Loading...
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
39#include "internal.h"
40
41#define CREATE_TRACE_POINTS
42#include <trace/events/pagemap.h>
43
44/* How many pages do we try to swap or page in/out together? */
45int page_cluster;
46
47static DEFINE_PER_CPU(struct pagevec, lru_add_pvec);
48static DEFINE_PER_CPU(struct pagevec, lru_rotate_pvecs);
49static DEFINE_PER_CPU(struct pagevec, lru_deactivate_file_pvecs);
50static DEFINE_PER_CPU(struct pagevec, lru_deactivate_pvecs);
51static DEFINE_PER_CPU(struct pagevec, lru_lazyfree_pvecs);
52#ifdef CONFIG_SMP
53static DEFINE_PER_CPU(struct pagevec, activate_page_pvecs);
54#endif
55
56/*
57 * This path almost never happens for VM activity - pages are normally
58 * freed via pagevecs. But it gets used by networking.
59 */
60static void __page_cache_release(struct page *page)
61{
62 if (PageLRU(page)) {
63 pg_data_t *pgdat = page_pgdat(page);
64 struct lruvec *lruvec;
65 unsigned long flags;
66
67 spin_lock_irqsave(&pgdat->lru_lock, flags);
68 lruvec = mem_cgroup_page_lruvec(page, pgdat);
69 VM_BUG_ON_PAGE(!PageLRU(page), page);
70 __ClearPageLRU(page);
71 del_page_from_lru_list(page, lruvec, page_off_lru(page));
72 spin_unlock_irqrestore(&pgdat->lru_lock, flags);
73 }
74 __ClearPageWaiters(page);
75}
76
77static void __put_single_page(struct page *page)
78{
79 __page_cache_release(page);
80 mem_cgroup_uncharge(page);
81 free_unref_page(page);
82}
83
84static void __put_compound_page(struct page *page)
85{
86 compound_page_dtor *dtor;
87
88 /*
89 * __page_cache_release() is supposed to be called for thp, not for
90 * hugetlb. This is because hugetlb page does never have PageLRU set
91 * (it's never listed to any LRU lists) and no memcg routines should
92 * be called for hugetlb (it has a separate hugetlb_cgroup.)
93 */
94 if (!PageHuge(page))
95 __page_cache_release(page);
96 dtor = get_compound_page_dtor(page);
97 (*dtor)(page);
98}
99
100void __put_page(struct page *page)
101{
102 if (is_zone_device_page(page)) {
103 put_dev_pagemap(page->pgmap);
104
105 /*
106 * The page belongs to the device that created pgmap. Do
107 * not return it to page allocator.
108 */
109 return;
110 }
111
112 if (unlikely(PageCompound(page)))
113 __put_compound_page(page);
114 else
115 __put_single_page(page);
116}
117EXPORT_SYMBOL(__put_page);
118
119/**
120 * put_pages_list() - release a list of pages
121 * @pages: list of pages threaded on page->lru
122 *
123 * Release a list of pages which are strung together on page.lru. Currently
124 * used by read_cache_pages() and related error recovery code.
125 */
126void put_pages_list(struct list_head *pages)
127{
128 while (!list_empty(pages)) {
129 struct page *victim;
130
131 victim = lru_to_page(pages);
132 list_del(&victim->lru);
133 put_page(victim);
134 }
135}
136EXPORT_SYMBOL(put_pages_list);
137
138/*
139 * get_kernel_pages() - pin kernel pages in memory
140 * @kiov: An array of struct kvec structures
141 * @nr_segs: number of segments to pin
142 * @write: pinning for read/write, currently ignored
143 * @pages: array that receives pointers to the pages pinned.
144 * Should be at least nr_segs long.
145 *
146 * Returns number of pages pinned. This may be fewer than the number
147 * requested. If nr_pages is 0 or negative, returns 0. If no pages
148 * were pinned, returns -errno. Each page returned must be released
149 * with a put_page() call when it is finished with.
150 */
151int get_kernel_pages(const struct kvec *kiov, int nr_segs, int write,
152 struct page **pages)
153{
154 int seg;
155
156 for (seg = 0; seg < nr_segs; seg++) {
157 if (WARN_ON(kiov[seg].iov_len != PAGE_SIZE))
158 return seg;
159
160 pages[seg] = kmap_to_page(kiov[seg].iov_base);
161 get_page(pages[seg]);
162 }
163
164 return seg;
165}
166EXPORT_SYMBOL_GPL(get_kernel_pages);
167
168/*
169 * get_kernel_page() - pin a kernel page in memory
170 * @start: starting kernel address
171 * @write: pinning for read/write, currently ignored
172 * @pages: array that receives pointer to the page pinned.
173 * Must be at least nr_segs long.
174 *
175 * Returns 1 if page is pinned. If the page was not pinned, returns
176 * -errno. The page returned must be released with a put_page() call
177 * when it is finished with.
178 */
179int get_kernel_page(unsigned long start, int write, struct page **pages)
180{
181 const struct kvec kiov = {
182 .iov_base = (void *)start,
183 .iov_len = PAGE_SIZE
184 };
185
186 return get_kernel_pages(&kiov, 1, write, pages);
187}
188EXPORT_SYMBOL_GPL(get_kernel_page);
189
190static void pagevec_lru_move_fn(struct pagevec *pvec,
191 void (*move_fn)(struct page *page, struct lruvec *lruvec, void *arg),
192 void *arg)
193{
194 int i;
195 struct pglist_data *pgdat = NULL;
196 struct lruvec *lruvec;
197 unsigned long flags = 0;
198
199 for (i = 0; i < pagevec_count(pvec); i++) {
200 struct page *page = pvec->pages[i];
201 struct pglist_data *pagepgdat = page_pgdat(page);
202
203 if (pagepgdat != pgdat) {
204 if (pgdat)
205 spin_unlock_irqrestore(&pgdat->lru_lock, flags);
206 pgdat = pagepgdat;
207 spin_lock_irqsave(&pgdat->lru_lock, flags);
208 }
209
210 lruvec = mem_cgroup_page_lruvec(page, pgdat);
211 (*move_fn)(page, lruvec, arg);
212 }
213 if (pgdat)
214 spin_unlock_irqrestore(&pgdat->lru_lock, flags);
215 release_pages(pvec->pages, pvec->nr);
216 pagevec_reinit(pvec);
217}
218
219static void pagevec_move_tail_fn(struct page *page, struct lruvec *lruvec,
220 void *arg)
221{
222 int *pgmoved = arg;
223
224 if (PageLRU(page) && !PageUnevictable(page)) {
225 del_page_from_lru_list(page, lruvec, page_lru(page));
226 ClearPageActive(page);
227 add_page_to_lru_list_tail(page, lruvec, page_lru(page));
228 (*pgmoved)++;
229 }
230}
231
232/*
233 * pagevec_move_tail() must be called with IRQ disabled.
234 * Otherwise this may cause nasty races.
235 */
236static void pagevec_move_tail(struct pagevec *pvec)
237{
238 int pgmoved = 0;
239
240 pagevec_lru_move_fn(pvec, pagevec_move_tail_fn, &pgmoved);
241 __count_vm_events(PGROTATED, pgmoved);
242}
243
244/*
245 * Writeback is about to end against a page which has been marked for immediate
246 * reclaim. If it still appears to be reclaimable, move it to the tail of the
247 * inactive list.
248 */
249void rotate_reclaimable_page(struct page *page)
250{
251 if (!PageLocked(page) && !PageDirty(page) &&
252 !PageUnevictable(page) && PageLRU(page)) {
253 struct pagevec *pvec;
254 unsigned long flags;
255
256 get_page(page);
257 local_irq_save(flags);
258 pvec = this_cpu_ptr(&lru_rotate_pvecs);
259 if (!pagevec_add(pvec, page) || PageCompound(page))
260 pagevec_move_tail(pvec);
261 local_irq_restore(flags);
262 }
263}
264
265static void update_page_reclaim_stat(struct lruvec *lruvec,
266 int file, int rotated)
267{
268 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
269
270 reclaim_stat->recent_scanned[file]++;
271 if (rotated)
272 reclaim_stat->recent_rotated[file]++;
273}
274
275static void __activate_page(struct page *page, struct lruvec *lruvec,
276 void *arg)
277{
278 if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
279 int file = page_is_file_cache(page);
280 int lru = page_lru_base_type(page);
281
282 del_page_from_lru_list(page, lruvec, lru);
283 SetPageActive(page);
284 lru += LRU_ACTIVE;
285 add_page_to_lru_list(page, lruvec, lru);
286 trace_mm_lru_activate(page);
287
288 __count_vm_event(PGACTIVATE);
289 update_page_reclaim_stat(lruvec, file, 1);
290 }
291}
292
293#ifdef CONFIG_SMP
294static void activate_page_drain(int cpu)
295{
296 struct pagevec *pvec = &per_cpu(activate_page_pvecs, cpu);
297
298 if (pagevec_count(pvec))
299 pagevec_lru_move_fn(pvec, __activate_page, NULL);
300}
301
302static bool need_activate_page_drain(int cpu)
303{
304 return pagevec_count(&per_cpu(activate_page_pvecs, cpu)) != 0;
305}
306
307void activate_page(struct page *page)
308{
309 page = compound_head(page);
310 if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
311 struct pagevec *pvec = &get_cpu_var(activate_page_pvecs);
312
313 get_page(page);
314 if (!pagevec_add(pvec, page) || PageCompound(page))
315 pagevec_lru_move_fn(pvec, __activate_page, NULL);
316 put_cpu_var(activate_page_pvecs);
317 }
318}
319
320#else
321static inline void activate_page_drain(int cpu)
322{
323}
324
325void activate_page(struct page *page)
326{
327 pg_data_t *pgdat = page_pgdat(page);
328
329 page = compound_head(page);
330 spin_lock_irq(&pgdat->lru_lock);
331 __activate_page(page, mem_cgroup_page_lruvec(page, pgdat), NULL);
332 spin_unlock_irq(&pgdat->lru_lock);
333}
334#endif
335
336static void __lru_cache_activate_page(struct page *page)
337{
338 struct pagevec *pvec = &get_cpu_var(lru_add_pvec);
339 int i;
340
341 /*
342 * Search backwards on the optimistic assumption that the page being
343 * activated has just been added to this pagevec. Note that only
344 * the local pagevec is examined as a !PageLRU page could be in the
345 * process of being released, reclaimed, migrated or on a remote
346 * pagevec that is currently being drained. Furthermore, marking
347 * a remote pagevec's page PageActive potentially hits a race where
348 * a page is marked PageActive just after it is added to the inactive
349 * list causing accounting errors and BUG_ON checks to trigger.
350 */
351 for (i = pagevec_count(pvec) - 1; i >= 0; i--) {
352 struct page *pagevec_page = pvec->pages[i];
353
354 if (pagevec_page == page) {
355 SetPageActive(page);
356 break;
357 }
358 }
359
360 put_cpu_var(lru_add_pvec);
361}
362
363/*
364 * Mark a page as having seen activity.
365 *
366 * inactive,unreferenced -> inactive,referenced
367 * inactive,referenced -> active,unreferenced
368 * active,unreferenced -> active,referenced
369 *
370 * When a newly allocated page is not yet visible, so safe for non-atomic ops,
371 * __SetPageReferenced(page) may be substituted for mark_page_accessed(page).
372 */
373void mark_page_accessed(struct page *page)
374{
375 page = compound_head(page);
376 if (!PageActive(page) && !PageUnevictable(page) &&
377 PageReferenced(page)) {
378
379 /*
380 * If the page is on the LRU, queue it for activation via
381 * activate_page_pvecs. Otherwise, assume the page is on a
382 * pagevec, mark it active and it'll be moved to the active
383 * LRU on the next drain.
384 */
385 if (PageLRU(page))
386 activate_page(page);
387 else
388 __lru_cache_activate_page(page);
389 ClearPageReferenced(page);
390 if (page_is_file_cache(page))
391 workingset_activation(page);
392 } else if (!PageReferenced(page)) {
393 SetPageReferenced(page);
394 }
395 if (page_is_idle(page))
396 clear_page_idle(page);
397}
398EXPORT_SYMBOL(mark_page_accessed);
399
400static void __lru_cache_add(struct page *page)
401{
402 struct pagevec *pvec = &get_cpu_var(lru_add_pvec);
403
404 get_page(page);
405 if (!pagevec_add(pvec, page) || PageCompound(page))
406 __pagevec_lru_add(pvec);
407 put_cpu_var(lru_add_pvec);
408}
409
410/**
411 * lru_cache_add_anon - add a page to the page lists
412 * @page: the page to add
413 */
414void lru_cache_add_anon(struct page *page)
415{
416 if (PageActive(page))
417 ClearPageActive(page);
418 __lru_cache_add(page);
419}
420
421void lru_cache_add_file(struct page *page)
422{
423 if (PageActive(page))
424 ClearPageActive(page);
425 __lru_cache_add(page);
426}
427EXPORT_SYMBOL(lru_cache_add_file);
428
429/**
430 * lru_cache_add - add a page to a page list
431 * @page: the page to be added to the LRU.
432 *
433 * Queue the page for addition to the LRU via pagevec. The decision on whether
434 * to add the page to the [in]active [file|anon] list is deferred until the
435 * pagevec is drained. This gives a chance for the caller of lru_cache_add()
436 * have the page added to the active list using mark_page_accessed().
437 */
438void lru_cache_add(struct page *page)
439{
440 VM_BUG_ON_PAGE(PageActive(page) && PageUnevictable(page), page);
441 VM_BUG_ON_PAGE(PageLRU(page), page);
442 __lru_cache_add(page);
443}
444
445/**
446 * lru_cache_add_active_or_unevictable
447 * @page: the page to be added to LRU
448 * @vma: vma in which page is mapped for determining reclaimability
449 *
450 * Place @page on the active or unevictable LRU list, depending on its
451 * evictability. Note that if the page is not evictable, it goes
452 * directly back onto it's zone's unevictable list, it does NOT use a
453 * per cpu pagevec.
454 */
455void lru_cache_add_active_or_unevictable(struct page *page,
456 struct vm_area_struct *vma)
457{
458 VM_BUG_ON_PAGE(PageLRU(page), page);
459
460 if (likely((vma->vm_flags & (VM_LOCKED | VM_SPECIAL)) != VM_LOCKED))
461 SetPageActive(page);
462 else if (!TestSetPageMlocked(page)) {
463 /*
464 * We use the irq-unsafe __mod_zone_page_stat because this
465 * counter is not modified from interrupt context, and the pte
466 * lock is held(spinlock), which implies preemption disabled.
467 */
468 __mod_zone_page_state(page_zone(page), NR_MLOCK,
469 hpage_nr_pages(page));
470 count_vm_event(UNEVICTABLE_PGMLOCKED);
471 }
472 lru_cache_add(page);
473}
474
475/*
476 * If the page can not be invalidated, it is moved to the
477 * inactive list to speed up its reclaim. It is moved to the
478 * head of the list, rather than the tail, to give the flusher
479 * threads some time to write it out, as this is much more
480 * effective than the single-page writeout from reclaim.
481 *
482 * If the page isn't page_mapped and dirty/writeback, the page
483 * could reclaim asap using PG_reclaim.
484 *
485 * 1. active, mapped page -> none
486 * 2. active, dirty/writeback page -> inactive, head, PG_reclaim
487 * 3. inactive, mapped page -> none
488 * 4. inactive, dirty/writeback page -> inactive, head, PG_reclaim
489 * 5. inactive, clean -> inactive, tail
490 * 6. Others -> none
491 *
492 * In 4, why it moves inactive's head, the VM expects the page would
493 * be write it out by flusher threads as this is much more effective
494 * than the single-page writeout from reclaim.
495 */
496static void lru_deactivate_file_fn(struct page *page, struct lruvec *lruvec,
497 void *arg)
498{
499 int lru, file;
500 bool active;
501
502 if (!PageLRU(page))
503 return;
504
505 if (PageUnevictable(page))
506 return;
507
508 /* Some processes are using the page */
509 if (page_mapped(page))
510 return;
511
512 active = PageActive(page);
513 file = page_is_file_cache(page);
514 lru = page_lru_base_type(page);
515
516 del_page_from_lru_list(page, lruvec, lru + active);
517 ClearPageActive(page);
518 ClearPageReferenced(page);
519
520 if (PageWriteback(page) || PageDirty(page)) {
521 /*
522 * PG_reclaim could be raced with end_page_writeback
523 * It can make readahead confusing. But race window
524 * is _really_ small and it's non-critical problem.
525 */
526 add_page_to_lru_list(page, lruvec, lru);
527 SetPageReclaim(page);
528 } else {
529 /*
530 * The page's writeback ends up during pagevec
531 * We moves tha page into tail of inactive.
532 */
533 add_page_to_lru_list_tail(page, lruvec, lru);
534 __count_vm_event(PGROTATED);
535 }
536
537 if (active)
538 __count_vm_event(PGDEACTIVATE);
539 update_page_reclaim_stat(lruvec, file, 0);
540}
541
542static void lru_deactivate_fn(struct page *page, struct lruvec *lruvec,
543 void *arg)
544{
545 if (PageLRU(page) && PageActive(page) && !PageUnevictable(page)) {
546 int file = page_is_file_cache(page);
547 int lru = page_lru_base_type(page);
548
549 del_page_from_lru_list(page, lruvec, lru + LRU_ACTIVE);
550 ClearPageActive(page);
551 ClearPageReferenced(page);
552 add_page_to_lru_list(page, lruvec, lru);
553
554 __count_vm_events(PGDEACTIVATE, hpage_nr_pages(page));
555 update_page_reclaim_stat(lruvec, file, 0);
556 }
557}
558
559static void lru_lazyfree_fn(struct page *page, struct lruvec *lruvec,
560 void *arg)
561{
562 if (PageLRU(page) && PageAnon(page) && PageSwapBacked(page) &&
563 !PageSwapCache(page) && !PageUnevictable(page)) {
564 bool active = PageActive(page);
565
566 del_page_from_lru_list(page, lruvec,
567 LRU_INACTIVE_ANON + active);
568 ClearPageActive(page);
569 ClearPageReferenced(page);
570 /*
571 * lazyfree pages are clean anonymous pages. They have
572 * SwapBacked flag cleared to distinguish normal anonymous
573 * pages
574 */
575 ClearPageSwapBacked(page);
576 add_page_to_lru_list(page, lruvec, LRU_INACTIVE_FILE);
577
578 __count_vm_events(PGLAZYFREE, hpage_nr_pages(page));
579 count_memcg_page_event(page, PGLAZYFREE);
580 update_page_reclaim_stat(lruvec, 1, 0);
581 }
582}
583
584/*
585 * Drain pages out of the cpu's pagevecs.
586 * Either "cpu" is the current CPU, and preemption has already been
587 * disabled; or "cpu" is being hot-unplugged, and is already dead.
588 */
589void lru_add_drain_cpu(int cpu)
590{
591 struct pagevec *pvec = &per_cpu(lru_add_pvec, cpu);
592
593 if (pagevec_count(pvec))
594 __pagevec_lru_add(pvec);
595
596 pvec = &per_cpu(lru_rotate_pvecs, cpu);
597 if (pagevec_count(pvec)) {
598 unsigned long flags;
599
600 /* No harm done if a racing interrupt already did this */
601 local_irq_save(flags);
602 pagevec_move_tail(pvec);
603 local_irq_restore(flags);
604 }
605
606 pvec = &per_cpu(lru_deactivate_file_pvecs, cpu);
607 if (pagevec_count(pvec))
608 pagevec_lru_move_fn(pvec, lru_deactivate_file_fn, NULL);
609
610 pvec = &per_cpu(lru_deactivate_pvecs, cpu);
611 if (pagevec_count(pvec))
612 pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
613
614 pvec = &per_cpu(lru_lazyfree_pvecs, cpu);
615 if (pagevec_count(pvec))
616 pagevec_lru_move_fn(pvec, lru_lazyfree_fn, NULL);
617
618 activate_page_drain(cpu);
619}
620
621/**
622 * deactivate_file_page - forcefully deactivate a file page
623 * @page: page to deactivate
624 *
625 * This function hints the VM that @page is a good reclaim candidate,
626 * for example if its invalidation fails due to the page being dirty
627 * or under writeback.
628 */
629void deactivate_file_page(struct page *page)
630{
631 /*
632 * In a workload with many unevictable page such as mprotect,
633 * unevictable page deactivation for accelerating reclaim is pointless.
634 */
635 if (PageUnevictable(page))
636 return;
637
638 if (likely(get_page_unless_zero(page))) {
639 struct pagevec *pvec = &get_cpu_var(lru_deactivate_file_pvecs);
640
641 if (!pagevec_add(pvec, page) || PageCompound(page))
642 pagevec_lru_move_fn(pvec, lru_deactivate_file_fn, NULL);
643 put_cpu_var(lru_deactivate_file_pvecs);
644 }
645}
646
647/*
648 * deactivate_page - deactivate a page
649 * @page: page to deactivate
650 *
651 * deactivate_page() moves @page to the inactive list if @page was on the active
652 * list and was not an unevictable page. This is done to accelerate the reclaim
653 * of @page.
654 */
655void deactivate_page(struct page *page)
656{
657 if (PageLRU(page) && PageActive(page) && !PageUnevictable(page)) {
658 struct pagevec *pvec = &get_cpu_var(lru_deactivate_pvecs);
659
660 get_page(page);
661 if (!pagevec_add(pvec, page) || PageCompound(page))
662 pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
663 put_cpu_var(lru_deactivate_pvecs);
664 }
665}
666
667/**
668 * mark_page_lazyfree - make an anon page lazyfree
669 * @page: page to deactivate
670 *
671 * mark_page_lazyfree() moves @page to the inactive file list.
672 * This is done to accelerate the reclaim of @page.
673 */
674void mark_page_lazyfree(struct page *page)
675{
676 if (PageLRU(page) && PageAnon(page) && PageSwapBacked(page) &&
677 !PageSwapCache(page) && !PageUnevictable(page)) {
678 struct pagevec *pvec = &get_cpu_var(lru_lazyfree_pvecs);
679
680 get_page(page);
681 if (!pagevec_add(pvec, page) || PageCompound(page))
682 pagevec_lru_move_fn(pvec, lru_lazyfree_fn, NULL);
683 put_cpu_var(lru_lazyfree_pvecs);
684 }
685}
686
687void lru_add_drain(void)
688{
689 lru_add_drain_cpu(get_cpu());
690 put_cpu();
691}
692
693#ifdef CONFIG_SMP
694
695static DEFINE_PER_CPU(struct work_struct, lru_add_drain_work);
696
697static void lru_add_drain_per_cpu(struct work_struct *dummy)
698{
699 lru_add_drain();
700}
701
702/*
703 * Doesn't need any cpu hotplug locking because we do rely on per-cpu
704 * kworkers being shut down before our page_alloc_cpu_dead callback is
705 * executed on the offlined cpu.
706 * Calling this function with cpu hotplug locks held can actually lead
707 * to obscure indirect dependencies via WQ context.
708 */
709void lru_add_drain_all(void)
710{
711 static DEFINE_MUTEX(lock);
712 static struct cpumask has_work;
713 int cpu;
714
715 /*
716 * Make sure nobody triggers this path before mm_percpu_wq is fully
717 * initialized.
718 */
719 if (WARN_ON(!mm_percpu_wq))
720 return;
721
722 mutex_lock(&lock);
723 cpumask_clear(&has_work);
724
725 for_each_online_cpu(cpu) {
726 struct work_struct *work = &per_cpu(lru_add_drain_work, cpu);
727
728 if (pagevec_count(&per_cpu(lru_add_pvec, cpu)) ||
729 pagevec_count(&per_cpu(lru_rotate_pvecs, cpu)) ||
730 pagevec_count(&per_cpu(lru_deactivate_file_pvecs, cpu)) ||
731 pagevec_count(&per_cpu(lru_deactivate_pvecs, cpu)) ||
732 pagevec_count(&per_cpu(lru_lazyfree_pvecs, cpu)) ||
733 need_activate_page_drain(cpu)) {
734 INIT_WORK(work, lru_add_drain_per_cpu);
735 queue_work_on(cpu, mm_percpu_wq, work);
736 cpumask_set_cpu(cpu, &has_work);
737 }
738 }
739
740 for_each_cpu(cpu, &has_work)
741 flush_work(&per_cpu(lru_add_drain_work, cpu));
742
743 mutex_unlock(&lock);
744}
745#else
746void lru_add_drain_all(void)
747{
748 lru_add_drain();
749}
750#endif
751
752/**
753 * release_pages - batched put_page()
754 * @pages: array of pages to release
755 * @nr: number of pages
756 *
757 * Decrement the reference count on all the pages in @pages. If it
758 * fell to zero, remove the page from the LRU and free it.
759 */
760void release_pages(struct page **pages, int nr)
761{
762 int i;
763 LIST_HEAD(pages_to_free);
764 struct pglist_data *locked_pgdat = NULL;
765 struct lruvec *lruvec;
766 unsigned long uninitialized_var(flags);
767 unsigned int uninitialized_var(lock_batch);
768
769 for (i = 0; i < nr; i++) {
770 struct page *page = pages[i];
771
772 /*
773 * Make sure the IRQ-safe lock-holding time does not get
774 * excessive with a continuous string of pages from the
775 * same pgdat. The lock is held only if pgdat != NULL.
776 */
777 if (locked_pgdat && ++lock_batch == SWAP_CLUSTER_MAX) {
778 spin_unlock_irqrestore(&locked_pgdat->lru_lock, flags);
779 locked_pgdat = NULL;
780 }
781
782 if (is_huge_zero_page(page))
783 continue;
784
785 if (is_zone_device_page(page)) {
786 if (locked_pgdat) {
787 spin_unlock_irqrestore(&locked_pgdat->lru_lock,
788 flags);
789 locked_pgdat = NULL;
790 }
791 /*
792 * ZONE_DEVICE pages that return 'false' from
793 * put_devmap_managed_page() do not require special
794 * processing, and instead, expect a call to
795 * put_page_testzero().
796 */
797 if (put_devmap_managed_page(page))
798 continue;
799 }
800
801 page = compound_head(page);
802 if (!put_page_testzero(page))
803 continue;
804
805 if (PageCompound(page)) {
806 if (locked_pgdat) {
807 spin_unlock_irqrestore(&locked_pgdat->lru_lock, flags);
808 locked_pgdat = NULL;
809 }
810 __put_compound_page(page);
811 continue;
812 }
813
814 if (PageLRU(page)) {
815 struct pglist_data *pgdat = page_pgdat(page);
816
817 if (pgdat != locked_pgdat) {
818 if (locked_pgdat)
819 spin_unlock_irqrestore(&locked_pgdat->lru_lock,
820 flags);
821 lock_batch = 0;
822 locked_pgdat = pgdat;
823 spin_lock_irqsave(&locked_pgdat->lru_lock, flags);
824 }
825
826 lruvec = mem_cgroup_page_lruvec(page, locked_pgdat);
827 VM_BUG_ON_PAGE(!PageLRU(page), page);
828 __ClearPageLRU(page);
829 del_page_from_lru_list(page, lruvec, page_off_lru(page));
830 }
831
832 /* Clear Active bit in case of parallel mark_page_accessed */
833 __ClearPageActive(page);
834 __ClearPageWaiters(page);
835
836 list_add(&page->lru, &pages_to_free);
837 }
838 if (locked_pgdat)
839 spin_unlock_irqrestore(&locked_pgdat->lru_lock, flags);
840
841 mem_cgroup_uncharge_list(&pages_to_free);
842 free_unref_page_list(&pages_to_free);
843}
844EXPORT_SYMBOL(release_pages);
845
846/*
847 * The pages which we're about to release may be in the deferred lru-addition
848 * queues. That would prevent them from really being freed right now. That's
849 * OK from a correctness point of view but is inefficient - those pages may be
850 * cache-warm and we want to give them back to the page allocator ASAP.
851 *
852 * So __pagevec_release() will drain those queues here. __pagevec_lru_add()
853 * and __pagevec_lru_add_active() call release_pages() directly to avoid
854 * mutual recursion.
855 */
856void __pagevec_release(struct pagevec *pvec)
857{
858 if (!pvec->percpu_pvec_drained) {
859 lru_add_drain();
860 pvec->percpu_pvec_drained = true;
861 }
862 release_pages(pvec->pages, pagevec_count(pvec));
863 pagevec_reinit(pvec);
864}
865EXPORT_SYMBOL(__pagevec_release);
866
867#ifdef CONFIG_TRANSPARENT_HUGEPAGE
868/* used by __split_huge_page_refcount() */
869void lru_add_page_tail(struct page *page, struct page *page_tail,
870 struct lruvec *lruvec, struct list_head *list)
871{
872 const int file = 0;
873
874 VM_BUG_ON_PAGE(!PageHead(page), page);
875 VM_BUG_ON_PAGE(PageCompound(page_tail), page);
876 VM_BUG_ON_PAGE(PageLRU(page_tail), page);
877 lockdep_assert_held(&lruvec_pgdat(lruvec)->lru_lock);
878
879 if (!list)
880 SetPageLRU(page_tail);
881
882 if (likely(PageLRU(page)))
883 list_add_tail(&page_tail->lru, &page->lru);
884 else if (list) {
885 /* page reclaim is reclaiming a huge page */
886 get_page(page_tail);
887 list_add_tail(&page_tail->lru, list);
888 } else {
889 /*
890 * Head page has not yet been counted, as an hpage,
891 * so we must account for each subpage individually.
892 *
893 * Put page_tail on the list at the correct position
894 * so they all end up in order.
895 */
896 add_page_to_lru_list_tail(page_tail, lruvec,
897 page_lru(page_tail));
898 }
899
900 if (!PageUnevictable(page))
901 update_page_reclaim_stat(lruvec, file, PageActive(page_tail));
902}
903#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
904
905static void __pagevec_lru_add_fn(struct page *page, struct lruvec *lruvec,
906 void *arg)
907{
908 enum lru_list lru;
909 int was_unevictable = TestClearPageUnevictable(page);
910
911 VM_BUG_ON_PAGE(PageLRU(page), page);
912
913 SetPageLRU(page);
914 /*
915 * Page becomes evictable in two ways:
916 * 1) Within LRU lock [munlock_vma_page() and __munlock_pagevec()].
917 * 2) Before acquiring LRU lock to put the page to correct LRU and then
918 * a) do PageLRU check with lock [check_move_unevictable_pages]
919 * b) do PageLRU check before lock [clear_page_mlock]
920 *
921 * (1) & (2a) are ok as LRU lock will serialize them. For (2b), we need
922 * following strict ordering:
923 *
924 * #0: __pagevec_lru_add_fn #1: clear_page_mlock
925 *
926 * SetPageLRU() TestClearPageMlocked()
927 * smp_mb() // explicit ordering // above provides strict
928 * // ordering
929 * PageMlocked() PageLRU()
930 *
931 *
932 * if '#1' does not observe setting of PG_lru by '#0' and fails
933 * isolation, the explicit barrier will make sure that page_evictable
934 * check will put the page in correct LRU. Without smp_mb(), SetPageLRU
935 * can be reordered after PageMlocked check and can make '#1' to fail
936 * the isolation of the page whose Mlocked bit is cleared (#0 is also
937 * looking at the same page) and the evictable page will be stranded
938 * in an unevictable LRU.
939 */
940 smp_mb();
941
942 if (page_evictable(page)) {
943 lru = page_lru(page);
944 update_page_reclaim_stat(lruvec, page_is_file_cache(page),
945 PageActive(page));
946 if (was_unevictable)
947 count_vm_event(UNEVICTABLE_PGRESCUED);
948 } else {
949 lru = LRU_UNEVICTABLE;
950 ClearPageActive(page);
951 SetPageUnevictable(page);
952 if (!was_unevictable)
953 count_vm_event(UNEVICTABLE_PGCULLED);
954 }
955
956 add_page_to_lru_list(page, lruvec, lru);
957 trace_mm_lru_insertion(page, lru);
958}
959
960/*
961 * Add the passed pages to the LRU, then drop the caller's refcount
962 * on them. Reinitialises the caller's pagevec.
963 */
964void __pagevec_lru_add(struct pagevec *pvec)
965{
966 pagevec_lru_move_fn(pvec, __pagevec_lru_add_fn, NULL);
967}
968EXPORT_SYMBOL(__pagevec_lru_add);
969
970/**
971 * pagevec_lookup_entries - gang pagecache lookup
972 * @pvec: Where the resulting entries are placed
973 * @mapping: The address_space to search
974 * @start: The starting entry index
975 * @nr_entries: The maximum number of pages
976 * @indices: The cache indices corresponding to the entries in @pvec
977 *
978 * pagevec_lookup_entries() will search for and return a group of up
979 * to @nr_pages pages and shadow entries in the mapping. All
980 * entries are placed in @pvec. pagevec_lookup_entries() takes a
981 * reference against actual pages in @pvec.
982 *
983 * The search returns a group of mapping-contiguous entries with
984 * ascending indexes. There may be holes in the indices due to
985 * not-present entries.
986 *
987 * pagevec_lookup_entries() returns the number of entries which were
988 * found.
989 */
990unsigned pagevec_lookup_entries(struct pagevec *pvec,
991 struct address_space *mapping,
992 pgoff_t start, unsigned nr_entries,
993 pgoff_t *indices)
994{
995 pvec->nr = find_get_entries(mapping, start, nr_entries,
996 pvec->pages, indices);
997 return pagevec_count(pvec);
998}
999
1000/**
1001 * pagevec_remove_exceptionals - pagevec exceptionals pruning
1002 * @pvec: The pagevec to prune
1003 *
1004 * pagevec_lookup_entries() fills both pages and exceptional radix
1005 * tree entries into the pagevec. This function prunes all
1006 * exceptionals from @pvec without leaving holes, so that it can be
1007 * passed on to page-only pagevec operations.
1008 */
1009void pagevec_remove_exceptionals(struct pagevec *pvec)
1010{
1011 int i, j;
1012
1013 for (i = 0, j = 0; i < pagevec_count(pvec); i++) {
1014 struct page *page = pvec->pages[i];
1015 if (!xa_is_value(page))
1016 pvec->pages[j++] = page;
1017 }
1018 pvec->nr = j;
1019}
1020
1021/**
1022 * pagevec_lookup_range - gang pagecache lookup
1023 * @pvec: Where the resulting pages are placed
1024 * @mapping: The address_space to search
1025 * @start: The starting page index
1026 * @end: The final page index
1027 *
1028 * pagevec_lookup_range() will search for & return a group of up to PAGEVEC_SIZE
1029 * pages in the mapping starting from index @start and upto index @end
1030 * (inclusive). The pages are placed in @pvec. pagevec_lookup() takes a
1031 * reference against the pages in @pvec.
1032 *
1033 * The search returns a group of mapping-contiguous pages with ascending
1034 * indexes. There may be holes in the indices due to not-present pages. We
1035 * also update @start to index the next page for the traversal.
1036 *
1037 * pagevec_lookup_range() returns the number of pages which were found. If this
1038 * number is smaller than PAGEVEC_SIZE, the end of specified range has been
1039 * reached.
1040 */
1041unsigned pagevec_lookup_range(struct pagevec *pvec,
1042 struct address_space *mapping, pgoff_t *start, pgoff_t end)
1043{
1044 pvec->nr = find_get_pages_range(mapping, start, end, PAGEVEC_SIZE,
1045 pvec->pages);
1046 return pagevec_count(pvec);
1047}
1048EXPORT_SYMBOL(pagevec_lookup_range);
1049
1050unsigned pagevec_lookup_range_tag(struct pagevec *pvec,
1051 struct address_space *mapping, pgoff_t *index, pgoff_t end,
1052 xa_mark_t tag)
1053{
1054 pvec->nr = find_get_pages_range_tag(mapping, index, end, tag,
1055 PAGEVEC_SIZE, pvec->pages);
1056 return pagevec_count(pvec);
1057}
1058EXPORT_SYMBOL(pagevec_lookup_range_tag);
1059
1060unsigned pagevec_lookup_range_nr_tag(struct pagevec *pvec,
1061 struct address_space *mapping, pgoff_t *index, pgoff_t end,
1062 xa_mark_t tag, unsigned max_pages)
1063{
1064 pvec->nr = find_get_pages_range_tag(mapping, index, end, tag,
1065 min_t(unsigned int, max_pages, PAGEVEC_SIZE), pvec->pages);
1066 return pagevec_count(pvec);
1067}
1068EXPORT_SYMBOL(pagevec_lookup_range_nr_tag);
1069/*
1070 * Perform any setup for the swap system
1071 */
1072void __init swap_setup(void)
1073{
1074 unsigned long megs = totalram_pages() >> (20 - PAGE_SHIFT);
1075
1076 /* Use a smaller cluster for small-memory machines */
1077 if (megs < 16)
1078 page_cluster = 2;
1079 else
1080 page_cluster = 3;
1081 /*
1082 * Right now other parts of the system means that we
1083 * _really_ don't want to cluster much more
1084 */
1085}
1/*
2 * linux/mm/swap.c
3 *
4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 */
6
7/*
8 * This file contains the default values for the operation of the
9 * Linux VM subsystem. Fine-tuning documentation can be found in
10 * Documentation/sysctl/vm.txt.
11 * Started 18.12.91
12 * Swap aging added 23.2.95, Stephen Tweedie.
13 * Buffermem limits added 12.3.98, Rik van Riel.
14 */
15
16#include <linux/mm.h>
17#include <linux/sched.h>
18#include <linux/kernel_stat.h>
19#include <linux/swap.h>
20#include <linux/mman.h>
21#include <linux/pagemap.h>
22#include <linux/pagevec.h>
23#include <linux/init.h>
24#include <linux/export.h>
25#include <linux/mm_inline.h>
26#include <linux/percpu_counter.h>
27#include <linux/memremap.h>
28#include <linux/percpu.h>
29#include <linux/cpu.h>
30#include <linux/notifier.h>
31#include <linux/backing-dev.h>
32#include <linux/memcontrol.h>
33#include <linux/gfp.h>
34#include <linux/uio.h>
35#include <linux/hugetlb.h>
36#include <linux/page_idle.h>
37
38#include "internal.h"
39
40#define CREATE_TRACE_POINTS
41#include <trace/events/pagemap.h>
42
43/* How many pages do we try to swap or page in/out together? */
44int page_cluster;
45
46static DEFINE_PER_CPU(struct pagevec, lru_add_pvec);
47static DEFINE_PER_CPU(struct pagevec, lru_rotate_pvecs);
48static DEFINE_PER_CPU(struct pagevec, lru_deactivate_file_pvecs);
49static DEFINE_PER_CPU(struct pagevec, lru_deactivate_pvecs);
50
51/*
52 * This path almost never happens for VM activity - pages are normally
53 * freed via pagevecs. But it gets used by networking.
54 */
55static void __page_cache_release(struct page *page)
56{
57 if (PageLRU(page)) {
58 struct zone *zone = page_zone(page);
59 struct lruvec *lruvec;
60 unsigned long flags;
61
62 spin_lock_irqsave(&zone->lru_lock, flags);
63 lruvec = mem_cgroup_page_lruvec(page, zone);
64 VM_BUG_ON_PAGE(!PageLRU(page), page);
65 __ClearPageLRU(page);
66 del_page_from_lru_list(page, lruvec, page_off_lru(page));
67 spin_unlock_irqrestore(&zone->lru_lock, flags);
68 }
69 mem_cgroup_uncharge(page);
70}
71
72static void __put_single_page(struct page *page)
73{
74 __page_cache_release(page);
75 free_hot_cold_page(page, false);
76}
77
78static void __put_compound_page(struct page *page)
79{
80 compound_page_dtor *dtor;
81
82 /*
83 * __page_cache_release() is supposed to be called for thp, not for
84 * hugetlb. This is because hugetlb page does never have PageLRU set
85 * (it's never listed to any LRU lists) and no memcg routines should
86 * be called for hugetlb (it has a separate hugetlb_cgroup.)
87 */
88 if (!PageHuge(page))
89 __page_cache_release(page);
90 dtor = get_compound_page_dtor(page);
91 (*dtor)(page);
92}
93
94void __put_page(struct page *page)
95{
96 if (unlikely(PageCompound(page)))
97 __put_compound_page(page);
98 else
99 __put_single_page(page);
100}
101EXPORT_SYMBOL(__put_page);
102
103/**
104 * put_pages_list() - release a list of pages
105 * @pages: list of pages threaded on page->lru
106 *
107 * Release a list of pages which are strung together on page.lru. Currently
108 * used by read_cache_pages() and related error recovery code.
109 */
110void put_pages_list(struct list_head *pages)
111{
112 while (!list_empty(pages)) {
113 struct page *victim;
114
115 victim = list_entry(pages->prev, struct page, lru);
116 list_del(&victim->lru);
117 put_page(victim);
118 }
119}
120EXPORT_SYMBOL(put_pages_list);
121
122/*
123 * get_kernel_pages() - pin kernel pages in memory
124 * @kiov: An array of struct kvec structures
125 * @nr_segs: number of segments to pin
126 * @write: pinning for read/write, currently ignored
127 * @pages: array that receives pointers to the pages pinned.
128 * Should be at least nr_segs long.
129 *
130 * Returns number of pages pinned. This may be fewer than the number
131 * requested. If nr_pages is 0 or negative, returns 0. If no pages
132 * were pinned, returns -errno. Each page returned must be released
133 * with a put_page() call when it is finished with.
134 */
135int get_kernel_pages(const struct kvec *kiov, int nr_segs, int write,
136 struct page **pages)
137{
138 int seg;
139
140 for (seg = 0; seg < nr_segs; seg++) {
141 if (WARN_ON(kiov[seg].iov_len != PAGE_SIZE))
142 return seg;
143
144 pages[seg] = kmap_to_page(kiov[seg].iov_base);
145 get_page(pages[seg]);
146 }
147
148 return seg;
149}
150EXPORT_SYMBOL_GPL(get_kernel_pages);
151
152/*
153 * get_kernel_page() - pin a kernel page in memory
154 * @start: starting kernel address
155 * @write: pinning for read/write, currently ignored
156 * @pages: array that receives pointer to the page pinned.
157 * Must be at least nr_segs long.
158 *
159 * Returns 1 if page is pinned. If the page was not pinned, returns
160 * -errno. The page returned must be released with a put_page() call
161 * when it is finished with.
162 */
163int get_kernel_page(unsigned long start, int write, struct page **pages)
164{
165 const struct kvec kiov = {
166 .iov_base = (void *)start,
167 .iov_len = PAGE_SIZE
168 };
169
170 return get_kernel_pages(&kiov, 1, write, pages);
171}
172EXPORT_SYMBOL_GPL(get_kernel_page);
173
174static void pagevec_lru_move_fn(struct pagevec *pvec,
175 void (*move_fn)(struct page *page, struct lruvec *lruvec, void *arg),
176 void *arg)
177{
178 int i;
179 struct zone *zone = NULL;
180 struct lruvec *lruvec;
181 unsigned long flags = 0;
182
183 for (i = 0; i < pagevec_count(pvec); i++) {
184 struct page *page = pvec->pages[i];
185 struct zone *pagezone = page_zone(page);
186
187 if (pagezone != zone) {
188 if (zone)
189 spin_unlock_irqrestore(&zone->lru_lock, flags);
190 zone = pagezone;
191 spin_lock_irqsave(&zone->lru_lock, flags);
192 }
193
194 lruvec = mem_cgroup_page_lruvec(page, zone);
195 (*move_fn)(page, lruvec, arg);
196 }
197 if (zone)
198 spin_unlock_irqrestore(&zone->lru_lock, flags);
199 release_pages(pvec->pages, pvec->nr, pvec->cold);
200 pagevec_reinit(pvec);
201}
202
203static void pagevec_move_tail_fn(struct page *page, struct lruvec *lruvec,
204 void *arg)
205{
206 int *pgmoved = arg;
207
208 if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
209 enum lru_list lru = page_lru_base_type(page);
210 list_move_tail(&page->lru, &lruvec->lists[lru]);
211 (*pgmoved)++;
212 }
213}
214
215/*
216 * pagevec_move_tail() must be called with IRQ disabled.
217 * Otherwise this may cause nasty races.
218 */
219static void pagevec_move_tail(struct pagevec *pvec)
220{
221 int pgmoved = 0;
222
223 pagevec_lru_move_fn(pvec, pagevec_move_tail_fn, &pgmoved);
224 __count_vm_events(PGROTATED, pgmoved);
225}
226
227/*
228 * Writeback is about to end against a page which has been marked for immediate
229 * reclaim. If it still appears to be reclaimable, move it to the tail of the
230 * inactive list.
231 */
232void rotate_reclaimable_page(struct page *page)
233{
234 if (!PageLocked(page) && !PageDirty(page) && !PageActive(page) &&
235 !PageUnevictable(page) && PageLRU(page)) {
236 struct pagevec *pvec;
237 unsigned long flags;
238
239 get_page(page);
240 local_irq_save(flags);
241 pvec = this_cpu_ptr(&lru_rotate_pvecs);
242 if (!pagevec_add(pvec, page))
243 pagevec_move_tail(pvec);
244 local_irq_restore(flags);
245 }
246}
247
248static void update_page_reclaim_stat(struct lruvec *lruvec,
249 int file, int rotated)
250{
251 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
252
253 reclaim_stat->recent_scanned[file]++;
254 if (rotated)
255 reclaim_stat->recent_rotated[file]++;
256}
257
258static void __activate_page(struct page *page, struct lruvec *lruvec,
259 void *arg)
260{
261 if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
262 int file = page_is_file_cache(page);
263 int lru = page_lru_base_type(page);
264
265 del_page_from_lru_list(page, lruvec, lru);
266 SetPageActive(page);
267 lru += LRU_ACTIVE;
268 add_page_to_lru_list(page, lruvec, lru);
269 trace_mm_lru_activate(page);
270
271 __count_vm_event(PGACTIVATE);
272 update_page_reclaim_stat(lruvec, file, 1);
273 }
274}
275
276#ifdef CONFIG_SMP
277static DEFINE_PER_CPU(struct pagevec, activate_page_pvecs);
278
279static void activate_page_drain(int cpu)
280{
281 struct pagevec *pvec = &per_cpu(activate_page_pvecs, cpu);
282
283 if (pagevec_count(pvec))
284 pagevec_lru_move_fn(pvec, __activate_page, NULL);
285}
286
287static bool need_activate_page_drain(int cpu)
288{
289 return pagevec_count(&per_cpu(activate_page_pvecs, cpu)) != 0;
290}
291
292void activate_page(struct page *page)
293{
294 if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
295 struct pagevec *pvec = &get_cpu_var(activate_page_pvecs);
296
297 get_page(page);
298 if (!pagevec_add(pvec, page))
299 pagevec_lru_move_fn(pvec, __activate_page, NULL);
300 put_cpu_var(activate_page_pvecs);
301 }
302}
303
304#else
305static inline void activate_page_drain(int cpu)
306{
307}
308
309static bool need_activate_page_drain(int cpu)
310{
311 return false;
312}
313
314void activate_page(struct page *page)
315{
316 struct zone *zone = page_zone(page);
317
318 spin_lock_irq(&zone->lru_lock);
319 __activate_page(page, mem_cgroup_page_lruvec(page, zone), NULL);
320 spin_unlock_irq(&zone->lru_lock);
321}
322#endif
323
324static void __lru_cache_activate_page(struct page *page)
325{
326 struct pagevec *pvec = &get_cpu_var(lru_add_pvec);
327 int i;
328
329 /*
330 * Search backwards on the optimistic assumption that the page being
331 * activated has just been added to this pagevec. Note that only
332 * the local pagevec is examined as a !PageLRU page could be in the
333 * process of being released, reclaimed, migrated or on a remote
334 * pagevec that is currently being drained. Furthermore, marking
335 * a remote pagevec's page PageActive potentially hits a race where
336 * a page is marked PageActive just after it is added to the inactive
337 * list causing accounting errors and BUG_ON checks to trigger.
338 */
339 for (i = pagevec_count(pvec) - 1; i >= 0; i--) {
340 struct page *pagevec_page = pvec->pages[i];
341
342 if (pagevec_page == page) {
343 SetPageActive(page);
344 break;
345 }
346 }
347
348 put_cpu_var(lru_add_pvec);
349}
350
351/*
352 * Mark a page as having seen activity.
353 *
354 * inactive,unreferenced -> inactive,referenced
355 * inactive,referenced -> active,unreferenced
356 * active,unreferenced -> active,referenced
357 *
358 * When a newly allocated page is not yet visible, so safe for non-atomic ops,
359 * __SetPageReferenced(page) may be substituted for mark_page_accessed(page).
360 */
361void mark_page_accessed(struct page *page)
362{
363 page = compound_head(page);
364 if (!PageActive(page) && !PageUnevictable(page) &&
365 PageReferenced(page)) {
366
367 /*
368 * If the page is on the LRU, queue it for activation via
369 * activate_page_pvecs. Otherwise, assume the page is on a
370 * pagevec, mark it active and it'll be moved to the active
371 * LRU on the next drain.
372 */
373 if (PageLRU(page))
374 activate_page(page);
375 else
376 __lru_cache_activate_page(page);
377 ClearPageReferenced(page);
378 if (page_is_file_cache(page))
379 workingset_activation(page);
380 } else if (!PageReferenced(page)) {
381 SetPageReferenced(page);
382 }
383 if (page_is_idle(page))
384 clear_page_idle(page);
385}
386EXPORT_SYMBOL(mark_page_accessed);
387
388static void __lru_cache_add(struct page *page)
389{
390 struct pagevec *pvec = &get_cpu_var(lru_add_pvec);
391
392 get_page(page);
393 if (!pagevec_space(pvec))
394 __pagevec_lru_add(pvec);
395 pagevec_add(pvec, page);
396 put_cpu_var(lru_add_pvec);
397}
398
399/**
400 * lru_cache_add: add a page to the page lists
401 * @page: the page to add
402 */
403void lru_cache_add_anon(struct page *page)
404{
405 if (PageActive(page))
406 ClearPageActive(page);
407 __lru_cache_add(page);
408}
409
410void lru_cache_add_file(struct page *page)
411{
412 if (PageActive(page))
413 ClearPageActive(page);
414 __lru_cache_add(page);
415}
416EXPORT_SYMBOL(lru_cache_add_file);
417
418/**
419 * lru_cache_add - add a page to a page list
420 * @page: the page to be added to the LRU.
421 *
422 * Queue the page for addition to the LRU via pagevec. The decision on whether
423 * to add the page to the [in]active [file|anon] list is deferred until the
424 * pagevec is drained. This gives a chance for the caller of lru_cache_add()
425 * have the page added to the active list using mark_page_accessed().
426 */
427void lru_cache_add(struct page *page)
428{
429 VM_BUG_ON_PAGE(PageActive(page) && PageUnevictable(page), page);
430 VM_BUG_ON_PAGE(PageLRU(page), page);
431 __lru_cache_add(page);
432}
433
434/**
435 * add_page_to_unevictable_list - add a page to the unevictable list
436 * @page: the page to be added to the unevictable list
437 *
438 * Add page directly to its zone's unevictable list. To avoid races with
439 * tasks that might be making the page evictable, through eg. munlock,
440 * munmap or exit, while it's not on the lru, we want to add the page
441 * while it's locked or otherwise "invisible" to other tasks. This is
442 * difficult to do when using the pagevec cache, so bypass that.
443 */
444void add_page_to_unevictable_list(struct page *page)
445{
446 struct zone *zone = page_zone(page);
447 struct lruvec *lruvec;
448
449 spin_lock_irq(&zone->lru_lock);
450 lruvec = mem_cgroup_page_lruvec(page, zone);
451 ClearPageActive(page);
452 SetPageUnevictable(page);
453 SetPageLRU(page);
454 add_page_to_lru_list(page, lruvec, LRU_UNEVICTABLE);
455 spin_unlock_irq(&zone->lru_lock);
456}
457
458/**
459 * lru_cache_add_active_or_unevictable
460 * @page: the page to be added to LRU
461 * @vma: vma in which page is mapped for determining reclaimability
462 *
463 * Place @page on the active or unevictable LRU list, depending on its
464 * evictability. Note that if the page is not evictable, it goes
465 * directly back onto it's zone's unevictable list, it does NOT use a
466 * per cpu pagevec.
467 */
468void lru_cache_add_active_or_unevictable(struct page *page,
469 struct vm_area_struct *vma)
470{
471 VM_BUG_ON_PAGE(PageLRU(page), page);
472
473 if (likely((vma->vm_flags & (VM_LOCKED | VM_SPECIAL)) != VM_LOCKED)) {
474 SetPageActive(page);
475 lru_cache_add(page);
476 return;
477 }
478
479 if (!TestSetPageMlocked(page)) {
480 /*
481 * We use the irq-unsafe __mod_zone_page_stat because this
482 * counter is not modified from interrupt context, and the pte
483 * lock is held(spinlock), which implies preemption disabled.
484 */
485 __mod_zone_page_state(page_zone(page), NR_MLOCK,
486 hpage_nr_pages(page));
487 count_vm_event(UNEVICTABLE_PGMLOCKED);
488 }
489 add_page_to_unevictable_list(page);
490}
491
492/*
493 * If the page can not be invalidated, it is moved to the
494 * inactive list to speed up its reclaim. It is moved to the
495 * head of the list, rather than the tail, to give the flusher
496 * threads some time to write it out, as this is much more
497 * effective than the single-page writeout from reclaim.
498 *
499 * If the page isn't page_mapped and dirty/writeback, the page
500 * could reclaim asap using PG_reclaim.
501 *
502 * 1. active, mapped page -> none
503 * 2. active, dirty/writeback page -> inactive, head, PG_reclaim
504 * 3. inactive, mapped page -> none
505 * 4. inactive, dirty/writeback page -> inactive, head, PG_reclaim
506 * 5. inactive, clean -> inactive, tail
507 * 6. Others -> none
508 *
509 * In 4, why it moves inactive's head, the VM expects the page would
510 * be write it out by flusher threads as this is much more effective
511 * than the single-page writeout from reclaim.
512 */
513static void lru_deactivate_file_fn(struct page *page, struct lruvec *lruvec,
514 void *arg)
515{
516 int lru, file;
517 bool active;
518
519 if (!PageLRU(page))
520 return;
521
522 if (PageUnevictable(page))
523 return;
524
525 /* Some processes are using the page */
526 if (page_mapped(page))
527 return;
528
529 active = PageActive(page);
530 file = page_is_file_cache(page);
531 lru = page_lru_base_type(page);
532
533 del_page_from_lru_list(page, lruvec, lru + active);
534 ClearPageActive(page);
535 ClearPageReferenced(page);
536 add_page_to_lru_list(page, lruvec, lru);
537
538 if (PageWriteback(page) || PageDirty(page)) {
539 /*
540 * PG_reclaim could be raced with end_page_writeback
541 * It can make readahead confusing. But race window
542 * is _really_ small and it's non-critical problem.
543 */
544 SetPageReclaim(page);
545 } else {
546 /*
547 * The page's writeback ends up during pagevec
548 * We moves tha page into tail of inactive.
549 */
550 list_move_tail(&page->lru, &lruvec->lists[lru]);
551 __count_vm_event(PGROTATED);
552 }
553
554 if (active)
555 __count_vm_event(PGDEACTIVATE);
556 update_page_reclaim_stat(lruvec, file, 0);
557}
558
559
560static void lru_deactivate_fn(struct page *page, struct lruvec *lruvec,
561 void *arg)
562{
563 if (PageLRU(page) && PageActive(page) && !PageUnevictable(page)) {
564 int file = page_is_file_cache(page);
565 int lru = page_lru_base_type(page);
566
567 del_page_from_lru_list(page, lruvec, lru + LRU_ACTIVE);
568 ClearPageActive(page);
569 ClearPageReferenced(page);
570 add_page_to_lru_list(page, lruvec, lru);
571
572 __count_vm_event(PGDEACTIVATE);
573 update_page_reclaim_stat(lruvec, file, 0);
574 }
575}
576
577/*
578 * Drain pages out of the cpu's pagevecs.
579 * Either "cpu" is the current CPU, and preemption has already been
580 * disabled; or "cpu" is being hot-unplugged, and is already dead.
581 */
582void lru_add_drain_cpu(int cpu)
583{
584 struct pagevec *pvec = &per_cpu(lru_add_pvec, cpu);
585
586 if (pagevec_count(pvec))
587 __pagevec_lru_add(pvec);
588
589 pvec = &per_cpu(lru_rotate_pvecs, cpu);
590 if (pagevec_count(pvec)) {
591 unsigned long flags;
592
593 /* No harm done if a racing interrupt already did this */
594 local_irq_save(flags);
595 pagevec_move_tail(pvec);
596 local_irq_restore(flags);
597 }
598
599 pvec = &per_cpu(lru_deactivate_file_pvecs, cpu);
600 if (pagevec_count(pvec))
601 pagevec_lru_move_fn(pvec, lru_deactivate_file_fn, NULL);
602
603 pvec = &per_cpu(lru_deactivate_pvecs, cpu);
604 if (pagevec_count(pvec))
605 pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
606
607 activate_page_drain(cpu);
608}
609
610/**
611 * deactivate_file_page - forcefully deactivate a file page
612 * @page: page to deactivate
613 *
614 * This function hints the VM that @page is a good reclaim candidate,
615 * for example if its invalidation fails due to the page being dirty
616 * or under writeback.
617 */
618void deactivate_file_page(struct page *page)
619{
620 /*
621 * In a workload with many unevictable page such as mprotect,
622 * unevictable page deactivation for accelerating reclaim is pointless.
623 */
624 if (PageUnevictable(page))
625 return;
626
627 if (likely(get_page_unless_zero(page))) {
628 struct pagevec *pvec = &get_cpu_var(lru_deactivate_file_pvecs);
629
630 if (!pagevec_add(pvec, page))
631 pagevec_lru_move_fn(pvec, lru_deactivate_file_fn, NULL);
632 put_cpu_var(lru_deactivate_file_pvecs);
633 }
634}
635
636/**
637 * deactivate_page - deactivate a page
638 * @page: page to deactivate
639 *
640 * deactivate_page() moves @page to the inactive list if @page was on the active
641 * list and was not an unevictable page. This is done to accelerate the reclaim
642 * of @page.
643 */
644void deactivate_page(struct page *page)
645{
646 if (PageLRU(page) && PageActive(page) && !PageUnevictable(page)) {
647 struct pagevec *pvec = &get_cpu_var(lru_deactivate_pvecs);
648
649 get_page(page);
650 if (!pagevec_add(pvec, page))
651 pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
652 put_cpu_var(lru_deactivate_pvecs);
653 }
654}
655
656void lru_add_drain(void)
657{
658 lru_add_drain_cpu(get_cpu());
659 put_cpu();
660}
661
662static void lru_add_drain_per_cpu(struct work_struct *dummy)
663{
664 lru_add_drain();
665}
666
667static DEFINE_PER_CPU(struct work_struct, lru_add_drain_work);
668
669void lru_add_drain_all(void)
670{
671 static DEFINE_MUTEX(lock);
672 static struct cpumask has_work;
673 int cpu;
674
675 mutex_lock(&lock);
676 get_online_cpus();
677 cpumask_clear(&has_work);
678
679 for_each_online_cpu(cpu) {
680 struct work_struct *work = &per_cpu(lru_add_drain_work, cpu);
681
682 if (pagevec_count(&per_cpu(lru_add_pvec, cpu)) ||
683 pagevec_count(&per_cpu(lru_rotate_pvecs, cpu)) ||
684 pagevec_count(&per_cpu(lru_deactivate_file_pvecs, cpu)) ||
685 pagevec_count(&per_cpu(lru_deactivate_pvecs, cpu)) ||
686 need_activate_page_drain(cpu)) {
687 INIT_WORK(work, lru_add_drain_per_cpu);
688 schedule_work_on(cpu, work);
689 cpumask_set_cpu(cpu, &has_work);
690 }
691 }
692
693 for_each_cpu(cpu, &has_work)
694 flush_work(&per_cpu(lru_add_drain_work, cpu));
695
696 put_online_cpus();
697 mutex_unlock(&lock);
698}
699
700/**
701 * release_pages - batched put_page()
702 * @pages: array of pages to release
703 * @nr: number of pages
704 * @cold: whether the pages are cache cold
705 *
706 * Decrement the reference count on all the pages in @pages. If it
707 * fell to zero, remove the page from the LRU and free it.
708 */
709void release_pages(struct page **pages, int nr, bool cold)
710{
711 int i;
712 LIST_HEAD(pages_to_free);
713 struct zone *zone = NULL;
714 struct lruvec *lruvec;
715 unsigned long uninitialized_var(flags);
716 unsigned int uninitialized_var(lock_batch);
717
718 for (i = 0; i < nr; i++) {
719 struct page *page = pages[i];
720
721 /*
722 * Make sure the IRQ-safe lock-holding time does not get
723 * excessive with a continuous string of pages from the
724 * same zone. The lock is held only if zone != NULL.
725 */
726 if (zone && ++lock_batch == SWAP_CLUSTER_MAX) {
727 spin_unlock_irqrestore(&zone->lru_lock, flags);
728 zone = NULL;
729 }
730
731 if (is_huge_zero_page(page)) {
732 put_huge_zero_page();
733 continue;
734 }
735
736 page = compound_head(page);
737 if (!put_page_testzero(page))
738 continue;
739
740 if (PageCompound(page)) {
741 if (zone) {
742 spin_unlock_irqrestore(&zone->lru_lock, flags);
743 zone = NULL;
744 }
745 __put_compound_page(page);
746 continue;
747 }
748
749 if (PageLRU(page)) {
750 struct zone *pagezone = page_zone(page);
751
752 if (pagezone != zone) {
753 if (zone)
754 spin_unlock_irqrestore(&zone->lru_lock,
755 flags);
756 lock_batch = 0;
757 zone = pagezone;
758 spin_lock_irqsave(&zone->lru_lock, flags);
759 }
760
761 lruvec = mem_cgroup_page_lruvec(page, zone);
762 VM_BUG_ON_PAGE(!PageLRU(page), page);
763 __ClearPageLRU(page);
764 del_page_from_lru_list(page, lruvec, page_off_lru(page));
765 }
766
767 /* Clear Active bit in case of parallel mark_page_accessed */
768 __ClearPageActive(page);
769
770 list_add(&page->lru, &pages_to_free);
771 }
772 if (zone)
773 spin_unlock_irqrestore(&zone->lru_lock, flags);
774
775 mem_cgroup_uncharge_list(&pages_to_free);
776 free_hot_cold_page_list(&pages_to_free, cold);
777}
778EXPORT_SYMBOL(release_pages);
779
780/*
781 * The pages which we're about to release may be in the deferred lru-addition
782 * queues. That would prevent them from really being freed right now. That's
783 * OK from a correctness point of view but is inefficient - those pages may be
784 * cache-warm and we want to give them back to the page allocator ASAP.
785 *
786 * So __pagevec_release() will drain those queues here. __pagevec_lru_add()
787 * and __pagevec_lru_add_active() call release_pages() directly to avoid
788 * mutual recursion.
789 */
790void __pagevec_release(struct pagevec *pvec)
791{
792 lru_add_drain();
793 release_pages(pvec->pages, pagevec_count(pvec), pvec->cold);
794 pagevec_reinit(pvec);
795}
796EXPORT_SYMBOL(__pagevec_release);
797
798#ifdef CONFIG_TRANSPARENT_HUGEPAGE
799/* used by __split_huge_page_refcount() */
800void lru_add_page_tail(struct page *page, struct page *page_tail,
801 struct lruvec *lruvec, struct list_head *list)
802{
803 const int file = 0;
804
805 VM_BUG_ON_PAGE(!PageHead(page), page);
806 VM_BUG_ON_PAGE(PageCompound(page_tail), page);
807 VM_BUG_ON_PAGE(PageLRU(page_tail), page);
808 VM_BUG_ON(NR_CPUS != 1 &&
809 !spin_is_locked(&lruvec_zone(lruvec)->lru_lock));
810
811 if (!list)
812 SetPageLRU(page_tail);
813
814 if (likely(PageLRU(page)))
815 list_add_tail(&page_tail->lru, &page->lru);
816 else if (list) {
817 /* page reclaim is reclaiming a huge page */
818 get_page(page_tail);
819 list_add_tail(&page_tail->lru, list);
820 } else {
821 struct list_head *list_head;
822 /*
823 * Head page has not yet been counted, as an hpage,
824 * so we must account for each subpage individually.
825 *
826 * Use the standard add function to put page_tail on the list,
827 * but then correct its position so they all end up in order.
828 */
829 add_page_to_lru_list(page_tail, lruvec, page_lru(page_tail));
830 list_head = page_tail->lru.prev;
831 list_move_tail(&page_tail->lru, list_head);
832 }
833
834 if (!PageUnevictable(page))
835 update_page_reclaim_stat(lruvec, file, PageActive(page_tail));
836}
837#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
838
839static void __pagevec_lru_add_fn(struct page *page, struct lruvec *lruvec,
840 void *arg)
841{
842 int file = page_is_file_cache(page);
843 int active = PageActive(page);
844 enum lru_list lru = page_lru(page);
845
846 VM_BUG_ON_PAGE(PageLRU(page), page);
847
848 SetPageLRU(page);
849 add_page_to_lru_list(page, lruvec, lru);
850 update_page_reclaim_stat(lruvec, file, active);
851 trace_mm_lru_insertion(page, lru);
852}
853
854/*
855 * Add the passed pages to the LRU, then drop the caller's refcount
856 * on them. Reinitialises the caller's pagevec.
857 */
858void __pagevec_lru_add(struct pagevec *pvec)
859{
860 pagevec_lru_move_fn(pvec, __pagevec_lru_add_fn, NULL);
861}
862EXPORT_SYMBOL(__pagevec_lru_add);
863
864/**
865 * pagevec_lookup_entries - gang pagecache lookup
866 * @pvec: Where the resulting entries are placed
867 * @mapping: The address_space to search
868 * @start: The starting entry index
869 * @nr_entries: The maximum number of entries
870 * @indices: The cache indices corresponding to the entries in @pvec
871 *
872 * pagevec_lookup_entries() will search for and return a group of up
873 * to @nr_entries pages and shadow entries in the mapping. All
874 * entries are placed in @pvec. pagevec_lookup_entries() takes a
875 * reference against actual pages in @pvec.
876 *
877 * The search returns a group of mapping-contiguous entries with
878 * ascending indexes. There may be holes in the indices due to
879 * not-present entries.
880 *
881 * pagevec_lookup_entries() returns the number of entries which were
882 * found.
883 */
884unsigned pagevec_lookup_entries(struct pagevec *pvec,
885 struct address_space *mapping,
886 pgoff_t start, unsigned nr_pages,
887 pgoff_t *indices)
888{
889 pvec->nr = find_get_entries(mapping, start, nr_pages,
890 pvec->pages, indices);
891 return pagevec_count(pvec);
892}
893
894/**
895 * pagevec_remove_exceptionals - pagevec exceptionals pruning
896 * @pvec: The pagevec to prune
897 *
898 * pagevec_lookup_entries() fills both pages and exceptional radix
899 * tree entries into the pagevec. This function prunes all
900 * exceptionals from @pvec without leaving holes, so that it can be
901 * passed on to page-only pagevec operations.
902 */
903void pagevec_remove_exceptionals(struct pagevec *pvec)
904{
905 int i, j;
906
907 for (i = 0, j = 0; i < pagevec_count(pvec); i++) {
908 struct page *page = pvec->pages[i];
909 if (!radix_tree_exceptional_entry(page))
910 pvec->pages[j++] = page;
911 }
912 pvec->nr = j;
913}
914
915/**
916 * pagevec_lookup - gang pagecache lookup
917 * @pvec: Where the resulting pages are placed
918 * @mapping: The address_space to search
919 * @start: The starting page index
920 * @nr_pages: The maximum number of pages
921 *
922 * pagevec_lookup() will search for and return a group of up to @nr_pages pages
923 * in the mapping. The pages are placed in @pvec. pagevec_lookup() takes a
924 * reference against the pages in @pvec.
925 *
926 * The search returns a group of mapping-contiguous pages with ascending
927 * indexes. There may be holes in the indices due to not-present pages.
928 *
929 * pagevec_lookup() returns the number of pages which were found.
930 */
931unsigned pagevec_lookup(struct pagevec *pvec, struct address_space *mapping,
932 pgoff_t start, unsigned nr_pages)
933{
934 pvec->nr = find_get_pages(mapping, start, nr_pages, pvec->pages);
935 return pagevec_count(pvec);
936}
937EXPORT_SYMBOL(pagevec_lookup);
938
939unsigned pagevec_lookup_tag(struct pagevec *pvec, struct address_space *mapping,
940 pgoff_t *index, int tag, unsigned nr_pages)
941{
942 pvec->nr = find_get_pages_tag(mapping, index, tag,
943 nr_pages, pvec->pages);
944 return pagevec_count(pvec);
945}
946EXPORT_SYMBOL(pagevec_lookup_tag);
947
948/*
949 * Perform any setup for the swap system
950 */
951void __init swap_setup(void)
952{
953 unsigned long megs = totalram_pages >> (20 - PAGE_SHIFT);
954#ifdef CONFIG_SWAP
955 int i;
956
957 for (i = 0; i < MAX_SWAPFILES; i++)
958 spin_lock_init(&swapper_spaces[i].tree_lock);
959#endif
960
961 /* Use a smaller cluster for small-memory machines */
962 if (megs < 16)
963 page_cluster = 2;
964 else
965 page_cluster = 3;
966 /*
967 * Right now other parts of the system means that we
968 * _really_ don't want to cluster much more
969 */
970}