vmscan: narrow the scenarios in whcih lumpy reclaim uses synchrounous reclaim
[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 #define CREATE_TRACE_POINTS
52 #include <trace/events/vmscan.h>
53
54 enum lumpy_mode {
55         LUMPY_MODE_NONE,
56         LUMPY_MODE_ASYNC,
57         LUMPY_MODE_SYNC,
58 };
59
60 struct scan_control {
61         /* Incremented by the number of inactive pages that were scanned */
62         unsigned long nr_scanned;
63
64         /* Number of pages freed so far during a call to shrink_zones() */
65         unsigned long nr_reclaimed;
66
67         /* How many pages shrink_list() should reclaim */
68         unsigned long nr_to_reclaim;
69
70         unsigned long hibernation_mode;
71
72         /* This context's GFP mask */
73         gfp_t gfp_mask;
74
75         int may_writepage;
76
77         /* Can mapped pages be reclaimed? */
78         int may_unmap;
79
80         /* Can pages be swapped as part of reclaim? */
81         int may_swap;
82
83         int swappiness;
84
85         int order;
86
87         /*
88          * Intend to reclaim enough continuous memory rather than reclaim
89          * enough amount of memory. i.e, mode for high order allocation.
90          */
91         enum lumpy_mode lumpy_reclaim_mode;
92
93         /* Which cgroup do we reclaim from */
94         struct mem_cgroup *mem_cgroup;
95
96         /*
97          * Nodemask of nodes allowed by the caller. If NULL, all nodes
98          * are scanned.
99          */
100         nodemask_t      *nodemask;
101 };
102
103 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
104
105 #ifdef ARCH_HAS_PREFETCH
106 #define prefetch_prev_lru_page(_page, _base, _field)                    \
107         do {                                                            \
108                 if ((_page)->lru.prev != _base) {                       \
109                         struct page *prev;                              \
110                                                                         \
111                         prev = lru_to_page(&(_page->lru));              \
112                         prefetch(&prev->_field);                        \
113                 }                                                       \
114         } while (0)
115 #else
116 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
117 #endif
118
119 #ifdef ARCH_HAS_PREFETCHW
120 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
121         do {                                                            \
122                 if ((_page)->lru.prev != _base) {                       \
123                         struct page *prev;                              \
124                                                                         \
125                         prev = lru_to_page(&(_page->lru));              \
126                         prefetchw(&prev->_field);                       \
127                 }                                                       \
128         } while (0)
129 #else
130 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
131 #endif
132
133 /*
134  * From 0 .. 100.  Higher means more swappy.
135  */
136 int vm_swappiness = 60;
137 long vm_total_pages;    /* The total number of pages which the VM controls */
138
139 static LIST_HEAD(shrinker_list);
140 static DECLARE_RWSEM(shrinker_rwsem);
141
142 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
143 #define scanning_global_lru(sc) (!(sc)->mem_cgroup)
144 #else
145 #define scanning_global_lru(sc) (1)
146 #endif
147
148 static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
149                                                   struct scan_control *sc)
150 {
151         if (!scanning_global_lru(sc))
152                 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
153
154         return &zone->reclaim_stat;
155 }
156
157 static unsigned long zone_nr_lru_pages(struct zone *zone,
158                                 struct scan_control *sc, enum lru_list lru)
159 {
160         if (!scanning_global_lru(sc))
161                 return mem_cgroup_zone_nr_pages(sc->mem_cgroup, zone, lru);
162
163         return zone_page_state(zone, NR_LRU_BASE + lru);
164 }
165
166
167 /*
168  * Add a shrinker callback to be called from the vm
169  */
170 void register_shrinker(struct shrinker *shrinker)
171 {
172         shrinker->nr = 0;
173         down_write(&shrinker_rwsem);
174         list_add_tail(&shrinker->list, &shrinker_list);
175         up_write(&shrinker_rwsem);
176 }
177 EXPORT_SYMBOL(register_shrinker);
178
179 /*
180  * Remove one
181  */
182 void unregister_shrinker(struct shrinker *shrinker)
183 {
184         down_write(&shrinker_rwsem);
185         list_del(&shrinker->list);
186         up_write(&shrinker_rwsem);
187 }
188 EXPORT_SYMBOL(unregister_shrinker);
189
190 #define SHRINK_BATCH 128
191 /*
192  * Call the shrink functions to age shrinkable caches
193  *
194  * Here we assume it costs one seek to replace a lru page and that it also
195  * takes a seek to recreate a cache object.  With this in mind we age equal
196  * percentages of the lru and ageable caches.  This should balance the seeks
197  * generated by these structures.
198  *
199  * If the vm encountered mapped pages on the LRU it increase the pressure on
200  * slab to avoid swapping.
201  *
202  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
203  *
204  * `lru_pages' represents the number of on-LRU pages in all the zones which
205  * are eligible for the caller's allocation attempt.  It is used for balancing
206  * slab reclaim versus page reclaim.
207  *
208  * Returns the number of slab objects which we shrunk.
209  */
210 unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
211                         unsigned long lru_pages)
212 {
213         struct shrinker *shrinker;
214         unsigned long ret = 0;
215
216         if (scanned == 0)
217                 scanned = SWAP_CLUSTER_MAX;
218
219         if (!down_read_trylock(&shrinker_rwsem))
220                 return 1;       /* Assume we'll be able to shrink next time */
221
222         list_for_each_entry(shrinker, &shrinker_list, list) {
223                 unsigned long long delta;
224                 unsigned long total_scan;
225                 unsigned long max_pass;
226
227                 max_pass = (*shrinker->shrink)(shrinker, 0, gfp_mask);
228                 delta = (4 * scanned) / shrinker->seeks;
229                 delta *= max_pass;
230                 do_div(delta, lru_pages + 1);
231                 shrinker->nr += delta;
232                 if (shrinker->nr < 0) {
233                         printk(KERN_ERR "shrink_slab: %pF negative objects to "
234                                "delete nr=%ld\n",
235                                shrinker->shrink, shrinker->nr);
236                         shrinker->nr = max_pass;
237                 }
238
239                 /*
240                  * Avoid risking looping forever due to too large nr value:
241                  * never try to free more than twice the estimate number of
242                  * freeable entries.
243                  */
244                 if (shrinker->nr > max_pass * 2)
245                         shrinker->nr = max_pass * 2;
246
247                 total_scan = shrinker->nr;
248                 shrinker->nr = 0;
249
250                 while (total_scan >= SHRINK_BATCH) {
251                         long this_scan = SHRINK_BATCH;
252                         int shrink_ret;
253                         int nr_before;
254
255                         nr_before = (*shrinker->shrink)(shrinker, 0, gfp_mask);
256                         shrink_ret = (*shrinker->shrink)(shrinker, this_scan,
257                                                                 gfp_mask);
258                         if (shrink_ret == -1)
259                                 break;
260                         if (shrink_ret < nr_before)
261                                 ret += nr_before - shrink_ret;
262                         count_vm_events(SLABS_SCANNED, this_scan);
263                         total_scan -= this_scan;
264
265                         cond_resched();
266                 }
267
268                 shrinker->nr += total_scan;
269         }
270         up_read(&shrinker_rwsem);
271         return ret;
272 }
273
274 static void set_lumpy_reclaim_mode(int priority, struct scan_control *sc,
275                                    bool sync)
276 {
277         enum lumpy_mode mode = sync ? LUMPY_MODE_SYNC : LUMPY_MODE_ASYNC;
278
279         /*
280          * Some reclaim have alredy been failed. No worth to try synchronous
281          * lumpy reclaim.
282          */
283         if (sync && sc->lumpy_reclaim_mode == LUMPY_MODE_NONE)
284                 return;
285
286         /*
287          * If we need a large contiguous chunk of memory, or have
288          * trouble getting a small set of contiguous pages, we
289          * will reclaim both active and inactive pages.
290          */
291         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
292                 sc->lumpy_reclaim_mode = mode;
293         else if (sc->order && priority < DEF_PRIORITY - 2)
294                 sc->lumpy_reclaim_mode = mode;
295         else
296                 sc->lumpy_reclaim_mode = LUMPY_MODE_NONE;
297 }
298
299 static void disable_lumpy_reclaim_mode(struct scan_control *sc)
300 {
301         sc->lumpy_reclaim_mode = LUMPY_MODE_NONE;
302 }
303
304 static inline int is_page_cache_freeable(struct page *page)
305 {
306         /*
307          * A freeable page cache page is referenced only by the caller
308          * that isolated the page, the page cache radix tree and
309          * optional buffer heads at page->private.
310          */
311         return page_count(page) - page_has_private(page) == 2;
312 }
313
314 static int may_write_to_queue(struct backing_dev_info *bdi,
315                               struct scan_control *sc)
316 {
317         if (current->flags & PF_SWAPWRITE)
318                 return 1;
319         if (!bdi_write_congested(bdi))
320                 return 1;
321         if (bdi == current->backing_dev_info)
322                 return 1;
323
324         /* lumpy reclaim for hugepage often need a lot of write */
325         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
326                 return 1;
327         return 0;
328 }
329
330 /*
331  * We detected a synchronous write error writing a page out.  Probably
332  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
333  * fsync(), msync() or close().
334  *
335  * The tricky part is that after writepage we cannot touch the mapping: nothing
336  * prevents it from being freed up.  But we have a ref on the page and once
337  * that page is locked, the mapping is pinned.
338  *
339  * We're allowed to run sleeping lock_page() here because we know the caller has
340  * __GFP_FS.
341  */
342 static void handle_write_error(struct address_space *mapping,
343                                 struct page *page, int error)
344 {
345         lock_page_nosync(page);
346         if (page_mapping(page) == mapping)
347                 mapping_set_error(mapping, error);
348         unlock_page(page);
349 }
350
351 /* possible outcome of pageout() */
352 typedef enum {
353         /* failed to write page out, page is locked */
354         PAGE_KEEP,
355         /* move page to the active list, page is locked */
356         PAGE_ACTIVATE,
357         /* page has been sent to the disk successfully, page is unlocked */
358         PAGE_SUCCESS,
359         /* page is clean and locked */
360         PAGE_CLEAN,
361 } pageout_t;
362
363 /*
364  * pageout is called by shrink_page_list() for each dirty page.
365  * Calls ->writepage().
366  */
367 static pageout_t pageout(struct page *page, struct address_space *mapping,
368                          struct scan_control *sc)
369 {
370         /*
371          * If the page is dirty, only perform writeback if that write
372          * will be non-blocking.  To prevent this allocation from being
373          * stalled by pagecache activity.  But note that there may be
374          * stalls if we need to run get_block().  We could test
375          * PagePrivate for that.
376          *
377          * If this process is currently in __generic_file_aio_write() against
378          * this page's queue, we can perform writeback even if that
379          * will block.
380          *
381          * If the page is swapcache, write it back even if that would
382          * block, for some throttling. This happens by accident, because
383          * swap_backing_dev_info is bust: it doesn't reflect the
384          * congestion state of the swapdevs.  Easy to fix, if needed.
385          */
386         if (!is_page_cache_freeable(page))
387                 return PAGE_KEEP;
388         if (!mapping) {
389                 /*
390                  * Some data journaling orphaned pages can have
391                  * page->mapping == NULL while being dirty with clean buffers.
392                  */
393                 if (page_has_private(page)) {
394                         if (try_to_free_buffers(page)) {
395                                 ClearPageDirty(page);
396                                 printk("%s: orphaned page\n", __func__);
397                                 return PAGE_CLEAN;
398                         }
399                 }
400                 return PAGE_KEEP;
401         }
402         if (mapping->a_ops->writepage == NULL)
403                 return PAGE_ACTIVATE;
404         if (!may_write_to_queue(mapping->backing_dev_info, sc)) {
405                 disable_lumpy_reclaim_mode(sc);
406                 return PAGE_KEEP;
407         }
408
409         if (clear_page_dirty_for_io(page)) {
410                 int res;
411                 struct writeback_control wbc = {
412                         .sync_mode = WB_SYNC_NONE,
413                         .nr_to_write = SWAP_CLUSTER_MAX,
414                         .range_start = 0,
415                         .range_end = LLONG_MAX,
416                         .for_reclaim = 1,
417                 };
418
419                 SetPageReclaim(page);
420                 res = mapping->a_ops->writepage(page, &wbc);
421                 if (res < 0)
422                         handle_write_error(mapping, page, res);
423                 if (res == AOP_WRITEPAGE_ACTIVATE) {
424                         ClearPageReclaim(page);
425                         return PAGE_ACTIVATE;
426                 }
427
428                 /*
429                  * Wait on writeback if requested to. This happens when
430                  * direct reclaiming a large contiguous area and the
431                  * first attempt to free a range of pages fails.
432                  */
433                 if (PageWriteback(page) &&
434                     sc->lumpy_reclaim_mode == LUMPY_MODE_SYNC)
435                         wait_on_page_writeback(page);
436
437                 if (!PageWriteback(page)) {
438                         /* synchronous write or broken a_ops? */
439                         ClearPageReclaim(page);
440                 }
441                 trace_mm_vmscan_writepage(page,
442                         trace_reclaim_flags(page, sc->lumpy_reclaim_mode));
443                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
444                 return PAGE_SUCCESS;
445         }
446
447         return PAGE_CLEAN;
448 }
449
450 /*
451  * Same as remove_mapping, but if the page is removed from the mapping, it
452  * gets returned with a refcount of 0.
453  */
454 static int __remove_mapping(struct address_space *mapping, struct page *page)
455 {
456         BUG_ON(!PageLocked(page));
457         BUG_ON(mapping != page_mapping(page));
458
459         spin_lock_irq(&mapping->tree_lock);
460         /*
461          * The non racy check for a busy page.
462          *
463          * Must be careful with the order of the tests. When someone has
464          * a ref to the page, it may be possible that they dirty it then
465          * drop the reference. So if PageDirty is tested before page_count
466          * here, then the following race may occur:
467          *
468          * get_user_pages(&page);
469          * [user mapping goes away]
470          * write_to(page);
471          *                              !PageDirty(page)    [good]
472          * SetPageDirty(page);
473          * put_page(page);
474          *                              !page_count(page)   [good, discard it]
475          *
476          * [oops, our write_to data is lost]
477          *
478          * Reversing the order of the tests ensures such a situation cannot
479          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
480          * load is not satisfied before that of page->_count.
481          *
482          * Note that if SetPageDirty is always performed via set_page_dirty,
483          * and thus under tree_lock, then this ordering is not required.
484          */
485         if (!page_freeze_refs(page, 2))
486                 goto cannot_free;
487         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
488         if (unlikely(PageDirty(page))) {
489                 page_unfreeze_refs(page, 2);
490                 goto cannot_free;
491         }
492
493         if (PageSwapCache(page)) {
494                 swp_entry_t swap = { .val = page_private(page) };
495                 __delete_from_swap_cache(page);
496                 spin_unlock_irq(&mapping->tree_lock);
497                 swapcache_free(swap, page);
498         } else {
499                 __remove_from_page_cache(page);
500                 spin_unlock_irq(&mapping->tree_lock);
501                 mem_cgroup_uncharge_cache_page(page);
502         }
503
504         return 1;
505
506 cannot_free:
507         spin_unlock_irq(&mapping->tree_lock);
508         return 0;
509 }
510
511 /*
512  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
513  * someone else has a ref on the page, abort and return 0.  If it was
514  * successfully detached, return 1.  Assumes the caller has a single ref on
515  * this page.
516  */
517 int remove_mapping(struct address_space *mapping, struct page *page)
518 {
519         if (__remove_mapping(mapping, page)) {
520                 /*
521                  * Unfreezing the refcount with 1 rather than 2 effectively
522                  * drops the pagecache ref for us without requiring another
523                  * atomic operation.
524                  */
525                 page_unfreeze_refs(page, 1);
526                 return 1;
527         }
528         return 0;
529 }
530
531 /**
532  * putback_lru_page - put previously isolated page onto appropriate LRU list
533  * @page: page to be put back to appropriate lru list
534  *
535  * Add previously isolated @page to appropriate LRU list.
536  * Page may still be unevictable for other reasons.
537  *
538  * lru_lock must not be held, interrupts must be enabled.
539  */
540 void putback_lru_page(struct page *page)
541 {
542         int lru;
543         int active = !!TestClearPageActive(page);
544         int was_unevictable = PageUnevictable(page);
545
546         VM_BUG_ON(PageLRU(page));
547
548 redo:
549         ClearPageUnevictable(page);
550
551         if (page_evictable(page, NULL)) {
552                 /*
553                  * For evictable pages, we can use the cache.
554                  * In event of a race, worst case is we end up with an
555                  * unevictable page on [in]active list.
556                  * We know how to handle that.
557                  */
558                 lru = active + page_lru_base_type(page);
559                 lru_cache_add_lru(page, lru);
560         } else {
561                 /*
562                  * Put unevictable pages directly on zone's unevictable
563                  * list.
564                  */
565                 lru = LRU_UNEVICTABLE;
566                 add_page_to_unevictable_list(page);
567                 /*
568                  * When racing with an mlock clearing (page is
569                  * unlocked), make sure that if the other thread does
570                  * not observe our setting of PG_lru and fails
571                  * isolation, we see PG_mlocked cleared below and move
572                  * the page back to the evictable list.
573                  *
574                  * The other side is TestClearPageMlocked().
575                  */
576                 smp_mb();
577         }
578
579         /*
580          * page's status can change while we move it among lru. If an evictable
581          * page is on unevictable list, it never be freed. To avoid that,
582          * check after we added it to the list, again.
583          */
584         if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
585                 if (!isolate_lru_page(page)) {
586                         put_page(page);
587                         goto redo;
588                 }
589                 /* This means someone else dropped this page from LRU
590                  * So, it will be freed or putback to LRU again. There is
591                  * nothing to do here.
592                  */
593         }
594
595         if (was_unevictable && lru != LRU_UNEVICTABLE)
596                 count_vm_event(UNEVICTABLE_PGRESCUED);
597         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
598                 count_vm_event(UNEVICTABLE_PGCULLED);
599
600         put_page(page);         /* drop ref from isolate */
601 }
602
603 enum page_references {
604         PAGEREF_RECLAIM,
605         PAGEREF_RECLAIM_CLEAN,
606         PAGEREF_KEEP,
607         PAGEREF_ACTIVATE,
608 };
609
610 static enum page_references page_check_references(struct page *page,
611                                                   struct scan_control *sc)
612 {
613         int referenced_ptes, referenced_page;
614         unsigned long vm_flags;
615
616         referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
617         referenced_page = TestClearPageReferenced(page);
618
619         /* Lumpy reclaim - ignore references */
620         if (sc->lumpy_reclaim_mode != LUMPY_MODE_NONE)
621                 return PAGEREF_RECLAIM;
622
623         /*
624          * Mlock lost the isolation race with us.  Let try_to_unmap()
625          * move the page to the unevictable list.
626          */
627         if (vm_flags & VM_LOCKED)
628                 return PAGEREF_RECLAIM;
629
630         if (referenced_ptes) {
631                 if (PageAnon(page))
632                         return PAGEREF_ACTIVATE;
633                 /*
634                  * All mapped pages start out with page table
635                  * references from the instantiating fault, so we need
636                  * to look twice if a mapped file page is used more
637                  * than once.
638                  *
639                  * Mark it and spare it for another trip around the
640                  * inactive list.  Another page table reference will
641                  * lead to its activation.
642                  *
643                  * Note: the mark is set for activated pages as well
644                  * so that recently deactivated but used pages are
645                  * quickly recovered.
646                  */
647                 SetPageReferenced(page);
648
649                 if (referenced_page)
650                         return PAGEREF_ACTIVATE;
651
652                 return PAGEREF_KEEP;
653         }
654
655         /* Reclaim if clean, defer dirty pages to writeback */
656         if (referenced_page)
657                 return PAGEREF_RECLAIM_CLEAN;
658
659         return PAGEREF_RECLAIM;
660 }
661
662 static noinline_for_stack void free_page_list(struct list_head *free_pages)
663 {
664         struct pagevec freed_pvec;
665         struct page *page, *tmp;
666
667         pagevec_init(&freed_pvec, 1);
668
669         list_for_each_entry_safe(page, tmp, free_pages, lru) {
670                 list_del(&page->lru);
671                 if (!pagevec_add(&freed_pvec, page)) {
672                         __pagevec_free(&freed_pvec);
673                         pagevec_reinit(&freed_pvec);
674                 }
675         }
676
677         pagevec_free(&freed_pvec);
678 }
679
680 /*
681  * shrink_page_list() returns the number of reclaimed pages
682  */
683 static unsigned long shrink_page_list(struct list_head *page_list,
684                                       struct scan_control *sc)
685 {
686         LIST_HEAD(ret_pages);
687         LIST_HEAD(free_pages);
688         int pgactivate = 0;
689         unsigned long nr_reclaimed = 0;
690
691         cond_resched();
692
693         while (!list_empty(page_list)) {
694                 enum page_references references;
695                 struct address_space *mapping;
696                 struct page *page;
697                 int may_enter_fs;
698
699                 cond_resched();
700
701                 page = lru_to_page(page_list);
702                 list_del(&page->lru);
703
704                 if (!trylock_page(page))
705                         goto keep;
706
707                 VM_BUG_ON(PageActive(page));
708
709                 sc->nr_scanned++;
710
711                 if (unlikely(!page_evictable(page, NULL)))
712                         goto cull_mlocked;
713
714                 if (!sc->may_unmap && page_mapped(page))
715                         goto keep_locked;
716
717                 /* Double the slab pressure for mapped and swapcache pages */
718                 if (page_mapped(page) || PageSwapCache(page))
719                         sc->nr_scanned++;
720
721                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
722                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
723
724                 if (PageWriteback(page)) {
725                         /*
726                          * Synchronous reclaim is performed in two passes,
727                          * first an asynchronous pass over the list to
728                          * start parallel writeback, and a second synchronous
729                          * pass to wait for the IO to complete.  Wait here
730                          * for any page for which writeback has already
731                          * started.
732                          */
733                         if (sc->lumpy_reclaim_mode == LUMPY_MODE_SYNC &&
734                             may_enter_fs)
735                                 wait_on_page_writeback(page);
736                         else {
737                                 unlock_page(page);
738                                 goto keep_lumpy;
739                         }
740                 }
741
742                 references = page_check_references(page, sc);
743                 switch (references) {
744                 case PAGEREF_ACTIVATE:
745                         goto activate_locked;
746                 case PAGEREF_KEEP:
747                         goto keep_locked;
748                 case PAGEREF_RECLAIM:
749                 case PAGEREF_RECLAIM_CLEAN:
750                         ; /* try to reclaim the page below */
751                 }
752
753                 /*
754                  * Anonymous process memory has backing store?
755                  * Try to allocate it some swap space here.
756                  */
757                 if (PageAnon(page) && !PageSwapCache(page)) {
758                         if (!(sc->gfp_mask & __GFP_IO))
759                                 goto keep_locked;
760                         if (!add_to_swap(page))
761                                 goto activate_locked;
762                         may_enter_fs = 1;
763                 }
764
765                 mapping = page_mapping(page);
766
767                 /*
768                  * The page is mapped into the page tables of one or more
769                  * processes. Try to unmap it here.
770                  */
771                 if (page_mapped(page) && mapping) {
772                         switch (try_to_unmap(page, TTU_UNMAP)) {
773                         case SWAP_FAIL:
774                                 goto activate_locked;
775                         case SWAP_AGAIN:
776                                 goto keep_locked;
777                         case SWAP_MLOCK:
778                                 goto cull_mlocked;
779                         case SWAP_SUCCESS:
780                                 ; /* try to free the page below */
781                         }
782                 }
783
784                 if (PageDirty(page)) {
785                         if (references == PAGEREF_RECLAIM_CLEAN)
786                                 goto keep_locked;
787                         if (!may_enter_fs)
788                                 goto keep_locked;
789                         if (!sc->may_writepage)
790                                 goto keep_locked;
791
792                         /* Page is dirty, try to write it out here */
793                         switch (pageout(page, mapping, sc)) {
794                         case PAGE_KEEP:
795                                 goto keep_locked;
796                         case PAGE_ACTIVATE:
797                                 goto activate_locked;
798                         case PAGE_SUCCESS:
799                                 if (PageWriteback(page))
800                                         goto keep_lumpy;
801                                 if (PageDirty(page))
802                                         goto keep;
803
804                                 /*
805                                  * A synchronous write - probably a ramdisk.  Go
806                                  * ahead and try to reclaim the page.
807                                  */
808                                 if (!trylock_page(page))
809                                         goto keep;
810                                 if (PageDirty(page) || PageWriteback(page))
811                                         goto keep_locked;
812                                 mapping = page_mapping(page);
813                         case PAGE_CLEAN:
814                                 ; /* try to free the page below */
815                         }
816                 }
817
818                 /*
819                  * If the page has buffers, try to free the buffer mappings
820                  * associated with this page. If we succeed we try to free
821                  * the page as well.
822                  *
823                  * We do this even if the page is PageDirty().
824                  * try_to_release_page() does not perform I/O, but it is
825                  * possible for a page to have PageDirty set, but it is actually
826                  * clean (all its buffers are clean).  This happens if the
827                  * buffers were written out directly, with submit_bh(). ext3
828                  * will do this, as well as the blockdev mapping.
829                  * try_to_release_page() will discover that cleanness and will
830                  * drop the buffers and mark the page clean - it can be freed.
831                  *
832                  * Rarely, pages can have buffers and no ->mapping.  These are
833                  * the pages which were not successfully invalidated in
834                  * truncate_complete_page().  We try to drop those buffers here
835                  * and if that worked, and the page is no longer mapped into
836                  * process address space (page_count == 1) it can be freed.
837                  * Otherwise, leave the page on the LRU so it is swappable.
838                  */
839                 if (page_has_private(page)) {
840                         if (!try_to_release_page(page, sc->gfp_mask))
841                                 goto activate_locked;
842                         if (!mapping && page_count(page) == 1) {
843                                 unlock_page(page);
844                                 if (put_page_testzero(page))
845                                         goto free_it;
846                                 else {
847                                         /*
848                                          * rare race with speculative reference.
849                                          * the speculative reference will free
850                                          * this page shortly, so we may
851                                          * increment nr_reclaimed here (and
852                                          * leave it off the LRU).
853                                          */
854                                         nr_reclaimed++;
855                                         continue;
856                                 }
857                         }
858                 }
859
860                 if (!mapping || !__remove_mapping(mapping, page))
861                         goto keep_locked;
862
863                 /*
864                  * At this point, we have no other references and there is
865                  * no way to pick any more up (removed from LRU, removed
866                  * from pagecache). Can use non-atomic bitops now (and
867                  * we obviously don't have to worry about waking up a process
868                  * waiting on the page lock, because there are no references.
869                  */
870                 __clear_page_locked(page);
871 free_it:
872                 nr_reclaimed++;
873
874                 /*
875                  * Is there need to periodically free_page_list? It would
876                  * appear not as the counts should be low
877                  */
878                 list_add(&page->lru, &free_pages);
879                 continue;
880
881 cull_mlocked:
882                 if (PageSwapCache(page))
883                         try_to_free_swap(page);
884                 unlock_page(page);
885                 putback_lru_page(page);
886                 disable_lumpy_reclaim_mode(sc);
887                 continue;
888
889 activate_locked:
890                 /* Not a candidate for swapping, so reclaim swap space. */
891                 if (PageSwapCache(page) && vm_swap_full())
892                         try_to_free_swap(page);
893                 VM_BUG_ON(PageActive(page));
894                 SetPageActive(page);
895                 pgactivate++;
896 keep_locked:
897                 unlock_page(page);
898 keep:
899                 disable_lumpy_reclaim_mode(sc);
900 keep_lumpy:
901                 list_add(&page->lru, &ret_pages);
902                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
903         }
904
905         free_page_list(&free_pages);
906
907         list_splice(&ret_pages, page_list);
908         count_vm_events(PGACTIVATE, pgactivate);
909         return nr_reclaimed;
910 }
911
912 /*
913  * Attempt to remove the specified page from its LRU.  Only take this page
914  * if it is of the appropriate PageActive status.  Pages which are being
915  * freed elsewhere are also ignored.
916  *
917  * page:        page to consider
918  * mode:        one of the LRU isolation modes defined above
919  *
920  * returns 0 on success, -ve errno on failure.
921  */
922 int __isolate_lru_page(struct page *page, int mode, int file)
923 {
924         int ret = -EINVAL;
925
926         /* Only take pages on the LRU. */
927         if (!PageLRU(page))
928                 return ret;
929
930         /*
931          * When checking the active state, we need to be sure we are
932          * dealing with comparible boolean values.  Take the logical not
933          * of each.
934          */
935         if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
936                 return ret;
937
938         if (mode != ISOLATE_BOTH && page_is_file_cache(page) != file)
939                 return ret;
940
941         /*
942          * When this function is being called for lumpy reclaim, we
943          * initially look into all LRU pages, active, inactive and
944          * unevictable; only give shrink_page_list evictable pages.
945          */
946         if (PageUnevictable(page))
947                 return ret;
948
949         ret = -EBUSY;
950
951         if (likely(get_page_unless_zero(page))) {
952                 /*
953                  * Be careful not to clear PageLRU until after we're
954                  * sure the page is not being freed elsewhere -- the
955                  * page release code relies on it.
956                  */
957                 ClearPageLRU(page);
958                 ret = 0;
959         }
960
961         return ret;
962 }
963
964 /*
965  * zone->lru_lock is heavily contended.  Some of the functions that
966  * shrink the lists perform better by taking out a batch of pages
967  * and working on them outside the LRU lock.
968  *
969  * For pagecache intensive workloads, this function is the hottest
970  * spot in the kernel (apart from copy_*_user functions).
971  *
972  * Appropriate locks must be held before calling this function.
973  *
974  * @nr_to_scan: The number of pages to look through on the list.
975  * @src:        The LRU list to pull pages off.
976  * @dst:        The temp list to put pages on to.
977  * @scanned:    The number of pages that were scanned.
978  * @order:      The caller's attempted allocation order
979  * @mode:       One of the LRU isolation modes
980  * @file:       True [1] if isolating file [!anon] pages
981  *
982  * returns how many pages were moved onto *@dst.
983  */
984 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
985                 struct list_head *src, struct list_head *dst,
986                 unsigned long *scanned, int order, int mode, int file)
987 {
988         unsigned long nr_taken = 0;
989         unsigned long nr_lumpy_taken = 0;
990         unsigned long nr_lumpy_dirty = 0;
991         unsigned long nr_lumpy_failed = 0;
992         unsigned long scan;
993
994         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
995                 struct page *page;
996                 unsigned long pfn;
997                 unsigned long end_pfn;
998                 unsigned long page_pfn;
999                 int zone_id;
1000
1001                 page = lru_to_page(src);
1002                 prefetchw_prev_lru_page(page, src, flags);
1003
1004                 VM_BUG_ON(!PageLRU(page));
1005
1006                 switch (__isolate_lru_page(page, mode, file)) {
1007                 case 0:
1008                         list_move(&page->lru, dst);
1009                         mem_cgroup_del_lru(page);
1010                         nr_taken++;
1011                         break;
1012
1013                 case -EBUSY:
1014                         /* else it is being freed elsewhere */
1015                         list_move(&page->lru, src);
1016                         mem_cgroup_rotate_lru_list(page, page_lru(page));
1017                         continue;
1018
1019                 default:
1020                         BUG();
1021                 }
1022
1023                 if (!order)
1024                         continue;
1025
1026                 /*
1027                  * Attempt to take all pages in the order aligned region
1028                  * surrounding the tag page.  Only take those pages of
1029                  * the same active state as that tag page.  We may safely
1030                  * round the target page pfn down to the requested order
1031                  * as the mem_map is guarenteed valid out to MAX_ORDER,
1032                  * where that page is in a different zone we will detect
1033                  * it from its zone id and abort this block scan.
1034                  */
1035                 zone_id = page_zone_id(page);
1036                 page_pfn = page_to_pfn(page);
1037                 pfn = page_pfn & ~((1 << order) - 1);
1038                 end_pfn = pfn + (1 << order);
1039                 for (; pfn < end_pfn; pfn++) {
1040                         struct page *cursor_page;
1041
1042                         /* The target page is in the block, ignore it. */
1043                         if (unlikely(pfn == page_pfn))
1044                                 continue;
1045
1046                         /* Avoid holes within the zone. */
1047                         if (unlikely(!pfn_valid_within(pfn)))
1048                                 break;
1049
1050                         cursor_page = pfn_to_page(pfn);
1051
1052                         /* Check that we have not crossed a zone boundary. */
1053                         if (unlikely(page_zone_id(cursor_page) != zone_id))
1054                                 continue;
1055
1056                         /*
1057                          * If we don't have enough swap space, reclaiming of
1058                          * anon page which don't already have a swap slot is
1059                          * pointless.
1060                          */
1061                         if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
1062                                         !PageSwapCache(cursor_page))
1063                                 continue;
1064
1065                         if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1066                                 list_move(&cursor_page->lru, dst);
1067                                 mem_cgroup_del_lru(cursor_page);
1068                                 nr_taken++;
1069                                 nr_lumpy_taken++;
1070                                 if (PageDirty(cursor_page))
1071                                         nr_lumpy_dirty++;
1072                                 scan++;
1073                         } else {
1074                                 if (mode == ISOLATE_BOTH &&
1075                                                 page_count(cursor_page))
1076                                         nr_lumpy_failed++;
1077                         }
1078                 }
1079         }
1080
1081         *scanned = scan;
1082
1083         trace_mm_vmscan_lru_isolate(order,
1084                         nr_to_scan, scan,
1085                         nr_taken,
1086                         nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1087                         mode);
1088         return nr_taken;
1089 }
1090
1091 static unsigned long isolate_pages_global(unsigned long nr,
1092                                         struct list_head *dst,
1093                                         unsigned long *scanned, int order,
1094                                         int mode, struct zone *z,
1095                                         int active, int file)
1096 {
1097         int lru = LRU_BASE;
1098         if (active)
1099                 lru += LRU_ACTIVE;
1100         if (file)
1101                 lru += LRU_FILE;
1102         return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1103                                                                 mode, file);
1104 }
1105
1106 /*
1107  * clear_active_flags() is a helper for shrink_active_list(), clearing
1108  * any active bits from the pages in the list.
1109  */
1110 static unsigned long clear_active_flags(struct list_head *page_list,
1111                                         unsigned int *count)
1112 {
1113         int nr_active = 0;
1114         int lru;
1115         struct page *page;
1116
1117         list_for_each_entry(page, page_list, lru) {
1118                 lru = page_lru_base_type(page);
1119                 if (PageActive(page)) {
1120                         lru += LRU_ACTIVE;
1121                         ClearPageActive(page);
1122                         nr_active++;
1123                 }
1124                 if (count)
1125                         count[lru]++;
1126         }
1127
1128         return nr_active;
1129 }
1130
1131 /**
1132  * isolate_lru_page - tries to isolate a page from its LRU list
1133  * @page: page to isolate from its LRU list
1134  *
1135  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1136  * vmstat statistic corresponding to whatever LRU list the page was on.
1137  *
1138  * Returns 0 if the page was removed from an LRU list.
1139  * Returns -EBUSY if the page was not on an LRU list.
1140  *
1141  * The returned page will have PageLRU() cleared.  If it was found on
1142  * the active list, it will have PageActive set.  If it was found on
1143  * the unevictable list, it will have the PageUnevictable bit set. That flag
1144  * may need to be cleared by the caller before letting the page go.
1145  *
1146  * The vmstat statistic corresponding to the list on which the page was
1147  * found will be decremented.
1148  *
1149  * Restrictions:
1150  * (1) Must be called with an elevated refcount on the page. This is a
1151  *     fundamentnal difference from isolate_lru_pages (which is called
1152  *     without a stable reference).
1153  * (2) the lru_lock must not be held.
1154  * (3) interrupts must be enabled.
1155  */
1156 int isolate_lru_page(struct page *page)
1157 {
1158         int ret = -EBUSY;
1159
1160         if (PageLRU(page)) {
1161                 struct zone *zone = page_zone(page);
1162
1163                 spin_lock_irq(&zone->lru_lock);
1164                 if (PageLRU(page) && get_page_unless_zero(page)) {
1165                         int lru = page_lru(page);
1166                         ret = 0;
1167                         ClearPageLRU(page);
1168
1169                         del_page_from_lru_list(zone, page, lru);
1170                 }
1171                 spin_unlock_irq(&zone->lru_lock);
1172         }
1173         return ret;
1174 }
1175
1176 /*
1177  * Are there way too many processes in the direct reclaim path already?
1178  */
1179 static int too_many_isolated(struct zone *zone, int file,
1180                 struct scan_control *sc)
1181 {
1182         unsigned long inactive, isolated;
1183
1184         if (current_is_kswapd())
1185                 return 0;
1186
1187         if (!scanning_global_lru(sc))
1188                 return 0;
1189
1190         if (file) {
1191                 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1192                 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1193         } else {
1194                 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1195                 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1196         }
1197
1198         return isolated > inactive;
1199 }
1200
1201 /*
1202  * TODO: Try merging with migrations version of putback_lru_pages
1203  */
1204 static noinline_for_stack void
1205 putback_lru_pages(struct zone *zone, struct scan_control *sc,
1206                                 unsigned long nr_anon, unsigned long nr_file,
1207                                 struct list_head *page_list)
1208 {
1209         struct page *page;
1210         struct pagevec pvec;
1211         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1212
1213         pagevec_init(&pvec, 1);
1214
1215         /*
1216          * Put back any unfreeable pages.
1217          */
1218         spin_lock(&zone->lru_lock);
1219         while (!list_empty(page_list)) {
1220                 int lru;
1221                 page = lru_to_page(page_list);
1222                 VM_BUG_ON(PageLRU(page));
1223                 list_del(&page->lru);
1224                 if (unlikely(!page_evictable(page, NULL))) {
1225                         spin_unlock_irq(&zone->lru_lock);
1226                         putback_lru_page(page);
1227                         spin_lock_irq(&zone->lru_lock);
1228                         continue;
1229                 }
1230                 SetPageLRU(page);
1231                 lru = page_lru(page);
1232                 add_page_to_lru_list(zone, page, lru);
1233                 if (is_active_lru(lru)) {
1234                         int file = is_file_lru(lru);
1235                         reclaim_stat->recent_rotated[file]++;
1236                 }
1237                 if (!pagevec_add(&pvec, page)) {
1238                         spin_unlock_irq(&zone->lru_lock);
1239                         __pagevec_release(&pvec);
1240                         spin_lock_irq(&zone->lru_lock);
1241                 }
1242         }
1243         __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1244         __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1245
1246         spin_unlock_irq(&zone->lru_lock);
1247         pagevec_release(&pvec);
1248 }
1249
1250 static noinline_for_stack void update_isolated_counts(struct zone *zone,
1251                                         struct scan_control *sc,
1252                                         unsigned long *nr_anon,
1253                                         unsigned long *nr_file,
1254                                         struct list_head *isolated_list)
1255 {
1256         unsigned long nr_active;
1257         unsigned int count[NR_LRU_LISTS] = { 0, };
1258         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1259
1260         nr_active = clear_active_flags(isolated_list, count);
1261         __count_vm_events(PGDEACTIVATE, nr_active);
1262
1263         __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1264                               -count[LRU_ACTIVE_FILE]);
1265         __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1266                               -count[LRU_INACTIVE_FILE]);
1267         __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1268                               -count[LRU_ACTIVE_ANON]);
1269         __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1270                               -count[LRU_INACTIVE_ANON]);
1271
1272         *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1273         *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1274         __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1275         __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1276
1277         reclaim_stat->recent_scanned[0] += *nr_anon;
1278         reclaim_stat->recent_scanned[1] += *nr_file;
1279 }
1280
1281 /*
1282  * Returns true if the caller should wait to clean dirty/writeback pages.
1283  *
1284  * If we are direct reclaiming for contiguous pages and we do not reclaim
1285  * everything in the list, try again and wait for writeback IO to complete.
1286  * This will stall high-order allocations noticeably. Only do that when really
1287  * need to free the pages under high memory pressure.
1288  */
1289 static inline bool should_reclaim_stall(unsigned long nr_taken,
1290                                         unsigned long nr_freed,
1291                                         int priority,
1292                                         struct scan_control *sc)
1293 {
1294         int lumpy_stall_priority;
1295
1296         /* kswapd should not stall on sync IO */
1297         if (current_is_kswapd())
1298                 return false;
1299
1300         /* Only stall on lumpy reclaim */
1301         if (sc->lumpy_reclaim_mode == LUMPY_MODE_NONE)
1302                 return false;
1303
1304         /* If we have relaimed everything on the isolated list, no stall */
1305         if (nr_freed == nr_taken)
1306                 return false;
1307
1308         /*
1309          * For high-order allocations, there are two stall thresholds.
1310          * High-cost allocations stall immediately where as lower
1311          * order allocations such as stacks require the scanning
1312          * priority to be much higher before stalling.
1313          */
1314         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1315                 lumpy_stall_priority = DEF_PRIORITY;
1316         else
1317                 lumpy_stall_priority = DEF_PRIORITY / 3;
1318
1319         return priority <= lumpy_stall_priority;
1320 }
1321
1322 /*
1323  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1324  * of reclaimed pages
1325  */
1326 static noinline_for_stack unsigned long
1327 shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1328                         struct scan_control *sc, int priority, int file)
1329 {
1330         LIST_HEAD(page_list);
1331         unsigned long nr_scanned;
1332         unsigned long nr_reclaimed = 0;
1333         unsigned long nr_taken;
1334         unsigned long nr_active;
1335         unsigned long nr_anon;
1336         unsigned long nr_file;
1337
1338         while (unlikely(too_many_isolated(zone, file, sc))) {
1339                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1340
1341                 /* We are about to die and free our memory. Return now. */
1342                 if (fatal_signal_pending(current))
1343                         return SWAP_CLUSTER_MAX;
1344         }
1345
1346         set_lumpy_reclaim_mode(priority, sc, false);
1347         lru_add_drain();
1348         spin_lock_irq(&zone->lru_lock);
1349
1350         if (scanning_global_lru(sc)) {
1351                 nr_taken = isolate_pages_global(nr_to_scan,
1352                         &page_list, &nr_scanned, sc->order,
1353                         sc->lumpy_reclaim_mode == LUMPY_MODE_NONE ?
1354                                         ISOLATE_INACTIVE : ISOLATE_BOTH,
1355                         zone, 0, file);
1356                 zone->pages_scanned += nr_scanned;
1357                 if (current_is_kswapd())
1358                         __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1359                                                nr_scanned);
1360                 else
1361                         __count_zone_vm_events(PGSCAN_DIRECT, zone,
1362                                                nr_scanned);
1363         } else {
1364                 nr_taken = mem_cgroup_isolate_pages(nr_to_scan,
1365                         &page_list, &nr_scanned, sc->order,
1366                         sc->lumpy_reclaim_mode == LUMPY_MODE_NONE ?
1367                                         ISOLATE_INACTIVE : ISOLATE_BOTH,
1368                         zone, sc->mem_cgroup,
1369                         0, file);
1370                 /*
1371                  * mem_cgroup_isolate_pages() keeps track of
1372                  * scanned pages on its own.
1373                  */
1374         }
1375
1376         if (nr_taken == 0) {
1377                 spin_unlock_irq(&zone->lru_lock);
1378                 return 0;
1379         }
1380
1381         update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
1382
1383         spin_unlock_irq(&zone->lru_lock);
1384
1385         nr_reclaimed = shrink_page_list(&page_list, sc);
1386
1387         /* Check if we should syncronously wait for writeback */
1388         if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1389                 /*
1390                  * The attempt at page out may have made some
1391                  * of the pages active, mark them inactive again.
1392                  */
1393                 nr_active = clear_active_flags(&page_list, NULL);
1394                 count_vm_events(PGDEACTIVATE, nr_active);
1395
1396                 set_lumpy_reclaim_mode(priority, sc, true);
1397                 nr_reclaimed += shrink_page_list(&page_list, sc);
1398         }
1399
1400         local_irq_disable();
1401         if (current_is_kswapd())
1402                 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1403         __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1404
1405         putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
1406
1407         trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1408                 zone_idx(zone),
1409                 nr_scanned, nr_reclaimed,
1410                 priority,
1411                 trace_shrink_flags(file, sc->lumpy_reclaim_mode));
1412         return nr_reclaimed;
1413 }
1414
1415 /*
1416  * This moves pages from the active list to the inactive list.
1417  *
1418  * We move them the other way if the page is referenced by one or more
1419  * processes, from rmap.
1420  *
1421  * If the pages are mostly unmapped, the processing is fast and it is
1422  * appropriate to hold zone->lru_lock across the whole operation.  But if
1423  * the pages are mapped, the processing is slow (page_referenced()) so we
1424  * should drop zone->lru_lock around each page.  It's impossible to balance
1425  * this, so instead we remove the pages from the LRU while processing them.
1426  * It is safe to rely on PG_active against the non-LRU pages in here because
1427  * nobody will play with that bit on a non-LRU page.
1428  *
1429  * The downside is that we have to touch page->_count against each page.
1430  * But we had to alter page->flags anyway.
1431  */
1432
1433 static void move_active_pages_to_lru(struct zone *zone,
1434                                      struct list_head *list,
1435                                      enum lru_list lru)
1436 {
1437         unsigned long pgmoved = 0;
1438         struct pagevec pvec;
1439         struct page *page;
1440
1441         pagevec_init(&pvec, 1);
1442
1443         while (!list_empty(list)) {
1444                 page = lru_to_page(list);
1445
1446                 VM_BUG_ON(PageLRU(page));
1447                 SetPageLRU(page);
1448
1449                 list_move(&page->lru, &zone->lru[lru].list);
1450                 mem_cgroup_add_lru_list(page, lru);
1451                 pgmoved++;
1452
1453                 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1454                         spin_unlock_irq(&zone->lru_lock);
1455                         if (buffer_heads_over_limit)
1456                                 pagevec_strip(&pvec);
1457                         __pagevec_release(&pvec);
1458                         spin_lock_irq(&zone->lru_lock);
1459                 }
1460         }
1461         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1462         if (!is_active_lru(lru))
1463                 __count_vm_events(PGDEACTIVATE, pgmoved);
1464 }
1465
1466 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1467                         struct scan_control *sc, int priority, int file)
1468 {
1469         unsigned long nr_taken;
1470         unsigned long pgscanned;
1471         unsigned long vm_flags;
1472         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1473         LIST_HEAD(l_active);
1474         LIST_HEAD(l_inactive);
1475         struct page *page;
1476         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1477         unsigned long nr_rotated = 0;
1478
1479         lru_add_drain();
1480         spin_lock_irq(&zone->lru_lock);
1481         if (scanning_global_lru(sc)) {
1482                 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1483                                                 &pgscanned, sc->order,
1484                                                 ISOLATE_ACTIVE, zone,
1485                                                 1, file);
1486                 zone->pages_scanned += pgscanned;
1487         } else {
1488                 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1489                                                 &pgscanned, sc->order,
1490                                                 ISOLATE_ACTIVE, zone,
1491                                                 sc->mem_cgroup, 1, file);
1492                 /*
1493                  * mem_cgroup_isolate_pages() keeps track of
1494                  * scanned pages on its own.
1495                  */
1496         }
1497
1498         reclaim_stat->recent_scanned[file] += nr_taken;
1499
1500         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1501         if (file)
1502                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1503         else
1504                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1505         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1506         spin_unlock_irq(&zone->lru_lock);
1507
1508         while (!list_empty(&l_hold)) {
1509                 cond_resched();
1510                 page = lru_to_page(&l_hold);
1511                 list_del(&page->lru);
1512
1513                 if (unlikely(!page_evictable(page, NULL))) {
1514                         putback_lru_page(page);
1515                         continue;
1516                 }
1517
1518                 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1519                         nr_rotated++;
1520                         /*
1521                          * Identify referenced, file-backed active pages and
1522                          * give them one more trip around the active list. So
1523                          * that executable code get better chances to stay in
1524                          * memory under moderate memory pressure.  Anon pages
1525                          * are not likely to be evicted by use-once streaming
1526                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1527                          * so we ignore them here.
1528                          */
1529                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1530                                 list_add(&page->lru, &l_active);
1531                                 continue;
1532                         }
1533                 }
1534
1535                 ClearPageActive(page);  /* we are de-activating */
1536                 list_add(&page->lru, &l_inactive);
1537         }
1538
1539         /*
1540          * Move pages back to the lru list.
1541          */
1542         spin_lock_irq(&zone->lru_lock);
1543         /*
1544          * Count referenced pages from currently used mappings as rotated,
1545          * even though only some of them are actually re-activated.  This
1546          * helps balance scan pressure between file and anonymous pages in
1547          * get_scan_ratio.
1548          */
1549         reclaim_stat->recent_rotated[file] += nr_rotated;
1550
1551         move_active_pages_to_lru(zone, &l_active,
1552                                                 LRU_ACTIVE + file * LRU_FILE);
1553         move_active_pages_to_lru(zone, &l_inactive,
1554                                                 LRU_BASE   + file * LRU_FILE);
1555         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1556         spin_unlock_irq(&zone->lru_lock);
1557 }
1558
1559 #ifdef CONFIG_SWAP
1560 static int inactive_anon_is_low_global(struct zone *zone)
1561 {
1562         unsigned long active, inactive;
1563
1564         active = zone_page_state(zone, NR_ACTIVE_ANON);
1565         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1566
1567         if (inactive * zone->inactive_ratio < active)
1568                 return 1;
1569
1570         return 0;
1571 }
1572
1573 /**
1574  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1575  * @zone: zone to check
1576  * @sc:   scan control of this context
1577  *
1578  * Returns true if the zone does not have enough inactive anon pages,
1579  * meaning some active anon pages need to be deactivated.
1580  */
1581 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1582 {
1583         int low;
1584
1585         /*
1586          * If we don't have swap space, anonymous page deactivation
1587          * is pointless.
1588          */
1589         if (!total_swap_pages)
1590                 return 0;
1591
1592         if (scanning_global_lru(sc))
1593                 low = inactive_anon_is_low_global(zone);
1594         else
1595                 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1596         return low;
1597 }
1598 #else
1599 static inline int inactive_anon_is_low(struct zone *zone,
1600                                         struct scan_control *sc)
1601 {
1602         return 0;
1603 }
1604 #endif
1605
1606 static int inactive_file_is_low_global(struct zone *zone)
1607 {
1608         unsigned long active, inactive;
1609
1610         active = zone_page_state(zone, NR_ACTIVE_FILE);
1611         inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1612
1613         return (active > inactive);
1614 }
1615
1616 /**
1617  * inactive_file_is_low - check if file pages need to be deactivated
1618  * @zone: zone to check
1619  * @sc:   scan control of this context
1620  *
1621  * When the system is doing streaming IO, memory pressure here
1622  * ensures that active file pages get deactivated, until more
1623  * than half of the file pages are on the inactive list.
1624  *
1625  * Once we get to that situation, protect the system's working
1626  * set from being evicted by disabling active file page aging.
1627  *
1628  * This uses a different ratio than the anonymous pages, because
1629  * the page cache uses a use-once replacement algorithm.
1630  */
1631 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1632 {
1633         int low;
1634
1635         if (scanning_global_lru(sc))
1636                 low = inactive_file_is_low_global(zone);
1637         else
1638                 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1639         return low;
1640 }
1641
1642 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1643                                 int file)
1644 {
1645         if (file)
1646                 return inactive_file_is_low(zone, sc);
1647         else
1648                 return inactive_anon_is_low(zone, sc);
1649 }
1650
1651 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1652         struct zone *zone, struct scan_control *sc, int priority)
1653 {
1654         int file = is_file_lru(lru);
1655
1656         if (is_active_lru(lru)) {
1657                 if (inactive_list_is_low(zone, sc, file))
1658                     shrink_active_list(nr_to_scan, zone, sc, priority, file);
1659                 return 0;
1660         }
1661
1662         return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1663 }
1664
1665 /*
1666  * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1667  * until we collected @swap_cluster_max pages to scan.
1668  */
1669 static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
1670                                        unsigned long *nr_saved_scan)
1671 {
1672         unsigned long nr;
1673
1674         *nr_saved_scan += nr_to_scan;
1675         nr = *nr_saved_scan;
1676
1677         if (nr >= SWAP_CLUSTER_MAX)
1678                 *nr_saved_scan = 0;
1679         else
1680                 nr = 0;
1681
1682         return nr;
1683 }
1684
1685 /*
1686  * Determine how aggressively the anon and file LRU lists should be
1687  * scanned.  The relative value of each set of LRU lists is determined
1688  * by looking at the fraction of the pages scanned we did rotate back
1689  * onto the active list instead of evict.
1690  *
1691  * nr[0] = anon pages to scan; nr[1] = file pages to scan
1692  */
1693 static void get_scan_count(struct zone *zone, struct scan_control *sc,
1694                                         unsigned long *nr, int priority)
1695 {
1696         unsigned long anon, file, free;
1697         unsigned long anon_prio, file_prio;
1698         unsigned long ap, fp;
1699         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1700         u64 fraction[2], denominator;
1701         enum lru_list l;
1702         int noswap = 0;
1703
1704         /* If we have no swap space, do not bother scanning anon pages. */
1705         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1706                 noswap = 1;
1707                 fraction[0] = 0;
1708                 fraction[1] = 1;
1709                 denominator = 1;
1710                 goto out;
1711         }
1712
1713         anon  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1714                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1715         file  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1716                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1717
1718         if (scanning_global_lru(sc)) {
1719                 free  = zone_page_state(zone, NR_FREE_PAGES);
1720                 /* If we have very few page cache pages,
1721                    force-scan anon pages. */
1722                 if (unlikely(file + free <= high_wmark_pages(zone))) {
1723                         fraction[0] = 1;
1724                         fraction[1] = 0;
1725                         denominator = 1;
1726                         goto out;
1727                 }
1728         }
1729
1730         /*
1731          * With swappiness at 100, anonymous and file have the same priority.
1732          * This scanning priority is essentially the inverse of IO cost.
1733          */
1734         anon_prio = sc->swappiness;
1735         file_prio = 200 - sc->swappiness;
1736
1737         /*
1738          * OK, so we have swap space and a fair amount of page cache
1739          * pages.  We use the recently rotated / recently scanned
1740          * ratios to determine how valuable each cache is.
1741          *
1742          * Because workloads change over time (and to avoid overflow)
1743          * we keep these statistics as a floating average, which ends
1744          * up weighing recent references more than old ones.
1745          *
1746          * anon in [0], file in [1]
1747          */
1748         spin_lock_irq(&zone->lru_lock);
1749         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1750                 reclaim_stat->recent_scanned[0] /= 2;
1751                 reclaim_stat->recent_rotated[0] /= 2;
1752         }
1753
1754         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1755                 reclaim_stat->recent_scanned[1] /= 2;
1756                 reclaim_stat->recent_rotated[1] /= 2;
1757         }
1758
1759         /*
1760          * The amount of pressure on anon vs file pages is inversely
1761          * proportional to the fraction of recently scanned pages on
1762          * each list that were recently referenced and in active use.
1763          */
1764         ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1765         ap /= reclaim_stat->recent_rotated[0] + 1;
1766
1767         fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1768         fp /= reclaim_stat->recent_rotated[1] + 1;
1769         spin_unlock_irq(&zone->lru_lock);
1770
1771         fraction[0] = ap;
1772         fraction[1] = fp;
1773         denominator = ap + fp + 1;
1774 out:
1775         for_each_evictable_lru(l) {
1776                 int file = is_file_lru(l);
1777                 unsigned long scan;
1778
1779                 scan = zone_nr_lru_pages(zone, sc, l);
1780                 if (priority || noswap) {
1781                         scan >>= priority;
1782                         scan = div64_u64(scan * fraction[file], denominator);
1783                 }
1784                 nr[l] = nr_scan_try_batch(scan,
1785                                           &reclaim_stat->nr_saved_scan[l]);
1786         }
1787 }
1788
1789 /*
1790  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1791  */
1792 static void shrink_zone(int priority, struct zone *zone,
1793                                 struct scan_control *sc)
1794 {
1795         unsigned long nr[NR_LRU_LISTS];
1796         unsigned long nr_to_scan;
1797         enum lru_list l;
1798         unsigned long nr_reclaimed = sc->nr_reclaimed;
1799         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1800
1801         get_scan_count(zone, sc, nr, priority);
1802
1803         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1804                                         nr[LRU_INACTIVE_FILE]) {
1805                 for_each_evictable_lru(l) {
1806                         if (nr[l]) {
1807                                 nr_to_scan = min_t(unsigned long,
1808                                                    nr[l], SWAP_CLUSTER_MAX);
1809                                 nr[l] -= nr_to_scan;
1810
1811                                 nr_reclaimed += shrink_list(l, nr_to_scan,
1812                                                             zone, sc, priority);
1813                         }
1814                 }
1815                 /*
1816                  * On large memory systems, scan >> priority can become
1817                  * really large. This is fine for the starting priority;
1818                  * we want to put equal scanning pressure on each zone.
1819                  * However, if the VM has a harder time of freeing pages,
1820                  * with multiple processes reclaiming pages, the total
1821                  * freeing target can get unreasonably large.
1822                  */
1823                 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
1824                         break;
1825         }
1826
1827         sc->nr_reclaimed = nr_reclaimed;
1828
1829         /*
1830          * Even if we did not try to evict anon pages at all, we want to
1831          * rebalance the anon lru active/inactive ratio.
1832          */
1833         if (inactive_anon_is_low(zone, sc))
1834                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1835
1836         throttle_vm_writeout(sc->gfp_mask);
1837 }
1838
1839 /*
1840  * This is the direct reclaim path, for page-allocating processes.  We only
1841  * try to reclaim pages from zones which will satisfy the caller's allocation
1842  * request.
1843  *
1844  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1845  * Because:
1846  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1847  *    allocation or
1848  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1849  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1850  *    zone defense algorithm.
1851  *
1852  * If a zone is deemed to be full of pinned pages then just give it a light
1853  * scan then give up on it.
1854  */
1855 static void shrink_zones(int priority, struct zonelist *zonelist,
1856                                         struct scan_control *sc)
1857 {
1858         struct zoneref *z;
1859         struct zone *zone;
1860
1861         for_each_zone_zonelist_nodemask(zone, z, zonelist,
1862                                         gfp_zone(sc->gfp_mask), sc->nodemask) {
1863                 if (!populated_zone(zone))
1864                         continue;
1865                 /*
1866                  * Take care memory controller reclaiming has small influence
1867                  * to global LRU.
1868                  */
1869                 if (scanning_global_lru(sc)) {
1870                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1871                                 continue;
1872                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1873                                 continue;       /* Let kswapd poll it */
1874                 }
1875
1876                 shrink_zone(priority, zone, sc);
1877         }
1878 }
1879
1880 static bool zone_reclaimable(struct zone *zone)
1881 {
1882         return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
1883 }
1884
1885 /*
1886  * As hibernation is going on, kswapd is freezed so that it can't mark
1887  * the zone into all_unreclaimable. It can't handle OOM during hibernation.
1888  * So let's check zone's unreclaimable in direct reclaim as well as kswapd.
1889  */
1890 static bool all_unreclaimable(struct zonelist *zonelist,
1891                 struct scan_control *sc)
1892 {
1893         struct zoneref *z;
1894         struct zone *zone;
1895         bool all_unreclaimable = true;
1896
1897         for_each_zone_zonelist_nodemask(zone, z, zonelist,
1898                         gfp_zone(sc->gfp_mask), sc->nodemask) {
1899                 if (!populated_zone(zone))
1900                         continue;
1901                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1902                         continue;
1903                 if (zone_reclaimable(zone)) {
1904                         all_unreclaimable = false;
1905                         break;
1906                 }
1907         }
1908
1909         return all_unreclaimable;
1910 }
1911
1912 /*
1913  * This is the main entry point to direct page reclaim.
1914  *
1915  * If a full scan of the inactive list fails to free enough memory then we
1916  * are "out of memory" and something needs to be killed.
1917  *
1918  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1919  * high - the zone may be full of dirty or under-writeback pages, which this
1920  * caller can't do much about.  We kick the writeback threads and take explicit
1921  * naps in the hope that some of these pages can be written.  But if the
1922  * allocating task holds filesystem locks which prevent writeout this might not
1923  * work, and the allocation attempt will fail.
1924  *
1925  * returns:     0, if no pages reclaimed
1926  *              else, the number of pages reclaimed
1927  */
1928 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
1929                                         struct scan_control *sc)
1930 {
1931         int priority;
1932         unsigned long total_scanned = 0;
1933         struct reclaim_state *reclaim_state = current->reclaim_state;
1934         struct zoneref *z;
1935         struct zone *zone;
1936         unsigned long writeback_threshold;
1937
1938         get_mems_allowed();
1939         delayacct_freepages_start();
1940
1941         if (scanning_global_lru(sc))
1942                 count_vm_event(ALLOCSTALL);
1943
1944         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1945                 sc->nr_scanned = 0;
1946                 if (!priority)
1947                         disable_swap_token();
1948                 shrink_zones(priority, zonelist, sc);
1949                 /*
1950                  * Don't shrink slabs when reclaiming memory from
1951                  * over limit cgroups
1952                  */
1953                 if (scanning_global_lru(sc)) {
1954                         unsigned long lru_pages = 0;
1955                         for_each_zone_zonelist(zone, z, zonelist,
1956                                         gfp_zone(sc->gfp_mask)) {
1957                                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1958                                         continue;
1959
1960                                 lru_pages += zone_reclaimable_pages(zone);
1961                         }
1962
1963                         shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
1964                         if (reclaim_state) {
1965                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
1966                                 reclaim_state->reclaimed_slab = 0;
1967                         }
1968                 }
1969                 total_scanned += sc->nr_scanned;
1970                 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
1971                         goto out;
1972
1973                 /*
1974                  * Try to write back as many pages as we just scanned.  This
1975                  * tends to cause slow streaming writers to write data to the
1976                  * disk smoothly, at the dirtying rate, which is nice.   But
1977                  * that's undesirable in laptop mode, where we *want* lumpy
1978                  * writeout.  So in laptop mode, write out the whole world.
1979                  */
1980                 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
1981                 if (total_scanned > writeback_threshold) {
1982                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
1983                         sc->may_writepage = 1;
1984                 }
1985
1986                 /* Take a nap, wait for some writeback to complete */
1987                 if (!sc->hibernation_mode && sc->nr_scanned &&
1988                     priority < DEF_PRIORITY - 2)
1989                         congestion_wait(BLK_RW_ASYNC, HZ/10);
1990         }
1991
1992 out:
1993         delayacct_freepages_end();
1994         put_mems_allowed();
1995
1996         if (sc->nr_reclaimed)
1997                 return sc->nr_reclaimed;
1998
1999         /* top priority shrink_zones still had more to do? don't OOM, then */
2000         if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
2001                 return 1;
2002
2003         return 0;
2004 }
2005
2006 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2007                                 gfp_t gfp_mask, nodemask_t *nodemask)
2008 {
2009         unsigned long nr_reclaimed;
2010         struct scan_control sc = {
2011                 .gfp_mask = gfp_mask,
2012                 .may_writepage = !laptop_mode,
2013                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2014                 .may_unmap = 1,
2015                 .may_swap = 1,
2016                 .swappiness = vm_swappiness,
2017                 .order = order,
2018                 .mem_cgroup = NULL,
2019                 .nodemask = nodemask,
2020         };
2021
2022         trace_mm_vmscan_direct_reclaim_begin(order,
2023                                 sc.may_writepage,
2024                                 gfp_mask);
2025
2026         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2027
2028         trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2029
2030         return nr_reclaimed;
2031 }
2032
2033 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2034
2035 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2036                                                 gfp_t gfp_mask, bool noswap,
2037                                                 unsigned int swappiness,
2038                                                 struct zone *zone)
2039 {
2040         struct scan_control sc = {
2041                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2042                 .may_writepage = !laptop_mode,
2043                 .may_unmap = 1,
2044                 .may_swap = !noswap,
2045                 .swappiness = swappiness,
2046                 .order = 0,
2047                 .mem_cgroup = mem,
2048         };
2049         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2050                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2051
2052         trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2053                                                       sc.may_writepage,
2054                                                       sc.gfp_mask);
2055
2056         /*
2057          * NOTE: Although we can get the priority field, using it
2058          * here is not a good idea, since it limits the pages we can scan.
2059          * if we don't reclaim here, the shrink_zone from balance_pgdat
2060          * will pick up pages from other mem cgroup's as well. We hack
2061          * the priority and make it zero.
2062          */
2063         shrink_zone(0, zone, &sc);
2064
2065         trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2066
2067         return sc.nr_reclaimed;
2068 }
2069
2070 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
2071                                            gfp_t gfp_mask,
2072                                            bool noswap,
2073                                            unsigned int swappiness)
2074 {
2075         struct zonelist *zonelist;
2076         unsigned long nr_reclaimed;
2077         struct scan_control sc = {
2078                 .may_writepage = !laptop_mode,
2079                 .may_unmap = 1,
2080                 .may_swap = !noswap,
2081                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2082                 .swappiness = swappiness,
2083                 .order = 0,
2084                 .mem_cgroup = mem_cont,
2085                 .nodemask = NULL, /* we don't care the placement */
2086         };
2087
2088         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2089                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2090         zonelist = NODE_DATA(numa_node_id())->node_zonelists;
2091
2092         trace_mm_vmscan_memcg_reclaim_begin(0,
2093                                             sc.may_writepage,
2094                                             sc.gfp_mask);
2095
2096         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2097
2098         trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2099
2100         return nr_reclaimed;
2101 }
2102 #endif
2103
2104 /* is kswapd sleeping prematurely? */
2105 static int sleeping_prematurely(pg_data_t *pgdat, int order, long remaining)
2106 {
2107         int i;
2108
2109         /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2110         if (remaining)
2111                 return 1;
2112
2113         /* If after HZ/10, a zone is below the high mark, it's premature */
2114         for (i = 0; i < pgdat->nr_zones; i++) {
2115                 struct zone *zone = pgdat->node_zones + i;
2116
2117                 if (!populated_zone(zone))
2118                         continue;
2119
2120                 if (zone->all_unreclaimable)
2121                         continue;
2122
2123                 if (!zone_watermark_ok(zone, order, high_wmark_pages(zone),
2124                                                                 0, 0))
2125                         return 1;
2126         }
2127
2128         return 0;
2129 }
2130
2131 /*
2132  * For kswapd, balance_pgdat() will work across all this node's zones until
2133  * they are all at high_wmark_pages(zone).
2134  *
2135  * Returns the number of pages which were actually freed.
2136  *
2137  * There is special handling here for zones which are full of pinned pages.
2138  * This can happen if the pages are all mlocked, or if they are all used by
2139  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
2140  * What we do is to detect the case where all pages in the zone have been
2141  * scanned twice and there has been zero successful reclaim.  Mark the zone as
2142  * dead and from now on, only perform a short scan.  Basically we're polling
2143  * the zone for when the problem goes away.
2144  *
2145  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
2146  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2147  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2148  * lower zones regardless of the number of free pages in the lower zones. This
2149  * interoperates with the page allocator fallback scheme to ensure that aging
2150  * of pages is balanced across the zones.
2151  */
2152 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
2153 {
2154         int all_zones_ok;
2155         int priority;
2156         int i;
2157         unsigned long total_scanned;
2158         struct reclaim_state *reclaim_state = current->reclaim_state;
2159         struct scan_control sc = {
2160                 .gfp_mask = GFP_KERNEL,
2161                 .may_unmap = 1,
2162                 .may_swap = 1,
2163                 /*
2164                  * kswapd doesn't want to be bailed out while reclaim. because
2165                  * we want to put equal scanning pressure on each zone.
2166                  */
2167                 .nr_to_reclaim = ULONG_MAX,
2168                 .swappiness = vm_swappiness,
2169                 .order = order,
2170                 .mem_cgroup = NULL,
2171         };
2172 loop_again:
2173         total_scanned = 0;
2174         sc.nr_reclaimed = 0;
2175         sc.may_writepage = !laptop_mode;
2176         count_vm_event(PAGEOUTRUN);
2177
2178         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2179                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
2180                 unsigned long lru_pages = 0;
2181                 int has_under_min_watermark_zone = 0;
2182
2183                 /* The swap token gets in the way of swapout... */
2184                 if (!priority)
2185                         disable_swap_token();
2186
2187                 all_zones_ok = 1;
2188
2189                 /*
2190                  * Scan in the highmem->dma direction for the highest
2191                  * zone which needs scanning
2192                  */
2193                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2194                         struct zone *zone = pgdat->node_zones + i;
2195
2196                         if (!populated_zone(zone))
2197                                 continue;
2198
2199                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2200                                 continue;
2201
2202                         /*
2203                          * Do some background aging of the anon list, to give
2204                          * pages a chance to be referenced before reclaiming.
2205                          */
2206                         if (inactive_anon_is_low(zone, &sc))
2207                                 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2208                                                         &sc, priority, 0);
2209
2210                         if (!zone_watermark_ok(zone, order,
2211                                         high_wmark_pages(zone), 0, 0)) {
2212                                 end_zone = i;
2213                                 break;
2214                         }
2215                 }
2216                 if (i < 0)
2217                         goto out;
2218
2219                 for (i = 0; i <= end_zone; i++) {
2220                         struct zone *zone = pgdat->node_zones + i;
2221
2222                         lru_pages += zone_reclaimable_pages(zone);
2223                 }
2224
2225                 /*
2226                  * Now scan the zone in the dma->highmem direction, stopping
2227                  * at the last zone which needs scanning.
2228                  *
2229                  * We do this because the page allocator works in the opposite
2230                  * direction.  This prevents the page allocator from allocating
2231                  * pages behind kswapd's direction of progress, which would
2232                  * cause too much scanning of the lower zones.
2233                  */
2234                 for (i = 0; i <= end_zone; i++) {
2235                         struct zone *zone = pgdat->node_zones + i;
2236                         int nr_slab;
2237
2238                         if (!populated_zone(zone))
2239                                 continue;
2240
2241                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2242                                 continue;
2243
2244                         sc.nr_scanned = 0;
2245
2246                         /*
2247                          * Call soft limit reclaim before calling shrink_zone.
2248                          * For now we ignore the return value
2249                          */
2250                         mem_cgroup_soft_limit_reclaim(zone, order, sc.gfp_mask);
2251
2252                         /*
2253                          * We put equal pressure on every zone, unless one
2254                          * zone has way too many pages free already.
2255                          */
2256                         if (!zone_watermark_ok(zone, order,
2257                                         8*high_wmark_pages(zone), end_zone, 0))
2258                                 shrink_zone(priority, zone, &sc);
2259                         reclaim_state->reclaimed_slab = 0;
2260                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
2261                                                 lru_pages);
2262                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2263                         total_scanned += sc.nr_scanned;
2264                         if (zone->all_unreclaimable)
2265                                 continue;
2266                         if (nr_slab == 0 && !zone_reclaimable(zone))
2267                                 zone->all_unreclaimable = 1;
2268                         /*
2269                          * If we've done a decent amount of scanning and
2270                          * the reclaim ratio is low, start doing writepage
2271                          * even in laptop mode
2272                          */
2273                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2274                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2275                                 sc.may_writepage = 1;
2276
2277                         if (!zone_watermark_ok(zone, order,
2278                                         high_wmark_pages(zone), end_zone, 0)) {
2279                                 all_zones_ok = 0;
2280                                 /*
2281                                  * We are still under min water mark.  This
2282                                  * means that we have a GFP_ATOMIC allocation
2283                                  * failure risk. Hurry up!
2284                                  */
2285                                 if (!zone_watermark_ok(zone, order,
2286                                             min_wmark_pages(zone), end_zone, 0))
2287                                         has_under_min_watermark_zone = 1;
2288                         }
2289
2290                 }
2291                 if (all_zones_ok)
2292                         break;          /* kswapd: all done */
2293                 /*
2294                  * OK, kswapd is getting into trouble.  Take a nap, then take
2295                  * another pass across the zones.
2296                  */
2297                 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2298                         if (has_under_min_watermark_zone)
2299                                 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2300                         else
2301                                 congestion_wait(BLK_RW_ASYNC, HZ/10);
2302                 }
2303
2304                 /*
2305                  * We do this so kswapd doesn't build up large priorities for
2306                  * example when it is freeing in parallel with allocators. It
2307                  * matches the direct reclaim path behaviour in terms of impact
2308                  * on zone->*_priority.
2309                  */
2310                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2311                         break;
2312         }
2313 out:
2314         if (!all_zones_ok) {
2315                 cond_resched();
2316
2317                 try_to_freeze();
2318
2319                 /*
2320                  * Fragmentation may mean that the system cannot be
2321                  * rebalanced for high-order allocations in all zones.
2322                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2323                  * it means the zones have been fully scanned and are still
2324                  * not balanced. For high-order allocations, there is
2325                  * little point trying all over again as kswapd may
2326                  * infinite loop.
2327                  *
2328                  * Instead, recheck all watermarks at order-0 as they
2329                  * are the most important. If watermarks are ok, kswapd will go
2330                  * back to sleep. High-order users can still perform direct
2331                  * reclaim if they wish.
2332                  */
2333                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2334                         order = sc.order = 0;
2335
2336                 goto loop_again;
2337         }
2338
2339         return sc.nr_reclaimed;
2340 }
2341
2342 /*
2343  * The background pageout daemon, started as a kernel thread
2344  * from the init process.
2345  *
2346  * This basically trickles out pages so that we have _some_
2347  * free memory available even if there is no other activity
2348  * that frees anything up. This is needed for things like routing
2349  * etc, where we otherwise might have all activity going on in
2350  * asynchronous contexts that cannot page things out.
2351  *
2352  * If there are applications that are active memory-allocators
2353  * (most normal use), this basically shouldn't matter.
2354  */
2355 static int kswapd(void *p)
2356 {
2357         unsigned long order;
2358         pg_data_t *pgdat = (pg_data_t*)p;
2359         struct task_struct *tsk = current;
2360         DEFINE_WAIT(wait);
2361         struct reclaim_state reclaim_state = {
2362                 .reclaimed_slab = 0,
2363         };
2364         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2365
2366         lockdep_set_current_reclaim_state(GFP_KERNEL);
2367
2368         if (!cpumask_empty(cpumask))
2369                 set_cpus_allowed_ptr(tsk, cpumask);
2370         current->reclaim_state = &reclaim_state;
2371
2372         /*
2373          * Tell the memory management that we're a "memory allocator",
2374          * and that if we need more memory we should get access to it
2375          * regardless (see "__alloc_pages()"). "kswapd" should
2376          * never get caught in the normal page freeing logic.
2377          *
2378          * (Kswapd normally doesn't need memory anyway, but sometimes
2379          * you need a small amount of memory in order to be able to
2380          * page out something else, and this flag essentially protects
2381          * us from recursively trying to free more memory as we're
2382          * trying to free the first piece of memory in the first place).
2383          */
2384         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2385         set_freezable();
2386
2387         order = 0;
2388         for ( ; ; ) {
2389                 unsigned long new_order;
2390                 int ret;
2391
2392                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2393                 new_order = pgdat->kswapd_max_order;
2394                 pgdat->kswapd_max_order = 0;
2395                 if (order < new_order) {
2396                         /*
2397                          * Don't sleep if someone wants a larger 'order'
2398                          * allocation
2399                          */
2400                         order = new_order;
2401                 } else {
2402                         if (!freezing(current) && !kthread_should_stop()) {
2403                                 long remaining = 0;
2404
2405                                 /* Try to sleep for a short interval */
2406                                 if (!sleeping_prematurely(pgdat, order, remaining)) {
2407                                         remaining = schedule_timeout(HZ/10);
2408                                         finish_wait(&pgdat->kswapd_wait, &wait);
2409                                         prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2410                                 }
2411
2412                                 /*
2413                                  * After a short sleep, check if it was a
2414                                  * premature sleep. If not, then go fully
2415                                  * to sleep until explicitly woken up
2416                                  */
2417                                 if (!sleeping_prematurely(pgdat, order, remaining)) {
2418                                         trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2419                                         schedule();
2420                                 } else {
2421                                         if (remaining)
2422                                                 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2423                                         else
2424                                                 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2425                                 }
2426                         }
2427
2428                         order = pgdat->kswapd_max_order;
2429                 }
2430                 finish_wait(&pgdat->kswapd_wait, &wait);
2431
2432                 ret = try_to_freeze();
2433                 if (kthread_should_stop())
2434                         break;
2435
2436                 /*
2437                  * We can speed up thawing tasks if we don't call balance_pgdat
2438                  * after returning from the refrigerator
2439                  */
2440                 if (!ret) {
2441                         trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2442                         balance_pgdat(pgdat, order);
2443                 }
2444         }
2445         return 0;
2446 }
2447
2448 /*
2449  * A zone is low on free memory, so wake its kswapd task to service it.
2450  */
2451 void wakeup_kswapd(struct zone *zone, int order)
2452 {
2453         pg_data_t *pgdat;
2454
2455         if (!populated_zone(zone))
2456                 return;
2457
2458         pgdat = zone->zone_pgdat;
2459         if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
2460                 return;
2461         if (pgdat->kswapd_max_order < order)
2462                 pgdat->kswapd_max_order = order;
2463         trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
2464         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2465                 return;
2466         if (!waitqueue_active(&pgdat->kswapd_wait))
2467                 return;
2468         wake_up_interruptible(&pgdat->kswapd_wait);
2469 }
2470
2471 /*
2472  * The reclaimable count would be mostly accurate.
2473  * The less reclaimable pages may be
2474  * - mlocked pages, which will be moved to unevictable list when encountered
2475  * - mapped pages, which may require several travels to be reclaimed
2476  * - dirty pages, which is not "instantly" reclaimable
2477  */
2478 unsigned long global_reclaimable_pages(void)
2479 {
2480         int nr;
2481
2482         nr = global_page_state(NR_ACTIVE_FILE) +
2483              global_page_state(NR_INACTIVE_FILE);
2484
2485         if (nr_swap_pages > 0)
2486                 nr += global_page_state(NR_ACTIVE_ANON) +
2487                       global_page_state(NR_INACTIVE_ANON);
2488
2489         return nr;
2490 }
2491
2492 unsigned long zone_reclaimable_pages(struct zone *zone)
2493 {
2494         int nr;
2495
2496         nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2497              zone_page_state(zone, NR_INACTIVE_FILE);
2498
2499         if (nr_swap_pages > 0)
2500                 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2501                       zone_page_state(zone, NR_INACTIVE_ANON);
2502
2503         return nr;
2504 }
2505
2506 #ifdef CONFIG_HIBERNATION
2507 /*
2508  * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2509  * freed pages.
2510  *
2511  * Rather than trying to age LRUs the aim is to preserve the overall
2512  * LRU order by reclaiming preferentially
2513  * inactive > active > active referenced > active mapped
2514  */
2515 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
2516 {
2517         struct reclaim_state reclaim_state;
2518         struct scan_control sc = {
2519                 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2520                 .may_swap = 1,
2521                 .may_unmap = 1,
2522                 .may_writepage = 1,
2523                 .nr_to_reclaim = nr_to_reclaim,
2524                 .hibernation_mode = 1,
2525                 .swappiness = vm_swappiness,
2526                 .order = 0,
2527         };
2528         struct zonelist * zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2529         struct task_struct *p = current;
2530         unsigned long nr_reclaimed;
2531
2532         p->flags |= PF_MEMALLOC;
2533         lockdep_set_current_reclaim_state(sc.gfp_mask);
2534         reclaim_state.reclaimed_slab = 0;
2535         p->reclaim_state = &reclaim_state;
2536
2537         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2538
2539         p->reclaim_state = NULL;
2540         lockdep_clear_current_reclaim_state();
2541         p->flags &= ~PF_MEMALLOC;
2542
2543         return nr_reclaimed;
2544 }
2545 #endif /* CONFIG_HIBERNATION */
2546
2547 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2548    not required for correctness.  So if the last cpu in a node goes
2549    away, we get changed to run anywhere: as the first one comes back,
2550    restore their cpu bindings. */
2551 static int __devinit cpu_callback(struct notifier_block *nfb,
2552                                   unsigned long action, void *hcpu)
2553 {
2554         int nid;
2555
2556         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2557                 for_each_node_state(nid, N_HIGH_MEMORY) {
2558                         pg_data_t *pgdat = NODE_DATA(nid);
2559                         const struct cpumask *mask;
2560
2561                         mask = cpumask_of_node(pgdat->node_id);
2562
2563                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2564                                 /* One of our CPUs online: restore mask */
2565                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2566                 }
2567         }
2568         return NOTIFY_OK;
2569 }
2570
2571 /*
2572  * This kswapd start function will be called by init and node-hot-add.
2573  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2574  */
2575 int kswapd_run(int nid)
2576 {
2577         pg_data_t *pgdat = NODE_DATA(nid);
2578         int ret = 0;
2579
2580         if (pgdat->kswapd)
2581                 return 0;
2582
2583         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2584         if (IS_ERR(pgdat->kswapd)) {
2585                 /* failure at boot is fatal */
2586                 BUG_ON(system_state == SYSTEM_BOOTING);
2587                 printk("Failed to start kswapd on node %d\n",nid);
2588                 ret = -1;
2589         }
2590         return ret;
2591 }
2592
2593 /*
2594  * Called by memory hotplug when all memory in a node is offlined.
2595  */
2596 void kswapd_stop(int nid)
2597 {
2598         struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
2599
2600         if (kswapd)
2601                 kthread_stop(kswapd);
2602 }
2603
2604 static int __init kswapd_init(void)
2605 {
2606         int nid;
2607
2608         swap_setup();
2609         for_each_node_state(nid, N_HIGH_MEMORY)
2610                 kswapd_run(nid);
2611         hotcpu_notifier(cpu_callback, 0);
2612         return 0;
2613 }
2614
2615 module_init(kswapd_init)
2616
2617 #ifdef CONFIG_NUMA
2618 /*
2619  * Zone reclaim mode
2620  *
2621  * If non-zero call zone_reclaim when the number of free pages falls below
2622  * the watermarks.
2623  */
2624 int zone_reclaim_mode __read_mostly;
2625
2626 #define RECLAIM_OFF 0
2627 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
2628 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
2629 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
2630
2631 /*
2632  * Priority for ZONE_RECLAIM. This determines the fraction of pages
2633  * of a node considered for each zone_reclaim. 4 scans 1/16th of
2634  * a zone.
2635  */
2636 #define ZONE_RECLAIM_PRIORITY 4
2637
2638 /*
2639  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2640  * occur.
2641  */
2642 int sysctl_min_unmapped_ratio = 1;
2643
2644 /*
2645  * If the number of slab pages in a zone grows beyond this percentage then
2646  * slab reclaim needs to occur.
2647  */
2648 int sysctl_min_slab_ratio = 5;
2649
2650 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
2651 {
2652         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
2653         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
2654                 zone_page_state(zone, NR_ACTIVE_FILE);
2655
2656         /*
2657          * It's possible for there to be more file mapped pages than
2658          * accounted for by the pages on the file LRU lists because
2659          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2660          */
2661         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
2662 }
2663
2664 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
2665 static long zone_pagecache_reclaimable(struct zone *zone)
2666 {
2667         long nr_pagecache_reclaimable;
2668         long delta = 0;
2669
2670         /*
2671          * If RECLAIM_SWAP is set, then all file pages are considered
2672          * potentially reclaimable. Otherwise, we have to worry about
2673          * pages like swapcache and zone_unmapped_file_pages() provides
2674          * a better estimate
2675          */
2676         if (zone_reclaim_mode & RECLAIM_SWAP)
2677                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
2678         else
2679                 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
2680
2681         /* If we can't clean pages, remove dirty pages from consideration */
2682         if (!(zone_reclaim_mode & RECLAIM_WRITE))
2683                 delta += zone_page_state(zone, NR_FILE_DIRTY);
2684
2685         /* Watch for any possible underflows due to delta */
2686         if (unlikely(delta > nr_pagecache_reclaimable))
2687                 delta = nr_pagecache_reclaimable;
2688
2689         return nr_pagecache_reclaimable - delta;
2690 }
2691
2692 /*
2693  * Try to free up some pages from this zone through reclaim.
2694  */
2695 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2696 {
2697         /* Minimum pages needed in order to stay on node */
2698         const unsigned long nr_pages = 1 << order;
2699         struct task_struct *p = current;
2700         struct reclaim_state reclaim_state;
2701         int priority;
2702         struct scan_control sc = {
2703                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2704                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2705                 .may_swap = 1,
2706                 .nr_to_reclaim = max_t(unsigned long, nr_pages,
2707                                        SWAP_CLUSTER_MAX),
2708                 .gfp_mask = gfp_mask,
2709                 .swappiness = vm_swappiness,
2710                 .order = order,
2711         };
2712         unsigned long nr_slab_pages0, nr_slab_pages1;
2713
2714         cond_resched();
2715         /*
2716          * We need to be able to allocate from the reserves for RECLAIM_SWAP
2717          * and we also need to be able to write out pages for RECLAIM_WRITE
2718          * and RECLAIM_SWAP.
2719          */
2720         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2721         lockdep_set_current_reclaim_state(gfp_mask);
2722         reclaim_state.reclaimed_slab = 0;
2723         p->reclaim_state = &reclaim_state;
2724
2725         if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
2726                 /*
2727                  * Free memory by calling shrink zone with increasing
2728                  * priorities until we have enough memory freed.
2729                  */
2730                 priority = ZONE_RECLAIM_PRIORITY;
2731                 do {
2732                         shrink_zone(priority, zone, &sc);
2733                         priority--;
2734                 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
2735         }
2736
2737         nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2738         if (nr_slab_pages0 > zone->min_slab_pages) {
2739                 /*
2740                  * shrink_slab() does not currently allow us to determine how
2741                  * many pages were freed in this zone. So we take the current
2742                  * number of slab pages and shake the slab until it is reduced
2743                  * by the same nr_pages that we used for reclaiming unmapped
2744                  * pages.
2745                  *
2746                  * Note that shrink_slab will free memory on all zones and may
2747                  * take a long time.
2748                  */
2749                 for (;;) {
2750                         unsigned long lru_pages = zone_reclaimable_pages(zone);
2751
2752                         /* No reclaimable slab or very low memory pressure */
2753                         if (!shrink_slab(sc.nr_scanned, gfp_mask, lru_pages))
2754                                 break;
2755
2756                         /* Freed enough memory */
2757                         nr_slab_pages1 = zone_page_state(zone,
2758                                                         NR_SLAB_RECLAIMABLE);
2759                         if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
2760                                 break;
2761                 }
2762
2763                 /*
2764                  * Update nr_reclaimed by the number of slab pages we
2765                  * reclaimed from this zone.
2766                  */
2767                 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2768                 if (nr_slab_pages1 < nr_slab_pages0)
2769                         sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
2770         }
2771
2772         p->reclaim_state = NULL;
2773         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
2774         lockdep_clear_current_reclaim_state();
2775         return sc.nr_reclaimed >= nr_pages;
2776 }
2777
2778 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2779 {
2780         int node_id;
2781         int ret;
2782
2783         /*
2784          * Zone reclaim reclaims unmapped file backed pages and
2785          * slab pages if we are over the defined limits.
2786          *
2787          * A small portion of unmapped file backed pages is needed for
2788          * file I/O otherwise pages read by file I/O will be immediately
2789          * thrown out if the zone is overallocated. So we do not reclaim
2790          * if less than a specified percentage of the zone is used by
2791          * unmapped file backed pages.
2792          */
2793         if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
2794             zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
2795                 return ZONE_RECLAIM_FULL;
2796
2797         if (zone->all_unreclaimable)
2798                 return ZONE_RECLAIM_FULL;
2799
2800         /*
2801          * Do not scan if the allocation should not be delayed.
2802          */
2803         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
2804                 return ZONE_RECLAIM_NOSCAN;
2805
2806         /*
2807          * Only run zone reclaim on the local zone or on zones that do not
2808          * have associated processors. This will favor the local processor
2809          * over remote processors and spread off node memory allocations
2810          * as wide as possible.
2811          */
2812         node_id = zone_to_nid(zone);
2813         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
2814                 return ZONE_RECLAIM_NOSCAN;
2815
2816         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2817                 return ZONE_RECLAIM_NOSCAN;
2818
2819         ret = __zone_reclaim(zone, gfp_mask, order);
2820         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2821
2822         if (!ret)
2823                 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
2824
2825         return ret;
2826 }
2827 #endif
2828
2829 /*
2830  * page_evictable - test whether a page is evictable
2831  * @page: the page to test
2832  * @vma: the VMA in which the page is or will be mapped, may be NULL
2833  *
2834  * Test whether page is evictable--i.e., should be placed on active/inactive
2835  * lists vs unevictable list.  The vma argument is !NULL when called from the
2836  * fault path to determine how to instantate a new page.
2837  *
2838  * Reasons page might not be evictable:
2839  * (1) page's mapping marked unevictable
2840  * (2) page is part of an mlocked VMA
2841  *
2842  */
2843 int page_evictable(struct page *page, struct vm_area_struct *vma)
2844 {
2845
2846         if (mapping_unevictable(page_mapping(page)))
2847                 return 0;
2848
2849         if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2850                 return 0;
2851
2852         return 1;
2853 }
2854
2855 /**
2856  * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2857  * @page: page to check evictability and move to appropriate lru list
2858  * @zone: zone page is in
2859  *
2860  * Checks a page for evictability and moves the page to the appropriate
2861  * zone lru list.
2862  *
2863  * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2864  * have PageUnevictable set.
2865  */
2866 static void check_move_unevictable_page(struct page *page, struct zone *zone)
2867 {
2868         VM_BUG_ON(PageActive(page));
2869
2870 retry:
2871         ClearPageUnevictable(page);
2872         if (page_evictable(page, NULL)) {
2873                 enum lru_list l = page_lru_base_type(page);
2874
2875                 __dec_zone_state(zone, NR_UNEVICTABLE);
2876                 list_move(&page->lru, &zone->lru[l].list);
2877                 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
2878                 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
2879                 __count_vm_event(UNEVICTABLE_PGRESCUED);
2880         } else {
2881                 /*
2882                  * rotate unevictable list
2883                  */
2884                 SetPageUnevictable(page);
2885                 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
2886                 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
2887                 if (page_evictable(page, NULL))
2888                         goto retry;
2889         }
2890 }
2891
2892 /**
2893  * scan_mapping_unevictable_pages - scan an address space for evictable pages
2894  * @mapping: struct address_space to scan for evictable pages
2895  *
2896  * Scan all pages in mapping.  Check unevictable pages for
2897  * evictability and move them to the appropriate zone lru list.
2898  */
2899 void scan_mapping_unevictable_pages(struct address_space *mapping)
2900 {
2901         pgoff_t next = 0;
2902         pgoff_t end   = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
2903                          PAGE_CACHE_SHIFT;
2904         struct zone *zone;
2905         struct pagevec pvec;
2906
2907         if (mapping->nrpages == 0)
2908                 return;
2909
2910         pagevec_init(&pvec, 0);
2911         while (next < end &&
2912                 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
2913                 int i;
2914                 int pg_scanned = 0;
2915
2916                 zone = NULL;
2917
2918                 for (i = 0; i < pagevec_count(&pvec); i++) {
2919                         struct page *page = pvec.pages[i];
2920                         pgoff_t page_index = page->index;
2921                         struct zone *pagezone = page_zone(page);
2922
2923                         pg_scanned++;
2924                         if (page_index > next)
2925                                 next = page_index;
2926                         next++;
2927
2928                         if (pagezone != zone) {
2929                                 if (zone)
2930                                         spin_unlock_irq(&zone->lru_lock);
2931                                 zone = pagezone;
2932                                 spin_lock_irq(&zone->lru_lock);
2933                         }
2934
2935                         if (PageLRU(page) && PageUnevictable(page))
2936                                 check_move_unevictable_page(page, zone);
2937                 }
2938                 if (zone)
2939                         spin_unlock_irq(&zone->lru_lock);
2940                 pagevec_release(&pvec);
2941
2942                 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
2943         }
2944
2945 }
2946
2947 /**
2948  * scan_zone_unevictable_pages - check unevictable list for evictable pages
2949  * @zone - zone of which to scan the unevictable list
2950  *
2951  * Scan @zone's unevictable LRU lists to check for pages that have become
2952  * evictable.  Move those that have to @zone's inactive list where they
2953  * become candidates for reclaim, unless shrink_inactive_zone() decides
2954  * to reactivate them.  Pages that are still unevictable are rotated
2955  * back onto @zone's unevictable list.
2956  */
2957 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
2958 static void scan_zone_unevictable_pages(struct zone *zone)
2959 {
2960         struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
2961         unsigned long scan;
2962         unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
2963
2964         while (nr_to_scan > 0) {
2965                 unsigned long batch_size = min(nr_to_scan,
2966                                                 SCAN_UNEVICTABLE_BATCH_SIZE);
2967
2968                 spin_lock_irq(&zone->lru_lock);
2969                 for (scan = 0;  scan < batch_size; scan++) {
2970                         struct page *page = lru_to_page(l_unevictable);
2971
2972                         if (!trylock_page(page))
2973                                 continue;
2974
2975                         prefetchw_prev_lru_page(page, l_unevictable, flags);
2976
2977                         if (likely(PageLRU(page) && PageUnevictable(page)))
2978                                 check_move_unevictable_page(page, zone);
2979
2980                         unlock_page(page);
2981                 }
2982                 spin_unlock_irq(&zone->lru_lock);
2983
2984                 nr_to_scan -= batch_size;
2985         }
2986 }
2987
2988
2989 /**
2990  * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
2991  *
2992  * A really big hammer:  scan all zones' unevictable LRU lists to check for
2993  * pages that have become evictable.  Move those back to the zones'
2994  * inactive list where they become candidates for reclaim.
2995  * This occurs when, e.g., we have unswappable pages on the unevictable lists,
2996  * and we add swap to the system.  As such, it runs in the context of a task
2997  * that has possibly/probably made some previously unevictable pages
2998  * evictable.
2999  */
3000 static void scan_all_zones_unevictable_pages(void)
3001 {
3002         struct zone *zone;
3003
3004         for_each_zone(zone) {
3005                 scan_zone_unevictable_pages(zone);
3006         }
3007 }
3008
3009 /*
3010  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
3011  * all nodes' unevictable lists for evictable pages
3012  */
3013 unsigned long scan_unevictable_pages;
3014
3015 int scan_unevictable_handler(struct ctl_table *table, int write,
3016                            void __user *buffer,
3017                            size_t *length, loff_t *ppos)
3018 {
3019         proc_doulongvec_minmax(table, write, buffer, length, ppos);
3020
3021         if (write && *(unsigned long *)table->data)
3022                 scan_all_zones_unevictable_pages();
3023
3024         scan_unevictable_pages = 0;
3025         return 0;
3026 }
3027
3028 #ifdef CONFIG_NUMA
3029 /*
3030  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
3031  * a specified node's per zone unevictable lists for evictable pages.
3032  */
3033
3034 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
3035                                           struct sysdev_attribute *attr,
3036                                           char *buf)
3037 {
3038         return sprintf(buf, "0\n");     /* always zero; should fit... */
3039 }
3040
3041 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
3042                                            struct sysdev_attribute *attr,
3043                                         const char *buf, size_t count)
3044 {
3045         struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
3046         struct zone *zone;
3047         unsigned long res;
3048         unsigned long req = strict_strtoul(buf, 10, &res);
3049
3050         if (!req)
3051                 return 1;       /* zero is no-op */
3052
3053         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
3054                 if (!populated_zone(zone))
3055                         continue;
3056                 scan_zone_unevictable_pages(zone);
3057         }
3058         return 1;
3059 }
3060
3061
3062 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3063                         read_scan_unevictable_node,
3064                         write_scan_unevictable_node);
3065
3066 int scan_unevictable_register_node(struct node *node)
3067 {
3068         return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
3069 }
3070
3071 void scan_unevictable_unregister_node(struct node *node)
3072 {
3073         sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
3074 }
3075 #endif