thp: transparent hugepage core
[linux-2.6.git] / mm / swap.c
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 #include <linux/gfp.h>
34
35 #include "internal.h"
36
37 /* How many pages do we try to swap or page in/out together? */
38 int page_cluster;
39
40 static DEFINE_PER_CPU(struct pagevec[NR_LRU_LISTS], lru_add_pvecs);
41 static DEFINE_PER_CPU(struct pagevec, lru_rotate_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  */
47 static void __page_cache_release(struct page *page)
48 {
49         if (PageLRU(page)) {
50                 unsigned long flags;
51                 struct zone *zone = page_zone(page);
52
53                 spin_lock_irqsave(&zone->lru_lock, flags);
54                 VM_BUG_ON(!PageLRU(page));
55                 __ClearPageLRU(page);
56                 del_page_from_lru(zone, page);
57                 spin_unlock_irqrestore(&zone->lru_lock, flags);
58         }
59 }
60
61 static void __put_single_page(struct page *page)
62 {
63         __page_cache_release(page);
64         free_hot_cold_page(page, 0);
65 }
66
67 static void __put_compound_page(struct page *page)
68 {
69         compound_page_dtor *dtor;
70
71         __page_cache_release(page);
72         dtor = get_compound_page_dtor(page);
73         (*dtor)(page);
74 }
75
76 static void put_compound_page(struct page *page)
77 {
78         if (unlikely(PageTail(page))) {
79                 /* __split_huge_page_refcount can run under us */
80                 struct page *page_head = page->first_page;
81                 smp_rmb();
82                 /*
83                  * If PageTail is still set after smp_rmb() we can be sure
84                  * that the page->first_page we read wasn't a dangling pointer.
85                  * See __split_huge_page_refcount() smp_wmb().
86                  */
87                 if (likely(PageTail(page) && get_page_unless_zero(page_head))) {
88                         unsigned long flags;
89                         /*
90                          * Verify that our page_head wasn't converted
91                          * to a a regular page before we got a
92                          * reference on it.
93                          */
94                         if (unlikely(!PageHead(page_head))) {
95                                 /* PageHead is cleared after PageTail */
96                                 smp_rmb();
97                                 VM_BUG_ON(PageTail(page));
98                                 goto out_put_head;
99                         }
100                         /*
101                          * Only run compound_lock on a valid PageHead,
102                          * after having it pinned with
103                          * get_page_unless_zero() above.
104                          */
105                         smp_mb();
106                         /* page_head wasn't a dangling pointer */
107                         flags = compound_lock_irqsave(page_head);
108                         if (unlikely(!PageTail(page))) {
109                                 /* __split_huge_page_refcount run before us */
110                                 compound_unlock_irqrestore(page_head, flags);
111                                 VM_BUG_ON(PageHead(page_head));
112                         out_put_head:
113                                 if (put_page_testzero(page_head))
114                                         __put_single_page(page_head);
115                         out_put_single:
116                                 if (put_page_testzero(page))
117                                         __put_single_page(page);
118                                 return;
119                         }
120                         VM_BUG_ON(page_head != page->first_page);
121                         /*
122                          * We can release the refcount taken by
123                          * get_page_unless_zero now that
124                          * split_huge_page_refcount is blocked on the
125                          * compound_lock.
126                          */
127                         if (put_page_testzero(page_head))
128                                 VM_BUG_ON(1);
129                         /* __split_huge_page_refcount will wait now */
130                         VM_BUG_ON(atomic_read(&page->_count) <= 0);
131                         atomic_dec(&page->_count);
132                         VM_BUG_ON(atomic_read(&page_head->_count) <= 0);
133                         compound_unlock_irqrestore(page_head, flags);
134                         if (put_page_testzero(page_head)) {
135                                 if (PageHead(page_head))
136                                         __put_compound_page(page_head);
137                                 else
138                                         __put_single_page(page_head);
139                         }
140                 } else {
141                         /* page_head is a dangling pointer */
142                         VM_BUG_ON(PageTail(page));
143                         goto out_put_single;
144                 }
145         } else if (put_page_testzero(page)) {
146                 if (PageHead(page))
147                         __put_compound_page(page);
148                 else
149                         __put_single_page(page);
150         }
151 }
152
153 void put_page(struct page *page)
154 {
155         if (unlikely(PageCompound(page)))
156                 put_compound_page(page);
157         else if (put_page_testzero(page))
158                 __put_single_page(page);
159 }
160 EXPORT_SYMBOL(put_page);
161
162 /**
163  * put_pages_list() - release a list of pages
164  * @pages: list of pages threaded on page->lru
165  *
166  * Release a list of pages which are strung together on page.lru.  Currently
167  * used by read_cache_pages() and related error recovery code.
168  */
169 void put_pages_list(struct list_head *pages)
170 {
171         while (!list_empty(pages)) {
172                 struct page *victim;
173
174                 victim = list_entry(pages->prev, struct page, lru);
175                 list_del(&victim->lru);
176                 page_cache_release(victim);
177         }
178 }
179 EXPORT_SYMBOL(put_pages_list);
180
181 /*
182  * pagevec_move_tail() must be called with IRQ disabled.
183  * Otherwise this may cause nasty races.
184  */
185 static void pagevec_move_tail(struct pagevec *pvec)
186 {
187         int i;
188         int pgmoved = 0;
189         struct zone *zone = NULL;
190
191         for (i = 0; i < pagevec_count(pvec); i++) {
192                 struct page *page = pvec->pages[i];
193                 struct zone *pagezone = page_zone(page);
194
195                 if (pagezone != zone) {
196                         if (zone)
197                                 spin_unlock(&zone->lru_lock);
198                         zone = pagezone;
199                         spin_lock(&zone->lru_lock);
200                 }
201                 if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
202                         int lru = page_lru_base_type(page);
203                         list_move_tail(&page->lru, &zone->lru[lru].list);
204                         pgmoved++;
205                 }
206         }
207         if (zone)
208                 spin_unlock(&zone->lru_lock);
209         __count_vm_events(PGROTATED, pgmoved);
210         release_pages(pvec->pages, pvec->nr, pvec->cold);
211         pagevec_reinit(pvec);
212 }
213
214 /*
215  * Writeback is about to end against a page which has been marked for immediate
216  * reclaim.  If it still appears to be reclaimable, move it to the tail of the
217  * inactive list.
218  */
219 void  rotate_reclaimable_page(struct page *page)
220 {
221         if (!PageLocked(page) && !PageDirty(page) && !PageActive(page) &&
222             !PageUnevictable(page) && PageLRU(page)) {
223                 struct pagevec *pvec;
224                 unsigned long flags;
225
226                 page_cache_get(page);
227                 local_irq_save(flags);
228                 pvec = &__get_cpu_var(lru_rotate_pvecs);
229                 if (!pagevec_add(pvec, page))
230                         pagevec_move_tail(pvec);
231                 local_irq_restore(flags);
232         }
233 }
234
235 static void update_page_reclaim_stat(struct zone *zone, struct page *page,
236                                      int file, int rotated)
237 {
238         struct zone_reclaim_stat *reclaim_stat = &zone->reclaim_stat;
239         struct zone_reclaim_stat *memcg_reclaim_stat;
240
241         memcg_reclaim_stat = mem_cgroup_get_reclaim_stat_from_page(page);
242
243         reclaim_stat->recent_scanned[file]++;
244         if (rotated)
245                 reclaim_stat->recent_rotated[file]++;
246
247         if (!memcg_reclaim_stat)
248                 return;
249
250         memcg_reclaim_stat->recent_scanned[file]++;
251         if (rotated)
252                 memcg_reclaim_stat->recent_rotated[file]++;
253 }
254
255 /*
256  * FIXME: speed this up?
257  */
258 void activate_page(struct page *page)
259 {
260         struct zone *zone = page_zone(page);
261
262         spin_lock_irq(&zone->lru_lock);
263         if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
264                 int file = page_is_file_cache(page);
265                 int lru = page_lru_base_type(page);
266                 del_page_from_lru_list(zone, page, lru);
267
268                 SetPageActive(page);
269                 lru += LRU_ACTIVE;
270                 add_page_to_lru_list(zone, page, lru);
271                 __count_vm_event(PGACTIVATE);
272
273                 update_page_reclaim_stat(zone, page, file, 1);
274         }
275         spin_unlock_irq(&zone->lru_lock);
276 }
277
278 /*
279  * Mark a page as having seen activity.
280  *
281  * inactive,unreferenced        ->      inactive,referenced
282  * inactive,referenced          ->      active,unreferenced
283  * active,unreferenced          ->      active,referenced
284  */
285 void mark_page_accessed(struct page *page)
286 {
287         if (!PageActive(page) && !PageUnevictable(page) &&
288                         PageReferenced(page) && PageLRU(page)) {
289                 activate_page(page);
290                 ClearPageReferenced(page);
291         } else if (!PageReferenced(page)) {
292                 SetPageReferenced(page);
293         }
294 }
295
296 EXPORT_SYMBOL(mark_page_accessed);
297
298 void __lru_cache_add(struct page *page, enum lru_list lru)
299 {
300         struct pagevec *pvec = &get_cpu_var(lru_add_pvecs)[lru];
301
302         page_cache_get(page);
303         if (!pagevec_add(pvec, page))
304                 ____pagevec_lru_add(pvec, lru);
305         put_cpu_var(lru_add_pvecs);
306 }
307 EXPORT_SYMBOL(__lru_cache_add);
308
309 /**
310  * lru_cache_add_lru - add a page to a page list
311  * @page: the page to be added to the LRU.
312  * @lru: the LRU list to which the page is added.
313  */
314 void lru_cache_add_lru(struct page *page, enum lru_list lru)
315 {
316         if (PageActive(page)) {
317                 VM_BUG_ON(PageUnevictable(page));
318                 ClearPageActive(page);
319         } else if (PageUnevictable(page)) {
320                 VM_BUG_ON(PageActive(page));
321                 ClearPageUnevictable(page);
322         }
323
324         VM_BUG_ON(PageLRU(page) || PageActive(page) || PageUnevictable(page));
325         __lru_cache_add(page, lru);
326 }
327
328 /**
329  * add_page_to_unevictable_list - add a page to the unevictable list
330  * @page:  the page to be added to the unevictable list
331  *
332  * Add page directly to its zone's unevictable list.  To avoid races with
333  * tasks that might be making the page evictable, through eg. munlock,
334  * munmap or exit, while it's not on the lru, we want to add the page
335  * while it's locked or otherwise "invisible" to other tasks.  This is
336  * difficult to do when using the pagevec cache, so bypass that.
337  */
338 void add_page_to_unevictable_list(struct page *page)
339 {
340         struct zone *zone = page_zone(page);
341
342         spin_lock_irq(&zone->lru_lock);
343         SetPageUnevictable(page);
344         SetPageLRU(page);
345         add_page_to_lru_list(zone, page, LRU_UNEVICTABLE);
346         spin_unlock_irq(&zone->lru_lock);
347 }
348
349 /*
350  * Drain pages out of the cpu's pagevecs.
351  * Either "cpu" is the current CPU, and preemption has already been
352  * disabled; or "cpu" is being hot-unplugged, and is already dead.
353  */
354 static void drain_cpu_pagevecs(int cpu)
355 {
356         struct pagevec *pvecs = per_cpu(lru_add_pvecs, cpu);
357         struct pagevec *pvec;
358         int lru;
359
360         for_each_lru(lru) {
361                 pvec = &pvecs[lru - LRU_BASE];
362                 if (pagevec_count(pvec))
363                         ____pagevec_lru_add(pvec, lru);
364         }
365
366         pvec = &per_cpu(lru_rotate_pvecs, cpu);
367         if (pagevec_count(pvec)) {
368                 unsigned long flags;
369
370                 /* No harm done if a racing interrupt already did this */
371                 local_irq_save(flags);
372                 pagevec_move_tail(pvec);
373                 local_irq_restore(flags);
374         }
375 }
376
377 void lru_add_drain(void)
378 {
379         drain_cpu_pagevecs(get_cpu());
380         put_cpu();
381 }
382
383 static void lru_add_drain_per_cpu(struct work_struct *dummy)
384 {
385         lru_add_drain();
386 }
387
388 /*
389  * Returns 0 for success
390  */
391 int lru_add_drain_all(void)
392 {
393         return schedule_on_each_cpu(lru_add_drain_per_cpu);
394 }
395
396 /*
397  * Batched page_cache_release().  Decrement the reference count on all the
398  * passed pages.  If it fell to zero then remove the page from the LRU and
399  * free it.
400  *
401  * Avoid taking zone->lru_lock if possible, but if it is taken, retain it
402  * for the remainder of the operation.
403  *
404  * The locking in this function is against shrink_inactive_list(): we recheck
405  * the page count inside the lock to see whether shrink_inactive_list()
406  * grabbed the page via the LRU.  If it did, give up: shrink_inactive_list()
407  * will free it.
408  */
409 void release_pages(struct page **pages, int nr, int cold)
410 {
411         int i;
412         struct pagevec pages_to_free;
413         struct zone *zone = NULL;
414         unsigned long uninitialized_var(flags);
415
416         pagevec_init(&pages_to_free, cold);
417         for (i = 0; i < nr; i++) {
418                 struct page *page = pages[i];
419
420                 if (unlikely(PageCompound(page))) {
421                         if (zone) {
422                                 spin_unlock_irqrestore(&zone->lru_lock, flags);
423                                 zone = NULL;
424                         }
425                         put_compound_page(page);
426                         continue;
427                 }
428
429                 if (!put_page_testzero(page))
430                         continue;
431
432                 if (PageLRU(page)) {
433                         struct zone *pagezone = page_zone(page);
434
435                         if (pagezone != zone) {
436                                 if (zone)
437                                         spin_unlock_irqrestore(&zone->lru_lock,
438                                                                         flags);
439                                 zone = pagezone;
440                                 spin_lock_irqsave(&zone->lru_lock, flags);
441                         }
442                         VM_BUG_ON(!PageLRU(page));
443                         __ClearPageLRU(page);
444                         del_page_from_lru(zone, page);
445                 }
446
447                 if (!pagevec_add(&pages_to_free, page)) {
448                         if (zone) {
449                                 spin_unlock_irqrestore(&zone->lru_lock, flags);
450                                 zone = NULL;
451                         }
452                         __pagevec_free(&pages_to_free);
453                         pagevec_reinit(&pages_to_free);
454                 }
455         }
456         if (zone)
457                 spin_unlock_irqrestore(&zone->lru_lock, flags);
458
459         pagevec_free(&pages_to_free);
460 }
461 EXPORT_SYMBOL(release_pages);
462
463 /*
464  * The pages which we're about to release may be in the deferred lru-addition
465  * queues.  That would prevent them from really being freed right now.  That's
466  * OK from a correctness point of view but is inefficient - those pages may be
467  * cache-warm and we want to give them back to the page allocator ASAP.
468  *
469  * So __pagevec_release() will drain those queues here.  __pagevec_lru_add()
470  * and __pagevec_lru_add_active() call release_pages() directly to avoid
471  * mutual recursion.
472  */
473 void __pagevec_release(struct pagevec *pvec)
474 {
475         lru_add_drain();
476         release_pages(pvec->pages, pagevec_count(pvec), pvec->cold);
477         pagevec_reinit(pvec);
478 }
479
480 EXPORT_SYMBOL(__pagevec_release);
481
482 /* used by __split_huge_page_refcount() */
483 void lru_add_page_tail(struct zone* zone,
484                        struct page *page, struct page *page_tail)
485 {
486         int active;
487         enum lru_list lru;
488         const int file = 0;
489         struct list_head *head;
490
491         VM_BUG_ON(!PageHead(page));
492         VM_BUG_ON(PageCompound(page_tail));
493         VM_BUG_ON(PageLRU(page_tail));
494         VM_BUG_ON(!spin_is_locked(&zone->lru_lock));
495
496         SetPageLRU(page_tail);
497
498         if (page_evictable(page_tail, NULL)) {
499                 if (PageActive(page)) {
500                         SetPageActive(page_tail);
501                         active = 1;
502                         lru = LRU_ACTIVE_ANON;
503                 } else {
504                         active = 0;
505                         lru = LRU_INACTIVE_ANON;
506                 }
507                 update_page_reclaim_stat(zone, page_tail, file, active);
508                 if (likely(PageLRU(page)))
509                         head = page->lru.prev;
510                 else
511                         head = &zone->lru[lru].list;
512                 __add_page_to_lru_list(zone, page_tail, lru, head);
513         } else {
514                 SetPageUnevictable(page_tail);
515                 add_page_to_lru_list(zone, page_tail, LRU_UNEVICTABLE);
516         }
517 }
518
519 /*
520  * Add the passed pages to the LRU, then drop the caller's refcount
521  * on them.  Reinitialises the caller's pagevec.
522  */
523 void ____pagevec_lru_add(struct pagevec *pvec, enum lru_list lru)
524 {
525         int i;
526         struct zone *zone = NULL;
527
528         VM_BUG_ON(is_unevictable_lru(lru));
529
530         for (i = 0; i < pagevec_count(pvec); i++) {
531                 struct page *page = pvec->pages[i];
532                 struct zone *pagezone = page_zone(page);
533                 int file;
534                 int active;
535
536                 if (pagezone != zone) {
537                         if (zone)
538                                 spin_unlock_irq(&zone->lru_lock);
539                         zone = pagezone;
540                         spin_lock_irq(&zone->lru_lock);
541                 }
542                 VM_BUG_ON(PageActive(page));
543                 VM_BUG_ON(PageUnevictable(page));
544                 VM_BUG_ON(PageLRU(page));
545                 SetPageLRU(page);
546                 active = is_active_lru(lru);
547                 file = is_file_lru(lru);
548                 if (active)
549                         SetPageActive(page);
550                 update_page_reclaim_stat(zone, page, file, active);
551                 add_page_to_lru_list(zone, page, lru);
552         }
553         if (zone)
554                 spin_unlock_irq(&zone->lru_lock);
555         release_pages(pvec->pages, pvec->nr, pvec->cold);
556         pagevec_reinit(pvec);
557 }
558
559 EXPORT_SYMBOL(____pagevec_lru_add);
560
561 /*
562  * Try to drop buffers from the pages in a pagevec
563  */
564 void pagevec_strip(struct pagevec *pvec)
565 {
566         int i;
567
568         for (i = 0; i < pagevec_count(pvec); i++) {
569                 struct page *page = pvec->pages[i];
570
571                 if (page_has_private(page) && trylock_page(page)) {
572                         if (page_has_private(page))
573                                 try_to_release_page(page, 0);
574                         unlock_page(page);
575                 }
576         }
577 }
578
579 /**
580  * pagevec_lookup - gang pagecache lookup
581  * @pvec:       Where the resulting pages are placed
582  * @mapping:    The address_space to search
583  * @start:      The starting page index
584  * @nr_pages:   The maximum number of pages
585  *
586  * pagevec_lookup() will search for and return a group of up to @nr_pages pages
587  * in the mapping.  The pages are placed in @pvec.  pagevec_lookup() takes a
588  * reference against the pages in @pvec.
589  *
590  * The search returns a group of mapping-contiguous pages with ascending
591  * indexes.  There may be holes in the indices due to not-present pages.
592  *
593  * pagevec_lookup() returns the number of pages which were found.
594  */
595 unsigned pagevec_lookup(struct pagevec *pvec, struct address_space *mapping,
596                 pgoff_t start, unsigned nr_pages)
597 {
598         pvec->nr = find_get_pages(mapping, start, nr_pages, pvec->pages);
599         return pagevec_count(pvec);
600 }
601
602 EXPORT_SYMBOL(pagevec_lookup);
603
604 unsigned pagevec_lookup_tag(struct pagevec *pvec, struct address_space *mapping,
605                 pgoff_t *index, int tag, unsigned nr_pages)
606 {
607         pvec->nr = find_get_pages_tag(mapping, index, tag,
608                                         nr_pages, pvec->pages);
609         return pagevec_count(pvec);
610 }
611
612 EXPORT_SYMBOL(pagevec_lookup_tag);
613
614 /*
615  * Perform any setup for the swap system
616  */
617 void __init swap_setup(void)
618 {
619         unsigned long megs = totalram_pages >> (20 - PAGE_SHIFT);
620
621 #ifdef CONFIG_SWAP
622         bdi_init(swapper_space.backing_dev_info);
623 #endif
624
625         /* Use a smaller cluster for small-memory machines */
626         if (megs < 16)
627                 page_cluster = 2;
628         else
629                 page_cluster = 3;
630         /*
631          * Right now other parts of the system means that we
632          * _really_ don't want to cluster much more
633          */
634 }