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