9c7e57cc63a34f7231b77a7d8b395d3157a34ba6
[linux-2.6.git] / mm / vmscan.c
1 /*
2  *  linux/mm/vmscan.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  *
6  *  Swap reorganised 29.12.95, Stephen Tweedie.
7  *  kswapd added: 7.1.96  sct
8  *  Removed kswapd_ctl limits, and swap out as many pages as needed
9  *  to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10  *  Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11  *  Multiqueue VM started 5.8.00, Rik van Riel.
12  */
13
14 #include <linux/mm.h>
15 #include <linux/module.h>
16 #include <linux/gfp.h>
17 #include <linux/kernel_stat.h>
18 #include <linux/swap.h>
19 #include <linux/pagemap.h>
20 #include <linux/init.h>
21 #include <linux/highmem.h>
22 #include <linux/vmstat.h>
23 #include <linux/file.h>
24 #include <linux/writeback.h>
25 #include <linux/blkdev.h>
26 #include <linux/buffer_head.h>  /* for try_to_release_page(),
27                                         buffer_heads_over_limit */
28 #include <linux/mm_inline.h>
29 #include <linux/pagevec.h>
30 #include <linux/backing-dev.h>
31 #include <linux/rmap.h>
32 #include <linux/topology.h>
33 #include <linux/cpu.h>
34 #include <linux/cpuset.h>
35 #include <linux/notifier.h>
36 #include <linux/rwsem.h>
37 #include <linux/delay.h>
38 #include <linux/kthread.h>
39 #include <linux/freezer.h>
40 #include <linux/memcontrol.h>
41 #include <linux/delayacct.h>
42 #include <linux/sysctl.h>
43
44 #include <asm/tlbflush.h>
45 #include <asm/div64.h>
46
47 #include <linux/swapops.h>
48
49 #include "internal.h"
50
51 struct scan_control {
52         /* Incremented by the number of inactive pages that were scanned */
53         unsigned long nr_scanned;
54
55         /* Number of pages freed so far during a call to shrink_zones() */
56         unsigned long nr_reclaimed;
57
58         /* How many pages shrink_list() should reclaim */
59         unsigned long nr_to_reclaim;
60
61         unsigned long hibernation_mode;
62
63         /* This context's GFP mask */
64         gfp_t gfp_mask;
65
66         int may_writepage;
67
68         /* Can mapped pages be reclaimed? */
69         int may_unmap;
70
71         /* Can pages be swapped as part of reclaim? */
72         int may_swap;
73
74         int swappiness;
75
76         int order;
77
78         /*
79          * Intend to reclaim enough contenious memory rather than to reclaim
80          * enough amount memory. I.e, it's the mode for high order allocation.
81          */
82         bool lumpy_reclaim_mode;
83
84         /* Which cgroup do we reclaim from */
85         struct mem_cgroup *mem_cgroup;
86
87         /*
88          * Nodemask of nodes allowed by the caller. If NULL, all nodes
89          * are scanned.
90          */
91         nodemask_t      *nodemask;
92 };
93
94 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
95
96 #ifdef ARCH_HAS_PREFETCH
97 #define prefetch_prev_lru_page(_page, _base, _field)                    \
98         do {                                                            \
99                 if ((_page)->lru.prev != _base) {                       \
100                         struct page *prev;                              \
101                                                                         \
102                         prev = lru_to_page(&(_page->lru));              \
103                         prefetch(&prev->_field);                        \
104                 }                                                       \
105         } while (0)
106 #else
107 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
108 #endif
109
110 #ifdef ARCH_HAS_PREFETCHW
111 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
112         do {                                                            \
113                 if ((_page)->lru.prev != _base) {                       \
114                         struct page *prev;                              \
115                                                                         \
116                         prev = lru_to_page(&(_page->lru));              \
117                         prefetchw(&prev->_field);                       \
118                 }                                                       \
119         } while (0)
120 #else
121 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
122 #endif
123
124 /*
125  * From 0 .. 100.  Higher means more swappy.
126  */
127 int vm_swappiness = 60;
128 long vm_total_pages;    /* The total number of pages which the VM controls */
129
130 static LIST_HEAD(shrinker_list);
131 static DECLARE_RWSEM(shrinker_rwsem);
132
133 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
134 #define scanning_global_lru(sc) (!(sc)->mem_cgroup)
135 #else
136 #define scanning_global_lru(sc) (1)
137 #endif
138
139 static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
140                                                   struct scan_control *sc)
141 {
142         if (!scanning_global_lru(sc))
143                 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
144
145         return &zone->reclaim_stat;
146 }
147
148 static unsigned long zone_nr_lru_pages(struct zone *zone,
149                                 struct scan_control *sc, enum lru_list lru)
150 {
151         if (!scanning_global_lru(sc))
152                 return mem_cgroup_zone_nr_pages(sc->mem_cgroup, zone, lru);
153
154         return zone_page_state(zone, NR_LRU_BASE + lru);
155 }
156
157
158 /*
159  * Add a shrinker callback to be called from the vm
160  */
161 void register_shrinker(struct shrinker *shrinker)
162 {
163         shrinker->nr = 0;
164         down_write(&shrinker_rwsem);
165         list_add_tail(&shrinker->list, &shrinker_list);
166         up_write(&shrinker_rwsem);
167 }
168 EXPORT_SYMBOL(register_shrinker);
169
170 /*
171  * Remove one
172  */
173 void unregister_shrinker(struct shrinker *shrinker)
174 {
175         down_write(&shrinker_rwsem);
176         list_del(&shrinker->list);
177         up_write(&shrinker_rwsem);
178 }
179 EXPORT_SYMBOL(unregister_shrinker);
180
181 #define SHRINK_BATCH 128
182 /*
183  * Call the shrink functions to age shrinkable caches
184  *
185  * Here we assume it costs one seek to replace a lru page and that it also
186  * takes a seek to recreate a cache object.  With this in mind we age equal
187  * percentages of the lru and ageable caches.  This should balance the seeks
188  * generated by these structures.
189  *
190  * If the vm encountered mapped pages on the LRU it increase the pressure on
191  * slab to avoid swapping.
192  *
193  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
194  *
195  * `lru_pages' represents the number of on-LRU pages in all the zones which
196  * are eligible for the caller's allocation attempt.  It is used for balancing
197  * slab reclaim versus page reclaim.
198  *
199  * Returns the number of slab objects which we shrunk.
200  */
201 unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
202                         unsigned long lru_pages)
203 {
204         struct shrinker *shrinker;
205         unsigned long ret = 0;
206
207         if (scanned == 0)
208                 scanned = SWAP_CLUSTER_MAX;
209
210         if (!down_read_trylock(&shrinker_rwsem))
211                 return 1;       /* Assume we'll be able to shrink next time */
212
213         list_for_each_entry(shrinker, &shrinker_list, list) {
214                 unsigned long long delta;
215                 unsigned long total_scan;
216                 unsigned long max_pass = (*shrinker->shrink)(0, gfp_mask);
217
218                 delta = (4 * scanned) / shrinker->seeks;
219                 delta *= max_pass;
220                 do_div(delta, lru_pages + 1);
221                 shrinker->nr += delta;
222                 if (shrinker->nr < 0) {
223                         printk(KERN_ERR "shrink_slab: %pF negative objects to "
224                                "delete nr=%ld\n",
225                                shrinker->shrink, shrinker->nr);
226                         shrinker->nr = max_pass;
227                 }
228
229                 /*
230                  * Avoid risking looping forever due to too large nr value:
231                  * never try to free more than twice the estimate number of
232                  * freeable entries.
233                  */
234                 if (shrinker->nr > max_pass * 2)
235                         shrinker->nr = max_pass * 2;
236
237                 total_scan = shrinker->nr;
238                 shrinker->nr = 0;
239
240                 while (total_scan >= SHRINK_BATCH) {
241                         long this_scan = SHRINK_BATCH;
242                         int shrink_ret;
243                         int nr_before;
244
245                         nr_before = (*shrinker->shrink)(0, gfp_mask);
246                         shrink_ret = (*shrinker->shrink)(this_scan, gfp_mask);
247                         if (shrink_ret == -1)
248                                 break;
249                         if (shrink_ret < nr_before)
250                                 ret += nr_before - shrink_ret;
251                         count_vm_events(SLABS_SCANNED, this_scan);
252                         total_scan -= this_scan;
253
254                         cond_resched();
255                 }
256
257                 shrinker->nr += total_scan;
258         }
259         up_read(&shrinker_rwsem);
260         return ret;
261 }
262
263 static inline int is_page_cache_freeable(struct page *page)
264 {
265         /*
266          * A freeable page cache page is referenced only by the caller
267          * that isolated the page, the page cache radix tree and
268          * optional buffer heads at page->private.
269          */
270         return page_count(page) - page_has_private(page) == 2;
271 }
272
273 static int may_write_to_queue(struct backing_dev_info *bdi)
274 {
275         if (current->flags & PF_SWAPWRITE)
276                 return 1;
277         if (!bdi_write_congested(bdi))
278                 return 1;
279         if (bdi == current->backing_dev_info)
280                 return 1;
281         return 0;
282 }
283
284 /*
285  * We detected a synchronous write error writing a page out.  Probably
286  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
287  * fsync(), msync() or close().
288  *
289  * The tricky part is that after writepage we cannot touch the mapping: nothing
290  * prevents it from being freed up.  But we have a ref on the page and once
291  * that page is locked, the mapping is pinned.
292  *
293  * We're allowed to run sleeping lock_page() here because we know the caller has
294  * __GFP_FS.
295  */
296 static void handle_write_error(struct address_space *mapping,
297                                 struct page *page, int error)
298 {
299         lock_page(page);
300         if (page_mapping(page) == mapping)
301                 mapping_set_error(mapping, error);
302         unlock_page(page);
303 }
304
305 /* Request for sync pageout. */
306 enum pageout_io {
307         PAGEOUT_IO_ASYNC,
308         PAGEOUT_IO_SYNC,
309 };
310
311 /* possible outcome of pageout() */
312 typedef enum {
313         /* failed to write page out, page is locked */
314         PAGE_KEEP,
315         /* move page to the active list, page is locked */
316         PAGE_ACTIVATE,
317         /* page has been sent to the disk successfully, page is unlocked */
318         PAGE_SUCCESS,
319         /* page is clean and locked */
320         PAGE_CLEAN,
321 } pageout_t;
322
323 /*
324  * pageout is called by shrink_page_list() for each dirty page.
325  * Calls ->writepage().
326  */
327 static pageout_t pageout(struct page *page, struct address_space *mapping,
328                                                 enum pageout_io sync_writeback)
329 {
330         /*
331          * If the page is dirty, only perform writeback if that write
332          * will be non-blocking.  To prevent this allocation from being
333          * stalled by pagecache activity.  But note that there may be
334          * stalls if we need to run get_block().  We could test
335          * PagePrivate for that.
336          *
337          * If this process is currently in __generic_file_aio_write() against
338          * this page's queue, we can perform writeback even if that
339          * will block.
340          *
341          * If the page is swapcache, write it back even if that would
342          * block, for some throttling. This happens by accident, because
343          * swap_backing_dev_info is bust: it doesn't reflect the
344          * congestion state of the swapdevs.  Easy to fix, if needed.
345          */
346         if (!is_page_cache_freeable(page))
347                 return PAGE_KEEP;
348         if (!mapping) {
349                 /*
350                  * Some data journaling orphaned pages can have
351                  * page->mapping == NULL while being dirty with clean buffers.
352                  */
353                 if (page_has_private(page)) {
354                         if (try_to_free_buffers(page)) {
355                                 ClearPageDirty(page);
356                                 printk("%s: orphaned page\n", __func__);
357                                 return PAGE_CLEAN;
358                         }
359                 }
360                 return PAGE_KEEP;
361         }
362         if (mapping->a_ops->writepage == NULL)
363                 return PAGE_ACTIVATE;
364         if (!may_write_to_queue(mapping->backing_dev_info))
365                 return PAGE_KEEP;
366
367         if (clear_page_dirty_for_io(page)) {
368                 int res;
369                 struct writeback_control wbc = {
370                         .sync_mode = WB_SYNC_NONE,
371                         .nr_to_write = SWAP_CLUSTER_MAX,
372                         .range_start = 0,
373                         .range_end = LLONG_MAX,
374                         .nonblocking = 1,
375                         .for_reclaim = 1,
376                 };
377
378                 SetPageReclaim(page);
379                 res = mapping->a_ops->writepage(page, &wbc);
380                 if (res < 0)
381                         handle_write_error(mapping, page, res);
382                 if (res == AOP_WRITEPAGE_ACTIVATE) {
383                         ClearPageReclaim(page);
384                         return PAGE_ACTIVATE;
385                 }
386
387                 /*
388                  * Wait on writeback if requested to. This happens when
389                  * direct reclaiming a large contiguous area and the
390                  * first attempt to free a range of pages fails.
391                  */
392                 if (PageWriteback(page) && sync_writeback == PAGEOUT_IO_SYNC)
393                         wait_on_page_writeback(page);
394
395                 if (!PageWriteback(page)) {
396                         /* synchronous write or broken a_ops? */
397                         ClearPageReclaim(page);
398                 }
399                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
400                 return PAGE_SUCCESS;
401         }
402
403         return PAGE_CLEAN;
404 }
405
406 /*
407  * Same as remove_mapping, but if the page is removed from the mapping, it
408  * gets returned with a refcount of 0.
409  */
410 static int __remove_mapping(struct address_space *mapping, struct page *page)
411 {
412         BUG_ON(!PageLocked(page));
413         BUG_ON(mapping != page_mapping(page));
414
415         spin_lock_irq(&mapping->tree_lock);
416         /*
417          * The non racy check for a busy page.
418          *
419          * Must be careful with the order of the tests. When someone has
420          * a ref to the page, it may be possible that they dirty it then
421          * drop the reference. So if PageDirty is tested before page_count
422          * here, then the following race may occur:
423          *
424          * get_user_pages(&page);
425          * [user mapping goes away]
426          * write_to(page);
427          *                              !PageDirty(page)    [good]
428          * SetPageDirty(page);
429          * put_page(page);
430          *                              !page_count(page)   [good, discard it]
431          *
432          * [oops, our write_to data is lost]
433          *
434          * Reversing the order of the tests ensures such a situation cannot
435          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
436          * load is not satisfied before that of page->_count.
437          *
438          * Note that if SetPageDirty is always performed via set_page_dirty,
439          * and thus under tree_lock, then this ordering is not required.
440          */
441         if (!page_freeze_refs(page, 2))
442                 goto cannot_free;
443         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
444         if (unlikely(PageDirty(page))) {
445                 page_unfreeze_refs(page, 2);
446                 goto cannot_free;
447         }
448
449         if (PageSwapCache(page)) {
450                 swp_entry_t swap = { .val = page_private(page) };
451                 __delete_from_swap_cache(page);
452                 spin_unlock_irq(&mapping->tree_lock);
453                 swapcache_free(swap, page);
454         } else {
455                 __remove_from_page_cache(page);
456                 spin_unlock_irq(&mapping->tree_lock);
457                 mem_cgroup_uncharge_cache_page(page);
458         }
459
460         return 1;
461
462 cannot_free:
463         spin_unlock_irq(&mapping->tree_lock);
464         return 0;
465 }
466
467 /*
468  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
469  * someone else has a ref on the page, abort and return 0.  If it was
470  * successfully detached, return 1.  Assumes the caller has a single ref on
471  * this page.
472  */
473 int remove_mapping(struct address_space *mapping, struct page *page)
474 {
475         if (__remove_mapping(mapping, page)) {
476                 /*
477                  * Unfreezing the refcount with 1 rather than 2 effectively
478                  * drops the pagecache ref for us without requiring another
479                  * atomic operation.
480                  */
481                 page_unfreeze_refs(page, 1);
482                 return 1;
483         }
484         return 0;
485 }
486
487 /**
488  * putback_lru_page - put previously isolated page onto appropriate LRU list
489  * @page: page to be put back to appropriate lru list
490  *
491  * Add previously isolated @page to appropriate LRU list.
492  * Page may still be unevictable for other reasons.
493  *
494  * lru_lock must not be held, interrupts must be enabled.
495  */
496 void putback_lru_page(struct page *page)
497 {
498         int lru;
499         int active = !!TestClearPageActive(page);
500         int was_unevictable = PageUnevictable(page);
501
502         VM_BUG_ON(PageLRU(page));
503
504 redo:
505         ClearPageUnevictable(page);
506
507         if (page_evictable(page, NULL)) {
508                 /*
509                  * For evictable pages, we can use the cache.
510                  * In event of a race, worst case is we end up with an
511                  * unevictable page on [in]active list.
512                  * We know how to handle that.
513                  */
514                 lru = active + page_lru_base_type(page);
515                 lru_cache_add_lru(page, lru);
516         } else {
517                 /*
518                  * Put unevictable pages directly on zone's unevictable
519                  * list.
520                  */
521                 lru = LRU_UNEVICTABLE;
522                 add_page_to_unevictable_list(page);
523                 /*
524                  * When racing with an mlock clearing (page is
525                  * unlocked), make sure that if the other thread does
526                  * not observe our setting of PG_lru and fails
527                  * isolation, we see PG_mlocked cleared below and move
528                  * the page back to the evictable list.
529                  *
530                  * The other side is TestClearPageMlocked().
531                  */
532                 smp_mb();
533         }
534
535         /*
536          * page's status can change while we move it among lru. If an evictable
537          * page is on unevictable list, it never be freed. To avoid that,
538          * check after we added it to the list, again.
539          */
540         if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
541                 if (!isolate_lru_page(page)) {
542                         put_page(page);
543                         goto redo;
544                 }
545                 /* This means someone else dropped this page from LRU
546                  * So, it will be freed or putback to LRU again. There is
547                  * nothing to do here.
548                  */
549         }
550
551         if (was_unevictable && lru != LRU_UNEVICTABLE)
552                 count_vm_event(UNEVICTABLE_PGRESCUED);
553         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
554                 count_vm_event(UNEVICTABLE_PGCULLED);
555
556         put_page(page);         /* drop ref from isolate */
557 }
558
559 enum page_references {
560         PAGEREF_RECLAIM,
561         PAGEREF_RECLAIM_CLEAN,
562         PAGEREF_KEEP,
563         PAGEREF_ACTIVATE,
564 };
565
566 static enum page_references page_check_references(struct page *page,
567                                                   struct scan_control *sc)
568 {
569         int referenced_ptes, referenced_page;
570         unsigned long vm_flags;
571
572         referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
573         referenced_page = TestClearPageReferenced(page);
574
575         /* Lumpy reclaim - ignore references */
576         if (sc->lumpy_reclaim_mode)
577                 return PAGEREF_RECLAIM;
578
579         /*
580          * Mlock lost the isolation race with us.  Let try_to_unmap()
581          * move the page to the unevictable list.
582          */
583         if (vm_flags & VM_LOCKED)
584                 return PAGEREF_RECLAIM;
585
586         if (referenced_ptes) {
587                 if (PageAnon(page))
588                         return PAGEREF_ACTIVATE;
589                 /*
590                  * All mapped pages start out with page table
591                  * references from the instantiating fault, so we need
592                  * to look twice if a mapped file page is used more
593                  * than once.
594                  *
595                  * Mark it and spare it for another trip around the
596                  * inactive list.  Another page table reference will
597                  * lead to its activation.
598                  *
599                  * Note: the mark is set for activated pages as well
600                  * so that recently deactivated but used pages are
601                  * quickly recovered.
602                  */
603                 SetPageReferenced(page);
604
605                 if (referenced_page)
606                         return PAGEREF_ACTIVATE;
607
608                 return PAGEREF_KEEP;
609         }
610
611         /* Reclaim if clean, defer dirty pages to writeback */
612         if (referenced_page)
613                 return PAGEREF_RECLAIM_CLEAN;
614
615         return PAGEREF_RECLAIM;
616 }
617
618 /*
619  * shrink_page_list() returns the number of reclaimed pages
620  */
621 static unsigned long shrink_page_list(struct list_head *page_list,
622                                         struct scan_control *sc,
623                                         enum pageout_io sync_writeback)
624 {
625         LIST_HEAD(ret_pages);
626         struct pagevec freed_pvec;
627         int pgactivate = 0;
628         unsigned long nr_reclaimed = 0;
629
630         cond_resched();
631
632         pagevec_init(&freed_pvec, 1);
633         while (!list_empty(page_list)) {
634                 enum page_references references;
635                 struct address_space *mapping;
636                 struct page *page;
637                 int may_enter_fs;
638
639                 cond_resched();
640
641                 page = lru_to_page(page_list);
642                 list_del(&page->lru);
643
644                 if (!trylock_page(page))
645                         goto keep;
646
647                 VM_BUG_ON(PageActive(page));
648
649                 sc->nr_scanned++;
650
651                 if (unlikely(!page_evictable(page, NULL)))
652                         goto cull_mlocked;
653
654                 if (!sc->may_unmap && page_mapped(page))
655                         goto keep_locked;
656
657                 /* Double the slab pressure for mapped and swapcache pages */
658                 if (page_mapped(page) || PageSwapCache(page))
659                         sc->nr_scanned++;
660
661                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
662                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
663
664                 if (PageWriteback(page)) {
665                         /*
666                          * Synchronous reclaim is performed in two passes,
667                          * first an asynchronous pass over the list to
668                          * start parallel writeback, and a second synchronous
669                          * pass to wait for the IO to complete.  Wait here
670                          * for any page for which writeback has already
671                          * started.
672                          */
673                         if (sync_writeback == PAGEOUT_IO_SYNC && may_enter_fs)
674                                 wait_on_page_writeback(page);
675                         else
676                                 goto keep_locked;
677                 }
678
679                 references = page_check_references(page, sc);
680                 switch (references) {
681                 case PAGEREF_ACTIVATE:
682                         goto activate_locked;
683                 case PAGEREF_KEEP:
684                         goto keep_locked;
685                 case PAGEREF_RECLAIM:
686                 case PAGEREF_RECLAIM_CLEAN:
687                         ; /* try to reclaim the page below */
688                 }
689
690                 /*
691                  * Anonymous process memory has backing store?
692                  * Try to allocate it some swap space here.
693                  */
694                 if (PageAnon(page) && !PageSwapCache(page)) {
695                         if (!(sc->gfp_mask & __GFP_IO))
696                                 goto keep_locked;
697                         if (!add_to_swap(page))
698                                 goto activate_locked;
699                         may_enter_fs = 1;
700                 }
701
702                 mapping = page_mapping(page);
703
704                 /*
705                  * The page is mapped into the page tables of one or more
706                  * processes. Try to unmap it here.
707                  */
708                 if (page_mapped(page) && mapping) {
709                         switch (try_to_unmap(page, TTU_UNMAP)) {
710                         case SWAP_FAIL:
711                                 goto activate_locked;
712                         case SWAP_AGAIN:
713                                 goto keep_locked;
714                         case SWAP_MLOCK:
715                                 goto cull_mlocked;
716                         case SWAP_SUCCESS:
717                                 ; /* try to free the page below */
718                         }
719                 }
720
721                 if (PageDirty(page)) {
722                         if (references == PAGEREF_RECLAIM_CLEAN)
723                                 goto keep_locked;
724                         if (!may_enter_fs)
725                                 goto keep_locked;
726                         if (!sc->may_writepage)
727                                 goto keep_locked;
728
729                         /* Page is dirty, try to write it out here */
730                         switch (pageout(page, mapping, sync_writeback)) {
731                         case PAGE_KEEP:
732                                 goto keep_locked;
733                         case PAGE_ACTIVATE:
734                                 goto activate_locked;
735                         case PAGE_SUCCESS:
736                                 if (PageWriteback(page) || PageDirty(page))
737                                         goto keep;
738                                 /*
739                                  * A synchronous write - probably a ramdisk.  Go
740                                  * ahead and try to reclaim the page.
741                                  */
742                                 if (!trylock_page(page))
743                                         goto keep;
744                                 if (PageDirty(page) || PageWriteback(page))
745                                         goto keep_locked;
746                                 mapping = page_mapping(page);
747                         case PAGE_CLEAN:
748                                 ; /* try to free the page below */
749                         }
750                 }
751
752                 /*
753                  * If the page has buffers, try to free the buffer mappings
754                  * associated with this page. If we succeed we try to free
755                  * the page as well.
756                  *
757                  * We do this even if the page is PageDirty().
758                  * try_to_release_page() does not perform I/O, but it is
759                  * possible for a page to have PageDirty set, but it is actually
760                  * clean (all its buffers are clean).  This happens if the
761                  * buffers were written out directly, with submit_bh(). ext3
762                  * will do this, as well as the blockdev mapping.
763                  * try_to_release_page() will discover that cleanness and will
764                  * drop the buffers and mark the page clean - it can be freed.
765                  *
766                  * Rarely, pages can have buffers and no ->mapping.  These are
767                  * the pages which were not successfully invalidated in
768                  * truncate_complete_page().  We try to drop those buffers here
769                  * and if that worked, and the page is no longer mapped into
770                  * process address space (page_count == 1) it can be freed.
771                  * Otherwise, leave the page on the LRU so it is swappable.
772                  */
773                 if (page_has_private(page)) {
774                         if (!try_to_release_page(page, sc->gfp_mask))
775                                 goto activate_locked;
776                         if (!mapping && page_count(page) == 1) {
777                                 unlock_page(page);
778                                 if (put_page_testzero(page))
779                                         goto free_it;
780                                 else {
781                                         /*
782                                          * rare race with speculative reference.
783                                          * the speculative reference will free
784                                          * this page shortly, so we may
785                                          * increment nr_reclaimed here (and
786                                          * leave it off the LRU).
787                                          */
788                                         nr_reclaimed++;
789                                         continue;
790                                 }
791                         }
792                 }
793
794                 if (!mapping || !__remove_mapping(mapping, page))
795                         goto keep_locked;
796
797                 /*
798                  * At this point, we have no other references and there is
799                  * no way to pick any more up (removed from LRU, removed
800                  * from pagecache). Can use non-atomic bitops now (and
801                  * we obviously don't have to worry about waking up a process
802                  * waiting on the page lock, because there are no references.
803                  */
804                 __clear_page_locked(page);
805 free_it:
806                 nr_reclaimed++;
807                 if (!pagevec_add(&freed_pvec, page)) {
808                         __pagevec_free(&freed_pvec);
809                         pagevec_reinit(&freed_pvec);
810                 }
811                 continue;
812
813 cull_mlocked:
814                 if (PageSwapCache(page))
815                         try_to_free_swap(page);
816                 unlock_page(page);
817                 putback_lru_page(page);
818                 continue;
819
820 activate_locked:
821                 /* Not a candidate for swapping, so reclaim swap space. */
822                 if (PageSwapCache(page) && vm_swap_full())
823                         try_to_free_swap(page);
824                 VM_BUG_ON(PageActive(page));
825                 SetPageActive(page);
826                 pgactivate++;
827 keep_locked:
828                 unlock_page(page);
829 keep:
830                 list_add(&page->lru, &ret_pages);
831                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
832         }
833         list_splice(&ret_pages, page_list);
834         if (pagevec_count(&freed_pvec))
835                 __pagevec_free(&freed_pvec);
836         count_vm_events(PGACTIVATE, pgactivate);
837         return nr_reclaimed;
838 }
839
840 /*
841  * Attempt to remove the specified page from its LRU.  Only take this page
842  * if it is of the appropriate PageActive status.  Pages which are being
843  * freed elsewhere are also ignored.
844  *
845  * page:        page to consider
846  * mode:        one of the LRU isolation modes defined above
847  *
848  * returns 0 on success, -ve errno on failure.
849  */
850 int __isolate_lru_page(struct page *page, int mode, int file)
851 {
852         int ret = -EINVAL;
853
854         /* Only take pages on the LRU. */
855         if (!PageLRU(page))
856                 return ret;
857
858         /*
859          * When checking the active state, we need to be sure we are
860          * dealing with comparible boolean values.  Take the logical not
861          * of each.
862          */
863         if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
864                 return ret;
865
866         if (mode != ISOLATE_BOTH && page_is_file_cache(page) != file)
867                 return ret;
868
869         /*
870          * When this function is being called for lumpy reclaim, we
871          * initially look into all LRU pages, active, inactive and
872          * unevictable; only give shrink_page_list evictable pages.
873          */
874         if (PageUnevictable(page))
875                 return ret;
876
877         ret = -EBUSY;
878
879         if (likely(get_page_unless_zero(page))) {
880                 /*
881                  * Be careful not to clear PageLRU until after we're
882                  * sure the page is not being freed elsewhere -- the
883                  * page release code relies on it.
884                  */
885                 ClearPageLRU(page);
886                 ret = 0;
887         }
888
889         return ret;
890 }
891
892 /*
893  * zone->lru_lock is heavily contended.  Some of the functions that
894  * shrink the lists perform better by taking out a batch of pages
895  * and working on them outside the LRU lock.
896  *
897  * For pagecache intensive workloads, this function is the hottest
898  * spot in the kernel (apart from copy_*_user functions).
899  *
900  * Appropriate locks must be held before calling this function.
901  *
902  * @nr_to_scan: The number of pages to look through on the list.
903  * @src:        The LRU list to pull pages off.
904  * @dst:        The temp list to put pages on to.
905  * @scanned:    The number of pages that were scanned.
906  * @order:      The caller's attempted allocation order
907  * @mode:       One of the LRU isolation modes
908  * @file:       True [1] if isolating file [!anon] pages
909  *
910  * returns how many pages were moved onto *@dst.
911  */
912 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
913                 struct list_head *src, struct list_head *dst,
914                 unsigned long *scanned, int order, int mode, int file)
915 {
916         unsigned long nr_taken = 0;
917         unsigned long scan;
918
919         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
920                 struct page *page;
921                 unsigned long pfn;
922                 unsigned long end_pfn;
923                 unsigned long page_pfn;
924                 int zone_id;
925
926                 page = lru_to_page(src);
927                 prefetchw_prev_lru_page(page, src, flags);
928
929                 VM_BUG_ON(!PageLRU(page));
930
931                 switch (__isolate_lru_page(page, mode, file)) {
932                 case 0:
933                         list_move(&page->lru, dst);
934                         mem_cgroup_del_lru(page);
935                         nr_taken++;
936                         break;
937
938                 case -EBUSY:
939                         /* else it is being freed elsewhere */
940                         list_move(&page->lru, src);
941                         mem_cgroup_rotate_lru_list(page, page_lru(page));
942                         continue;
943
944                 default:
945                         BUG();
946                 }
947
948                 if (!order)
949                         continue;
950
951                 /*
952                  * Attempt to take all pages in the order aligned region
953                  * surrounding the tag page.  Only take those pages of
954                  * the same active state as that tag page.  We may safely
955                  * round the target page pfn down to the requested order
956                  * as the mem_map is guarenteed valid out to MAX_ORDER,
957                  * where that page is in a different zone we will detect
958                  * it from its zone id and abort this block scan.
959                  */
960                 zone_id = page_zone_id(page);
961                 page_pfn = page_to_pfn(page);
962                 pfn = page_pfn & ~((1 << order) - 1);
963                 end_pfn = pfn + (1 << order);
964                 for (; pfn < end_pfn; pfn++) {
965                         struct page *cursor_page;
966
967                         /* The target page is in the block, ignore it. */
968                         if (unlikely(pfn == page_pfn))
969                                 continue;
970
971                         /* Avoid holes within the zone. */
972                         if (unlikely(!pfn_valid_within(pfn)))
973                                 break;
974
975                         cursor_page = pfn_to_page(pfn);
976
977                         /* Check that we have not crossed a zone boundary. */
978                         if (unlikely(page_zone_id(cursor_page) != zone_id))
979                                 continue;
980
981                         /*
982                          * If we don't have enough swap space, reclaiming of
983                          * anon page which don't already have a swap slot is
984                          * pointless.
985                          */
986                         if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
987                                         !PageSwapCache(cursor_page))
988                                 continue;
989
990                         if (__isolate_lru_page(cursor_page, mode, file) == 0) {
991                                 list_move(&cursor_page->lru, dst);
992                                 mem_cgroup_del_lru(cursor_page);
993                                 nr_taken++;
994                                 scan++;
995                         }
996                 }
997         }
998
999         *scanned = scan;
1000         return nr_taken;
1001 }
1002
1003 static unsigned long isolate_pages_global(unsigned long nr,
1004                                         struct list_head *dst,
1005                                         unsigned long *scanned, int order,
1006                                         int mode, struct zone *z,
1007                                         int active, int file)
1008 {
1009         int lru = LRU_BASE;
1010         if (active)
1011                 lru += LRU_ACTIVE;
1012         if (file)
1013                 lru += LRU_FILE;
1014         return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1015                                                                 mode, file);
1016 }
1017
1018 /*
1019  * clear_active_flags() is a helper for shrink_active_list(), clearing
1020  * any active bits from the pages in the list.
1021  */
1022 static unsigned long clear_active_flags(struct list_head *page_list,
1023                                         unsigned int *count)
1024 {
1025         int nr_active = 0;
1026         int lru;
1027         struct page *page;
1028
1029         list_for_each_entry(page, page_list, lru) {
1030                 lru = page_lru_base_type(page);
1031                 if (PageActive(page)) {
1032                         lru += LRU_ACTIVE;
1033                         ClearPageActive(page);
1034                         nr_active++;
1035                 }
1036                 count[lru]++;
1037         }
1038
1039         return nr_active;
1040 }
1041
1042 /**
1043  * isolate_lru_page - tries to isolate a page from its LRU list
1044  * @page: page to isolate from its LRU list
1045  *
1046  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1047  * vmstat statistic corresponding to whatever LRU list the page was on.
1048  *
1049  * Returns 0 if the page was removed from an LRU list.
1050  * Returns -EBUSY if the page was not on an LRU list.
1051  *
1052  * The returned page will have PageLRU() cleared.  If it was found on
1053  * the active list, it will have PageActive set.  If it was found on
1054  * the unevictable list, it will have the PageUnevictable bit set. That flag
1055  * may need to be cleared by the caller before letting the page go.
1056  *
1057  * The vmstat statistic corresponding to the list on which the page was
1058  * found will be decremented.
1059  *
1060  * Restrictions:
1061  * (1) Must be called with an elevated refcount on the page. This is a
1062  *     fundamentnal difference from isolate_lru_pages (which is called
1063  *     without a stable reference).
1064  * (2) the lru_lock must not be held.
1065  * (3) interrupts must be enabled.
1066  */
1067 int isolate_lru_page(struct page *page)
1068 {
1069         int ret = -EBUSY;
1070
1071         if (PageLRU(page)) {
1072                 struct zone *zone = page_zone(page);
1073
1074                 spin_lock_irq(&zone->lru_lock);
1075                 if (PageLRU(page) && get_page_unless_zero(page)) {
1076                         int lru = page_lru(page);
1077                         ret = 0;
1078                         ClearPageLRU(page);
1079
1080                         del_page_from_lru_list(zone, page, lru);
1081                 }
1082                 spin_unlock_irq(&zone->lru_lock);
1083         }
1084         return ret;
1085 }
1086
1087 /*
1088  * Are there way too many processes in the direct reclaim path already?
1089  */
1090 static int too_many_isolated(struct zone *zone, int file,
1091                 struct scan_control *sc)
1092 {
1093         unsigned long inactive, isolated;
1094
1095         if (current_is_kswapd())
1096                 return 0;
1097
1098         if (!scanning_global_lru(sc))
1099                 return 0;
1100
1101         if (file) {
1102                 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1103                 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1104         } else {
1105                 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1106                 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1107         }
1108
1109         return isolated > inactive;
1110 }
1111
1112 /*
1113  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1114  * of reclaimed pages
1115  */
1116 static unsigned long shrink_inactive_list(unsigned long max_scan,
1117                         struct zone *zone, struct scan_control *sc,
1118                         int priority, int file)
1119 {
1120         LIST_HEAD(page_list);
1121         struct pagevec pvec;
1122         unsigned long nr_scanned = 0;
1123         unsigned long nr_reclaimed = 0;
1124         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1125
1126         while (unlikely(too_many_isolated(zone, file, sc))) {
1127                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1128
1129                 /* We are about to die and free our memory. Return now. */
1130                 if (fatal_signal_pending(current))
1131                         return SWAP_CLUSTER_MAX;
1132         }
1133
1134
1135         pagevec_init(&pvec, 1);
1136
1137         lru_add_drain();
1138         spin_lock_irq(&zone->lru_lock);
1139         do {
1140                 struct page *page;
1141                 unsigned long nr_taken;
1142                 unsigned long nr_scan;
1143                 unsigned long nr_freed;
1144                 unsigned long nr_active;
1145                 unsigned int count[NR_LRU_LISTS] = { 0, };
1146                 int mode = sc->lumpy_reclaim_mode ? ISOLATE_BOTH : ISOLATE_INACTIVE;
1147                 unsigned long nr_anon;
1148                 unsigned long nr_file;
1149
1150                 if (scanning_global_lru(sc)) {
1151                         nr_taken = isolate_pages_global(SWAP_CLUSTER_MAX,
1152                                                         &page_list, &nr_scan,
1153                                                         sc->order, mode,
1154                                                         zone, 0, file);
1155                         zone->pages_scanned += nr_scan;
1156                         if (current_is_kswapd())
1157                                 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1158                                                        nr_scan);
1159                         else
1160                                 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1161                                                        nr_scan);
1162                 } else {
1163                         nr_taken = mem_cgroup_isolate_pages(SWAP_CLUSTER_MAX,
1164                                                         &page_list, &nr_scan,
1165                                                         sc->order, mode,
1166                                                         zone, sc->mem_cgroup,
1167                                                         0, file);
1168                         /*
1169                          * mem_cgroup_isolate_pages() keeps track of
1170                          * scanned pages on its own.
1171                          */
1172                 }
1173
1174                 if (nr_taken == 0)
1175                         goto done;
1176
1177                 nr_active = clear_active_flags(&page_list, count);
1178                 __count_vm_events(PGDEACTIVATE, nr_active);
1179
1180                 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1181                                                 -count[LRU_ACTIVE_FILE]);
1182                 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1183                                                 -count[LRU_INACTIVE_FILE]);
1184                 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1185                                                 -count[LRU_ACTIVE_ANON]);
1186                 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1187                                                 -count[LRU_INACTIVE_ANON]);
1188
1189                 nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1190                 nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1191                 __mod_zone_page_state(zone, NR_ISOLATED_ANON, nr_anon);
1192                 __mod_zone_page_state(zone, NR_ISOLATED_FILE, nr_file);
1193
1194                 reclaim_stat->recent_scanned[0] += nr_anon;
1195                 reclaim_stat->recent_scanned[1] += nr_file;
1196
1197                 spin_unlock_irq(&zone->lru_lock);
1198
1199                 nr_scanned += nr_scan;
1200                 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
1201
1202                 /*
1203                  * If we are direct reclaiming for contiguous pages and we do
1204                  * not reclaim everything in the list, try again and wait
1205                  * for IO to complete. This will stall high-order allocations
1206                  * but that should be acceptable to the caller
1207                  */
1208                 if (nr_freed < nr_taken && !current_is_kswapd() &&
1209                     sc->lumpy_reclaim_mode) {
1210                         congestion_wait(BLK_RW_ASYNC, HZ/10);
1211
1212                         /*
1213                          * The attempt at page out may have made some
1214                          * of the pages active, mark them inactive again.
1215                          */
1216                         nr_active = clear_active_flags(&page_list, count);
1217                         count_vm_events(PGDEACTIVATE, nr_active);
1218
1219                         nr_freed += shrink_page_list(&page_list, sc,
1220                                                         PAGEOUT_IO_SYNC);
1221                 }
1222
1223                 nr_reclaimed += nr_freed;
1224
1225                 local_irq_disable();
1226                 if (current_is_kswapd())
1227                         __count_vm_events(KSWAPD_STEAL, nr_freed);
1228                 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
1229
1230                 spin_lock(&zone->lru_lock);
1231                 /*
1232                  * Put back any unfreeable pages.
1233                  */
1234                 while (!list_empty(&page_list)) {
1235                         int lru;
1236                         page = lru_to_page(&page_list);
1237                         VM_BUG_ON(PageLRU(page));
1238                         list_del(&page->lru);
1239                         if (unlikely(!page_evictable(page, NULL))) {
1240                                 spin_unlock_irq(&zone->lru_lock);
1241                                 putback_lru_page(page);
1242                                 spin_lock_irq(&zone->lru_lock);
1243                                 continue;
1244                         }
1245                         SetPageLRU(page);
1246                         lru = page_lru(page);
1247                         add_page_to_lru_list(zone, page, lru);
1248                         if (is_active_lru(lru)) {
1249                                 int file = is_file_lru(lru);
1250                                 reclaim_stat->recent_rotated[file]++;
1251                         }
1252                         if (!pagevec_add(&pvec, page)) {
1253                                 spin_unlock_irq(&zone->lru_lock);
1254                                 __pagevec_release(&pvec);
1255                                 spin_lock_irq(&zone->lru_lock);
1256                         }
1257                 }
1258                 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1259                 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1260
1261         } while (nr_scanned < max_scan);
1262
1263 done:
1264         spin_unlock_irq(&zone->lru_lock);
1265         pagevec_release(&pvec);
1266         return nr_reclaimed;
1267 }
1268
1269 /*
1270  * We are about to scan this zone at a certain priority level.  If that priority
1271  * level is smaller (ie: more urgent) than the previous priority, then note
1272  * that priority level within the zone.  This is done so that when the next
1273  * process comes in to scan this zone, it will immediately start out at this
1274  * priority level rather than having to build up its own scanning priority.
1275  * Here, this priority affects only the reclaim-mapped threshold.
1276  */
1277 static inline void note_zone_scanning_priority(struct zone *zone, int priority)
1278 {
1279         if (priority < zone->prev_priority)
1280                 zone->prev_priority = priority;
1281 }
1282
1283 /*
1284  * This moves pages from the active list to the inactive list.
1285  *
1286  * We move them the other way if the page is referenced by one or more
1287  * processes, from rmap.
1288  *
1289  * If the pages are mostly unmapped, the processing is fast and it is
1290  * appropriate to hold zone->lru_lock across the whole operation.  But if
1291  * the pages are mapped, the processing is slow (page_referenced()) so we
1292  * should drop zone->lru_lock around each page.  It's impossible to balance
1293  * this, so instead we remove the pages from the LRU while processing them.
1294  * It is safe to rely on PG_active against the non-LRU pages in here because
1295  * nobody will play with that bit on a non-LRU page.
1296  *
1297  * The downside is that we have to touch page->_count against each page.
1298  * But we had to alter page->flags anyway.
1299  */
1300
1301 static void move_active_pages_to_lru(struct zone *zone,
1302                                      struct list_head *list,
1303                                      enum lru_list lru)
1304 {
1305         unsigned long pgmoved = 0;
1306         struct pagevec pvec;
1307         struct page *page;
1308
1309         pagevec_init(&pvec, 1);
1310
1311         while (!list_empty(list)) {
1312                 page = lru_to_page(list);
1313
1314                 VM_BUG_ON(PageLRU(page));
1315                 SetPageLRU(page);
1316
1317                 list_move(&page->lru, &zone->lru[lru].list);
1318                 mem_cgroup_add_lru_list(page, lru);
1319                 pgmoved++;
1320
1321                 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1322                         spin_unlock_irq(&zone->lru_lock);
1323                         if (buffer_heads_over_limit)
1324                                 pagevec_strip(&pvec);
1325                         __pagevec_release(&pvec);
1326                         spin_lock_irq(&zone->lru_lock);
1327                 }
1328         }
1329         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1330         if (!is_active_lru(lru))
1331                 __count_vm_events(PGDEACTIVATE, pgmoved);
1332 }
1333
1334 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1335                         struct scan_control *sc, int priority, int file)
1336 {
1337         unsigned long nr_taken;
1338         unsigned long pgscanned;
1339         unsigned long vm_flags;
1340         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1341         LIST_HEAD(l_active);
1342         LIST_HEAD(l_inactive);
1343         struct page *page;
1344         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1345         unsigned long nr_rotated = 0;
1346
1347         lru_add_drain();
1348         spin_lock_irq(&zone->lru_lock);
1349         if (scanning_global_lru(sc)) {
1350                 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1351                                                 &pgscanned, sc->order,
1352                                                 ISOLATE_ACTIVE, zone,
1353                                                 1, file);
1354                 zone->pages_scanned += pgscanned;
1355         } else {
1356                 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1357                                                 &pgscanned, sc->order,
1358                                                 ISOLATE_ACTIVE, zone,
1359                                                 sc->mem_cgroup, 1, file);
1360                 /*
1361                  * mem_cgroup_isolate_pages() keeps track of
1362                  * scanned pages on its own.
1363                  */
1364         }
1365
1366         reclaim_stat->recent_scanned[file] += nr_taken;
1367
1368         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1369         if (file)
1370                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1371         else
1372                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1373         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1374         spin_unlock_irq(&zone->lru_lock);
1375
1376         while (!list_empty(&l_hold)) {
1377                 cond_resched();
1378                 page = lru_to_page(&l_hold);
1379                 list_del(&page->lru);
1380
1381                 if (unlikely(!page_evictable(page, NULL))) {
1382                         putback_lru_page(page);
1383                         continue;
1384                 }
1385
1386                 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1387                         nr_rotated++;
1388                         /*
1389                          * Identify referenced, file-backed active pages and
1390                          * give them one more trip around the active list. So
1391                          * that executable code get better chances to stay in
1392                          * memory under moderate memory pressure.  Anon pages
1393                          * are not likely to be evicted by use-once streaming
1394                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1395                          * so we ignore them here.
1396                          */
1397                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1398                                 list_add(&page->lru, &l_active);
1399                                 continue;
1400                         }
1401                 }
1402
1403                 ClearPageActive(page);  /* we are de-activating */
1404                 list_add(&page->lru, &l_inactive);
1405         }
1406
1407         /*
1408          * Move pages back to the lru list.
1409          */
1410         spin_lock_irq(&zone->lru_lock);
1411         /*
1412          * Count referenced pages from currently used mappings as rotated,
1413          * even though only some of them are actually re-activated.  This
1414          * helps balance scan pressure between file and anonymous pages in
1415          * get_scan_ratio.
1416          */
1417         reclaim_stat->recent_rotated[file] += nr_rotated;
1418
1419         move_active_pages_to_lru(zone, &l_active,
1420                                                 LRU_ACTIVE + file * LRU_FILE);
1421         move_active_pages_to_lru(zone, &l_inactive,
1422                                                 LRU_BASE   + file * LRU_FILE);
1423         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1424         spin_unlock_irq(&zone->lru_lock);
1425 }
1426
1427 static int inactive_anon_is_low_global(struct zone *zone)
1428 {
1429         unsigned long active, inactive;
1430
1431         active = zone_page_state(zone, NR_ACTIVE_ANON);
1432         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1433
1434         if (inactive * zone->inactive_ratio < active)
1435                 return 1;
1436
1437         return 0;
1438 }
1439
1440 /**
1441  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1442  * @zone: zone to check
1443  * @sc:   scan control of this context
1444  *
1445  * Returns true if the zone does not have enough inactive anon pages,
1446  * meaning some active anon pages need to be deactivated.
1447  */
1448 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1449 {
1450         int low;
1451
1452         if (scanning_global_lru(sc))
1453                 low = inactive_anon_is_low_global(zone);
1454         else
1455                 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1456         return low;
1457 }
1458
1459 static int inactive_file_is_low_global(struct zone *zone)
1460 {
1461         unsigned long active, inactive;
1462
1463         active = zone_page_state(zone, NR_ACTIVE_FILE);
1464         inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1465
1466         return (active > inactive);
1467 }
1468
1469 /**
1470  * inactive_file_is_low - check if file pages need to be deactivated
1471  * @zone: zone to check
1472  * @sc:   scan control of this context
1473  *
1474  * When the system is doing streaming IO, memory pressure here
1475  * ensures that active file pages get deactivated, until more
1476  * than half of the file pages are on the inactive list.
1477  *
1478  * Once we get to that situation, protect the system's working
1479  * set from being evicted by disabling active file page aging.
1480  *
1481  * This uses a different ratio than the anonymous pages, because
1482  * the page cache uses a use-once replacement algorithm.
1483  */
1484 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1485 {
1486         int low;
1487
1488         if (scanning_global_lru(sc))
1489                 low = inactive_file_is_low_global(zone);
1490         else
1491                 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1492         return low;
1493 }
1494
1495 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1496                                 int file)
1497 {
1498         if (file)
1499                 return inactive_file_is_low(zone, sc);
1500         else
1501                 return inactive_anon_is_low(zone, sc);
1502 }
1503
1504 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1505         struct zone *zone, struct scan_control *sc, int priority)
1506 {
1507         int file = is_file_lru(lru);
1508
1509         if (is_active_lru(lru)) {
1510                 if (inactive_list_is_low(zone, sc, file))
1511                     shrink_active_list(nr_to_scan, zone, sc, priority, file);
1512                 return 0;
1513         }
1514
1515         return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1516 }
1517
1518 /*
1519  * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1520  * until we collected @swap_cluster_max pages to scan.
1521  */
1522 static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
1523                                        unsigned long *nr_saved_scan)
1524 {
1525         unsigned long nr;
1526
1527         *nr_saved_scan += nr_to_scan;
1528         nr = *nr_saved_scan;
1529
1530         if (nr >= SWAP_CLUSTER_MAX)
1531                 *nr_saved_scan = 0;
1532         else
1533                 nr = 0;
1534
1535         return nr;
1536 }
1537
1538 /*
1539  * Determine how aggressively the anon and file LRU lists should be
1540  * scanned.  The relative value of each set of LRU lists is determined
1541  * by looking at the fraction of the pages scanned we did rotate back
1542  * onto the active list instead of evict.
1543  *
1544  * nr[0] = anon pages to scan; nr[1] = file pages to scan
1545  */
1546 static void get_scan_count(struct zone *zone, struct scan_control *sc,
1547                                         unsigned long *nr, int priority)
1548 {
1549         unsigned long anon, file, free;
1550         unsigned long anon_prio, file_prio;
1551         unsigned long ap, fp;
1552         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1553         u64 fraction[2], denominator;
1554         enum lru_list l;
1555         int noswap = 0;
1556
1557         /* If we have no swap space, do not bother scanning anon pages. */
1558         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1559                 noswap = 1;
1560                 fraction[0] = 0;
1561                 fraction[1] = 1;
1562                 denominator = 1;
1563                 goto out;
1564         }
1565
1566         anon  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1567                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1568         file  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1569                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1570
1571         if (scanning_global_lru(sc)) {
1572                 free  = zone_page_state(zone, NR_FREE_PAGES);
1573                 /* If we have very few page cache pages,
1574                    force-scan anon pages. */
1575                 if (unlikely(file + free <= high_wmark_pages(zone))) {
1576                         fraction[0] = 1;
1577                         fraction[1] = 0;
1578                         denominator = 1;
1579                         goto out;
1580                 }
1581         }
1582
1583         /*
1584          * OK, so we have swap space and a fair amount of page cache
1585          * pages.  We use the recently rotated / recently scanned
1586          * ratios to determine how valuable each cache is.
1587          *
1588          * Because workloads change over time (and to avoid overflow)
1589          * we keep these statistics as a floating average, which ends
1590          * up weighing recent references more than old ones.
1591          *
1592          * anon in [0], file in [1]
1593          */
1594         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1595                 spin_lock_irq(&zone->lru_lock);
1596                 reclaim_stat->recent_scanned[0] /= 2;
1597                 reclaim_stat->recent_rotated[0] /= 2;
1598                 spin_unlock_irq(&zone->lru_lock);
1599         }
1600
1601         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1602                 spin_lock_irq(&zone->lru_lock);
1603                 reclaim_stat->recent_scanned[1] /= 2;
1604                 reclaim_stat->recent_rotated[1] /= 2;
1605                 spin_unlock_irq(&zone->lru_lock);
1606         }
1607
1608         /*
1609          * With swappiness at 100, anonymous and file have the same priority.
1610          * This scanning priority is essentially the inverse of IO cost.
1611          */
1612         anon_prio = sc->swappiness;
1613         file_prio = 200 - sc->swappiness;
1614
1615         /*
1616          * The amount of pressure on anon vs file pages is inversely
1617          * proportional to the fraction of recently scanned pages on
1618          * each list that were recently referenced and in active use.
1619          */
1620         ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1621         ap /= reclaim_stat->recent_rotated[0] + 1;
1622
1623         fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1624         fp /= reclaim_stat->recent_rotated[1] + 1;
1625
1626         fraction[0] = ap;
1627         fraction[1] = fp;
1628         denominator = ap + fp + 1;
1629 out:
1630         for_each_evictable_lru(l) {
1631                 int file = is_file_lru(l);
1632                 unsigned long scan;
1633
1634                 scan = zone_nr_lru_pages(zone, sc, l);
1635                 if (priority || noswap) {
1636                         scan >>= priority;
1637                         scan = div64_u64(scan * fraction[file], denominator);
1638                 }
1639                 nr[l] = nr_scan_try_batch(scan,
1640                                           &reclaim_stat->nr_saved_scan[l]);
1641         }
1642 }
1643
1644 static void set_lumpy_reclaim_mode(int priority, struct scan_control *sc)
1645 {
1646         /*
1647          * If we need a large contiguous chunk of memory, or have
1648          * trouble getting a small set of contiguous pages, we
1649          * will reclaim both active and inactive pages.
1650          */
1651         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1652                 sc->lumpy_reclaim_mode = 1;
1653         else if (sc->order && priority < DEF_PRIORITY - 2)
1654                 sc->lumpy_reclaim_mode = 1;
1655         else
1656                 sc->lumpy_reclaim_mode = 0;
1657 }
1658
1659 /*
1660  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1661  */
1662 static void shrink_zone(int priority, struct zone *zone,
1663                                 struct scan_control *sc)
1664 {
1665         unsigned long nr[NR_LRU_LISTS];
1666         unsigned long nr_to_scan;
1667         enum lru_list l;
1668         unsigned long nr_reclaimed = sc->nr_reclaimed;
1669         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1670
1671         get_scan_count(zone, sc, nr, priority);
1672
1673         set_lumpy_reclaim_mode(priority, sc);
1674
1675         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1676                                         nr[LRU_INACTIVE_FILE]) {
1677                 for_each_evictable_lru(l) {
1678                         if (nr[l]) {
1679                                 nr_to_scan = min_t(unsigned long,
1680                                                    nr[l], SWAP_CLUSTER_MAX);
1681                                 nr[l] -= nr_to_scan;
1682
1683                                 nr_reclaimed += shrink_list(l, nr_to_scan,
1684                                                             zone, sc, priority);
1685                         }
1686                 }
1687                 /*
1688                  * On large memory systems, scan >> priority can become
1689                  * really large. This is fine for the starting priority;
1690                  * we want to put equal scanning pressure on each zone.
1691                  * However, if the VM has a harder time of freeing pages,
1692                  * with multiple processes reclaiming pages, the total
1693                  * freeing target can get unreasonably large.
1694                  */
1695                 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
1696                         break;
1697         }
1698
1699         sc->nr_reclaimed = nr_reclaimed;
1700
1701         /*
1702          * Even if we did not try to evict anon pages at all, we want to
1703          * rebalance the anon lru active/inactive ratio.
1704          */
1705         if (inactive_anon_is_low(zone, sc) && nr_swap_pages > 0)
1706                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1707
1708         throttle_vm_writeout(sc->gfp_mask);
1709 }
1710
1711 /*
1712  * This is the direct reclaim path, for page-allocating processes.  We only
1713  * try to reclaim pages from zones which will satisfy the caller's allocation
1714  * request.
1715  *
1716  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1717  * Because:
1718  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1719  *    allocation or
1720  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1721  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1722  *    zone defense algorithm.
1723  *
1724  * If a zone is deemed to be full of pinned pages then just give it a light
1725  * scan then give up on it.
1726  */
1727 static bool shrink_zones(int priority, struct zonelist *zonelist,
1728                                         struct scan_control *sc)
1729 {
1730         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1731         struct zoneref *z;
1732         struct zone *zone;
1733         bool all_unreclaimable = true;
1734
1735         for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1736                                         sc->nodemask) {
1737                 if (!populated_zone(zone))
1738                         continue;
1739                 /*
1740                  * Take care memory controller reclaiming has small influence
1741                  * to global LRU.
1742                  */
1743                 if (scanning_global_lru(sc)) {
1744                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1745                                 continue;
1746                         note_zone_scanning_priority(zone, priority);
1747
1748                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1749                                 continue;       /* Let kswapd poll it */
1750                 } else {
1751                         /*
1752                          * Ignore cpuset limitation here. We just want to reduce
1753                          * # of used pages by us regardless of memory shortage.
1754                          */
1755                         mem_cgroup_note_reclaim_priority(sc->mem_cgroup,
1756                                                         priority);
1757                 }
1758
1759                 shrink_zone(priority, zone, sc);
1760                 all_unreclaimable = false;
1761         }
1762         return all_unreclaimable;
1763 }
1764
1765 /*
1766  * This is the main entry point to direct page reclaim.
1767  *
1768  * If a full scan of the inactive list fails to free enough memory then we
1769  * are "out of memory" and something needs to be killed.
1770  *
1771  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1772  * high - the zone may be full of dirty or under-writeback pages, which this
1773  * caller can't do much about.  We kick the writeback threads and take explicit
1774  * naps in the hope that some of these pages can be written.  But if the
1775  * allocating task holds filesystem locks which prevent writeout this might not
1776  * work, and the allocation attempt will fail.
1777  *
1778  * returns:     0, if no pages reclaimed
1779  *              else, the number of pages reclaimed
1780  */
1781 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
1782                                         struct scan_control *sc)
1783 {
1784         int priority;
1785         bool all_unreclaimable;
1786         unsigned long total_scanned = 0;
1787         struct reclaim_state *reclaim_state = current->reclaim_state;
1788         unsigned long lru_pages = 0;
1789         struct zoneref *z;
1790         struct zone *zone;
1791         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1792         unsigned long writeback_threshold;
1793
1794         get_mems_allowed();
1795         delayacct_freepages_start();
1796
1797         if (scanning_global_lru(sc))
1798                 count_vm_event(ALLOCSTALL);
1799         /*
1800          * mem_cgroup will not do shrink_slab.
1801          */
1802         if (scanning_global_lru(sc)) {
1803                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1804
1805                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1806                                 continue;
1807
1808                         lru_pages += zone_reclaimable_pages(zone);
1809                 }
1810         }
1811
1812         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1813                 sc->nr_scanned = 0;
1814                 if (!priority)
1815                         disable_swap_token();
1816                 all_unreclaimable = shrink_zones(priority, zonelist, sc);
1817                 /*
1818                  * Don't shrink slabs when reclaiming memory from
1819                  * over limit cgroups
1820                  */
1821                 if (scanning_global_lru(sc)) {
1822                         shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
1823                         if (reclaim_state) {
1824                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
1825                                 reclaim_state->reclaimed_slab = 0;
1826                         }
1827                 }
1828                 total_scanned += sc->nr_scanned;
1829                 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
1830                         goto out;
1831
1832                 /*
1833                  * Try to write back as many pages as we just scanned.  This
1834                  * tends to cause slow streaming writers to write data to the
1835                  * disk smoothly, at the dirtying rate, which is nice.   But
1836                  * that's undesirable in laptop mode, where we *want* lumpy
1837                  * writeout.  So in laptop mode, write out the whole world.
1838                  */
1839                 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
1840                 if (total_scanned > writeback_threshold) {
1841                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
1842                         sc->may_writepage = 1;
1843                 }
1844
1845                 /* Take a nap, wait for some writeback to complete */
1846                 if (!sc->hibernation_mode && sc->nr_scanned &&
1847                     priority < DEF_PRIORITY - 2)
1848                         congestion_wait(BLK_RW_ASYNC, HZ/10);
1849         }
1850
1851 out:
1852         /*
1853          * Now that we've scanned all the zones at this priority level, note
1854          * that level within the zone so that the next thread which performs
1855          * scanning of this zone will immediately start out at this priority
1856          * level.  This affects only the decision whether or not to bring
1857          * mapped pages onto the inactive list.
1858          */
1859         if (priority < 0)
1860                 priority = 0;
1861
1862         if (scanning_global_lru(sc)) {
1863                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1864
1865                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1866                                 continue;
1867
1868                         zone->prev_priority = priority;
1869                 }
1870         } else
1871                 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority);
1872
1873         delayacct_freepages_end();
1874         put_mems_allowed();
1875
1876         if (sc->nr_reclaimed)
1877                 return sc->nr_reclaimed;
1878
1879         /* top priority shrink_zones still had more to do? don't OOM, then */
1880         if (scanning_global_lru(sc) && !all_unreclaimable)
1881                 return 1;
1882
1883         return 0;
1884 }
1885
1886 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
1887                                 gfp_t gfp_mask, nodemask_t *nodemask)
1888 {
1889         struct scan_control sc = {
1890                 .gfp_mask = gfp_mask,
1891                 .may_writepage = !laptop_mode,
1892                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
1893                 .may_unmap = 1,
1894                 .may_swap = 1,
1895                 .swappiness = vm_swappiness,
1896                 .order = order,
1897                 .mem_cgroup = NULL,
1898                 .nodemask = nodemask,
1899         };
1900
1901         return do_try_to_free_pages(zonelist, &sc);
1902 }
1903
1904 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
1905
1906 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
1907                                                 gfp_t gfp_mask, bool noswap,
1908                                                 unsigned int swappiness,
1909                                                 struct zone *zone, int nid)
1910 {
1911         struct scan_control sc = {
1912                 .may_writepage = !laptop_mode,
1913                 .may_unmap = 1,
1914                 .may_swap = !noswap,
1915                 .swappiness = swappiness,
1916                 .order = 0,
1917                 .mem_cgroup = mem,
1918         };
1919         nodemask_t nm  = nodemask_of_node(nid);
1920
1921         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1922                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1923         sc.nodemask = &nm;
1924         sc.nr_reclaimed = 0;
1925         sc.nr_scanned = 0;
1926         /*
1927          * NOTE: Although we can get the priority field, using it
1928          * here is not a good idea, since it limits the pages we can scan.
1929          * if we don't reclaim here, the shrink_zone from balance_pgdat
1930          * will pick up pages from other mem cgroup's as well. We hack
1931          * the priority and make it zero.
1932          */
1933         shrink_zone(0, zone, &sc);
1934         return sc.nr_reclaimed;
1935 }
1936
1937 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
1938                                            gfp_t gfp_mask,
1939                                            bool noswap,
1940                                            unsigned int swappiness)
1941 {
1942         struct zonelist *zonelist;
1943         struct scan_control sc = {
1944                 .may_writepage = !laptop_mode,
1945                 .may_unmap = 1,
1946                 .may_swap = !noswap,
1947                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
1948                 .swappiness = swappiness,
1949                 .order = 0,
1950                 .mem_cgroup = mem_cont,
1951                 .nodemask = NULL, /* we don't care the placement */
1952         };
1953
1954         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1955                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1956         zonelist = NODE_DATA(numa_node_id())->node_zonelists;
1957         return do_try_to_free_pages(zonelist, &sc);
1958 }
1959 #endif
1960
1961 /* is kswapd sleeping prematurely? */
1962 static int sleeping_prematurely(pg_data_t *pgdat, int order, long remaining)
1963 {
1964         int i;
1965
1966         /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
1967         if (remaining)
1968                 return 1;
1969
1970         /* If after HZ/10, a zone is below the high mark, it's premature */
1971         for (i = 0; i < pgdat->nr_zones; i++) {
1972                 struct zone *zone = pgdat->node_zones + i;
1973
1974                 if (!populated_zone(zone))
1975                         continue;
1976
1977                 if (zone->all_unreclaimable)
1978                         continue;
1979
1980                 if (!zone_watermark_ok(zone, order, high_wmark_pages(zone),
1981                                                                 0, 0))
1982                         return 1;
1983         }
1984
1985         return 0;
1986 }
1987
1988 /*
1989  * For kswapd, balance_pgdat() will work across all this node's zones until
1990  * they are all at high_wmark_pages(zone).
1991  *
1992  * Returns the number of pages which were actually freed.
1993  *
1994  * There is special handling here for zones which are full of pinned pages.
1995  * This can happen if the pages are all mlocked, or if they are all used by
1996  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1997  * What we do is to detect the case where all pages in the zone have been
1998  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1999  * dead and from now on, only perform a short scan.  Basically we're polling
2000  * the zone for when the problem goes away.
2001  *
2002  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
2003  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2004  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2005  * lower zones regardless of the number of free pages in the lower zones. This
2006  * interoperates with the page allocator fallback scheme to ensure that aging
2007  * of pages is balanced across the zones.
2008  */
2009 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
2010 {
2011         int all_zones_ok;
2012         int priority;
2013         int i;
2014         unsigned long total_scanned;
2015         struct reclaim_state *reclaim_state = current->reclaim_state;
2016         struct scan_control sc = {
2017                 .gfp_mask = GFP_KERNEL,
2018                 .may_unmap = 1,
2019                 .may_swap = 1,
2020                 /*
2021                  * kswapd doesn't want to be bailed out while reclaim. because
2022                  * we want to put equal scanning pressure on each zone.
2023                  */
2024                 .nr_to_reclaim = ULONG_MAX,
2025                 .swappiness = vm_swappiness,
2026                 .order = order,
2027                 .mem_cgroup = NULL,
2028         };
2029         /*
2030          * temp_priority is used to remember the scanning priority at which
2031          * this zone was successfully refilled to
2032          * free_pages == high_wmark_pages(zone).
2033          */
2034         int temp_priority[MAX_NR_ZONES];
2035
2036 loop_again:
2037         total_scanned = 0;
2038         sc.nr_reclaimed = 0;
2039         sc.may_writepage = !laptop_mode;
2040         count_vm_event(PAGEOUTRUN);
2041
2042         for (i = 0; i < pgdat->nr_zones; i++)
2043                 temp_priority[i] = DEF_PRIORITY;
2044
2045         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2046                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
2047                 unsigned long lru_pages = 0;
2048                 int has_under_min_watermark_zone = 0;
2049
2050                 /* The swap token gets in the way of swapout... */
2051                 if (!priority)
2052                         disable_swap_token();
2053
2054                 all_zones_ok = 1;
2055
2056                 /*
2057                  * Scan in the highmem->dma direction for the highest
2058                  * zone which needs scanning
2059                  */
2060                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2061                         struct zone *zone = pgdat->node_zones + i;
2062
2063                         if (!populated_zone(zone))
2064                                 continue;
2065
2066                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2067                                 continue;
2068
2069                         /*
2070                          * Do some background aging of the anon list, to give
2071                          * pages a chance to be referenced before reclaiming.
2072                          */
2073                         if (inactive_anon_is_low(zone, &sc))
2074                                 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2075                                                         &sc, priority, 0);
2076
2077                         if (!zone_watermark_ok(zone, order,
2078                                         high_wmark_pages(zone), 0, 0)) {
2079                                 end_zone = i;
2080                                 break;
2081                         }
2082                 }
2083                 if (i < 0)
2084                         goto out;
2085
2086                 for (i = 0; i <= end_zone; i++) {
2087                         struct zone *zone = pgdat->node_zones + i;
2088
2089                         lru_pages += zone_reclaimable_pages(zone);
2090                 }
2091
2092                 /*
2093                  * Now scan the zone in the dma->highmem direction, stopping
2094                  * at the last zone which needs scanning.
2095                  *
2096                  * We do this because the page allocator works in the opposite
2097                  * direction.  This prevents the page allocator from allocating
2098                  * pages behind kswapd's direction of progress, which would
2099                  * cause too much scanning of the lower zones.
2100                  */
2101                 for (i = 0; i <= end_zone; i++) {
2102                         struct zone *zone = pgdat->node_zones + i;
2103                         int nr_slab;
2104                         int nid, zid;
2105
2106                         if (!populated_zone(zone))
2107                                 continue;
2108
2109                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2110                                 continue;
2111
2112                         temp_priority[i] = priority;
2113                         sc.nr_scanned = 0;
2114                         note_zone_scanning_priority(zone, priority);
2115
2116                         nid = pgdat->node_id;
2117                         zid = zone_idx(zone);
2118                         /*
2119                          * Call soft limit reclaim before calling shrink_zone.
2120                          * For now we ignore the return value
2121                          */
2122                         mem_cgroup_soft_limit_reclaim(zone, order, sc.gfp_mask,
2123                                                         nid, zid);
2124                         /*
2125                          * We put equal pressure on every zone, unless one
2126                          * zone has way too many pages free already.
2127                          */
2128                         if (!zone_watermark_ok(zone, order,
2129                                         8*high_wmark_pages(zone), end_zone, 0))
2130                                 shrink_zone(priority, zone, &sc);
2131                         reclaim_state->reclaimed_slab = 0;
2132                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
2133                                                 lru_pages);
2134                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2135                         total_scanned += sc.nr_scanned;
2136                         if (zone->all_unreclaimable)
2137                                 continue;
2138                         if (nr_slab == 0 &&
2139                             zone->pages_scanned >= (zone_reclaimable_pages(zone) * 6))
2140                                 zone->all_unreclaimable = 1;
2141                         /*
2142                          * If we've done a decent amount of scanning and
2143                          * the reclaim ratio is low, start doing writepage
2144                          * even in laptop mode
2145                          */
2146                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2147                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2148                                 sc.may_writepage = 1;
2149
2150                         if (!zone_watermark_ok(zone, order,
2151                                         high_wmark_pages(zone), end_zone, 0)) {
2152                                 all_zones_ok = 0;
2153                                 /*
2154                                  * We are still under min water mark.  This
2155                                  * means that we have a GFP_ATOMIC allocation
2156                                  * failure risk. Hurry up!
2157                                  */
2158                                 if (!zone_watermark_ok(zone, order,
2159                                             min_wmark_pages(zone), end_zone, 0))
2160                                         has_under_min_watermark_zone = 1;
2161                         }
2162
2163                 }
2164                 if (all_zones_ok)
2165                         break;          /* kswapd: all done */
2166                 /*
2167                  * OK, kswapd is getting into trouble.  Take a nap, then take
2168                  * another pass across the zones.
2169                  */
2170                 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2171                         if (has_under_min_watermark_zone)
2172                                 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2173                         else
2174                                 congestion_wait(BLK_RW_ASYNC, HZ/10);
2175                 }
2176
2177                 /*
2178                  * We do this so kswapd doesn't build up large priorities for
2179                  * example when it is freeing in parallel with allocators. It
2180                  * matches the direct reclaim path behaviour in terms of impact
2181                  * on zone->*_priority.
2182                  */
2183                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2184                         break;
2185         }
2186 out:
2187         /*
2188          * Note within each zone the priority level at which this zone was
2189          * brought into a happy state.  So that the next thread which scans this
2190          * zone will start out at that priority level.
2191          */
2192         for (i = 0; i < pgdat->nr_zones; i++) {
2193                 struct zone *zone = pgdat->node_zones + i;
2194
2195                 zone->prev_priority = temp_priority[i];
2196         }
2197         if (!all_zones_ok) {
2198                 cond_resched();
2199
2200                 try_to_freeze();
2201
2202                 /*
2203                  * Fragmentation may mean that the system cannot be
2204                  * rebalanced for high-order allocations in all zones.
2205                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2206                  * it means the zones have been fully scanned and are still
2207                  * not balanced. For high-order allocations, there is
2208                  * little point trying all over again as kswapd may
2209                  * infinite loop.
2210                  *
2211                  * Instead, recheck all watermarks at order-0 as they
2212                  * are the most important. If watermarks are ok, kswapd will go
2213                  * back to sleep. High-order users can still perform direct
2214                  * reclaim if they wish.
2215                  */
2216                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2217                         order = sc.order = 0;
2218
2219                 goto loop_again;
2220         }
2221
2222         return sc.nr_reclaimed;
2223 }
2224
2225 /*
2226  * The background pageout daemon, started as a kernel thread
2227  * from the init process.
2228  *
2229  * This basically trickles out pages so that we have _some_
2230  * free memory available even if there is no other activity
2231  * that frees anything up. This is needed for things like routing
2232  * etc, where we otherwise might have all activity going on in
2233  * asynchronous contexts that cannot page things out.
2234  *
2235  * If there are applications that are active memory-allocators
2236  * (most normal use), this basically shouldn't matter.
2237  */
2238 static int kswapd(void *p)
2239 {
2240         unsigned long order;
2241         pg_data_t *pgdat = (pg_data_t*)p;
2242         struct task_struct *tsk = current;
2243         DEFINE_WAIT(wait);
2244         struct reclaim_state reclaim_state = {
2245                 .reclaimed_slab = 0,
2246         };
2247         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2248
2249         lockdep_set_current_reclaim_state(GFP_KERNEL);
2250
2251         if (!cpumask_empty(cpumask))
2252                 set_cpus_allowed_ptr(tsk, cpumask);
2253         current->reclaim_state = &reclaim_state;
2254
2255         /*
2256          * Tell the memory management that we're a "memory allocator",
2257          * and that if we need more memory we should get access to it
2258          * regardless (see "__alloc_pages()"). "kswapd" should
2259          * never get caught in the normal page freeing logic.
2260          *
2261          * (Kswapd normally doesn't need memory anyway, but sometimes
2262          * you need a small amount of memory in order to be able to
2263          * page out something else, and this flag essentially protects
2264          * us from recursively trying to free more memory as we're
2265          * trying to free the first piece of memory in the first place).
2266          */
2267         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2268         set_freezable();
2269
2270         order = 0;
2271         for ( ; ; ) {
2272                 unsigned long new_order;
2273                 int ret;
2274
2275                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2276                 new_order = pgdat->kswapd_max_order;
2277                 pgdat->kswapd_max_order = 0;
2278                 if (order < new_order) {
2279                         /*
2280                          * Don't sleep if someone wants a larger 'order'
2281                          * allocation
2282                          */
2283                         order = new_order;
2284                 } else {
2285                         if (!freezing(current) && !kthread_should_stop()) {
2286                                 long remaining = 0;
2287
2288                                 /* Try to sleep for a short interval */
2289                                 if (!sleeping_prematurely(pgdat, order, remaining)) {
2290                                         remaining = schedule_timeout(HZ/10);
2291                                         finish_wait(&pgdat->kswapd_wait, &wait);
2292                                         prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2293                                 }
2294
2295                                 /*
2296                                  * After a short sleep, check if it was a
2297                                  * premature sleep. If not, then go fully
2298                                  * to sleep until explicitly woken up
2299                                  */
2300                                 if (!sleeping_prematurely(pgdat, order, remaining))
2301                                         schedule();
2302                                 else {
2303                                         if (remaining)
2304                                                 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2305                                         else
2306                                                 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2307                                 }
2308                         }
2309
2310                         order = pgdat->kswapd_max_order;
2311                 }
2312                 finish_wait(&pgdat->kswapd_wait, &wait);
2313
2314                 ret = try_to_freeze();
2315                 if (kthread_should_stop())
2316                         break;
2317
2318                 /*
2319                  * We can speed up thawing tasks if we don't call balance_pgdat
2320                  * after returning from the refrigerator
2321                  */
2322                 if (!ret)
2323                         balance_pgdat(pgdat, order);
2324         }
2325         return 0;
2326 }
2327
2328 /*
2329  * A zone is low on free memory, so wake its kswapd task to service it.
2330  */
2331 void wakeup_kswapd(struct zone *zone, int order)
2332 {
2333         pg_data_t *pgdat;
2334
2335         if (!populated_zone(zone))
2336                 return;
2337
2338         pgdat = zone->zone_pgdat;
2339         if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
2340                 return;
2341         if (pgdat->kswapd_max_order < order)
2342                 pgdat->kswapd_max_order = order;
2343         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2344                 return;
2345         if (!waitqueue_active(&pgdat->kswapd_wait))
2346                 return;
2347         wake_up_interruptible(&pgdat->kswapd_wait);
2348 }
2349
2350 /*
2351  * The reclaimable count would be mostly accurate.
2352  * The less reclaimable pages may be
2353  * - mlocked pages, which will be moved to unevictable list when encountered
2354  * - mapped pages, which may require several travels to be reclaimed
2355  * - dirty pages, which is not "instantly" reclaimable
2356  */
2357 unsigned long global_reclaimable_pages(void)
2358 {
2359         int nr;
2360
2361         nr = global_page_state(NR_ACTIVE_FILE) +
2362              global_page_state(NR_INACTIVE_FILE);
2363
2364         if (nr_swap_pages > 0)
2365                 nr += global_page_state(NR_ACTIVE_ANON) +
2366                       global_page_state(NR_INACTIVE_ANON);
2367
2368         return nr;
2369 }
2370
2371 unsigned long zone_reclaimable_pages(struct zone *zone)
2372 {
2373         int nr;
2374
2375         nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2376              zone_page_state(zone, NR_INACTIVE_FILE);
2377
2378         if (nr_swap_pages > 0)
2379                 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2380                       zone_page_state(zone, NR_INACTIVE_ANON);
2381
2382         return nr;
2383 }
2384
2385 #ifdef CONFIG_HIBERNATION
2386 /*
2387  * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2388  * freed pages.
2389  *
2390  * Rather than trying to age LRUs the aim is to preserve the overall
2391  * LRU order by reclaiming preferentially
2392  * inactive > active > active referenced > active mapped
2393  */
2394 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
2395 {
2396         struct reclaim_state reclaim_state;
2397         struct scan_control sc = {
2398                 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2399                 .may_swap = 1,
2400                 .may_unmap = 1,
2401                 .may_writepage = 1,
2402                 .nr_to_reclaim = nr_to_reclaim,
2403                 .hibernation_mode = 1,
2404                 .swappiness = vm_swappiness,
2405                 .order = 0,
2406         };
2407         struct zonelist * zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2408         struct task_struct *p = current;
2409         unsigned long nr_reclaimed;
2410
2411         p->flags |= PF_MEMALLOC;
2412         lockdep_set_current_reclaim_state(sc.gfp_mask);
2413         reclaim_state.reclaimed_slab = 0;
2414         p->reclaim_state = &reclaim_state;
2415
2416         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2417
2418         p->reclaim_state = NULL;
2419         lockdep_clear_current_reclaim_state();
2420         p->flags &= ~PF_MEMALLOC;
2421
2422         return nr_reclaimed;
2423 }
2424 #endif /* CONFIG_HIBERNATION */
2425
2426 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2427    not required for correctness.  So if the last cpu in a node goes
2428    away, we get changed to run anywhere: as the first one comes back,
2429    restore their cpu bindings. */
2430 static int __devinit cpu_callback(struct notifier_block *nfb,
2431                                   unsigned long action, void *hcpu)
2432 {
2433         int nid;
2434
2435         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2436                 for_each_node_state(nid, N_HIGH_MEMORY) {
2437                         pg_data_t *pgdat = NODE_DATA(nid);
2438                         const struct cpumask *mask;
2439
2440                         mask = cpumask_of_node(pgdat->node_id);
2441
2442                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2443                                 /* One of our CPUs online: restore mask */
2444                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2445                 }
2446         }
2447         return NOTIFY_OK;
2448 }
2449
2450 /*
2451  * This kswapd start function will be called by init and node-hot-add.
2452  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2453  */
2454 int kswapd_run(int nid)
2455 {
2456         pg_data_t *pgdat = NODE_DATA(nid);
2457         int ret = 0;
2458
2459         if (pgdat->kswapd)
2460                 return 0;
2461
2462         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2463         if (IS_ERR(pgdat->kswapd)) {
2464                 /* failure at boot is fatal */
2465                 BUG_ON(system_state == SYSTEM_BOOTING);
2466                 printk("Failed to start kswapd on node %d\n",nid);
2467                 ret = -1;
2468         }
2469         return ret;
2470 }
2471
2472 /*
2473  * Called by memory hotplug when all memory in a node is offlined.
2474  */
2475 void kswapd_stop(int nid)
2476 {
2477         struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
2478
2479         if (kswapd)
2480                 kthread_stop(kswapd);
2481 }
2482
2483 static int __init kswapd_init(void)
2484 {
2485         int nid;
2486
2487         swap_setup();
2488         for_each_node_state(nid, N_HIGH_MEMORY)
2489                 kswapd_run(nid);
2490         hotcpu_notifier(cpu_callback, 0);
2491         return 0;
2492 }
2493
2494 module_init(kswapd_init)
2495
2496 #ifdef CONFIG_NUMA
2497 /*
2498  * Zone reclaim mode
2499  *
2500  * If non-zero call zone_reclaim when the number of free pages falls below
2501  * the watermarks.
2502  */
2503 int zone_reclaim_mode __read_mostly;
2504
2505 #define RECLAIM_OFF 0
2506 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
2507 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
2508 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
2509
2510 /*
2511  * Priority for ZONE_RECLAIM. This determines the fraction of pages
2512  * of a node considered for each zone_reclaim. 4 scans 1/16th of
2513  * a zone.
2514  */
2515 #define ZONE_RECLAIM_PRIORITY 4
2516
2517 /*
2518  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2519  * occur.
2520  */
2521 int sysctl_min_unmapped_ratio = 1;
2522
2523 /*
2524  * If the number of slab pages in a zone grows beyond this percentage then
2525  * slab reclaim needs to occur.
2526  */
2527 int sysctl_min_slab_ratio = 5;
2528
2529 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
2530 {
2531         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
2532         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
2533                 zone_page_state(zone, NR_ACTIVE_FILE);
2534
2535         /*
2536          * It's possible for there to be more file mapped pages than
2537          * accounted for by the pages on the file LRU lists because
2538          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2539          */
2540         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
2541 }
2542
2543 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
2544 static long zone_pagecache_reclaimable(struct zone *zone)
2545 {
2546         long nr_pagecache_reclaimable;
2547         long delta = 0;
2548
2549         /*
2550          * If RECLAIM_SWAP is set, then all file pages are considered
2551          * potentially reclaimable. Otherwise, we have to worry about
2552          * pages like swapcache and zone_unmapped_file_pages() provides
2553          * a better estimate
2554          */
2555         if (zone_reclaim_mode & RECLAIM_SWAP)
2556                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
2557         else
2558                 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
2559
2560         /* If we can't clean pages, remove dirty pages from consideration */
2561         if (!(zone_reclaim_mode & RECLAIM_WRITE))
2562                 delta += zone_page_state(zone, NR_FILE_DIRTY);
2563
2564         /* Watch for any possible underflows due to delta */
2565         if (unlikely(delta > nr_pagecache_reclaimable))
2566                 delta = nr_pagecache_reclaimable;
2567
2568         return nr_pagecache_reclaimable - delta;
2569 }
2570
2571 /*
2572  * Try to free up some pages from this zone through reclaim.
2573  */
2574 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2575 {
2576         /* Minimum pages needed in order to stay on node */
2577         const unsigned long nr_pages = 1 << order;
2578         struct task_struct *p = current;
2579         struct reclaim_state reclaim_state;
2580         int priority;
2581         struct scan_control sc = {
2582                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2583                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2584                 .may_swap = 1,
2585                 .nr_to_reclaim = max_t(unsigned long, nr_pages,
2586                                        SWAP_CLUSTER_MAX),
2587                 .gfp_mask = gfp_mask,
2588                 .swappiness = vm_swappiness,
2589                 .order = order,
2590         };
2591         unsigned long slab_reclaimable;
2592
2593         disable_swap_token();
2594         cond_resched();
2595         /*
2596          * We need to be able to allocate from the reserves for RECLAIM_SWAP
2597          * and we also need to be able to write out pages for RECLAIM_WRITE
2598          * and RECLAIM_SWAP.
2599          */
2600         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2601         lockdep_set_current_reclaim_state(gfp_mask);
2602         reclaim_state.reclaimed_slab = 0;
2603         p->reclaim_state = &reclaim_state;
2604
2605         if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
2606                 /*
2607                  * Free memory by calling shrink zone with increasing
2608                  * priorities until we have enough memory freed.
2609                  */
2610                 priority = ZONE_RECLAIM_PRIORITY;
2611                 do {
2612                         note_zone_scanning_priority(zone, priority);
2613                         shrink_zone(priority, zone, &sc);
2614                         priority--;
2615                 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
2616         }
2617
2618         slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2619         if (slab_reclaimable > zone->min_slab_pages) {
2620                 /*
2621                  * shrink_slab() does not currently allow us to determine how
2622                  * many pages were freed in this zone. So we take the current
2623                  * number of slab pages and shake the slab until it is reduced
2624                  * by the same nr_pages that we used for reclaiming unmapped
2625                  * pages.
2626                  *
2627                  * Note that shrink_slab will free memory on all zones and may
2628                  * take a long time.
2629                  */
2630                 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
2631                         zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
2632                                 slab_reclaimable - nr_pages)
2633                         ;
2634
2635                 /*
2636                  * Update nr_reclaimed by the number of slab pages we
2637                  * reclaimed from this zone.
2638                  */
2639                 sc.nr_reclaimed += slab_reclaimable -
2640                         zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2641         }
2642
2643         p->reclaim_state = NULL;
2644         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
2645         lockdep_clear_current_reclaim_state();
2646         return sc.nr_reclaimed >= nr_pages;
2647 }
2648
2649 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2650 {
2651         int node_id;
2652         int ret;
2653
2654         /*
2655          * Zone reclaim reclaims unmapped file backed pages and
2656          * slab pages if we are over the defined limits.
2657          *
2658          * A small portion of unmapped file backed pages is needed for
2659          * file I/O otherwise pages read by file I/O will be immediately
2660          * thrown out if the zone is overallocated. So we do not reclaim
2661          * if less than a specified percentage of the zone is used by
2662          * unmapped file backed pages.
2663          */
2664         if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
2665             zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
2666                 return ZONE_RECLAIM_FULL;
2667
2668         if (zone->all_unreclaimable)
2669                 return ZONE_RECLAIM_FULL;
2670
2671         /*
2672          * Do not scan if the allocation should not be delayed.
2673          */
2674         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
2675                 return ZONE_RECLAIM_NOSCAN;
2676
2677         /*
2678          * Only run zone reclaim on the local zone or on zones that do not
2679          * have associated processors. This will favor the local processor
2680          * over remote processors and spread off node memory allocations
2681          * as wide as possible.
2682          */
2683         node_id = zone_to_nid(zone);
2684         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
2685                 return ZONE_RECLAIM_NOSCAN;
2686
2687         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2688                 return ZONE_RECLAIM_NOSCAN;
2689
2690         ret = __zone_reclaim(zone, gfp_mask, order);
2691         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2692
2693         if (!ret)
2694                 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
2695
2696         return ret;
2697 }
2698 #endif
2699
2700 /*
2701  * page_evictable - test whether a page is evictable
2702  * @page: the page to test
2703  * @vma: the VMA in which the page is or will be mapped, may be NULL
2704  *
2705  * Test whether page is evictable--i.e., should be placed on active/inactive
2706  * lists vs unevictable list.  The vma argument is !NULL when called from the
2707  * fault path to determine how to instantate a new page.
2708  *
2709  * Reasons page might not be evictable:
2710  * (1) page's mapping marked unevictable
2711  * (2) page is part of an mlocked VMA
2712  *
2713  */
2714 int page_evictable(struct page *page, struct vm_area_struct *vma)
2715 {
2716
2717         if (mapping_unevictable(page_mapping(page)))
2718                 return 0;
2719
2720         if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2721                 return 0;
2722
2723         return 1;
2724 }
2725
2726 /**
2727  * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2728  * @page: page to check evictability and move to appropriate lru list
2729  * @zone: zone page is in
2730  *
2731  * Checks a page for evictability and moves the page to the appropriate
2732  * zone lru list.
2733  *
2734  * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2735  * have PageUnevictable set.
2736  */
2737 static void check_move_unevictable_page(struct page *page, struct zone *zone)
2738 {
2739         VM_BUG_ON(PageActive(page));
2740
2741 retry:
2742         ClearPageUnevictable(page);
2743         if (page_evictable(page, NULL)) {
2744                 enum lru_list l = page_lru_base_type(page);
2745
2746                 __dec_zone_state(zone, NR_UNEVICTABLE);
2747                 list_move(&page->lru, &zone->lru[l].list);
2748                 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
2749                 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
2750                 __count_vm_event(UNEVICTABLE_PGRESCUED);
2751         } else {
2752                 /*
2753                  * rotate unevictable list
2754                  */
2755                 SetPageUnevictable(page);
2756                 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
2757                 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
2758                 if (page_evictable(page, NULL))
2759                         goto retry;
2760         }
2761 }
2762
2763 /**
2764  * scan_mapping_unevictable_pages - scan an address space for evictable pages
2765  * @mapping: struct address_space to scan for evictable pages
2766  *
2767  * Scan all pages in mapping.  Check unevictable pages for
2768  * evictability and move them to the appropriate zone lru list.
2769  */
2770 void scan_mapping_unevictable_pages(struct address_space *mapping)
2771 {
2772         pgoff_t next = 0;
2773         pgoff_t end   = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
2774                          PAGE_CACHE_SHIFT;
2775         struct zone *zone;
2776         struct pagevec pvec;
2777
2778         if (mapping->nrpages == 0)
2779                 return;
2780
2781         pagevec_init(&pvec, 0);
2782         while (next < end &&
2783                 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
2784                 int i;
2785                 int pg_scanned = 0;
2786
2787                 zone = NULL;
2788
2789                 for (i = 0; i < pagevec_count(&pvec); i++) {
2790                         struct page *page = pvec.pages[i];
2791                         pgoff_t page_index = page->index;
2792                         struct zone *pagezone = page_zone(page);
2793
2794                         pg_scanned++;
2795                         if (page_index > next)
2796                                 next = page_index;
2797                         next++;
2798
2799                         if (pagezone != zone) {
2800                                 if (zone)
2801                                         spin_unlock_irq(&zone->lru_lock);
2802                                 zone = pagezone;
2803                                 spin_lock_irq(&zone->lru_lock);
2804                         }
2805
2806                         if (PageLRU(page) && PageUnevictable(page))
2807                                 check_move_unevictable_page(page, zone);
2808                 }
2809                 if (zone)
2810                         spin_unlock_irq(&zone->lru_lock);
2811                 pagevec_release(&pvec);
2812
2813                 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
2814         }
2815
2816 }
2817
2818 /**
2819  * scan_zone_unevictable_pages - check unevictable list for evictable pages
2820  * @zone - zone of which to scan the unevictable list
2821  *
2822  * Scan @zone's unevictable LRU lists to check for pages that have become
2823  * evictable.  Move those that have to @zone's inactive list where they
2824  * become candidates for reclaim, unless shrink_inactive_zone() decides
2825  * to reactivate them.  Pages that are still unevictable are rotated
2826  * back onto @zone's unevictable list.
2827  */
2828 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
2829 static void scan_zone_unevictable_pages(struct zone *zone)
2830 {
2831         struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
2832         unsigned long scan;
2833         unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
2834
2835         while (nr_to_scan > 0) {
2836                 unsigned long batch_size = min(nr_to_scan,
2837                                                 SCAN_UNEVICTABLE_BATCH_SIZE);
2838
2839                 spin_lock_irq(&zone->lru_lock);
2840                 for (scan = 0;  scan < batch_size; scan++) {
2841                         struct page *page = lru_to_page(l_unevictable);
2842
2843                         if (!trylock_page(page))
2844                                 continue;
2845
2846                         prefetchw_prev_lru_page(page, l_unevictable, flags);
2847
2848                         if (likely(PageLRU(page) && PageUnevictable(page)))
2849                                 check_move_unevictable_page(page, zone);
2850
2851                         unlock_page(page);
2852                 }
2853                 spin_unlock_irq(&zone->lru_lock);
2854
2855                 nr_to_scan -= batch_size;
2856         }
2857 }
2858
2859
2860 /**
2861  * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
2862  *
2863  * A really big hammer:  scan all zones' unevictable LRU lists to check for
2864  * pages that have become evictable.  Move those back to the zones'
2865  * inactive list where they become candidates for reclaim.
2866  * This occurs when, e.g., we have unswappable pages on the unevictable lists,
2867  * and we add swap to the system.  As such, it runs in the context of a task
2868  * that has possibly/probably made some previously unevictable pages
2869  * evictable.
2870  */
2871 static void scan_all_zones_unevictable_pages(void)
2872 {
2873         struct zone *zone;
2874
2875         for_each_zone(zone) {
2876                 scan_zone_unevictable_pages(zone);
2877         }
2878 }
2879
2880 /*
2881  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
2882  * all nodes' unevictable lists for evictable pages
2883  */
2884 unsigned long scan_unevictable_pages;
2885
2886 int scan_unevictable_handler(struct ctl_table *table, int write,
2887                            void __user *buffer,
2888                            size_t *length, loff_t *ppos)
2889 {
2890         proc_doulongvec_minmax(table, write, buffer, length, ppos);
2891
2892         if (write && *(unsigned long *)table->data)
2893                 scan_all_zones_unevictable_pages();
2894
2895         scan_unevictable_pages = 0;
2896         return 0;
2897 }
2898
2899 /*
2900  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
2901  * a specified node's per zone unevictable lists for evictable pages.
2902  */
2903
2904 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
2905                                           struct sysdev_attribute *attr,
2906                                           char *buf)
2907 {
2908         return sprintf(buf, "0\n");     /* always zero; should fit... */
2909 }
2910
2911 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
2912                                            struct sysdev_attribute *attr,
2913                                         const char *buf, size_t count)
2914 {
2915         struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
2916         struct zone *zone;
2917         unsigned long res;
2918         unsigned long req = strict_strtoul(buf, 10, &res);
2919
2920         if (!req)
2921                 return 1;       /* zero is no-op */
2922
2923         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2924                 if (!populated_zone(zone))
2925                         continue;
2926                 scan_zone_unevictable_pages(zone);
2927         }
2928         return 1;
2929 }
2930
2931
2932 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
2933                         read_scan_unevictable_node,
2934                         write_scan_unevictable_node);
2935
2936 int scan_unevictable_register_node(struct node *node)
2937 {
2938         return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
2939 }
2940
2941 void scan_unevictable_unregister_node(struct node *node)
2942 {
2943         sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
2944 }
2945