swap.c revision 3e2f41f1f64744f7942980d93cc93dd3e5924560 
1/*
2 *  linux/mm/swap.c
3 *
4 *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5 */
6
7/*
8 * This file contains the default values for the operation of the
9 * Linux VM subsystem. Fine-tuning documentation can be found in
10 * Documentation/sysctl/vm.txt.
11 * Started 18.12.91
12 * Swap aging added 23.2.95, Stephen Tweedie.
13 * Buffermem limits added 12.3.98, Rik van Riel.
14 */
15
16#include <linux/mm.h>
17#include <linux/sched.h>
18#include <linux/kernel_stat.h>
19#include <linux/swap.h>
20#include <linux/mman.h>
21#include <linux/pagemap.h>
22#include <linux/pagevec.h>
23#include <linux/init.h>
24#include <linux/module.h>
25#include <linux/mm_inline.h>
26#include <linux/buffer_head.h>	/* for try_to_release_page() */
27#include <linux/percpu_counter.h>
28#include <linux/percpu.h>
29#include <linux/cpu.h>
30#include <linux/notifier.h>
31#include <linux/backing-dev.h>
32#include <linux/memcontrol.h>
33
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);
41
42/*
43 * This path almost never happens for VM activity - pages are normally
44 * freed via pagevecs.  But it gets used by networking.
45 */
46static void __page_cache_release(struct page *page)
47{
48	if (PageLRU(page)) {
49		unsigned long flags;
50		struct zone *zone = page_zone(page);
51
52		spin_lock_irqsave(&zone->lru_lock, flags);
53		VM_BUG_ON(!PageLRU(page));
54		__ClearPageLRU(page);
55		del_page_from_lru(zone, page);
56		spin_unlock_irqrestore(&zone->lru_lock, flags);
57	}
58	free_hot_page(page);
59}
60
61static void put_compound_page(struct page *page)
62{
63	page = compound_head(page);
64	if (put_page_testzero(page)) {
65		compound_page_dtor *dtor;
66
67		dtor = get_compound_page_dtor(page);
68		(*dtor)(page);
69	}
70}
71
72void put_page(struct page *page)
73{
74	if (unlikely(PageCompound(page)))
75		put_compound_page(page);
76	else if (put_page_testzero(page))
77		__page_cache_release(page);
78}
79EXPORT_SYMBOL(put_page);
80
81/**
82 * put_pages_list() - release a list of pages
83 * @pages: list of pages threaded on page->lru
84 *
85 * Release a list of pages which are strung together on page.lru.  Currently
86 * used by read_cache_pages() and related error recovery code.
87 */
88void put_pages_list(struct list_head *pages)
89{
90	while (!list_empty(pages)) {
91		struct page *victim;
92
93		victim = list_entry(pages->prev, struct page, lru);
94		list_del(&victim->lru);
95		page_cache_release(victim);
96	}
97}
98EXPORT_SYMBOL(put_pages_list);
99
100/*
101 * pagevec_move_tail() must be called with IRQ disabled.
102 * Otherwise this may cause nasty races.
103 */
104static void pagevec_move_tail(struct pagevec *pvec)
105{
106	int i;
107	int pgmoved = 0;
108	struct zone *zone = NULL;
109
110	for (i = 0; i < pagevec_count(pvec); i++) {
111		struct page *page = pvec->pages[i];
112		struct zone *pagezone = page_zone(page);
113
114		if (pagezone != zone) {
115			if (zone)
116				spin_unlock(&zone->lru_lock);
117			zone = pagezone;
118			spin_lock(&zone->lru_lock);
119		}
120		if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
121			int lru = page_is_file_cache(page);
122			list_move_tail(&page->lru, &zone->lru[lru].list);
123			pgmoved++;
124		}
125	}
126	if (zone)
127		spin_unlock(&zone->lru_lock);
128	__count_vm_events(PGROTATED, pgmoved);
129	release_pages(pvec->pages, pvec->nr, pvec->cold);
130	pagevec_reinit(pvec);
131}
132
133/*
134 * Writeback is about to end against a page which has been marked for immediate
135 * reclaim.  If it still appears to be reclaimable, move it to the tail of the
136 * inactive list.
137 */
138void  rotate_reclaimable_page(struct page *page)
139{
140	if (!PageLocked(page) && !PageDirty(page) && !PageActive(page) &&
141	    !PageUnevictable(page) && PageLRU(page)) {
142		struct pagevec *pvec;
143		unsigned long flags;
144
145		page_cache_get(page);
146		local_irq_save(flags);
147		pvec = &__get_cpu_var(lru_rotate_pvecs);
148		if (!pagevec_add(pvec, page))
149			pagevec_move_tail(pvec);
150		local_irq_restore(flags);
151	}
152}
153
154static void update_page_reclaim_stat(struct zone *zone, struct page *page,
155				     int file, int rotated)
156{
157	struct zone_reclaim_stat *reclaim_stat = &zone->reclaim_stat;
158	struct zone_reclaim_stat *memcg_reclaim_stat;
159
160	memcg_reclaim_stat = mem_cgroup_get_reclaim_stat_from_page(page);
161
162	reclaim_stat->recent_scanned[file]++;
163	if (rotated)
164		reclaim_stat->recent_rotated[file]++;
165
166	if (!memcg_reclaim_stat)
167		return;
168
169	memcg_reclaim_stat->recent_scanned[file]++;
170	if (rotated)
171		memcg_reclaim_stat->recent_rotated[file]++;
172}
173
174/*
175 * FIXME: speed this up?
176 */
177void activate_page(struct page *page)
178{
179	struct zone *zone = page_zone(page);
180
181	spin_lock_irq(&zone->lru_lock);
182	if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
183		int file = page_is_file_cache(page);
184		int lru = LRU_BASE + file;
185		del_page_from_lru_list(zone, page, lru);
186
187		SetPageActive(page);
188		lru += LRU_ACTIVE;
189		add_page_to_lru_list(zone, page, lru);
190		__count_vm_event(PGACTIVATE);
191
192		update_page_reclaim_stat(zone, page, !!file, 1);
193	}
194	spin_unlock_irq(&zone->lru_lock);
195}
196
197/*
198 * Mark a page as having seen activity.
199 *
200 * inactive,unreferenced	->	inactive,referenced
201 * inactive,referenced		->	active,unreferenced
202 * active,unreferenced		->	active,referenced
203 */
204void mark_page_accessed(struct page *page)
205{
206	if (!PageActive(page) && !PageUnevictable(page) &&
207			PageReferenced(page) && PageLRU(page)) {
208		activate_page(page);
209		ClearPageReferenced(page);
210	} else if (!PageReferenced(page)) {
211		SetPageReferenced(page);
212	}
213}
214
215EXPORT_SYMBOL(mark_page_accessed);
216
217void __lru_cache_add(struct page *page, enum lru_list lru)
218{
219	struct pagevec *pvec = &get_cpu_var(lru_add_pvecs)[lru];
220
221	page_cache_get(page);
222	if (!pagevec_add(pvec, page))
223		____pagevec_lru_add(pvec, lru);
224	put_cpu_var(lru_add_pvecs);
225}
226
227/**
228 * lru_cache_add_lru - add a page to a page list
229 * @page: the page to be added to the LRU.
230 * @lru: the LRU list to which the page is added.
231 */
232void lru_cache_add_lru(struct page *page, enum lru_list lru)
233{
234	if (PageActive(page)) {
235		VM_BUG_ON(PageUnevictable(page));
236		ClearPageActive(page);
237	} else if (PageUnevictable(page)) {
238		VM_BUG_ON(PageActive(page));
239		ClearPageUnevictable(page);
240	}
241
242	VM_BUG_ON(PageLRU(page) || PageActive(page) || PageUnevictable(page));
243	__lru_cache_add(page, lru);
244}
245
246/**
247 * add_page_to_unevictable_list - add a page to the unevictable list
248 * @page:  the page to be added to the unevictable list
249 *
250 * Add page directly to its zone's unevictable list.  To avoid races with
251 * tasks that might be making the page evictable, through eg. munlock,
252 * munmap or exit, while it's not on the lru, we want to add the page
253 * while it's locked or otherwise "invisible" to other tasks.  This is
254 * difficult to do when using the pagevec cache, so bypass that.
255 */
256void add_page_to_unevictable_list(struct page *page)
257{
258	struct zone *zone = page_zone(page);
259
260	spin_lock_irq(&zone->lru_lock);
261	SetPageUnevictable(page);
262	SetPageLRU(page);
263	add_page_to_lru_list(zone, page, LRU_UNEVICTABLE);
264	spin_unlock_irq(&zone->lru_lock);
265}
266
267/*
268 * Drain pages out of the cpu's pagevecs.
269 * Either "cpu" is the current CPU, and preemption has already been
270 * disabled; or "cpu" is being hot-unplugged, and is already dead.
271 */
272static void drain_cpu_pagevecs(int cpu)
273{
274	struct pagevec *pvecs = per_cpu(lru_add_pvecs, cpu);
275	struct pagevec *pvec;
276	int lru;
277
278	for_each_lru(lru) {
279		pvec = &pvecs[lru - LRU_BASE];
280		if (pagevec_count(pvec))
281			____pagevec_lru_add(pvec, lru);
282	}
283
284	pvec = &per_cpu(lru_rotate_pvecs, cpu);
285	if (pagevec_count(pvec)) {
286		unsigned long flags;
287
288		/* No harm done if a racing interrupt already did this */
289		local_irq_save(flags);
290		pagevec_move_tail(pvec);
291		local_irq_restore(flags);
292	}
293}
294
295void lru_add_drain(void)
296{
297	drain_cpu_pagevecs(get_cpu());
298	put_cpu();
299}
300
301static void lru_add_drain_per_cpu(struct work_struct *dummy)
302{
303	lru_add_drain();
304}
305
306/*
307 * Returns 0 for success
308 */
309int lru_add_drain_all(void)
310{
311	return schedule_on_each_cpu(lru_add_drain_per_cpu);
312}
313
314/*
315 * Batched page_cache_release().  Decrement the reference count on all the
316 * passed pages.  If it fell to zero then remove the page from the LRU and
317 * free it.
318 *
319 * Avoid taking zone->lru_lock if possible, but if it is taken, retain it
320 * for the remainder of the operation.
321 *
322 * The locking in this function is against shrink_inactive_list(): we recheck
323 * the page count inside the lock to see whether shrink_inactive_list()
324 * grabbed the page via the LRU.  If it did, give up: shrink_inactive_list()
325 * will free it.
326 */
327void release_pages(struct page **pages, int nr, int cold)
328{
329	int i;
330	struct pagevec pages_to_free;
331	struct zone *zone = NULL;
332	unsigned long uninitialized_var(flags);
333
334	pagevec_init(&pages_to_free, cold);
335	for (i = 0; i < nr; i++) {
336		struct page *page = pages[i];
337
338		if (unlikely(PageCompound(page))) {
339			if (zone) {
340				spin_unlock_irqrestore(&zone->lru_lock, flags);
341				zone = NULL;
342			}
343			put_compound_page(page);
344			continue;
345		}
346
347		if (!put_page_testzero(page))
348			continue;
349
350		if (PageLRU(page)) {
351			struct zone *pagezone = page_zone(page);
352
353			if (pagezone != zone) {
354				if (zone)
355					spin_unlock_irqrestore(&zone->lru_lock,
356									flags);
357				zone = pagezone;
358				spin_lock_irqsave(&zone->lru_lock, flags);
359			}
360			VM_BUG_ON(!PageLRU(page));
361			__ClearPageLRU(page);
362			del_page_from_lru(zone, page);
363		}
364
365		if (!pagevec_add(&pages_to_free, page)) {
366			if (zone) {
367				spin_unlock_irqrestore(&zone->lru_lock, flags);
368				zone = NULL;
369			}
370			__pagevec_free(&pages_to_free);
371			pagevec_reinit(&pages_to_free);
372  		}
373	}
374	if (zone)
375		spin_unlock_irqrestore(&zone->lru_lock, flags);
376
377	pagevec_free(&pages_to_free);
378}
379
380/*
381 * The pages which we're about to release may be in the deferred lru-addition
382 * queues.  That would prevent them from really being freed right now.  That's
383 * OK from a correctness point of view but is inefficient - those pages may be
384 * cache-warm and we want to give them back to the page allocator ASAP.
385 *
386 * So __pagevec_release() will drain those queues here.  __pagevec_lru_add()
387 * and __pagevec_lru_add_active() call release_pages() directly to avoid
388 * mutual recursion.
389 */
390void __pagevec_release(struct pagevec *pvec)
391{
392	lru_add_drain();
393	release_pages(pvec->pages, pagevec_count(pvec), pvec->cold);
394	pagevec_reinit(pvec);
395}
396
397EXPORT_SYMBOL(__pagevec_release);
398
399/*
400 * Add the passed pages to the LRU, then drop the caller's refcount
401 * on them.  Reinitialises the caller's pagevec.
402 */
403void ____pagevec_lru_add(struct pagevec *pvec, enum lru_list lru)
404{
405	int i;
406	struct zone *zone = NULL;
407
408	VM_BUG_ON(is_unevictable_lru(lru));
409
410	for (i = 0; i < pagevec_count(pvec); i++) {
411		struct page *page = pvec->pages[i];
412		struct zone *pagezone = page_zone(page);
413		int file;
414		int active;
415
416		if (pagezone != zone) {
417			if (zone)
418				spin_unlock_irq(&zone->lru_lock);
419			zone = pagezone;
420			spin_lock_irq(&zone->lru_lock);
421		}
422		VM_BUG_ON(PageActive(page));
423		VM_BUG_ON(PageUnevictable(page));
424		VM_BUG_ON(PageLRU(page));
425		SetPageLRU(page);
426		active = is_active_lru(lru);
427		file = is_file_lru(lru);
428		if (active)
429			SetPageActive(page);
430		update_page_reclaim_stat(zone, page, file, active);
431		add_page_to_lru_list(zone, page, lru);
432	}
433	if (zone)
434		spin_unlock_irq(&zone->lru_lock);
435	release_pages(pvec->pages, pvec->nr, pvec->cold);
436	pagevec_reinit(pvec);
437}
438
439EXPORT_SYMBOL(____pagevec_lru_add);
440
441/*
442 * Try to drop buffers from the pages in a pagevec
443 */
444void pagevec_strip(struct pagevec *pvec)
445{
446	int i;
447
448	for (i = 0; i < pagevec_count(pvec); i++) {
449		struct page *page = pvec->pages[i];
450
451		if (PagePrivate(page) && trylock_page(page)) {
452			if (PagePrivate(page))
453				try_to_release_page(page, 0);
454			unlock_page(page);
455		}
456	}
457}
458
459/**
460 * pagevec_swap_free - try to free swap space from the pages in a pagevec
461 * @pvec: pagevec with swapcache pages to free the swap space of
462 *
463 * The caller needs to hold an extra reference to each page and
464 * not hold the page lock on the pages.  This function uses a
465 * trylock on the page lock so it may not always free the swap
466 * space associated with a page.
467 */
468void pagevec_swap_free(struct pagevec *pvec)
469{
470	int i;
471
472	for (i = 0; i < pagevec_count(pvec); i++) {
473		struct page *page = pvec->pages[i];
474
475		if (PageSwapCache(page) && trylock_page(page)) {
476			try_to_free_swap(page);
477			unlock_page(page);
478		}
479	}
480}
481
482/**
483 * pagevec_lookup - gang pagecache lookup
484 * @pvec:	Where the resulting pages are placed
485 * @mapping:	The address_space to search
486 * @start:	The starting page index
487 * @nr_pages:	The maximum number of pages
488 *
489 * pagevec_lookup() will search for and return a group of up to @nr_pages pages
490 * in the mapping.  The pages are placed in @pvec.  pagevec_lookup() takes a
491 * reference against the pages in @pvec.
492 *
493 * The search returns a group of mapping-contiguous pages with ascending
494 * indexes.  There may be holes in the indices due to not-present pages.
495 *
496 * pagevec_lookup() returns the number of pages which were found.
497 */
498unsigned pagevec_lookup(struct pagevec *pvec, struct address_space *mapping,
499		pgoff_t start, unsigned nr_pages)
500{
501	pvec->nr = find_get_pages(mapping, start, nr_pages, pvec->pages);
502	return pagevec_count(pvec);
503}
504
505EXPORT_SYMBOL(pagevec_lookup);
506
507unsigned pagevec_lookup_tag(struct pagevec *pvec, struct address_space *mapping,
508		pgoff_t *index, int tag, unsigned nr_pages)
509{
510	pvec->nr = find_get_pages_tag(mapping, index, tag,
511					nr_pages, pvec->pages);
512	return pagevec_count(pvec);
513}
514
515EXPORT_SYMBOL(pagevec_lookup_tag);
516
517#ifdef CONFIG_SMP
518/*
519 * We tolerate a little inaccuracy to avoid ping-ponging the counter between
520 * CPUs
521 */
522#define ACCT_THRESHOLD	max(16, NR_CPUS * 2)
523
524static DEFINE_PER_CPU(long, committed_space);
525
526void vm_acct_memory(long pages)
527{
528	long *local;
529
530	preempt_disable();
531	local = &__get_cpu_var(committed_space);
532	*local += pages;
533	if (*local > ACCT_THRESHOLD || *local < -ACCT_THRESHOLD) {
534		atomic_long_add(*local, &vm_committed_space);
535		*local = 0;
536	}
537	preempt_enable();
538}
539
540#ifdef CONFIG_HOTPLUG_CPU
541
542/* Drop the CPU's cached committed space back into the central pool. */
543static int cpu_swap_callback(struct notifier_block *nfb,
544			     unsigned long action,
545			     void *hcpu)
546{
547	long *committed;
548
549	committed = &per_cpu(committed_space, (long)hcpu);
550	if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
551		atomic_long_add(*committed, &vm_committed_space);
552		*committed = 0;
553		drain_cpu_pagevecs((long)hcpu);
554	}
555	return NOTIFY_OK;
556}
557#endif /* CONFIG_HOTPLUG_CPU */
558#endif /* CONFIG_SMP */
559
560/*
561 * Perform any setup for the swap system
562 */
563void __init swap_setup(void)
564{
565	unsigned long megs = num_physpages >> (20 - PAGE_SHIFT);
566
567#ifdef CONFIG_SWAP
568	bdi_init(swapper_space.backing_dev_info);
569#endif
570
571	/* Use a smaller cluster for small-memory machines */
572	if (megs < 16)
573		page_cluster = 2;
574	else
575		page_cluster = 3;
576	/*
577	 * Right now other parts of the system means that we
578	 * _really_ don't want to cluster much more
579	 */
580#ifdef CONFIG_HOTPLUG_CPU
581	hotcpu_notifier(cpu_swap_callback, 0);
582#endif
583}
584