Merge git://git.kernel.org/pub/scm/linux/kernel/git/sfrench/cifs-2.6
[linux-2.6/next.git] / mm / swap.c
blob5602f1a1b1e724ba547bdee3ef21d773fd24ee04
1 /*
2 * linux/mm/swap.c
4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 */
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.
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 #include <linux/gfp.h>
35 #include "internal.h"
37 /* How many pages do we try to swap or page in/out together? */
38 int page_cluster;
40 static DEFINE_PER_CPU(struct pagevec[NR_LRU_LISTS], lru_add_pvecs);
41 static DEFINE_PER_CPU(struct pagevec, lru_rotate_pvecs);
42 static DEFINE_PER_CPU(struct pagevec, lru_deactivate_pvecs);
45 * This path almost never happens for VM activity - pages are normally
46 * freed via pagevecs. But it gets used by networking.
48 static void __page_cache_release(struct page *page)
50 if (PageLRU(page)) {
51 unsigned long flags;
52 struct zone *zone = page_zone(page);
54 spin_lock_irqsave(&zone->lru_lock, flags);
55 VM_BUG_ON(!PageLRU(page));
56 __ClearPageLRU(page);
57 del_page_from_lru(zone, page);
58 spin_unlock_irqrestore(&zone->lru_lock, flags);
62 static void __put_single_page(struct page *page)
64 __page_cache_release(page);
65 free_hot_cold_page(page, 0);
68 static void __put_compound_page(struct page *page)
70 compound_page_dtor *dtor;
72 __page_cache_release(page);
73 dtor = get_compound_page_dtor(page);
74 (*dtor)(page);
77 static void put_compound_page(struct page *page)
79 if (unlikely(PageTail(page))) {
80 /* __split_huge_page_refcount can run under us */
81 struct page *page_head = page->first_page;
82 smp_rmb();
84 * If PageTail is still set after smp_rmb() we can be sure
85 * that the page->first_page we read wasn't a dangling pointer.
86 * See __split_huge_page_refcount() smp_wmb().
88 if (likely(PageTail(page) && get_page_unless_zero(page_head))) {
89 unsigned long flags;
91 * Verify that our page_head wasn't converted
92 * to a a regular page before we got a
93 * reference on it.
95 if (unlikely(!PageHead(page_head))) {
96 /* PageHead is cleared after PageTail */
97 smp_rmb();
98 VM_BUG_ON(PageTail(page));
99 goto out_put_head;
102 * Only run compound_lock on a valid PageHead,
103 * after having it pinned with
104 * get_page_unless_zero() above.
106 smp_mb();
107 /* page_head wasn't a dangling pointer */
108 flags = compound_lock_irqsave(page_head);
109 if (unlikely(!PageTail(page))) {
110 /* __split_huge_page_refcount run before us */
111 compound_unlock_irqrestore(page_head, flags);
112 VM_BUG_ON(PageHead(page_head));
113 out_put_head:
114 if (put_page_testzero(page_head))
115 __put_single_page(page_head);
116 out_put_single:
117 if (put_page_testzero(page))
118 __put_single_page(page);
119 return;
121 VM_BUG_ON(page_head != page->first_page);
123 * We can release the refcount taken by
124 * get_page_unless_zero now that
125 * split_huge_page_refcount is blocked on the
126 * compound_lock.
128 if (put_page_testzero(page_head))
129 VM_BUG_ON(1);
130 /* __split_huge_page_refcount will wait now */
131 VM_BUG_ON(atomic_read(&page->_count) <= 0);
132 atomic_dec(&page->_count);
133 VM_BUG_ON(atomic_read(&page_head->_count) <= 0);
134 compound_unlock_irqrestore(page_head, flags);
135 if (put_page_testzero(page_head)) {
136 if (PageHead(page_head))
137 __put_compound_page(page_head);
138 else
139 __put_single_page(page_head);
141 } else {
142 /* page_head is a dangling pointer */
143 VM_BUG_ON(PageTail(page));
144 goto out_put_single;
146 } else if (put_page_testzero(page)) {
147 if (PageHead(page))
148 __put_compound_page(page);
149 else
150 __put_single_page(page);
154 void put_page(struct page *page)
156 if (unlikely(PageCompound(page)))
157 put_compound_page(page);
158 else if (put_page_testzero(page))
159 __put_single_page(page);
161 EXPORT_SYMBOL(put_page);
164 * put_pages_list() - release a list of pages
165 * @pages: list of pages threaded on page->lru
167 * Release a list of pages which are strung together on page.lru. Currently
168 * used by read_cache_pages() and related error recovery code.
170 void put_pages_list(struct list_head *pages)
172 while (!list_empty(pages)) {
173 struct page *victim;
175 victim = list_entry(pages->prev, struct page, lru);
176 list_del(&victim->lru);
177 page_cache_release(victim);
180 EXPORT_SYMBOL(put_pages_list);
182 static void pagevec_lru_move_fn(struct pagevec *pvec,
183 void (*move_fn)(struct page *page, void *arg),
184 void *arg)
186 int i;
187 struct zone *zone = NULL;
188 unsigned long flags = 0;
190 for (i = 0; i < pagevec_count(pvec); i++) {
191 struct page *page = pvec->pages[i];
192 struct zone *pagezone = page_zone(page);
194 if (pagezone != zone) {
195 if (zone)
196 spin_unlock_irqrestore(&zone->lru_lock, flags);
197 zone = pagezone;
198 spin_lock_irqsave(&zone->lru_lock, flags);
201 (*move_fn)(page, arg);
203 if (zone)
204 spin_unlock_irqrestore(&zone->lru_lock, flags);
205 release_pages(pvec->pages, pvec->nr, pvec->cold);
206 pagevec_reinit(pvec);
209 static void pagevec_move_tail_fn(struct page *page, void *arg)
211 int *pgmoved = arg;
212 struct zone *zone = page_zone(page);
214 if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
215 enum lru_list lru = page_lru_base_type(page);
216 list_move_tail(&page->lru, &zone->lru[lru].list);
217 mem_cgroup_rotate_reclaimable_page(page);
218 (*pgmoved)++;
223 * pagevec_move_tail() must be called with IRQ disabled.
224 * Otherwise this may cause nasty races.
226 static void pagevec_move_tail(struct pagevec *pvec)
228 int pgmoved = 0;
230 pagevec_lru_move_fn(pvec, pagevec_move_tail_fn, &pgmoved);
231 __count_vm_events(PGROTATED, pgmoved);
235 * Writeback is about to end against a page which has been marked for immediate
236 * reclaim. If it still appears to be reclaimable, move it to the tail of the
237 * inactive list.
239 void rotate_reclaimable_page(struct page *page)
241 if (!PageLocked(page) && !PageDirty(page) && !PageActive(page) &&
242 !PageUnevictable(page) && PageLRU(page)) {
243 struct pagevec *pvec;
244 unsigned long flags;
246 page_cache_get(page);
247 local_irq_save(flags);
248 pvec = &__get_cpu_var(lru_rotate_pvecs);
249 if (!pagevec_add(pvec, page))
250 pagevec_move_tail(pvec);
251 local_irq_restore(flags);
255 static void update_page_reclaim_stat(struct zone *zone, struct page *page,
256 int file, int rotated)
258 struct zone_reclaim_stat *reclaim_stat = &zone->reclaim_stat;
259 struct zone_reclaim_stat *memcg_reclaim_stat;
261 memcg_reclaim_stat = mem_cgroup_get_reclaim_stat_from_page(page);
263 reclaim_stat->recent_scanned[file]++;
264 if (rotated)
265 reclaim_stat->recent_rotated[file]++;
267 if (!memcg_reclaim_stat)
268 return;
270 memcg_reclaim_stat->recent_scanned[file]++;
271 if (rotated)
272 memcg_reclaim_stat->recent_rotated[file]++;
276 * FIXME: speed this up?
278 void activate_page(struct page *page)
280 struct zone *zone = page_zone(page);
282 spin_lock_irq(&zone->lru_lock);
283 if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
284 int file = page_is_file_cache(page);
285 int lru = page_lru_base_type(page);
286 del_page_from_lru_list(zone, page, lru);
288 SetPageActive(page);
289 lru += LRU_ACTIVE;
290 add_page_to_lru_list(zone, page, lru);
291 __count_vm_event(PGACTIVATE);
293 update_page_reclaim_stat(zone, page, file, 1);
295 spin_unlock_irq(&zone->lru_lock);
299 * Mark a page as having seen activity.
301 * inactive,unreferenced -> inactive,referenced
302 * inactive,referenced -> active,unreferenced
303 * active,unreferenced -> active,referenced
305 void mark_page_accessed(struct page *page)
307 if (!PageActive(page) && !PageUnevictable(page) &&
308 PageReferenced(page) && PageLRU(page)) {
309 activate_page(page);
310 ClearPageReferenced(page);
311 } else if (!PageReferenced(page)) {
312 SetPageReferenced(page);
316 EXPORT_SYMBOL(mark_page_accessed);
318 void __lru_cache_add(struct page *page, enum lru_list lru)
320 struct pagevec *pvec = &get_cpu_var(lru_add_pvecs)[lru];
322 page_cache_get(page);
323 if (!pagevec_add(pvec, page))
324 ____pagevec_lru_add(pvec, lru);
325 put_cpu_var(lru_add_pvecs);
327 EXPORT_SYMBOL(__lru_cache_add);
330 * lru_cache_add_lru - add a page to a page list
331 * @page: the page to be added to the LRU.
332 * @lru: the LRU list to which the page is added.
334 void lru_cache_add_lru(struct page *page, enum lru_list lru)
336 if (PageActive(page)) {
337 VM_BUG_ON(PageUnevictable(page));
338 ClearPageActive(page);
339 } else if (PageUnevictable(page)) {
340 VM_BUG_ON(PageActive(page));
341 ClearPageUnevictable(page);
344 VM_BUG_ON(PageLRU(page) || PageActive(page) || PageUnevictable(page));
345 __lru_cache_add(page, lru);
349 * add_page_to_unevictable_list - add a page to the unevictable list
350 * @page: the page to be added to the unevictable list
352 * Add page directly to its zone's unevictable list. To avoid races with
353 * tasks that might be making the page evictable, through eg. munlock,
354 * munmap or exit, while it's not on the lru, we want to add the page
355 * while it's locked or otherwise "invisible" to other tasks. This is
356 * difficult to do when using the pagevec cache, so bypass that.
358 void add_page_to_unevictable_list(struct page *page)
360 struct zone *zone = page_zone(page);
362 spin_lock_irq(&zone->lru_lock);
363 SetPageUnevictable(page);
364 SetPageLRU(page);
365 add_page_to_lru_list(zone, page, LRU_UNEVICTABLE);
366 spin_unlock_irq(&zone->lru_lock);
370 * If the page can not be invalidated, it is moved to the
371 * inactive list to speed up its reclaim. It is moved to the
372 * head of the list, rather than the tail, to give the flusher
373 * threads some time to write it out, as this is much more
374 * effective than the single-page writeout from reclaim.
376 * If the page isn't page_mapped and dirty/writeback, the page
377 * could reclaim asap using PG_reclaim.
379 * 1. active, mapped page -> none
380 * 2. active, dirty/writeback page -> inactive, head, PG_reclaim
381 * 3. inactive, mapped page -> none
382 * 4. inactive, dirty/writeback page -> inactive, head, PG_reclaim
383 * 5. inactive, clean -> inactive, tail
384 * 6. Others -> none
386 * In 4, why it moves inactive's head, the VM expects the page would
387 * be write it out by flusher threads as this is much more effective
388 * than the single-page writeout from reclaim.
390 static void lru_deactivate_fn(struct page *page, void *arg)
392 int lru, file;
393 bool active;
394 struct zone *zone = page_zone(page);
396 if (!PageLRU(page))
397 return;
399 if (PageUnevictable(page))
400 return;
402 /* Some processes are using the page */
403 if (page_mapped(page))
404 return;
406 active = PageActive(page);
408 file = page_is_file_cache(page);
409 lru = page_lru_base_type(page);
410 del_page_from_lru_list(zone, page, lru + active);
411 ClearPageActive(page);
412 ClearPageReferenced(page);
413 add_page_to_lru_list(zone, page, lru);
415 if (PageWriteback(page) || PageDirty(page)) {
417 * PG_reclaim could be raced with end_page_writeback
418 * It can make readahead confusing. But race window
419 * is _really_ small and it's non-critical problem.
421 SetPageReclaim(page);
422 } else {
424 * The page's writeback ends up during pagevec
425 * We moves tha page into tail of inactive.
427 list_move_tail(&page->lru, &zone->lru[lru].list);
428 mem_cgroup_rotate_reclaimable_page(page);
429 __count_vm_event(PGROTATED);
432 if (active)
433 __count_vm_event(PGDEACTIVATE);
434 update_page_reclaim_stat(zone, page, file, 0);
438 * Drain pages out of the cpu's pagevecs.
439 * Either "cpu" is the current CPU, and preemption has already been
440 * disabled; or "cpu" is being hot-unplugged, and is already dead.
442 static void drain_cpu_pagevecs(int cpu)
444 struct pagevec *pvecs = per_cpu(lru_add_pvecs, cpu);
445 struct pagevec *pvec;
446 int lru;
448 for_each_lru(lru) {
449 pvec = &pvecs[lru - LRU_BASE];
450 if (pagevec_count(pvec))
451 ____pagevec_lru_add(pvec, lru);
454 pvec = &per_cpu(lru_rotate_pvecs, cpu);
455 if (pagevec_count(pvec)) {
456 unsigned long flags;
458 /* No harm done if a racing interrupt already did this */
459 local_irq_save(flags);
460 pagevec_move_tail(pvec);
461 local_irq_restore(flags);
464 pvec = &per_cpu(lru_deactivate_pvecs, cpu);
465 if (pagevec_count(pvec))
466 pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
470 * deactivate_page - forcefully deactivate a page
471 * @page: page to deactivate
473 * This function hints the VM that @page is a good reclaim candidate,
474 * for example if its invalidation fails due to the page being dirty
475 * or under writeback.
477 void deactivate_page(struct page *page)
479 if (likely(get_page_unless_zero(page))) {
480 struct pagevec *pvec = &get_cpu_var(lru_deactivate_pvecs);
482 if (!pagevec_add(pvec, page))
483 pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
484 put_cpu_var(lru_deactivate_pvecs);
488 void lru_add_drain(void)
490 drain_cpu_pagevecs(get_cpu());
491 put_cpu();
494 static void lru_add_drain_per_cpu(struct work_struct *dummy)
496 lru_add_drain();
500 * Returns 0 for success
502 int lru_add_drain_all(void)
504 return schedule_on_each_cpu(lru_add_drain_per_cpu);
508 * Batched page_cache_release(). Decrement the reference count on all the
509 * passed pages. If it fell to zero then remove the page from the LRU and
510 * free it.
512 * Avoid taking zone->lru_lock if possible, but if it is taken, retain it
513 * for the remainder of the operation.
515 * The locking in this function is against shrink_inactive_list(): we recheck
516 * the page count inside the lock to see whether shrink_inactive_list()
517 * grabbed the page via the LRU. If it did, give up: shrink_inactive_list()
518 * will free it.
520 void release_pages(struct page **pages, int nr, int cold)
522 int i;
523 struct pagevec pages_to_free;
524 struct zone *zone = NULL;
525 unsigned long uninitialized_var(flags);
527 pagevec_init(&pages_to_free, cold);
528 for (i = 0; i < nr; i++) {
529 struct page *page = pages[i];
531 if (unlikely(PageCompound(page))) {
532 if (zone) {
533 spin_unlock_irqrestore(&zone->lru_lock, flags);
534 zone = NULL;
536 put_compound_page(page);
537 continue;
540 if (!put_page_testzero(page))
541 continue;
543 if (PageLRU(page)) {
544 struct zone *pagezone = page_zone(page);
546 if (pagezone != zone) {
547 if (zone)
548 spin_unlock_irqrestore(&zone->lru_lock,
549 flags);
550 zone = pagezone;
551 spin_lock_irqsave(&zone->lru_lock, flags);
553 VM_BUG_ON(!PageLRU(page));
554 __ClearPageLRU(page);
555 del_page_from_lru(zone, page);
558 if (!pagevec_add(&pages_to_free, page)) {
559 if (zone) {
560 spin_unlock_irqrestore(&zone->lru_lock, flags);
561 zone = NULL;
563 __pagevec_free(&pages_to_free);
564 pagevec_reinit(&pages_to_free);
567 if (zone)
568 spin_unlock_irqrestore(&zone->lru_lock, flags);
570 pagevec_free(&pages_to_free);
572 EXPORT_SYMBOL(release_pages);
575 * The pages which we're about to release may be in the deferred lru-addition
576 * queues. That would prevent them from really being freed right now. That's
577 * OK from a correctness point of view but is inefficient - those pages may be
578 * cache-warm and we want to give them back to the page allocator ASAP.
580 * So __pagevec_release() will drain those queues here. __pagevec_lru_add()
581 * and __pagevec_lru_add_active() call release_pages() directly to avoid
582 * mutual recursion.
584 void __pagevec_release(struct pagevec *pvec)
586 lru_add_drain();
587 release_pages(pvec->pages, pagevec_count(pvec), pvec->cold);
588 pagevec_reinit(pvec);
591 EXPORT_SYMBOL(__pagevec_release);
593 /* used by __split_huge_page_refcount() */
594 void lru_add_page_tail(struct zone* zone,
595 struct page *page, struct page *page_tail)
597 int active;
598 enum lru_list lru;
599 const int file = 0;
600 struct list_head *head;
602 VM_BUG_ON(!PageHead(page));
603 VM_BUG_ON(PageCompound(page_tail));
604 VM_BUG_ON(PageLRU(page_tail));
605 VM_BUG_ON(!spin_is_locked(&zone->lru_lock));
607 SetPageLRU(page_tail);
609 if (page_evictable(page_tail, NULL)) {
610 if (PageActive(page)) {
611 SetPageActive(page_tail);
612 active = 1;
613 lru = LRU_ACTIVE_ANON;
614 } else {
615 active = 0;
616 lru = LRU_INACTIVE_ANON;
618 update_page_reclaim_stat(zone, page_tail, file, active);
619 if (likely(PageLRU(page)))
620 head = page->lru.prev;
621 else
622 head = &zone->lru[lru].list;
623 __add_page_to_lru_list(zone, page_tail, lru, head);
624 } else {
625 SetPageUnevictable(page_tail);
626 add_page_to_lru_list(zone, page_tail, LRU_UNEVICTABLE);
630 static void ____pagevec_lru_add_fn(struct page *page, void *arg)
632 enum lru_list lru = (enum lru_list)arg;
633 struct zone *zone = page_zone(page);
634 int file = is_file_lru(lru);
635 int active = is_active_lru(lru);
637 VM_BUG_ON(PageActive(page));
638 VM_BUG_ON(PageUnevictable(page));
639 VM_BUG_ON(PageLRU(page));
641 SetPageLRU(page);
642 if (active)
643 SetPageActive(page);
644 update_page_reclaim_stat(zone, page, file, active);
645 add_page_to_lru_list(zone, page, lru);
649 * Add the passed pages to the LRU, then drop the caller's refcount
650 * on them. Reinitialises the caller's pagevec.
652 void ____pagevec_lru_add(struct pagevec *pvec, enum lru_list lru)
654 VM_BUG_ON(is_unevictable_lru(lru));
656 pagevec_lru_move_fn(pvec, ____pagevec_lru_add_fn, (void *)lru);
659 EXPORT_SYMBOL(____pagevec_lru_add);
662 * Try to drop buffers from the pages in a pagevec
664 void pagevec_strip(struct pagevec *pvec)
666 int i;
668 for (i = 0; i < pagevec_count(pvec); i++) {
669 struct page *page = pvec->pages[i];
671 if (page_has_private(page) && trylock_page(page)) {
672 if (page_has_private(page))
673 try_to_release_page(page, 0);
674 unlock_page(page);
680 * pagevec_lookup - gang pagecache lookup
681 * @pvec: Where the resulting pages are placed
682 * @mapping: The address_space to search
683 * @start: The starting page index
684 * @nr_pages: The maximum number of pages
686 * pagevec_lookup() will search for and return a group of up to @nr_pages pages
687 * in the mapping. The pages are placed in @pvec. pagevec_lookup() takes a
688 * reference against the pages in @pvec.
690 * The search returns a group of mapping-contiguous pages with ascending
691 * indexes. There may be holes in the indices due to not-present pages.
693 * pagevec_lookup() returns the number of pages which were found.
695 unsigned pagevec_lookup(struct pagevec *pvec, struct address_space *mapping,
696 pgoff_t start, unsigned nr_pages)
698 pvec->nr = find_get_pages(mapping, start, nr_pages, pvec->pages);
699 return pagevec_count(pvec);
702 EXPORT_SYMBOL(pagevec_lookup);
704 unsigned pagevec_lookup_tag(struct pagevec *pvec, struct address_space *mapping,
705 pgoff_t *index, int tag, unsigned nr_pages)
707 pvec->nr = find_get_pages_tag(mapping, index, tag,
708 nr_pages, pvec->pages);
709 return pagevec_count(pvec);
712 EXPORT_SYMBOL(pagevec_lookup_tag);
715 * Perform any setup for the swap system
717 void __init swap_setup(void)
719 unsigned long megs = totalram_pages >> (20 - PAGE_SHIFT);
721 #ifdef CONFIG_SWAP
722 bdi_init(swapper_space.backing_dev_info);
723 #endif
725 /* Use a smaller cluster for small-memory machines */
726 if (megs < 16)
727 page_cluster = 2;
728 else
729 page_cluster = 3;
731 * Right now other parts of the system means that we
732 * _really_ don't want to cluster much more