Linux Audio

Check our new training course

Loading...
v4.6
  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}
v3.5.6
  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/percpu.h>
 28#include <linux/cpu.h>
 29#include <linux/notifier.h>
 30#include <linux/backing-dev.h>
 31#include <linux/memcontrol.h>
 32#include <linux/gfp.h>
 
 
 
 33
 34#include "internal.h"
 35
 
 
 
 36/* How many pages do we try to swap or page in/out together? */
 37int page_cluster;
 38
 39static DEFINE_PER_CPU(struct pagevec[NR_LRU_LISTS], lru_add_pvecs);
 40static DEFINE_PER_CPU(struct pagevec, lru_rotate_pvecs);
 
 41static DEFINE_PER_CPU(struct pagevec, lru_deactivate_pvecs);
 42
 43/*
 44 * This path almost never happens for VM activity - pages are normally
 45 * freed via pagevecs.  But it gets used by networking.
 46 */
 47static void __page_cache_release(struct page *page)
 48{
 49	if (PageLRU(page)) {
 50		struct zone *zone = page_zone(page);
 51		struct lruvec *lruvec;
 52		unsigned long flags;
 53
 54		spin_lock_irqsave(&zone->lru_lock, flags);
 55		lruvec = mem_cgroup_page_lruvec(page, zone);
 56		VM_BUG_ON(!PageLRU(page));
 57		__ClearPageLRU(page);
 58		del_page_from_lru_list(page, lruvec, page_off_lru(page));
 59		spin_unlock_irqrestore(&zone->lru_lock, flags);
 60	}
 
 61}
 62
 63static void __put_single_page(struct page *page)
 64{
 65	__page_cache_release(page);
 66	free_hot_cold_page(page, 0);
 67}
 68
 69static void __put_compound_page(struct page *page)
 70{
 71	compound_page_dtor *dtor;
 72
 73	__page_cache_release(page);
 
 
 
 
 
 
 
 74	dtor = get_compound_page_dtor(page);
 75	(*dtor)(page);
 76}
 77
 78static void put_compound_page(struct page *page)
 79{
 80	if (unlikely(PageTail(page))) {
 81		/* __split_huge_page_refcount can run under us */
 82		struct page *page_head = compound_trans_head(page);
 83
 84		if (likely(page != page_head &&
 85			   get_page_unless_zero(page_head))) {
 86			unsigned long flags;
 87
 88			/*
 89			 * THP can not break up slab pages so avoid taking
 90			 * compound_lock().  Slab performs non-atomic bit ops
 91			 * on page->flags for better performance.  In particular
 92			 * slab_unlock() in slub used to be a hot path.  It is
 93			 * still hot on arches that do not support
 94			 * this_cpu_cmpxchg_double().
 95			 */
 96			if (PageSlab(page_head)) {
 97				if (PageTail(page)) {
 98					if (put_page_testzero(page_head))
 99						VM_BUG_ON(1);
100
101					atomic_dec(&page->_mapcount);
102					goto skip_lock_tail;
103				} else
104					goto skip_lock;
105			}
106			/*
107			 * page_head wasn't a dangling pointer but it
108			 * may not be a head page anymore by the time
109			 * we obtain the lock. That is ok as long as it
110			 * can't be freed from under us.
111			 */
112			flags = compound_lock_irqsave(page_head);
113			if (unlikely(!PageTail(page))) {
114				/* __split_huge_page_refcount run before us */
115				compound_unlock_irqrestore(page_head, flags);
116skip_lock:
117				if (put_page_testzero(page_head))
118					__put_single_page(page_head);
119out_put_single:
120				if (put_page_testzero(page))
121					__put_single_page(page);
122				return;
123			}
124			VM_BUG_ON(page_head != page->first_page);
125			/*
126			 * We can release the refcount taken by
127			 * get_page_unless_zero() now that
128			 * __split_huge_page_refcount() is blocked on
129			 * the compound_lock.
130			 */
131			if (put_page_testzero(page_head))
132				VM_BUG_ON(1);
133			/* __split_huge_page_refcount will wait now */
134			VM_BUG_ON(page_mapcount(page) <= 0);
135			atomic_dec(&page->_mapcount);
136			VM_BUG_ON(atomic_read(&page_head->_count) <= 0);
137			VM_BUG_ON(atomic_read(&page->_count) != 0);
138			compound_unlock_irqrestore(page_head, flags);
139
140skip_lock_tail:
141			if (put_page_testzero(page_head)) {
142				if (PageHead(page_head))
143					__put_compound_page(page_head);
144				else
145					__put_single_page(page_head);
146			}
147		} else {
148			/* page_head is a dangling pointer */
149			VM_BUG_ON(PageTail(page));
150			goto out_put_single;
151		}
152	} else if (put_page_testzero(page)) {
153		if (PageHead(page))
154			__put_compound_page(page);
155		else
156			__put_single_page(page);
157	}
158}
159
160void put_page(struct page *page)
161{
162	if (unlikely(PageCompound(page)))
163		put_compound_page(page);
164	else if (put_page_testzero(page))
165		__put_single_page(page);
166}
167EXPORT_SYMBOL(put_page);
168
169/*
170 * This function is exported but must not be called by anything other
171 * than get_page(). It implements the slow path of get_page().
172 */
173bool __get_page_tail(struct page *page)
174{
175	/*
176	 * This takes care of get_page() if run on a tail page
177	 * returned by one of the get_user_pages/follow_page variants.
178	 * get_user_pages/follow_page itself doesn't need the compound
179	 * lock because it runs __get_page_tail_foll() under the
180	 * proper PT lock that already serializes against
181	 * split_huge_page().
182	 */
183	unsigned long flags;
184	bool got = false;
185	struct page *page_head = compound_trans_head(page);
186
187	if (likely(page != page_head && get_page_unless_zero(page_head))) {
188
189		/* Ref to put_compound_page() comment. */
190		if (PageSlab(page_head)) {
191			if (likely(PageTail(page))) {
192				__get_page_tail_foll(page, false);
193				return true;
194			} else {
195				put_page(page_head);
196				return false;
197			}
198		}
199
200		/*
201		 * page_head wasn't a dangling pointer but it
202		 * may not be a head page anymore by the time
203		 * we obtain the lock. That is ok as long as it
204		 * can't be freed from under us.
205		 */
206		flags = compound_lock_irqsave(page_head);
207		/* here __split_huge_page_refcount won't run anymore */
208		if (likely(PageTail(page))) {
209			__get_page_tail_foll(page, false);
210			got = true;
211		}
212		compound_unlock_irqrestore(page_head, flags);
213		if (unlikely(!got))
214			put_page(page_head);
215	}
216	return got;
217}
218EXPORT_SYMBOL(__get_page_tail);
219
220/**
221 * put_pages_list() - release a list of pages
222 * @pages: list of pages threaded on page->lru
223 *
224 * Release a list of pages which are strung together on page.lru.  Currently
225 * used by read_cache_pages() and related error recovery code.
226 */
227void put_pages_list(struct list_head *pages)
228{
229	while (!list_empty(pages)) {
230		struct page *victim;
231
232		victim = list_entry(pages->prev, struct page, lru);
233		list_del(&victim->lru);
234		page_cache_release(victim);
235	}
236}
237EXPORT_SYMBOL(put_pages_list);
238
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
239static void pagevec_lru_move_fn(struct pagevec *pvec,
240	void (*move_fn)(struct page *page, struct lruvec *lruvec, void *arg),
241	void *arg)
242{
243	int i;
244	struct zone *zone = NULL;
245	struct lruvec *lruvec;
246	unsigned long flags = 0;
247
248	for (i = 0; i < pagevec_count(pvec); i++) {
249		struct page *page = pvec->pages[i];
250		struct zone *pagezone = page_zone(page);
251
252		if (pagezone != zone) {
253			if (zone)
254				spin_unlock_irqrestore(&zone->lru_lock, flags);
255			zone = pagezone;
256			spin_lock_irqsave(&zone->lru_lock, flags);
257		}
258
259		lruvec = mem_cgroup_page_lruvec(page, zone);
260		(*move_fn)(page, lruvec, arg);
261	}
262	if (zone)
263		spin_unlock_irqrestore(&zone->lru_lock, flags);
264	release_pages(pvec->pages, pvec->nr, pvec->cold);
265	pagevec_reinit(pvec);
266}
267
268static void pagevec_move_tail_fn(struct page *page, struct lruvec *lruvec,
269				 void *arg)
270{
271	int *pgmoved = arg;
272
273	if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
274		enum lru_list lru = page_lru_base_type(page);
275		list_move_tail(&page->lru, &lruvec->lists[lru]);
276		(*pgmoved)++;
277	}
278}
279
280/*
281 * pagevec_move_tail() must be called with IRQ disabled.
282 * Otherwise this may cause nasty races.
283 */
284static void pagevec_move_tail(struct pagevec *pvec)
285{
286	int pgmoved = 0;
287
288	pagevec_lru_move_fn(pvec, pagevec_move_tail_fn, &pgmoved);
289	__count_vm_events(PGROTATED, pgmoved);
290}
291
292/*
293 * Writeback is about to end against a page which has been marked for immediate
294 * reclaim.  If it still appears to be reclaimable, move it to the tail of the
295 * inactive list.
296 */
297void rotate_reclaimable_page(struct page *page)
298{
299	if (!PageLocked(page) && !PageDirty(page) && !PageActive(page) &&
300	    !PageUnevictable(page) && PageLRU(page)) {
301		struct pagevec *pvec;
302		unsigned long flags;
303
304		page_cache_get(page);
305		local_irq_save(flags);
306		pvec = &__get_cpu_var(lru_rotate_pvecs);
307		if (!pagevec_add(pvec, page))
308			pagevec_move_tail(pvec);
309		local_irq_restore(flags);
310	}
311}
312
313static void update_page_reclaim_stat(struct lruvec *lruvec,
314				     int file, int rotated)
315{
316	struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
317
318	reclaim_stat->recent_scanned[file]++;
319	if (rotated)
320		reclaim_stat->recent_rotated[file]++;
321}
322
323static void __activate_page(struct page *page, struct lruvec *lruvec,
324			    void *arg)
325{
326	if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
327		int file = page_is_file_cache(page);
328		int lru = page_lru_base_type(page);
329
330		del_page_from_lru_list(page, lruvec, lru);
331		SetPageActive(page);
332		lru += LRU_ACTIVE;
333		add_page_to_lru_list(page, lruvec, lru);
 
334
335		__count_vm_event(PGACTIVATE);
336		update_page_reclaim_stat(lruvec, file, 1);
337	}
338}
339
340#ifdef CONFIG_SMP
341static DEFINE_PER_CPU(struct pagevec, activate_page_pvecs);
342
343static void activate_page_drain(int cpu)
344{
345	struct pagevec *pvec = &per_cpu(activate_page_pvecs, cpu);
346
347	if (pagevec_count(pvec))
348		pagevec_lru_move_fn(pvec, __activate_page, NULL);
349}
350
 
 
 
 
 
351void activate_page(struct page *page)
352{
353	if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
354		struct pagevec *pvec = &get_cpu_var(activate_page_pvecs);
355
356		page_cache_get(page);
357		if (!pagevec_add(pvec, page))
358			pagevec_lru_move_fn(pvec, __activate_page, NULL);
359		put_cpu_var(activate_page_pvecs);
360	}
361}
362
363#else
364static inline void activate_page_drain(int cpu)
365{
366}
367
 
 
 
 
 
368void activate_page(struct page *page)
369{
370	struct zone *zone = page_zone(page);
371
372	spin_lock_irq(&zone->lru_lock);
373	__activate_page(page, mem_cgroup_page_lruvec(page, zone), NULL);
374	spin_unlock_irq(&zone->lru_lock);
375}
376#endif
377
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
378/*
379 * Mark a page as having seen activity.
380 *
381 * inactive,unreferenced	->	inactive,referenced
382 * inactive,referenced		->	active,unreferenced
383 * active,unreferenced		->	active,referenced
 
 
 
384 */
385void mark_page_accessed(struct page *page)
386{
 
387	if (!PageActive(page) && !PageUnevictable(page) &&
388			PageReferenced(page) && PageLRU(page)) {
389		activate_page(page);
 
 
 
 
 
 
 
 
 
 
390		ClearPageReferenced(page);
 
 
391	} else if (!PageReferenced(page)) {
392		SetPageReferenced(page);
393	}
 
 
394}
395EXPORT_SYMBOL(mark_page_accessed);
396
397void __lru_cache_add(struct page *page, enum lru_list lru)
398{
399	struct pagevec *pvec = &get_cpu_var(lru_add_pvecs)[lru];
400
401	page_cache_get(page);
402	if (!pagevec_add(pvec, page))
403		__pagevec_lru_add(pvec, lru);
404	put_cpu_var(lru_add_pvecs);
 
405}
406EXPORT_SYMBOL(__lru_cache_add);
407
408/**
409 * lru_cache_add_lru - add a page to a page list
410 * @page: the page to be added to the LRU.
411 * @lru: the LRU list to which the page is added.
412 */
413void lru_cache_add_lru(struct page *page, enum lru_list lru)
 
 
 
 
 
 
 
414{
415	if (PageActive(page)) {
416		VM_BUG_ON(PageUnevictable(page));
417		ClearPageActive(page);
418	} else if (PageUnevictable(page)) {
419		VM_BUG_ON(PageActive(page));
420		ClearPageUnevictable(page);
421	}
422
423	VM_BUG_ON(PageLRU(page) || PageActive(page) || PageUnevictable(page));
424	__lru_cache_add(page, lru);
 
 
 
 
 
 
 
 
 
 
 
 
425}
426
427/**
428 * add_page_to_unevictable_list - add a page to the unevictable list
429 * @page:  the page to be added to the unevictable list
430 *
431 * Add page directly to its zone's unevictable list.  To avoid races with
432 * tasks that might be making the page evictable, through eg. munlock,
433 * munmap or exit, while it's not on the lru, we want to add the page
434 * while it's locked or otherwise "invisible" to other tasks.  This is
435 * difficult to do when using the pagevec cache, so bypass that.
436 */
437void add_page_to_unevictable_list(struct page *page)
438{
439	struct zone *zone = page_zone(page);
440	struct lruvec *lruvec;
441
442	spin_lock_irq(&zone->lru_lock);
443	lruvec = mem_cgroup_page_lruvec(page, zone);
 
444	SetPageUnevictable(page);
445	SetPageLRU(page);
446	add_page_to_lru_list(page, lruvec, LRU_UNEVICTABLE);
447	spin_unlock_irq(&zone->lru_lock);
448}
449
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
450/*
451 * If the page can not be invalidated, it is moved to the
452 * inactive list to speed up its reclaim.  It is moved to the
453 * head of the list, rather than the tail, to give the flusher
454 * threads some time to write it out, as this is much more
455 * effective than the single-page writeout from reclaim.
456 *
457 * If the page isn't page_mapped and dirty/writeback, the page
458 * could reclaim asap using PG_reclaim.
459 *
460 * 1. active, mapped page -> none
461 * 2. active, dirty/writeback page -> inactive, head, PG_reclaim
462 * 3. inactive, mapped page -> none
463 * 4. inactive, dirty/writeback page -> inactive, head, PG_reclaim
464 * 5. inactive, clean -> inactive, tail
465 * 6. Others -> none
466 *
467 * In 4, why it moves inactive's head, the VM expects the page would
468 * be write it out by flusher threads as this is much more effective
469 * than the single-page writeout from reclaim.
470 */
471static void lru_deactivate_fn(struct page *page, struct lruvec *lruvec,
472			      void *arg)
473{
474	int lru, file;
475	bool active;
476
477	if (!PageLRU(page))
478		return;
479
480	if (PageUnevictable(page))
481		return;
482
483	/* Some processes are using the page */
484	if (page_mapped(page))
485		return;
486
487	active = PageActive(page);
488	file = page_is_file_cache(page);
489	lru = page_lru_base_type(page);
490
491	del_page_from_lru_list(page, lruvec, lru + active);
492	ClearPageActive(page);
493	ClearPageReferenced(page);
494	add_page_to_lru_list(page, lruvec, lru);
495
496	if (PageWriteback(page) || PageDirty(page)) {
497		/*
498		 * PG_reclaim could be raced with end_page_writeback
499		 * It can make readahead confusing.  But race window
500		 * is _really_ small and  it's non-critical problem.
501		 */
502		SetPageReclaim(page);
503	} else {
504		/*
505		 * The page's writeback ends up during pagevec
506		 * We moves tha page into tail of inactive.
507		 */
508		list_move_tail(&page->lru, &lruvec->lists[lru]);
509		__count_vm_event(PGROTATED);
510	}
511
512	if (active)
513		__count_vm_event(PGDEACTIVATE);
514	update_page_reclaim_stat(lruvec, file, 0);
515}
516
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
517/*
518 * Drain pages out of the cpu's pagevecs.
519 * Either "cpu" is the current CPU, and preemption has already been
520 * disabled; or "cpu" is being hot-unplugged, and is already dead.
521 */
522void lru_add_drain_cpu(int cpu)
523{
524	struct pagevec *pvecs = per_cpu(lru_add_pvecs, cpu);
525	struct pagevec *pvec;
526	int lru;
527
528	for_each_lru(lru) {
529		pvec = &pvecs[lru - LRU_BASE];
530		if (pagevec_count(pvec))
531			__pagevec_lru_add(pvec, lru);
532	}
533
534	pvec = &per_cpu(lru_rotate_pvecs, cpu);
535	if (pagevec_count(pvec)) {
536		unsigned long flags;
537
538		/* No harm done if a racing interrupt already did this */
539		local_irq_save(flags);
540		pagevec_move_tail(pvec);
541		local_irq_restore(flags);
542	}
543
 
 
 
 
544	pvec = &per_cpu(lru_deactivate_pvecs, cpu);
545	if (pagevec_count(pvec))
546		pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
547
548	activate_page_drain(cpu);
549}
550
551/**
552 * deactivate_page - forcefully deactivate a page
553 * @page: page to deactivate
554 *
555 * This function hints the VM that @page is a good reclaim candidate,
556 * for example if its invalidation fails due to the page being dirty
557 * or under writeback.
558 */
559void deactivate_page(struct page *page)
560{
561	/*
562	 * In a workload with many unevictable page such as mprotect, unevictable
563	 * page deactivation for accelerating reclaim is pointless.
564	 */
565	if (PageUnevictable(page))
566		return;
567
568	if (likely(get_page_unless_zero(page))) {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
569		struct pagevec *pvec = &get_cpu_var(lru_deactivate_pvecs);
570
 
571		if (!pagevec_add(pvec, page))
572			pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
573		put_cpu_var(lru_deactivate_pvecs);
574	}
575}
576
577void lru_add_drain(void)
578{
579	lru_add_drain_cpu(get_cpu());
580	put_cpu();
581}
582
583static void lru_add_drain_per_cpu(struct work_struct *dummy)
584{
585	lru_add_drain();
586}
587
588/*
589 * Returns 0 for success
590 */
591int lru_add_drain_all(void)
592{
593	return schedule_on_each_cpu(lru_add_drain_per_cpu);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
594}
595
596/*
597 * Batched page_cache_release().  Decrement the reference count on all the
598 * passed pages.  If it fell to zero then remove the page from the LRU and
599 * free it.
600 *
601 * Avoid taking zone->lru_lock if possible, but if it is taken, retain it
602 * for the remainder of the operation.
603 *
604 * The locking in this function is against shrink_inactive_list(): we recheck
605 * the page count inside the lock to see whether shrink_inactive_list()
606 * grabbed the page via the LRU.  If it did, give up: shrink_inactive_list()
607 * will free it.
608 */
609void release_pages(struct page **pages, int nr, int cold)
610{
611	int i;
612	LIST_HEAD(pages_to_free);
613	struct zone *zone = NULL;
614	struct lruvec *lruvec;
615	unsigned long uninitialized_var(flags);
 
616
617	for (i = 0; i < nr; i++) {
618		struct page *page = pages[i];
619
620		if (unlikely(PageCompound(page))) {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
621			if (zone) {
622				spin_unlock_irqrestore(&zone->lru_lock, flags);
623				zone = NULL;
624			}
625			put_compound_page(page);
626			continue;
627		}
628
629		if (!put_page_testzero(page))
630			continue;
631
632		if (PageLRU(page)) {
633			struct zone *pagezone = page_zone(page);
634
635			if (pagezone != zone) {
636				if (zone)
637					spin_unlock_irqrestore(&zone->lru_lock,
638									flags);
 
639				zone = pagezone;
640				spin_lock_irqsave(&zone->lru_lock, flags);
641			}
642
643			lruvec = mem_cgroup_page_lruvec(page, zone);
644			VM_BUG_ON(!PageLRU(page));
645			__ClearPageLRU(page);
646			del_page_from_lru_list(page, lruvec, page_off_lru(page));
647		}
648
 
 
 
649		list_add(&page->lru, &pages_to_free);
650	}
651	if (zone)
652		spin_unlock_irqrestore(&zone->lru_lock, flags);
653
 
654	free_hot_cold_page_list(&pages_to_free, cold);
655}
656EXPORT_SYMBOL(release_pages);
657
658/*
659 * The pages which we're about to release may be in the deferred lru-addition
660 * queues.  That would prevent them from really being freed right now.  That's
661 * OK from a correctness point of view but is inefficient - those pages may be
662 * cache-warm and we want to give them back to the page allocator ASAP.
663 *
664 * So __pagevec_release() will drain those queues here.  __pagevec_lru_add()
665 * and __pagevec_lru_add_active() call release_pages() directly to avoid
666 * mutual recursion.
667 */
668void __pagevec_release(struct pagevec *pvec)
669{
670	lru_add_drain();
671	release_pages(pvec->pages, pagevec_count(pvec), pvec->cold);
672	pagevec_reinit(pvec);
673}
674EXPORT_SYMBOL(__pagevec_release);
675
676#ifdef CONFIG_TRANSPARENT_HUGEPAGE
677/* used by __split_huge_page_refcount() */
678void lru_add_page_tail(struct page *page, struct page *page_tail,
679		       struct lruvec *lruvec)
680{
681	int uninitialized_var(active);
682	enum lru_list lru;
683	const int file = 0;
684
685	VM_BUG_ON(!PageHead(page));
686	VM_BUG_ON(PageCompound(page_tail));
687	VM_BUG_ON(PageLRU(page_tail));
688	VM_BUG_ON(NR_CPUS != 1 &&
689		  !spin_is_locked(&lruvec_zone(lruvec)->lru_lock));
690
691	SetPageLRU(page_tail);
692
693	if (page_evictable(page_tail, NULL)) {
694		if (PageActive(page)) {
695			SetPageActive(page_tail);
696			active = 1;
697			lru = LRU_ACTIVE_ANON;
698		} else {
699			active = 0;
700			lru = LRU_INACTIVE_ANON;
701		}
702	} else {
703		SetPageUnevictable(page_tail);
704		lru = LRU_UNEVICTABLE;
705	}
706
707	if (likely(PageLRU(page)))
708		list_add_tail(&page_tail->lru, &page->lru);
709	else {
 
 
 
 
710		struct list_head *list_head;
711		/*
712		 * Head page has not yet been counted, as an hpage,
713		 * so we must account for each subpage individually.
714		 *
715		 * Use the standard add function to put page_tail on the list,
716		 * but then correct its position so they all end up in order.
717		 */
718		add_page_to_lru_list(page_tail, lruvec, lru);
719		list_head = page_tail->lru.prev;
720		list_move_tail(&page_tail->lru, list_head);
721	}
722
723	if (!PageUnevictable(page))
724		update_page_reclaim_stat(lruvec, file, active);
725}
726#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
727
728static void __pagevec_lru_add_fn(struct page *page, struct lruvec *lruvec,
729				 void *arg)
730{
731	enum lru_list lru = (enum lru_list)arg;
732	int file = is_file_lru(lru);
733	int active = is_active_lru(lru);
734
735	VM_BUG_ON(PageActive(page));
736	VM_BUG_ON(PageUnevictable(page));
737	VM_BUG_ON(PageLRU(page));
738
739	SetPageLRU(page);
740	if (active)
741		SetPageActive(page);
742	add_page_to_lru_list(page, lruvec, lru);
743	update_page_reclaim_stat(lruvec, file, active);
 
744}
745
746/*
747 * Add the passed pages to the LRU, then drop the caller's refcount
748 * on them.  Reinitialises the caller's pagevec.
749 */
750void __pagevec_lru_add(struct pagevec *pvec, enum lru_list lru)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
751{
752	VM_BUG_ON(is_unevictable_lru(lru));
753
754	pagevec_lru_move_fn(pvec, __pagevec_lru_add_fn, (void *)lru);
 
 
 
 
 
755}
756EXPORT_SYMBOL(__pagevec_lru_add);
757
758/**
759 * pagevec_lookup - gang pagecache lookup
760 * @pvec:	Where the resulting pages are placed
761 * @mapping:	The address_space to search
762 * @start:	The starting page index
763 * @nr_pages:	The maximum number of pages
764 *
765 * pagevec_lookup() will search for and return a group of up to @nr_pages pages
766 * in the mapping.  The pages are placed in @pvec.  pagevec_lookup() takes a
767 * reference against the pages in @pvec.
768 *
769 * The search returns a group of mapping-contiguous pages with ascending
770 * indexes.  There may be holes in the indices due to not-present pages.
771 *
772 * pagevec_lookup() returns the number of pages which were found.
773 */
774unsigned pagevec_lookup(struct pagevec *pvec, struct address_space *mapping,
775		pgoff_t start, unsigned nr_pages)
776{
777	pvec->nr = find_get_pages(mapping, start, nr_pages, pvec->pages);
778	return pagevec_count(pvec);
779}
780EXPORT_SYMBOL(pagevec_lookup);
781
782unsigned pagevec_lookup_tag(struct pagevec *pvec, struct address_space *mapping,
783		pgoff_t *index, int tag, unsigned nr_pages)
784{
785	pvec->nr = find_get_pages_tag(mapping, index, tag,
786					nr_pages, pvec->pages);
787	return pagevec_count(pvec);
788}
789EXPORT_SYMBOL(pagevec_lookup_tag);
790
791/*
792 * Perform any setup for the swap system
793 */
794void __init swap_setup(void)
795{
796	unsigned long megs = totalram_pages >> (20 - PAGE_SHIFT);
 
 
797
798#ifdef CONFIG_SWAP
799	bdi_init(swapper_space.backing_dev_info);
800#endif
801
802	/* Use a smaller cluster for small-memory machines */
803	if (megs < 16)
804		page_cluster = 2;
805	else
806		page_cluster = 3;
807	/*
808	 * Right now other parts of the system means that we
809	 * _really_ don't want to cluster much more
810	 */
811}