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