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