(folio)) {
288 struct folio_batch *fbatch;
289 unsigned long flags;
290
291 folio_get(folio);
292 local_lock_irqsave(&lru_rotate.lock, flags);
293 fbatch = this_cpu_ptr(&lru_rotate.fbatch);
294 folio_batch_add_and_move(fbatch, folio, lru_move_tail_fn);
295 local_unlock_irqrestore(&lru_rotate.lock, flags);
296 }
297}
298
299void lru_note_cost(struct lruvec *lruvec, bool file,
300 unsigned int nr_io, unsigned int nr_rotated)
301{
302 unsigned long cost;
303
304 /*
305 * Reflect the relative cost of incurring IO and spending CPU
306 * time on rotations. This doesn't attempt to make a precise
307 * comparison, it just says: if reloads are about comparable
308 * between the LRU lists, or rotations are overwhelmingly
309 * different between them, adjust scan balance for CPU work.
310 */
311 cost = nr_io * SWAP_CLUSTER_MAX + nr_rotated;
312
313 do {
314 unsigned long lrusize;
315
316 /*
317 * Hold lruvec->lru_lock is safe here, since
318 * 1) The pinned lruvec in reclaim, or
319 * 2) From a pre-LRU page during refault (which also holds the
320 * rcu lock, so would be safe even if the page was on the LRU
321 * and could move simultaneously to a new lruvec).
322 */
323 spin_lock_irq(&lruvec->lru_lock);
324 /* Record cost event */
325 if (file)
326 lruvec->file_cost += cost;
327 else
328 lruvec->anon_cost += cost;
329
330 /*
331 * Decay previous events
332 *
333 * Because workloads change over time (and to avoid
334 * overflow) we keep these statistics as a floating
335 * average, which ends up weighing recent refaults
336 * more than old ones.
337 */
338 lrusize = lruvec_page_state(lruvec, NR_INACTIVE_ANON) +
339 lruvec_page_state(lruvec, NR_ACTIVE_ANON) +
340 lruvec_page_state(lruvec, NR_INACTIVE_FILE) +
341 lruvec_page_state(lruvec, NR_ACTIVE_FILE);
342
343 if (lruvec->file_cost + lruvec->anon_cost > lrusize / 4) {
344 lruvec->file_cost /= 2;
345 lruvec->anon_cost /= 2;
346 }
347 spin_unlock_irq(&lruvec->lru_lock);
348 } while ((lruvec = parent_lruvec(lruvec)));
349}
350
351void lru_note_cost_refault(struct folio *folio)
352{
353 lru_note_cost(folio_lruvec(folio), folio_is_file_lru(folio),
354 folio_nr_pages(folio), 0);
355}
356
357static void folio_activate_fn(struct lruvec *lruvec, struct folio *folio)
358{
359 if (!folio_test_active(folio) && !folio_test_unevictable(folio)) {
360 long nr_pages = folio_nr_pages(folio);
361
362 lruvec_del_folio(lruvec, folio);
363 folio_set_active(folio);
364 lruvec_add_folio(lruvec, folio);
365 trace_mm_lru_activate(folio);
366
367 __count_vm_events(PGACTIVATE, nr_pages);
368 __count_memcg_events(lruvec_memcg(lruvec), PGACTIVATE,
369 nr_pages);
370 }
371}
372
373#ifdef CONFIG_SMP
374static void folio_activate_drain(int cpu)
375{
376 struct folio_batch *fbatch = &per_cpu(cpu_fbatches.activate, cpu);
377
378 if (folio_batch_count(fbatch))
379 folio_batch_move_lru(fbatch, folio_activate_fn);
380}
381
382void folio_activate(struct folio *folio)
383{
384 if (folio_test_lru(folio) && !folio_test_active(folio) &&
385 !folio_test_unevictable(folio)) {
386 struct folio_batch *fbatch;
387
388 folio_get(folio);
389 local_lock(&cpu_fbatches.lock);
390 fbatch = this_cpu_ptr(&cpu_fbatches.activate);
391 folio_batch_add_and_move(fbatch, folio, folio_activate_fn);
392 local_unlock(&cpu_fbatches.lock);
393 }
394}
395
396#else
397static inline void folio_activate_drain(int cpu)
398{
399}
400
401void folio_activate(struct folio *folio)
402{
403 struct lruvec *lruvec;
404
405 if (folio_test_clear_lru(folio)) {
406 lruvec = folio_lruvec_lock_irq(folio);
407 folio_activate_fn(lruvec, folio);
408 unlock_page_lruvec_irq(lruvec);
409 folio_set_lru(folio);
410 }
411}
412#endif
413
414static void __lru_cache_activate_folio(struct folio *folio)
415{
416 struct folio_batch *fbatch;
417 int i;
418
419 local_lock(&cpu_fbatches.lock);
420 fbatch = this_cpu_ptr(&cpu_fbatches.lru_add);
421
422 /*
423 * Search backwards on the optimistic assumption that the folio being
424 * activated has just been added to this batch. Note that only
425 * the local batch is examined as a !LRU folio could be in the
426 * process of being released, reclaimed, migrated or on a remote
427 * batch that is currently being drained. Furthermore, marking
428 * a remote batch's folio active potentially hits a race where
429 * a folio is marked active just after it is added to the inactive
430 * list causing accounting errors and BUG_ON checks to trigger.
431 */
432 for (i = folio_batch_count(fbatch) - 1; i >= 0; i--) {
433 struct folio *batch_folio = fbatch->folios[i];
434
435 if (batch_folio == folio) {
436 folio_set_active(folio);
437 break;
438 }
439 }
440
441 local_unlock(&cpu_fbatches.lock);
442}
443
444#ifdef CONFIG_LRU_GEN
445static void folio_inc_refs(struct folio *folio)
446{
447 unsigned long new_flags, old_flags = READ_ONCE(folio->flags);
448
449 if (folio_test_unevictable(folio))
450 return;
451
452 if (!folio_test_referenced(folio)) {
453 folio_set_referenced(folio);
454 return;
455 }
456
457 if (!folio_test_workingset(folio)) {
458 folio_set_workingset(folio);
459 return;
460 }
461
462 /* see the comment on MAX_NR_TIERS */
463 do {
464 new_flags = old_flags & LRU_REFS_MASK;
465 if (new_flags == LRU_REFS_MASK)
466 break;
467
468 new_flags += BIT(LRU_REFS_PGOFF);
469 new_flags |= old_flags & ~LRU_REFS_MASK;
470 } while (!try_cmpxchg(&folio->flags, &old_flags, new_flags));
471}
472#else
473static void folio_inc_refs(struct folio *folio)
474{
475}
476#endif /* CONFIG_LRU_GEN */
477
478/*
479 * Mark a page as having seen activity.
480 *
481 * inactive,unreferenced -> inactive,referenced
482 * inactive,referenced -> active,unreferenced
483 * active,unreferenced -> active,referenced
484 *
485 * When a newly allocated page is not yet visible, so safe for non-atomic ops,
486 * __SetPageReferenced(page) may be substituted for mark_page_accessed(page).
487 */
488void folio_mark_accessed(struct folio *folio)
489{
490 if (lru_gen_enabled()) {
491 folio_inc_refs(folio);
492 return;
493 }
494
495 if (!folio_test_referenced(folio)) {
496 folio_set_referenced(folio);
497 } else if (folio_test_unevictable(folio)) {
498 /*
499 * Unevictable pages are on the "LRU_UNEVICTABLE" list. But,
500 * this list is never rotated or maintained, so marking an
501 * unevictable page accessed has no effect.
502 */
503 } else if (!folio_test_active(folio)) {
504 /*
505 * If the folio is on the LRU, queue it for activation via
506 * cpu_fbatches.activate. Otherwise, assume the folio is in a
507 * folio_batch, mark it active and it'll be moved to the active
508 * LRU on the next drain.
509 */
510 if (folio_test_lru(folio))
511 folio_activate(folio);
512 else
513 __lru_cache_activate_folio(folio);
514 folio_clear_referenced(folio);
515 workingset_activation(folio);
516 }
517 if (folio_test_idle(folio))
518 folio_clear_idle(folio);
519}
520EXPORT_SYMBOL(folio_mark_accessed);
521
522/**
523 * folio_add_lru - Add a folio to an LRU list.
524 * @folio: The folio to be added to the LRU.
525 *
526 * Queue the folio for addition to the LRU. The decision on whether
527 * to add the page to the [in]active [file|anon] list is deferred until the
528 * folio_batch is drained. This gives a chance for the caller of folio_add_lru()
529 * have the folio added to the active list using folio_mark_accessed().
530 */
531void folio_add_lru(struct folio *folio)
532{
533 struct folio_batch *fbatch;
534
535 VM_BUG_ON_FOLIO(folio_test_active(folio) &&
536 folio_test_unevictable(folio), folio);
537 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
538
539 /* see the comment in lru_gen_add_folio() */
540 if (lru_gen_enabled() && !folio_test_unevictable(folio) &&
541 lru_gen_in_fault() && !(current->flags & PF_MEMALLOC))
542 folio_set_active(folio);
543
544 folio_get(folio);
545 local_lock(&cpu_fbatches.lock);
546 fbatch = this_cpu_ptr(&cpu_fbatches.lru_add);
547 folio_batch_add_and_move(fbatch, folio, lru_add_fn);
548 local_unlock(&cpu_fbatches.lock);
549}
550EXPORT_SYMBOL(folio_add_lru);
551
552/**
553 * folio_add_lru_vma() - Add a folio to the appropate LRU list for this VMA.
554 * @folio: The folio to be added to the LRU.
555 * @vma: VMA in which the folio is mapped.
556 *
557 * If the VMA is mlocked, @folio is added to the unevictable list.
558 * Otherwise, it is treated the same way as folio_add_lru().
559 */
560void folio_add_lru_vma(struct folio *folio, struct vm_area_struct *vma)
561{
562 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
563
564 if (unlikely((vma->vm_flags & (VM_LOCKED | VM_SPECIAL)) == VM_LOCKED))
565 mlock_new_page(&folio->page);
566 else
567 folio_add_lru(folio);
568}
569
570/*
571 * If the folio cannot be invalidated, it is moved to the
572 * inactive list to speed up its reclaim. It is moved to the
573 * head of the list, rather than the tail, to give the flusher
574 * threads some time to write it out, as this is much more
575 * effective than the single-page writeout from reclaim.
576 *
577 * If the folio isn't mapped and dirty/writeback, the folio
578 * could be reclaimed asap using the reclaim flag.
579 *
580 * 1. active, mapped folio -> none
581 * 2. active, dirty/writeback folio -> inactive, head, reclaim
582 * 3. inactive, mapped folio -> none
583 * 4. inactive, dirty/writeback folio -> inactive, head, reclaim
584 * 5. inactive, clean -> inactive, tail
585 * 6. Others -> none
586 *
587 * In 4, it moves to the head of the inactive list so the folio is
588 * written out by flusher threads as this is much more efficient
589 * than the single-page writeout from reclaim.
590 */
591static void lru_deactivate_file_fn(struct lruvec *lruvec, struct folio *folio)
592{
593 bool active = folio_test_active(folio);
594 long nr_pages = folio_nr_pages(folio);
595
596 if (folio_test_unevictable(folio))
597 return;
598
599 /* Some processes are using the folio */
600 if (folio_mapped(folio))
601 return;
602
603 lruvec_del_folio(lruvec, folio);
604 folio_clear_active(folio);
605 folio_clear_referenced(folio);
606
607 if (folio_test_writeback(folio) || folio_test_dirty(folio)) {
608 /*
609 * Setting the reclaim flag could race with
610 * folio_end_writeback() and confuse readahead. But the
611 * race window is _really_ small and it's not a critical
612 * problem.
613 */
614 lruvec_add_folio(lruvec, folio);
615 folio_set_reclaim(folio);
616 } else {
617 /*
618 * The folio's writeback ended while it was in the batch.
619 * We move that folio to the tail of the inactive list.
620 */
621 lruvec_add_folio_tail(lruvec, folio);
622 __count_vm_events(PGROTATED, nr_pages);
623 }
624
625 if (active) {
626 __count_vm_events(PGDEACTIVATE, nr_pages);
627 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
628 nr_pages);
629 }
630}
631
632static void lru_deactivate_fn(struct lruvec *lruvec, struct folio *folio)
633{
634 if (!folio_test_unevictable(folio) && (folio_test_active(folio) || lru_gen_enabled())) {
635 long nr_pages = folio_nr_pages(folio);
636
637 lruvec_del_folio(lruvec, folio);
638 folio_clear_active(folio);
639 folio_clear_referenced(folio);
640 lruvec_add_folio(lruvec, folio);
641
642 __count_vm_events(PGDEACTIVATE, nr_pages);
643 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
644 nr_pages);
645 }
646}
647
648static void lru_lazyfree_fn(struct lruvec *lruvec, struct folio *folio)
649{
650 if (folio_test_anon(folio) && folio_test_swapbacked(folio) &&
651 !folio_test_swapcache(folio) && !folio_test_unevictable(folio)) {
652 long nr_pages = folio_nr_pages(folio);
653
654 lruvec_del_folio(lruvec, folio);
655 folio_clear_active(folio);
656 folio_clear_referenced(folio);
657 /*
658 * Lazyfree folios are clean anonymous folios. They have
659 * the swapbacked flag cleared, to distinguish them from normal
660 * anonymous folios
661 */
662 folio_clear_swapbacked(folio);
663 lruvec_add_folio(lruvec, folio);
664
665 __count_vm_events(PGLAZYFREE, nr_pages);
666 __count_memcg_events(lruvec_memcg(lruvec), PGLAZYFREE,
667 nr_pages);
668 }
669}
670
671/*
672 * Drain pages out of the cpu's folio_batch.
673 * Either "cpu" is the current CPU, and preemption has already been
674 * disabled; or "cpu" is being hot-unplugged, and is already dead.
675 */
676void lru_add_drain_cpu(int cpu)
677{
678 struct cpu_fbatches *fbatches = &per_cpu(cpu_fbatches, cpu);
679 struct folio_batch *fbatch = &fbatches->lru_add;
680
681 if (folio_batch_count(fbatch))
682 folio_batch_move_lru(fbatch, lru_add_fn);
683
684 fbatch = &per_cpu(lru_rotate.fbatch, cpu);
685 /* Disabling interrupts below acts as a compiler barrier. */
686 if (data_race(folio_batch_count(fbatch))) {
687 unsigned long flags;
688
689 /* No harm done if a racing interrupt already did this */
690 local_lock_irqsave(&lru_rotate.lock, flags);
691 folio_batch_move_lru(fbatch, lru_move_tail_fn);
692 local_unlock_irqrestore(&lru_rotate.lock, flags);
693 }
694
695 fbatch = &fbatches->lru_deactivate_file;
696 if (folio_batch_count(fbatch))
697 folio_batch_move_lru(fbatch, lru_deactivate_file_fn);
698
699 fbatch = &fbatches->lru_deactivate;
700 if (folio_batch_count(fbatch))
701 folio_batch_move_lru(fbatch, lru_deactivate_fn);
702
703 fbatch = &fbatches->lru_lazyfree;
704 if (folio_batch_count(fbatch))
705 folio_batch_move_lru(fbatch, lru_lazyfree_fn);
706
707 folio_activate_drain(cpu);
708}
709
710/**
711 * deactivate_file_folio() - Deactivate a file folio.
712 * @folio: Folio to deactivate.
713 *
714 * This function hints to the VM that @folio is a good reclaim candidate,
715 * for example if its invalidation fails due to the folio being dirty
716 * or under writeback.
717 *
718 * Context: Caller holds a reference on the folio.
719 */
720void deactivate_file_folio(struct folio *folio)
721{
722 struct folio_batch *fbatch;
723
724 /* Deactivating an unevictable folio will not accelerate reclaim */
725 if (folio_test_unevictable(folio))
726 return;
727
728 folio_get(folio);
729 local_lock(&cpu_fbatches.lock);
730 fbatch = this_cpu_ptr(&cpu_fbatches.lru_deactivate_file);
731 folio_batch_add_and_move(fbatch, folio, lru_deactivate_file_fn);
732 local_unlock(&cpu_fbatches.lock);
733}
734
735/*
736 * deactivate_page - deactivate a page
737 * @page: page to deactivate
738 *
739 * deactivate_page() moves @page to the inactive list if @page was on the active
740 * list and was not an unevictable page. This is done to accelerate the reclaim
741 * of @page.
742 */
743void deactivate_page(struct page *page)
744{
745 struct folio *folio = page_folio(page);
746
747 if (folio_test_lru(folio) && !folio_test_unevictable(folio) &&
748 (folio_test_active(folio) || lru_gen_enabled())) {
749 struct folio_batch *fbatch;
750
751 folio_get(folio);
752 local_lock(&cpu_fbatches.lock);
753 fbatch = this_cpu_ptr(&cpu_fbatches.lru_deactivate);
754 folio_batch_add_and_move(fbatch, folio, lru_deactivate_fn);
755 local_unlock(&cpu_fbatches.lock);
756 }
757}
758
759/**
760 * mark_page_lazyfree - make an anon page lazyfree
761 * @page: page to deactivate
762 *
763 * mark_page_lazyfree() moves @page to the inactive file list.
764 * This is done to accelerate the reclaim of @page.
765 */
766void mark_page_lazyfree(struct page *page)
767{
768 struct folio *folio = page_folio(page);
769
770 if (folio_test_lru(folio) && folio_test_anon(folio) &&
771 folio_test_swapbacked(folio) && !folio_test_swapcache(folio) &&
772 !folio_test_unevictable(folio)) {
773 struct folio_batch *fbatch;
774
775 folio_get(folio);
776 local_lock(&cpu_fbatches.lock);
777 fbatch = this_cpu_ptr(&cpu_fbatches.lru_lazyfree);
778 folio_batch_add_and_move(fbatch, folio, lru_lazyfree_fn);
779 local_unlock(&cpu_fbatches.lock);
780 }
781}
782
783void lru_add_drain(void)
784{
785 local_lock(&cpu_fbatches.lock);
786 lru_add_drain_cpu(smp_processor_id());
787 local_unlock(&cpu_fbatches.lock);
788 mlock_page_drain_local();
789}
790
791/*
792 * It's called from per-cpu workqueue context in SMP case so
793 * lru_add_drain_cpu and invalidate_bh_lrus_cpu should run on
794 * the same cpu. It shouldn't be a problem in !SMP case since
795 * the core is only one and the locks will disable preemption.
796 */
797static void lru_add_and_bh_lrus_drain(void)
798{
799 local_lock(&cpu_fbatches.lock);
800 lru_add_drain_cpu(smp_processor_id());
801 local_unlock(&cpu_fbatches.lock);
802 invalidate_bh_lrus_cpu();
803 mlock_page_drain_local();
804}
805
806void lru_add_drain_cpu_zone(struct zone *zone)
807{
808 local_lock(&cpu_fbatches.lock);
809 lru_add_drain_cpu(smp_processor_id());
810 drain_local_pages(zone);
811 local_unlock(&cpu_fbatches.lock);
812 mlock_page_drain_local();
813}
814
815#ifdef CONFIG_SMP
816
817static DEFINE_PER_CPU(struct work_struct, lru_add_drain_work);
818
819static void lru_add_drain_per_cpu(struct work_struct *dummy)
820{
821 lru_add_and_bh_lrus_drain();
822}
823
824static bool cpu_needs_drain(unsigned int cpu)
825{
826 struct cpu_fbatches *fbatches = &per_cpu(cpu_fbatches, cpu);
827
828 /* Check these in order of likelihood that they're not zero */
829 return folio_batch_count(&fbatches->lru_add) ||
830 data_race(folio_batch_count(&per_cpu(lru_rotate.fbatch, cpu))) ||
831 folio_batch_count(&fbatches->lru_deactivate_file) ||
832 folio_batch_count(&fbatches->lru_deactivate) ||
833 folio_batch_count(&fbatches->lru_lazyfree) ||
834 folio_batch_count(&fbatches->activate) ||
835 need_mlock_page_drain(cpu) ||
836 has_bh_in_lru(cpu, NULL);
837}
838
839/*
840 * Doesn't need any cpu hotplug locking because we do rely on per-cpu
841 * kworkers being shut down before our page_alloc_cpu_dead callback is
842 * executed on the offlined cpu.
843 * Calling this function with cpu hotplug locks held can actually lead
844 * to obscure indirect dependencies via WQ context.
845 */
846static inline void __lru_add_drain_all(bool force_all_cpus)
847{
848 /*
849 * lru_drain_gen - Global pages generation number
850 *
851 * (A) Definition: global lru_drain_gen = x implies that all generations
852 * 0 < n <= x are already *scheduled* for draining.
853 *
854 * This is an optimization for the highly-contended use case where a
855 * user space workload keeps constantly generating a flow of pages for
856 * each CPU.
857 */
858 static unsigned int lru_drain_gen;
859 static struct cpumask has_work;
860 static DEFINE_MUTEX(lock);
861 unsigned cpu, this_gen;
862
863 /*
864 * Make sure nobody triggers this path before mm_percpu_wq is fully
865 * initialized.
866 */
867 if (WARN_ON(!mm_percpu_wq))
868 return;
869
870 /*
871 * Guarantee folio_batch counter stores visible by this CPU
872 * are visible to other CPUs before loading the current drain
873 * generation.
874 */
875 smp_mb();
876
877 /*
878 * (B) Locally cache global LRU draining generation number
879 *
880 * The read barrier ensures that the counter is loaded before the mutex
881 * is taken. It pairs with smp_mb() inside the mutex critical section
882 * at (D).
883 */
884 this_gen = smp_load_acquire(&lru_drain_gen);
885
886 mutex_lock(&lock);
887
888 /*
889 * (C) Exit the draining operation if a newer generation, from another
890 * lru_add_drain_all(), was already scheduled for draining. Check (A).
891 */
892 if (unlikely(this_gen != lru_drain_gen && !force_all_cpus))
893 goto done;
894
895 /*
896 * (D) Increment global generation number
897 *
898 * Pairs with smp_load_acquire() at (B), outside of the critical
899 * section. Use a full memory barrier to guarantee that the
900 * new global drain generation number is stored before loading
901 * folio_batch counters.
902 *
903 * This pairing must be done here, before the for_each_online_cpu loop
904 * below which drains the page vectors.
905 *
906 * Let x, y, and z represent some system CPU numbers, where x < y < z.
907 * Assume CPU #z is in the middle of the for_each_online_cpu loop
908 * below and has already reached CPU #y's per-cpu data. CPU #x comes
909 * along, adds some pages to its per-cpu vectors, then calls
910 * lru_add_drain_all().
911 *
912 * If the paired barrier is done at any later step, e.g. after the
913 * loop, CPU #x will just exit at (C) and miss flushing out all of its
914 * added pages.
915 */
916 WRITE_ONCE(lru_drain_gen, lru_drain_gen + 1);
917 smp_mb();
918
919 cpumask_clear(&has_work);
920 for_each_online_cpu(cpu) {
921 struct work_struct *work = &per_cpu(lru_add_drain_work, cpu);
922
923 if (cpu_needs_drain(cpu)) {
924 INIT_WORK(work, lru_add_drain_per_cpu);
925 queue_work_on(cpu, mm_percpu_wq, work);
926 __cpumask_set_cpu(cpu, &has_work);
927 }
928 }
929
930 for_each_cpu(cpu, &has_work)
931 flush_work(&per_cpu(lru_add_drain_work, cpu));
932
933done:
934 mutex_unlock(&lock);
935}
936
937void lru_add_drain_all(void)
938{
939 __lru_add_drain_all(false);
940}
941#else
942void lru_add_drain_all(void)
943{
944 lru_add_drain();
945}
946#endif /* CONFIG_SMP */
947
948atomic_t lru_disable_count = ATOMIC_INIT(0);
949
950/*
951 * lru_cache_disable() needs to be called before we start compiling
952 * a list of pages to be migrated using isolate_lru_page().
953 * It drains pages on LRU cache and then disable on all cpus until
954 * lru_cache_enable is called.
955 *
956 * Must be paired with a call to lru_cache_enable().
957 */
958void lru_cache_disable(void)
959{
960 atomic_inc(&lru_disable_count);
961 /*
962 * Readers of lru_disable_count are protected by either disabling
963 * preemption or rcu_read_lock:
964 *
965 * preempt_disable, local_irq_disable [bh_lru_lock()]
966 * rcu_read_lock [rt_spin_lock CONFIG_PREEMPT_RT]
967 * preempt_disable [local_lock !CONFIG_PREEMPT_RT]
968 *
969 * Since v5.1 kernel, synchronize_rcu() is guaranteed to wait on
970 * preempt_disable() regions of code. So any CPU which sees
971 * lru_disable_count = 0 will have exited the critical
972 * section when synchronize_rcu() returns.
973 */
974 synchronize_rcu_expedited();
975#ifdef CONFIG_SMP
976 __lru_add_drain_all(true);
977#else
978 lru_add_and_bh_lrus_drain();
979#endif
980}
981
982/**
983 * release_pages - batched put_page()
984 * @arg: array of pages to release
985 * @nr: number of pages
986 *
987 * Decrement the reference count on all the pages in @arg. If it
988 * fell to zero, remove the page from the LRU and free it.
989 *
990 * Note that the argument can be an array of pages, encoded pages,
991 * or folio pointers. We ignore any encoded bits, and turn any of
992 * them into just a folio that gets free'd.
993 */
994void release_pages(release_pages_arg arg, int nr)
995{
996 int i;
997 struct encoded_page **encoded = arg.encoded_pages;
998 LIST_HEAD(pages_to_free);
999 struct lruvec *lruvec = NULL;
1000 unsigned long flags = 0;
1001 unsigned int lock_batch;
1002
1003 for (i = 0; i < nr; i++) {
1004 struct folio *folio;
1005
1006 /* Turn any of the argument types into a folio */
1007 folio = page_folio(encoded_page_ptr(encoded[i]));
1008
1009 /*
1010 * Make sure the IRQ-safe lock-holding time does not get
1011 * excessive with a continuous string of pages from the
1012 * same lruvec. The lock is held only if lruvec != NULL.
1013 */
1014 if (lruvec && ++lock_batch == SWAP_CLUSTER_MAX) {
1015 unlock_page_lruvec_irqrestore(lruvec, flags);
1016 lruvec = NULL;
1017 }
1018
1019 if (is_huge_zero_page(&folio->page))
1020 continue;
1021
1022 if (folio_is_zone_device(folio)) {
1023 if (lruvec) {
1024 unlock_page_lruvec_irqrestore(lruvec, flags);
1025 lruvec = NULL;
1026 }
1027 if (put_devmap_managed_page(&folio->page))
1028 continue;
1029 if (folio_put_testzero(folio))
1030 free_zone_device_page(&folio->page);
1031 continue;
1032 }
1033
1034 if (!folio_put_testzero(folio))
1035 continue;
1036
1037 if (folio_test_large(folio)) {
1038 if (lruvec) {
1039 unlock_page_lruvec_irqrestore(lruvec, flags);
1040 lruvec = NULL;
1041 }
1042 __folio_put_large(folio);
1043 continue;
1044 }
1045
1046 if (folio_test_lru(folio)) {
1047 struct lruvec *prev_lruvec = lruvec;
1048
1049 lruvec = folio_lruvec_relock_irqsave(folio, lruvec,
1050 &flags);
1051 if (prev_lruvec != lruvec)
1052 lock_batch = 0;
1053
1054 lruvec_del_folio(lruvec, folio);
1055 __folio_clear_lru_flags(folio);
1056 }
1057
1058 /*
1059 * In rare cases, when truncation or holepunching raced with
1060 * munlock after VM_LOCKED was cleared, Mlocked may still be
1061 * found set here. This does not indicate a problem, unless
1062 * "unevictable_pgs_cleared" appears worryingly large.
1063 */
1064 if (unlikely(folio_test_mlocked(folio))) {
1065 __folio_clear_mlocked(folio);
1066 zone_stat_sub_folio(folio, NR_MLOCK);
1067 count_vm_event(UNEVICTABLE_PGCLEARED);
1068 }
1069
1070 list_add(&folio->lru, &pages_to_free);
1071 }
1072 if (lruvec)
1073 unlock_page_lruvec_irqrestore(lruvec, flags);
1074
1075 mem_cgroup_uncharge_list(&pages_to_free);
1076 free_unref_page_list(&pages_to_free);
1077}
1078EXPORT_SYMBOL(release_pages);
1079
1080/*
1081 * The pages which we're about to release may be in the deferred lru-addition
1082 * queues. That would prevent them from really being freed right now. That's
1083 * OK from a correctness point of view but is inefficient - those pages may be
1084 * cache-warm and we want to give them back to the page allocator ASAP.
1085 *
1086 * So __pagevec_release() will drain those queues here.
1087 * folio_batch_move_lru() calls folios_put() directly to avoid
1088 * mutual recursion.
1089 */
1090void __pagevec_release(struct pagevec *pvec)
1091{
1092 if (!pvec->percpu_pvec_drained) {
1093 lru_add_drain();
1094 pvec->percpu_pvec_drained = true;
1095 }
1096 release_pages(pvec->pages, pagevec_count(pvec));
1097 pagevec_reinit(pvec);
1098}
1099EXPORT_SYMBOL(__pagevec_release);
1100
1101/**
1102 * folio_batch_remove_exceptionals() - Prune non-folios from a batch.
1103 * @fbatch: The batch to prune
1104 *
1105 * find_get_entries() fills a batch with both folios and shadow/swap/DAX
1106 * entries. This function prunes all the non-folio entries from @fbatch
1107 * without leaving holes, so that it can be passed on to folio-only batch
1108 * operations.
1109 */
1110void folio_batch_remove_exceptionals(struct folio_batch *fbatch)
1111{
1112 unsigned int i, j;
1113
1114 for (i = 0, j = 0; i < folio_batch_count(fbatch); i++) {
1115 struct folio *folio = fbatch->folios[i];
1116 if (!xa_is_value(folio))
1117 fbatch->folios[j++] = folio;
1118 }
1119 fbatch->nr = j;
1120}
1121
1122unsigned pagevec_lookup_range_tag(struct pagevec *pvec,
1123 struct address_space *mapping, pgoff_t *index, pgoff_t end,
1124 xa_mark_t tag)
1125{
1126 pvec->nr = find_get_pages_range_tag(mapping, index, end, tag,
1127 PAGEVEC_SIZE, pvec->pages);
1128 return pagevec_count(pvec);
1129}
1130EXPORT_SYMBOL(pagevec_lookup_range_tag);
1131
1132/*
1133 * Perform any setup for the swap system
1134 */
1135void __init swap_setup(void)
1136{
1137 unsigned long megs = totalram_pages() >> (20 - PAGE_SHIFT);
1138
1139 /* Use a smaller cluster for small-memory machines */
1140 if (megs < 16)
1141 page_cluster = 2;
1142 else
1143 page_cluster = 3;
1144 /*
1145 * Right now other parts of the system means that we
1146 * _really_ don't want to cluster much more
1147 */
1148}