4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
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.
12 * Swap aging added 23.2.95, Stephen Tweedie.
13 * Buffermem limits added 12.3.98, Rik van Riel.
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>
36 /* How many pages do we try to swap or page in/out together? */
39 static DEFINE_PER_CPU(struct pagevec[NR_LRU_LISTS], lru_add_pvecs);
40 static DEFINE_PER_CPU(struct pagevec, lru_rotate_pvecs);
41 static DEFINE_PER_CPU(struct pagevec, lru_deactivate_pvecs);
44 * This path almost never happens for VM activity - pages are normally
45 * freed via pagevecs. But it gets used by networking.
47 static void __page_cache_release(struct page *page)
51 struct zone *zone = page_zone(page);
53 spin_lock_irqsave(&zone->lru_lock, flags);
54 VM_BUG_ON(!PageLRU(page));
56 del_page_from_lru_list(zone, page, page_off_lru(page));
57 spin_unlock_irqrestore(&zone->lru_lock, flags);
61 static void __put_single_page(struct page *page)
63 __page_cache_release(page);
64 free_hot_cold_page(page, 0);
67 static void __put_compound_page(struct page *page)
69 compound_page_dtor *dtor;
71 __page_cache_release(page);
72 dtor = get_compound_page_dtor(page);
76 static void put_compound_page(struct page *page)
78 if (unlikely(PageTail(page))) {
79 /* __split_huge_page_refcount can run under us */
80 struct page *page_head = compound_trans_head(page);
82 if (likely(page != page_head &&
83 get_page_unless_zero(page_head))) {
87 * THP can not break up slab pages so avoid taking
88 * compound_lock(). Slab performs non-atomic bit ops
89 * on page->flags for better performance. In particular
90 * slab_unlock() in slub used to be a hot path. It is
91 * still hot on arches that do not support
92 * this_cpu_cmpxchg_double().
94 if (PageSlab(page_head)) {
96 if (put_page_testzero(page_head))
99 atomic_dec(&page->_mapcount);
105 * page_head wasn't a dangling pointer but it
106 * may not be a head page anymore by the time
107 * we obtain the lock. That is ok as long as it
108 * can't be freed from under us.
110 flags = compound_lock_irqsave(page_head);
111 if (unlikely(!PageTail(page))) {
112 /* __split_huge_page_refcount run before us */
113 compound_unlock_irqrestore(page_head, flags);
115 if (put_page_testzero(page_head))
116 __put_single_page(page_head);
118 if (put_page_testzero(page))
119 __put_single_page(page);
122 VM_BUG_ON(page_head != page->first_page);
124 * We can release the refcount taken by
125 * get_page_unless_zero() now that
126 * __split_huge_page_refcount() is blocked on
129 if (put_page_testzero(page_head))
131 /* __split_huge_page_refcount will wait now */
132 VM_BUG_ON(page_mapcount(page) <= 0);
133 atomic_dec(&page->_mapcount);
134 VM_BUG_ON(atomic_read(&page_head->_count) <= 0);
135 VM_BUG_ON(atomic_read(&page->_count) != 0);
136 compound_unlock_irqrestore(page_head, flags);
139 if (put_page_testzero(page_head)) {
140 if (PageHead(page_head))
141 __put_compound_page(page_head);
143 __put_single_page(page_head);
146 /* page_head is a dangling pointer */
147 VM_BUG_ON(PageTail(page));
150 } else if (put_page_testzero(page)) {
152 __put_compound_page(page);
154 __put_single_page(page);
158 void put_page(struct page *page)
160 if (unlikely(PageCompound(page)))
161 put_compound_page(page);
162 else if (put_page_testzero(page))
163 __put_single_page(page);
165 EXPORT_SYMBOL(put_page);
168 * This function is exported but must not be called by anything other
169 * than get_page(). It implements the slow path of get_page().
171 bool __get_page_tail(struct page *page)
174 * This takes care of get_page() if run on a tail page
175 * returned by one of the get_user_pages/follow_page variants.
176 * get_user_pages/follow_page itself doesn't need the compound
177 * lock because it runs __get_page_tail_foll() under the
178 * proper PT lock that already serializes against
183 struct page *page_head = compound_trans_head(page);
185 if (likely(page != page_head && get_page_unless_zero(page_head))) {
187 /* Ref to put_compound_page() comment. */
188 if (PageSlab(page_head)) {
189 if (likely(PageTail(page))) {
190 __get_page_tail_foll(page, false);
199 * page_head wasn't a dangling pointer but it
200 * may not be a head page anymore by the time
201 * we obtain the lock. That is ok as long as it
202 * can't be freed from under us.
204 flags = compound_lock_irqsave(page_head);
205 /* here __split_huge_page_refcount won't run anymore */
206 if (likely(PageTail(page))) {
207 __get_page_tail_foll(page, false);
210 compound_unlock_irqrestore(page_head, flags);
216 EXPORT_SYMBOL(__get_page_tail);
219 * put_pages_list() - release a list of pages
220 * @pages: list of pages threaded on page->lru
222 * Release a list of pages which are strung together on page.lru. Currently
223 * used by read_cache_pages() and related error recovery code.
225 void put_pages_list(struct list_head *pages)
227 while (!list_empty(pages)) {
230 victim = list_entry(pages->prev, struct page, lru);
231 list_del(&victim->lru);
232 page_cache_release(victim);
235 EXPORT_SYMBOL(put_pages_list);
237 static void pagevec_lru_move_fn(struct pagevec *pvec,
238 void (*move_fn)(struct page *page, void *arg),
242 struct zone *zone = NULL;
243 unsigned long flags = 0;
245 for (i = 0; i < pagevec_count(pvec); i++) {
246 struct page *page = pvec->pages[i];
247 struct zone *pagezone = page_zone(page);
249 if (pagezone != zone) {
251 spin_unlock_irqrestore(&zone->lru_lock, flags);
253 spin_lock_irqsave(&zone->lru_lock, flags);
256 (*move_fn)(page, arg);
259 spin_unlock_irqrestore(&zone->lru_lock, flags);
260 release_pages(pvec->pages, pvec->nr, pvec->cold);
261 pagevec_reinit(pvec);
264 static void pagevec_move_tail_fn(struct page *page, void *arg)
268 if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
269 enum lru_list lru = page_lru_base_type(page);
270 struct lruvec *lruvec;
272 lruvec = mem_cgroup_lru_move_lists(page_zone(page),
274 list_move_tail(&page->lru, &lruvec->lists[lru]);
280 * pagevec_move_tail() must be called with IRQ disabled.
281 * Otherwise this may cause nasty races.
283 static void pagevec_move_tail(struct pagevec *pvec)
287 pagevec_lru_move_fn(pvec, pagevec_move_tail_fn, &pgmoved);
288 __count_vm_events(PGROTATED, pgmoved);
292 * Writeback is about to end against a page which has been marked for immediate
293 * reclaim. If it still appears to be reclaimable, move it to the tail of the
296 void rotate_reclaimable_page(struct page *page)
298 if (!PageLocked(page) && !PageDirty(page) && !PageActive(page) &&
299 !PageUnevictable(page) && PageLRU(page)) {
300 struct pagevec *pvec;
303 page_cache_get(page);
304 local_irq_save(flags);
305 pvec = &__get_cpu_var(lru_rotate_pvecs);
306 if (!pagevec_add(pvec, page))
307 pagevec_move_tail(pvec);
308 local_irq_restore(flags);
312 static void update_page_reclaim_stat(struct zone *zone, struct page *page,
313 int file, int rotated)
315 struct zone_reclaim_stat *reclaim_stat;
317 reclaim_stat = mem_cgroup_get_reclaim_stat_from_page(page);
319 reclaim_stat = &zone->lruvec.reclaim_stat;
321 reclaim_stat->recent_scanned[file]++;
323 reclaim_stat->recent_rotated[file]++;
326 static void __activate_page(struct page *page, void *arg)
328 struct zone *zone = page_zone(page);
330 if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
331 int file = page_is_file_cache(page);
332 int lru = page_lru_base_type(page);
333 del_page_from_lru_list(zone, page, lru);
337 add_page_to_lru_list(zone, page, lru);
338 __count_vm_event(PGACTIVATE);
340 update_page_reclaim_stat(zone, page, file, 1);
345 static DEFINE_PER_CPU(struct pagevec, activate_page_pvecs);
347 static void activate_page_drain(int cpu)
349 struct pagevec *pvec = &per_cpu(activate_page_pvecs, cpu);
351 if (pagevec_count(pvec))
352 pagevec_lru_move_fn(pvec, __activate_page, NULL);
355 void activate_page(struct page *page)
357 if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
358 struct pagevec *pvec = &get_cpu_var(activate_page_pvecs);
360 page_cache_get(page);
361 if (!pagevec_add(pvec, page))
362 pagevec_lru_move_fn(pvec, __activate_page, NULL);
363 put_cpu_var(activate_page_pvecs);
368 static inline void activate_page_drain(int cpu)
372 void activate_page(struct page *page)
374 struct zone *zone = page_zone(page);
376 spin_lock_irq(&zone->lru_lock);
377 __activate_page(page, NULL);
378 spin_unlock_irq(&zone->lru_lock);
383 * Mark a page as having seen activity.
385 * inactive,unreferenced -> inactive,referenced
386 * inactive,referenced -> active,unreferenced
387 * active,unreferenced -> active,referenced
389 void mark_page_accessed(struct page *page)
391 if (!PageActive(page) && !PageUnevictable(page) &&
392 PageReferenced(page) && PageLRU(page)) {
394 ClearPageReferenced(page);
395 } else if (!PageReferenced(page)) {
396 SetPageReferenced(page);
399 EXPORT_SYMBOL(mark_page_accessed);
401 void __lru_cache_add(struct page *page, enum lru_list lru)
403 struct pagevec *pvec = &get_cpu_var(lru_add_pvecs)[lru];
405 page_cache_get(page);
406 if (!pagevec_add(pvec, page))
407 __pagevec_lru_add(pvec, lru);
408 put_cpu_var(lru_add_pvecs);
410 EXPORT_SYMBOL(__lru_cache_add);
413 * lru_cache_add_lru - add a page to a page list
414 * @page: the page to be added to the LRU.
415 * @lru: the LRU list to which the page is added.
417 void lru_cache_add_lru(struct page *page, enum lru_list lru)
419 if (PageActive(page)) {
420 VM_BUG_ON(PageUnevictable(page));
421 ClearPageActive(page);
422 } else if (PageUnevictable(page)) {
423 VM_BUG_ON(PageActive(page));
424 ClearPageUnevictable(page);
427 VM_BUG_ON(PageLRU(page) || PageActive(page) || PageUnevictable(page));
428 __lru_cache_add(page, lru);
432 * add_page_to_unevictable_list - add a page to the unevictable list
433 * @page: the page to be added to the unevictable list
435 * Add page directly to its zone's unevictable list. To avoid races with
436 * tasks that might be making the page evictable, through eg. munlock,
437 * munmap or exit, while it's not on the lru, we want to add the page
438 * while it's locked or otherwise "invisible" to other tasks. This is
439 * difficult to do when using the pagevec cache, so bypass that.
441 void add_page_to_unevictable_list(struct page *page)
443 struct zone *zone = page_zone(page);
445 spin_lock_irq(&zone->lru_lock);
446 SetPageUnevictable(page);
448 add_page_to_lru_list(zone, page, LRU_UNEVICTABLE);
449 spin_unlock_irq(&zone->lru_lock);
453 * If the page can not be invalidated, it is moved to the
454 * inactive list to speed up its reclaim. It is moved to the
455 * head of the list, rather than the tail, to give the flusher
456 * threads some time to write it out, as this is much more
457 * effective than the single-page writeout from reclaim.
459 * If the page isn't page_mapped and dirty/writeback, the page
460 * could reclaim asap using PG_reclaim.
462 * 1. active, mapped page -> none
463 * 2. active, dirty/writeback page -> inactive, head, PG_reclaim
464 * 3. inactive, mapped page -> none
465 * 4. inactive, dirty/writeback page -> inactive, head, PG_reclaim
466 * 5. inactive, clean -> inactive, tail
469 * In 4, why it moves inactive's head, the VM expects the page would
470 * be write it out by flusher threads as this is much more effective
471 * than the single-page writeout from reclaim.
473 static void lru_deactivate_fn(struct page *page, void *arg)
477 struct zone *zone = page_zone(page);
482 if (PageUnevictable(page))
485 /* Some processes are using the page */
486 if (page_mapped(page))
489 active = PageActive(page);
491 file = page_is_file_cache(page);
492 lru = page_lru_base_type(page);
493 del_page_from_lru_list(zone, page, lru + active);
494 ClearPageActive(page);
495 ClearPageReferenced(page);
496 add_page_to_lru_list(zone, page, lru);
498 if (PageWriteback(page) || PageDirty(page)) {
500 * PG_reclaim could be raced with end_page_writeback
501 * It can make readahead confusing. But race window
502 * is _really_ small and it's non-critical problem.
504 SetPageReclaim(page);
506 struct lruvec *lruvec;
508 * The page's writeback ends up during pagevec
509 * We moves tha page into tail of inactive.
511 lruvec = mem_cgroup_lru_move_lists(zone, page, lru, lru);
512 list_move_tail(&page->lru, &lruvec->lists[lru]);
513 __count_vm_event(PGROTATED);
517 __count_vm_event(PGDEACTIVATE);
518 update_page_reclaim_stat(zone, page, file, 0);
522 * Drain pages out of the cpu's pagevecs.
523 * Either "cpu" is the current CPU, and preemption has already been
524 * disabled; or "cpu" is being hot-unplugged, and is already dead.
526 void lru_add_drain_cpu(int cpu)
528 struct pagevec *pvecs = per_cpu(lru_add_pvecs, cpu);
529 struct pagevec *pvec;
533 pvec = &pvecs[lru - LRU_BASE];
534 if (pagevec_count(pvec))
535 __pagevec_lru_add(pvec, lru);
538 pvec = &per_cpu(lru_rotate_pvecs, cpu);
539 if (pagevec_count(pvec)) {
542 /* No harm done if a racing interrupt already did this */
543 local_irq_save(flags);
544 pagevec_move_tail(pvec);
545 local_irq_restore(flags);
548 pvec = &per_cpu(lru_deactivate_pvecs, cpu);
549 if (pagevec_count(pvec))
550 pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
552 activate_page_drain(cpu);
556 * deactivate_page - forcefully deactivate a page
557 * @page: page to deactivate
559 * This function hints the VM that @page is a good reclaim candidate,
560 * for example if its invalidation fails due to the page being dirty
561 * or under writeback.
563 void deactivate_page(struct page *page)
566 * In a workload with many unevictable page such as mprotect, unevictable
567 * page deactivation for accelerating reclaim is pointless.
569 if (PageUnevictable(page))
572 if (likely(get_page_unless_zero(page))) {
573 struct pagevec *pvec = &get_cpu_var(lru_deactivate_pvecs);
575 if (!pagevec_add(pvec, page))
576 pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
577 put_cpu_var(lru_deactivate_pvecs);
581 void lru_add_drain(void)
583 lru_add_drain_cpu(get_cpu());
587 static void lru_add_drain_per_cpu(struct work_struct *dummy)
593 * Returns 0 for success
595 int lru_add_drain_all(void)
597 return schedule_on_each_cpu(lru_add_drain_per_cpu);
601 * Batched page_cache_release(). Decrement the reference count on all the
602 * passed pages. If it fell to zero then remove the page from the LRU and
605 * Avoid taking zone->lru_lock if possible, but if it is taken, retain it
606 * for the remainder of the operation.
608 * The locking in this function is against shrink_inactive_list(): we recheck
609 * the page count inside the lock to see whether shrink_inactive_list()
610 * grabbed the page via the LRU. If it did, give up: shrink_inactive_list()
613 void release_pages(struct page **pages, int nr, int cold)
616 LIST_HEAD(pages_to_free);
617 struct zone *zone = NULL;
618 unsigned long uninitialized_var(flags);
620 for (i = 0; i < nr; i++) {
621 struct page *page = pages[i];
623 if (unlikely(PageCompound(page))) {
625 spin_unlock_irqrestore(&zone->lru_lock, flags);
628 put_compound_page(page);
632 if (!put_page_testzero(page))
636 struct zone *pagezone = page_zone(page);
638 if (pagezone != zone) {
640 spin_unlock_irqrestore(&zone->lru_lock,
643 spin_lock_irqsave(&zone->lru_lock, flags);
645 VM_BUG_ON(!PageLRU(page));
646 __ClearPageLRU(page);
647 del_page_from_lru_list(zone, page, page_off_lru(page));
650 list_add(&page->lru, &pages_to_free);
653 spin_unlock_irqrestore(&zone->lru_lock, flags);
655 free_hot_cold_page_list(&pages_to_free, cold);
657 EXPORT_SYMBOL(release_pages);
660 * The pages which we're about to release may be in the deferred lru-addition
661 * queues. That would prevent them from really being freed right now. That's
662 * OK from a correctness point of view but is inefficient - those pages may be
663 * cache-warm and we want to give them back to the page allocator ASAP.
665 * So __pagevec_release() will drain those queues here. __pagevec_lru_add()
666 * and __pagevec_lru_add_active() call release_pages() directly to avoid
669 void __pagevec_release(struct pagevec *pvec)
672 release_pages(pvec->pages, pagevec_count(pvec), pvec->cold);
673 pagevec_reinit(pvec);
675 EXPORT_SYMBOL(__pagevec_release);
677 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
678 /* used by __split_huge_page_refcount() */
679 void lru_add_page_tail(struct zone* zone,
680 struct page *page, struct page *page_tail)
682 int uninitialized_var(active);
686 VM_BUG_ON(!PageHead(page));
687 VM_BUG_ON(PageCompound(page_tail));
688 VM_BUG_ON(PageLRU(page_tail));
689 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&zone->lru_lock));
691 SetPageLRU(page_tail);
693 if (page_evictable(page_tail, NULL)) {
694 if (PageActive(page)) {
695 SetPageActive(page_tail);
697 lru = LRU_ACTIVE_ANON;
700 lru = LRU_INACTIVE_ANON;
703 SetPageUnevictable(page_tail);
704 lru = LRU_UNEVICTABLE;
707 if (likely(PageLRU(page)))
708 list_add_tail(&page_tail->lru, &page->lru);
710 struct list_head *list_head;
712 * Head page has not yet been counted, as an hpage,
713 * so we must account for each subpage individually.
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.
718 add_page_to_lru_list(zone, page_tail, lru);
719 list_head = page_tail->lru.prev;
720 list_move_tail(&page_tail->lru, list_head);
723 if (!PageUnevictable(page))
724 update_page_reclaim_stat(zone, page_tail, file, active);
726 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
728 static void __pagevec_lru_add_fn(struct page *page, void *arg)
730 enum lru_list lru = (enum lru_list)arg;
731 struct zone *zone = page_zone(page);
732 int file = is_file_lru(lru);
733 int active = is_active_lru(lru);
735 VM_BUG_ON(PageActive(page));
736 VM_BUG_ON(PageUnevictable(page));
737 VM_BUG_ON(PageLRU(page));
742 add_page_to_lru_list(zone, page, lru);
743 update_page_reclaim_stat(zone, page, file, active);
747 * Add the passed pages to the LRU, then drop the caller's refcount
748 * on them. Reinitialises the caller's pagevec.
750 void __pagevec_lru_add(struct pagevec *pvec, enum lru_list lru)
752 VM_BUG_ON(is_unevictable_lru(lru));
754 pagevec_lru_move_fn(pvec, __pagevec_lru_add_fn, (void *)lru);
756 EXPORT_SYMBOL(__pagevec_lru_add);
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
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.
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.
772 * pagevec_lookup() returns the number of pages which were found.
774 unsigned pagevec_lookup(struct pagevec *pvec, struct address_space *mapping,
775 pgoff_t start, unsigned nr_pages)
777 pvec->nr = find_get_pages(mapping, start, nr_pages, pvec->pages);
778 return pagevec_count(pvec);
780 EXPORT_SYMBOL(pagevec_lookup);
782 unsigned pagevec_lookup_tag(struct pagevec *pvec, struct address_space *mapping,
783 pgoff_t *index, int tag, unsigned nr_pages)
785 pvec->nr = find_get_pages_tag(mapping, index, tag,
786 nr_pages, pvec->pages);
787 return pagevec_count(pvec);
789 EXPORT_SYMBOL(pagevec_lookup_tag);
792 * Perform any setup for the swap system
794 void __init swap_setup(void)
796 unsigned long megs = totalram_pages >> (20 - PAGE_SHIFT);
799 bdi_init(swapper_space.backing_dev_info);
802 /* Use a smaller cluster for small-memory machines */
808 * Right now other parts of the system means that we
809 * _really_ don't want to cluster much more