c9177202c8ce2ce2f9bfe9045f2a40a8105de8e8
[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/compaction.h>
36 #include <linux/notifier.h>
37 #include <linux/rwsem.h>
38 #include <linux/delay.h>
39 #include <linux/kthread.h>
40 #include <linux/freezer.h>
41 #include <linux/memcontrol.h>
42 #include <linux/delayacct.h>
43 #include <linux/sysctl.h>
44 #include <linux/oom.h>
45 #include <linux/prefetch.h>
46
47 #include <asm/tlbflush.h>
48 #include <asm/div64.h>
49
50 #include <linux/swapops.h>
51
52 #include "internal.h"
53
54 #define CREATE_TRACE_POINTS
55 #include <trace/events/vmscan.h>
56
57 /*
58  * reclaim_mode determines how the inactive list is shrunk
59  * RECLAIM_MODE_SINGLE: Reclaim only order-0 pages
60  * RECLAIM_MODE_ASYNC:  Do not block
61  * RECLAIM_MODE_SYNC:   Allow blocking e.g. call wait_on_page_writeback
62  * RECLAIM_MODE_LUMPYRECLAIM: For high-order allocations, take a reference
63  *                      page from the LRU and reclaim all pages within a
64  *                      naturally aligned range
65  * RECLAIM_MODE_COMPACTION: For high-order allocations, reclaim a number of
66  *                      order-0 pages and then compact the zone
67  */
68 typedef unsigned __bitwise__ reclaim_mode_t;
69 #define RECLAIM_MODE_SINGLE             ((__force reclaim_mode_t)0x01u)
70 #define RECLAIM_MODE_ASYNC              ((__force reclaim_mode_t)0x02u)
71 #define RECLAIM_MODE_SYNC               ((__force reclaim_mode_t)0x04u)
72 #define RECLAIM_MODE_LUMPYRECLAIM       ((__force reclaim_mode_t)0x08u)
73 #define RECLAIM_MODE_COMPACTION         ((__force reclaim_mode_t)0x10u)
74
75 struct scan_control {
76         /* Incremented by the number of inactive pages that were scanned */
77         unsigned long nr_scanned;
78
79         /* Number of pages freed so far during a call to shrink_zones() */
80         unsigned long nr_reclaimed;
81
82         /* How many pages shrink_list() should reclaim */
83         unsigned long nr_to_reclaim;
84
85         unsigned long hibernation_mode;
86
87         /* This context's GFP mask */
88         gfp_t gfp_mask;
89
90         int may_writepage;
91
92         /* Can mapped pages be reclaimed? */
93         int may_unmap;
94
95         /* Can pages be swapped as part of reclaim? */
96         int may_swap;
97
98         int swappiness;
99
100         int order;
101
102         /*
103          * Intend to reclaim enough continuous memory rather than reclaim
104          * enough amount of memory. i.e, mode for high order allocation.
105          */
106         reclaim_mode_t reclaim_mode;
107
108         /* Which cgroup do we reclaim from */
109         struct mem_cgroup *mem_cgroup;
110
111         /*
112          * Nodemask of nodes allowed by the caller. If NULL, all nodes
113          * are scanned.
114          */
115         nodemask_t      *nodemask;
116 };
117
118 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
119
120 #ifdef ARCH_HAS_PREFETCH
121 #define prefetch_prev_lru_page(_page, _base, _field)                    \
122         do {                                                            \
123                 if ((_page)->lru.prev != _base) {                       \
124                         struct page *prev;                              \
125                                                                         \
126                         prev = lru_to_page(&(_page->lru));              \
127                         prefetch(&prev->_field);                        \
128                 }                                                       \
129         } while (0)
130 #else
131 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
132 #endif
133
134 #ifdef ARCH_HAS_PREFETCHW
135 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
136         do {                                                            \
137                 if ((_page)->lru.prev != _base) {                       \
138                         struct page *prev;                              \
139                                                                         \
140                         prev = lru_to_page(&(_page->lru));              \
141                         prefetchw(&prev->_field);                       \
142                 }                                                       \
143         } while (0)
144 #else
145 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
146 #endif
147
148 /*
149  * From 0 .. 100.  Higher means more swappy.
150  */
151 int vm_swappiness = 60;
152 long vm_total_pages;    /* The total number of pages which the VM controls */
153
154 static LIST_HEAD(shrinker_list);
155 static DECLARE_RWSEM(shrinker_rwsem);
156
157 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
158 #define scanning_global_lru(sc) (!(sc)->mem_cgroup)
159 #else
160 #define scanning_global_lru(sc) (1)
161 #endif
162
163 static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
164                                                   struct scan_control *sc)
165 {
166         if (!scanning_global_lru(sc))
167                 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
168
169         return &zone->reclaim_stat;
170 }
171
172 static unsigned long zone_nr_lru_pages(struct zone *zone,
173                                 struct scan_control *sc, enum lru_list lru)
174 {
175         if (!scanning_global_lru(sc))
176                 return mem_cgroup_zone_nr_pages(sc->mem_cgroup, zone, lru);
177
178         return zone_page_state(zone, NR_LRU_BASE + lru);
179 }
180
181
182 /*
183  * Add a shrinker callback to be called from the vm
184  */
185 void register_shrinker(struct shrinker *shrinker)
186 {
187         shrinker->nr = 0;
188         down_write(&shrinker_rwsem);
189         list_add_tail(&shrinker->list, &shrinker_list);
190         up_write(&shrinker_rwsem);
191 }
192 EXPORT_SYMBOL(register_shrinker);
193
194 /*
195  * Remove one
196  */
197 void unregister_shrinker(struct shrinker *shrinker)
198 {
199         down_write(&shrinker_rwsem);
200         list_del(&shrinker->list);
201         up_write(&shrinker_rwsem);
202 }
203 EXPORT_SYMBOL(unregister_shrinker);
204
205 #define SHRINK_BATCH 128
206 /*
207  * Call the shrink functions to age shrinkable caches
208  *
209  * Here we assume it costs one seek to replace a lru page and that it also
210  * takes a seek to recreate a cache object.  With this in mind we age equal
211  * percentages of the lru and ageable caches.  This should balance the seeks
212  * generated by these structures.
213  *
214  * If the vm encountered mapped pages on the LRU it increase the pressure on
215  * slab to avoid swapping.
216  *
217  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
218  *
219  * `lru_pages' represents the number of on-LRU pages in all the zones which
220  * are eligible for the caller's allocation attempt.  It is used for balancing
221  * slab reclaim versus page reclaim.
222  *
223  * Returns the number of slab objects which we shrunk.
224  */
225 unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
226                         unsigned long lru_pages)
227 {
228         struct shrinker *shrinker;
229         unsigned long ret = 0;
230
231         if (scanned == 0)
232                 scanned = SWAP_CLUSTER_MAX;
233
234         if (!down_read_trylock(&shrinker_rwsem))
235                 return 1;       /* Assume we'll be able to shrink next time */
236
237         list_for_each_entry(shrinker, &shrinker_list, list) {
238                 unsigned long long delta;
239                 unsigned long total_scan;
240                 unsigned long max_pass;
241
242                 max_pass = (*shrinker->shrink)(shrinker, 0, gfp_mask);
243                 delta = (4 * scanned) / shrinker->seeks;
244                 delta *= max_pass;
245                 do_div(delta, lru_pages + 1);
246                 shrinker->nr += delta;
247                 if (shrinker->nr < 0) {
248                         printk(KERN_ERR "shrink_slab: %pF negative objects to "
249                                "delete nr=%ld\n",
250                                shrinker->shrink, shrinker->nr);
251                         shrinker->nr = max_pass;
252                 }
253
254                 /*
255                  * Avoid risking looping forever due to too large nr value:
256                  * never try to free more than twice the estimate number of
257                  * freeable entries.
258                  */
259                 if (shrinker->nr > max_pass * 2)
260                         shrinker->nr = max_pass * 2;
261
262                 total_scan = shrinker->nr;
263                 shrinker->nr = 0;
264
265                 while (total_scan >= SHRINK_BATCH) {
266                         long this_scan = SHRINK_BATCH;
267                         int shrink_ret;
268                         int nr_before;
269
270                         nr_before = (*shrinker->shrink)(shrinker, 0, gfp_mask);
271                         shrink_ret = (*shrinker->shrink)(shrinker, this_scan,
272                                                                 gfp_mask);
273                         if (shrink_ret == -1)
274                                 break;
275                         if (shrink_ret < nr_before)
276                                 ret += nr_before - shrink_ret;
277                         count_vm_events(SLABS_SCANNED, this_scan);
278                         total_scan -= this_scan;
279
280                         cond_resched();
281                 }
282
283                 shrinker->nr += total_scan;
284         }
285         up_read(&shrinker_rwsem);
286         return ret;
287 }
288
289 static void set_reclaim_mode(int priority, struct scan_control *sc,
290                                    bool sync)
291 {
292         reclaim_mode_t syncmode = sync ? RECLAIM_MODE_SYNC : RECLAIM_MODE_ASYNC;
293
294         /*
295          * Initially assume we are entering either lumpy reclaim or
296          * reclaim/compaction.Depending on the order, we will either set the
297          * sync mode or just reclaim order-0 pages later.
298          */
299         if (COMPACTION_BUILD)
300                 sc->reclaim_mode = RECLAIM_MODE_COMPACTION;
301         else
302                 sc->reclaim_mode = RECLAIM_MODE_LUMPYRECLAIM;
303
304         /*
305          * Avoid using lumpy reclaim or reclaim/compaction if possible by
306          * restricting when its set to either costly allocations or when
307          * under memory pressure
308          */
309         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
310                 sc->reclaim_mode |= syncmode;
311         else if (sc->order && priority < DEF_PRIORITY - 2)
312                 sc->reclaim_mode |= syncmode;
313         else
314                 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
315 }
316
317 static void reset_reclaim_mode(struct scan_control *sc)
318 {
319         sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
320 }
321
322 static inline int is_page_cache_freeable(struct page *page)
323 {
324         /*
325          * A freeable page cache page is referenced only by the caller
326          * that isolated the page, the page cache radix tree and
327          * optional buffer heads at page->private.
328          */
329         return page_count(page) - page_has_private(page) == 2;
330 }
331
332 static int may_write_to_queue(struct backing_dev_info *bdi,
333                               struct scan_control *sc)
334 {
335         if (current->flags & PF_SWAPWRITE)
336                 return 1;
337         if (!bdi_write_congested(bdi))
338                 return 1;
339         if (bdi == current->backing_dev_info)
340                 return 1;
341
342         /* lumpy reclaim for hugepage often need a lot of write */
343         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
344                 return 1;
345         return 0;
346 }
347
348 /*
349  * We detected a synchronous write error writing a page out.  Probably
350  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
351  * fsync(), msync() or close().
352  *
353  * The tricky part is that after writepage we cannot touch the mapping: nothing
354  * prevents it from being freed up.  But we have a ref on the page and once
355  * that page is locked, the mapping is pinned.
356  *
357  * We're allowed to run sleeping lock_page() here because we know the caller has
358  * __GFP_FS.
359  */
360 static void handle_write_error(struct address_space *mapping,
361                                 struct page *page, int error)
362 {
363         lock_page(page);
364         if (page_mapping(page) == mapping)
365                 mapping_set_error(mapping, error);
366         unlock_page(page);
367 }
368
369 /* possible outcome of pageout() */
370 typedef enum {
371         /* failed to write page out, page is locked */
372         PAGE_KEEP,
373         /* move page to the active list, page is locked */
374         PAGE_ACTIVATE,
375         /* page has been sent to the disk successfully, page is unlocked */
376         PAGE_SUCCESS,
377         /* page is clean and locked */
378         PAGE_CLEAN,
379 } pageout_t;
380
381 /*
382  * pageout is called by shrink_page_list() for each dirty page.
383  * Calls ->writepage().
384  */
385 static pageout_t pageout(struct page *page, struct address_space *mapping,
386                          struct scan_control *sc)
387 {
388         /*
389          * If the page is dirty, only perform writeback if that write
390          * will be non-blocking.  To prevent this allocation from being
391          * stalled by pagecache activity.  But note that there may be
392          * stalls if we need to run get_block().  We could test
393          * PagePrivate for that.
394          *
395          * If this process is currently in __generic_file_aio_write() against
396          * this page's queue, we can perform writeback even if that
397          * will block.
398          *
399          * If the page is swapcache, write it back even if that would
400          * block, for some throttling. This happens by accident, because
401          * swap_backing_dev_info is bust: it doesn't reflect the
402          * congestion state of the swapdevs.  Easy to fix, if needed.
403          */
404         if (!is_page_cache_freeable(page))
405                 return PAGE_KEEP;
406         if (!mapping) {
407                 /*
408                  * Some data journaling orphaned pages can have
409                  * page->mapping == NULL while being dirty with clean buffers.
410                  */
411                 if (page_has_private(page)) {
412                         if (try_to_free_buffers(page)) {
413                                 ClearPageDirty(page);
414                                 printk("%s: orphaned page\n", __func__);
415                                 return PAGE_CLEAN;
416                         }
417                 }
418                 return PAGE_KEEP;
419         }
420         if (mapping->a_ops->writepage == NULL)
421                 return PAGE_ACTIVATE;
422         if (!may_write_to_queue(mapping->backing_dev_info, sc))
423                 return PAGE_KEEP;
424
425         if (clear_page_dirty_for_io(page)) {
426                 int res;
427                 struct writeback_control wbc = {
428                         .sync_mode = WB_SYNC_NONE,
429                         .nr_to_write = SWAP_CLUSTER_MAX,
430                         .range_start = 0,
431                         .range_end = LLONG_MAX,
432                         .for_reclaim = 1,
433                 };
434
435                 SetPageReclaim(page);
436                 res = mapping->a_ops->writepage(page, &wbc);
437                 if (res < 0)
438                         handle_write_error(mapping, page, res);
439                 if (res == AOP_WRITEPAGE_ACTIVATE) {
440                         ClearPageReclaim(page);
441                         return PAGE_ACTIVATE;
442                 }
443
444                 /*
445                  * Wait on writeback if requested to. This happens when
446                  * direct reclaiming a large contiguous area and the
447                  * first attempt to free a range of pages fails.
448                  */
449                 if (PageWriteback(page) &&
450                     (sc->reclaim_mode & RECLAIM_MODE_SYNC))
451                         wait_on_page_writeback(page);
452
453                 if (!PageWriteback(page)) {
454                         /* synchronous write or broken a_ops? */
455                         ClearPageReclaim(page);
456                 }
457                 trace_mm_vmscan_writepage(page,
458                         trace_reclaim_flags(page, sc->reclaim_mode));
459                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
460                 return PAGE_SUCCESS;
461         }
462
463         return PAGE_CLEAN;
464 }
465
466 /*
467  * Same as remove_mapping, but if the page is removed from the mapping, it
468  * gets returned with a refcount of 0.
469  */
470 static int __remove_mapping(struct address_space *mapping, struct page *page)
471 {
472         BUG_ON(!PageLocked(page));
473         BUG_ON(mapping != page_mapping(page));
474
475         spin_lock_irq(&mapping->tree_lock);
476         /*
477          * The non racy check for a busy page.
478          *
479          * Must be careful with the order of the tests. When someone has
480          * a ref to the page, it may be possible that they dirty it then
481          * drop the reference. So if PageDirty is tested before page_count
482          * here, then the following race may occur:
483          *
484          * get_user_pages(&page);
485          * [user mapping goes away]
486          * write_to(page);
487          *                              !PageDirty(page)    [good]
488          * SetPageDirty(page);
489          * put_page(page);
490          *                              !page_count(page)   [good, discard it]
491          *
492          * [oops, our write_to data is lost]
493          *
494          * Reversing the order of the tests ensures such a situation cannot
495          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
496          * load is not satisfied before that of page->_count.
497          *
498          * Note that if SetPageDirty is always performed via set_page_dirty,
499          * and thus under tree_lock, then this ordering is not required.
500          */
501         if (!page_freeze_refs(page, 2))
502                 goto cannot_free;
503         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
504         if (unlikely(PageDirty(page))) {
505                 page_unfreeze_refs(page, 2);
506                 goto cannot_free;
507         }
508
509         if (PageSwapCache(page)) {
510                 swp_entry_t swap = { .val = page_private(page) };
511                 __delete_from_swap_cache(page);
512                 spin_unlock_irq(&mapping->tree_lock);
513                 swapcache_free(swap, page);
514         } else {
515                 void (*freepage)(struct page *);
516
517                 freepage = mapping->a_ops->freepage;
518
519                 __delete_from_page_cache(page);
520                 spin_unlock_irq(&mapping->tree_lock);
521                 mem_cgroup_uncharge_cache_page(page);
522
523                 if (freepage != NULL)
524                         freepage(page);
525         }
526
527         return 1;
528
529 cannot_free:
530         spin_unlock_irq(&mapping->tree_lock);
531         return 0;
532 }
533
534 /*
535  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
536  * someone else has a ref on the page, abort and return 0.  If it was
537  * successfully detached, return 1.  Assumes the caller has a single ref on
538  * this page.
539  */
540 int remove_mapping(struct address_space *mapping, struct page *page)
541 {
542         if (__remove_mapping(mapping, page)) {
543                 /*
544                  * Unfreezing the refcount with 1 rather than 2 effectively
545                  * drops the pagecache ref for us without requiring another
546                  * atomic operation.
547                  */
548                 page_unfreeze_refs(page, 1);
549                 return 1;
550         }
551         return 0;
552 }
553
554 /**
555  * putback_lru_page - put previously isolated page onto appropriate LRU list
556  * @page: page to be put back to appropriate lru list
557  *
558  * Add previously isolated @page to appropriate LRU list.
559  * Page may still be unevictable for other reasons.
560  *
561  * lru_lock must not be held, interrupts must be enabled.
562  */
563 void putback_lru_page(struct page *page)
564 {
565         int lru;
566         int active = !!TestClearPageActive(page);
567         int was_unevictable = PageUnevictable(page);
568
569         VM_BUG_ON(PageLRU(page));
570
571 redo:
572         ClearPageUnevictable(page);
573
574         if (page_evictable(page, NULL)) {
575                 /*
576                  * For evictable pages, we can use the cache.
577                  * In event of a race, worst case is we end up with an
578                  * unevictable page on [in]active list.
579                  * We know how to handle that.
580                  */
581                 lru = active + page_lru_base_type(page);
582                 lru_cache_add_lru(page, lru);
583         } else {
584                 /*
585                  * Put unevictable pages directly on zone's unevictable
586                  * list.
587                  */
588                 lru = LRU_UNEVICTABLE;
589                 add_page_to_unevictable_list(page);
590                 /*
591                  * When racing with an mlock clearing (page is
592                  * unlocked), make sure that if the other thread does
593                  * not observe our setting of PG_lru and fails
594                  * isolation, we see PG_mlocked cleared below and move
595                  * the page back to the evictable list.
596                  *
597                  * The other side is TestClearPageMlocked().
598                  */
599                 smp_mb();
600         }
601
602         /*
603          * page's status can change while we move it among lru. If an evictable
604          * page is on unevictable list, it never be freed. To avoid that,
605          * check after we added it to the list, again.
606          */
607         if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
608                 if (!isolate_lru_page(page)) {
609                         put_page(page);
610                         goto redo;
611                 }
612                 /* This means someone else dropped this page from LRU
613                  * So, it will be freed or putback to LRU again. There is
614                  * nothing to do here.
615                  */
616         }
617
618         if (was_unevictable && lru != LRU_UNEVICTABLE)
619                 count_vm_event(UNEVICTABLE_PGRESCUED);
620         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
621                 count_vm_event(UNEVICTABLE_PGCULLED);
622
623         put_page(page);         /* drop ref from isolate */
624 }
625
626 enum page_references {
627         PAGEREF_RECLAIM,
628         PAGEREF_RECLAIM_CLEAN,
629         PAGEREF_KEEP,
630         PAGEREF_ACTIVATE,
631 };
632
633 static enum page_references page_check_references(struct page *page,
634                                                   struct scan_control *sc)
635 {
636         int referenced_ptes, referenced_page;
637         unsigned long vm_flags;
638
639         referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
640         referenced_page = TestClearPageReferenced(page);
641
642         /* Lumpy reclaim - ignore references */
643         if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
644                 return PAGEREF_RECLAIM;
645
646         /*
647          * Mlock lost the isolation race with us.  Let try_to_unmap()
648          * move the page to the unevictable list.
649          */
650         if (vm_flags & VM_LOCKED)
651                 return PAGEREF_RECLAIM;
652
653         if (referenced_ptes) {
654                 if (PageAnon(page))
655                         return PAGEREF_ACTIVATE;
656                 /*
657                  * All mapped pages start out with page table
658                  * references from the instantiating fault, so we need
659                  * to look twice if a mapped file page is used more
660                  * than once.
661                  *
662                  * Mark it and spare it for another trip around the
663                  * inactive list.  Another page table reference will
664                  * lead to its activation.
665                  *
666                  * Note: the mark is set for activated pages as well
667                  * so that recently deactivated but used pages are
668                  * quickly recovered.
669                  */
670                 SetPageReferenced(page);
671
672                 if (referenced_page)
673                         return PAGEREF_ACTIVATE;
674
675                 return PAGEREF_KEEP;
676         }
677
678         /* Reclaim if clean, defer dirty pages to writeback */
679         if (referenced_page && !PageSwapBacked(page))
680                 return PAGEREF_RECLAIM_CLEAN;
681
682         return PAGEREF_RECLAIM;
683 }
684
685 static noinline_for_stack void free_page_list(struct list_head *free_pages)
686 {
687         struct pagevec freed_pvec;
688         struct page *page, *tmp;
689
690         pagevec_init(&freed_pvec, 1);
691
692         list_for_each_entry_safe(page, tmp, free_pages, lru) {
693                 list_del(&page->lru);
694                 if (!pagevec_add(&freed_pvec, page)) {
695                         __pagevec_free(&freed_pvec);
696                         pagevec_reinit(&freed_pvec);
697                 }
698         }
699
700         pagevec_free(&freed_pvec);
701 }
702
703 /*
704  * shrink_page_list() returns the number of reclaimed pages
705  */
706 static unsigned long shrink_page_list(struct list_head *page_list,
707                                       struct zone *zone,
708                                       struct scan_control *sc)
709 {
710         LIST_HEAD(ret_pages);
711         LIST_HEAD(free_pages);
712         int pgactivate = 0;
713         unsigned long nr_dirty = 0;
714         unsigned long nr_congested = 0;
715         unsigned long nr_reclaimed = 0;
716
717         cond_resched();
718
719         while (!list_empty(page_list)) {
720                 enum page_references references;
721                 struct address_space *mapping;
722                 struct page *page;
723                 int may_enter_fs;
724
725                 cond_resched();
726
727                 page = lru_to_page(page_list);
728                 list_del(&page->lru);
729
730                 if (!trylock_page(page))
731                         goto keep;
732
733                 VM_BUG_ON(PageActive(page));
734                 VM_BUG_ON(page_zone(page) != zone);
735
736                 sc->nr_scanned++;
737
738                 if (unlikely(!page_evictable(page, NULL)))
739                         goto cull_mlocked;
740
741                 if (!sc->may_unmap && page_mapped(page))
742                         goto keep_locked;
743
744                 /* Double the slab pressure for mapped and swapcache pages */
745                 if (page_mapped(page) || PageSwapCache(page))
746                         sc->nr_scanned++;
747
748                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
749                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
750
751                 if (PageWriteback(page)) {
752                         /*
753                          * Synchronous reclaim is performed in two passes,
754                          * first an asynchronous pass over the list to
755                          * start parallel writeback, and a second synchronous
756                          * pass to wait for the IO to complete.  Wait here
757                          * for any page for which writeback has already
758                          * started.
759                          */
760                         if ((sc->reclaim_mode & RECLAIM_MODE_SYNC) &&
761                             may_enter_fs)
762                                 wait_on_page_writeback(page);
763                         else {
764                                 unlock_page(page);
765                                 goto keep_lumpy;
766                         }
767                 }
768
769                 references = page_check_references(page, sc);
770                 switch (references) {
771                 case PAGEREF_ACTIVATE:
772                         goto activate_locked;
773                 case PAGEREF_KEEP:
774                         goto keep_locked;
775                 case PAGEREF_RECLAIM:
776                 case PAGEREF_RECLAIM_CLEAN:
777                         ; /* try to reclaim the page below */
778                 }
779
780                 /*
781                  * Anonymous process memory has backing store?
782                  * Try to allocate it some swap space here.
783                  */
784                 if (PageAnon(page) && !PageSwapCache(page)) {
785                         if (!(sc->gfp_mask & __GFP_IO))
786                                 goto keep_locked;
787                         if (!add_to_swap(page))
788                                 goto activate_locked;
789                         may_enter_fs = 1;
790                 }
791
792                 mapping = page_mapping(page);
793
794                 /*
795                  * The page is mapped into the page tables of one or more
796                  * processes. Try to unmap it here.
797                  */
798                 if (page_mapped(page) && mapping) {
799                         switch (try_to_unmap(page, TTU_UNMAP)) {
800                         case SWAP_FAIL:
801                                 goto activate_locked;
802                         case SWAP_AGAIN:
803                                 goto keep_locked;
804                         case SWAP_MLOCK:
805                                 goto cull_mlocked;
806                         case SWAP_SUCCESS:
807                                 ; /* try to free the page below */
808                         }
809                 }
810
811                 if (PageDirty(page)) {
812                         nr_dirty++;
813
814                         if (references == PAGEREF_RECLAIM_CLEAN)
815                                 goto keep_locked;
816                         if (!may_enter_fs)
817                                 goto keep_locked;
818                         if (!sc->may_writepage)
819                                 goto keep_locked;
820
821                         /* Page is dirty, try to write it out here */
822                         switch (pageout(page, mapping, sc)) {
823                         case PAGE_KEEP:
824                                 nr_congested++;
825                                 goto keep_locked;
826                         case PAGE_ACTIVATE:
827                                 goto activate_locked;
828                         case PAGE_SUCCESS:
829                                 if (PageWriteback(page))
830                                         goto keep_lumpy;
831                                 if (PageDirty(page))
832                                         goto keep;
833
834                                 /*
835                                  * A synchronous write - probably a ramdisk.  Go
836                                  * ahead and try to reclaim the page.
837                                  */
838                                 if (!trylock_page(page))
839                                         goto keep;
840                                 if (PageDirty(page) || PageWriteback(page))
841                                         goto keep_locked;
842                                 mapping = page_mapping(page);
843                         case PAGE_CLEAN:
844                                 ; /* try to free the page below */
845                         }
846                 }
847
848                 /*
849                  * If the page has buffers, try to free the buffer mappings
850                  * associated with this page. If we succeed we try to free
851                  * the page as well.
852                  *
853                  * We do this even if the page is PageDirty().
854                  * try_to_release_page() does not perform I/O, but it is
855                  * possible for a page to have PageDirty set, but it is actually
856                  * clean (all its buffers are clean).  This happens if the
857                  * buffers were written out directly, with submit_bh(). ext3
858                  * will do this, as well as the blockdev mapping.
859                  * try_to_release_page() will discover that cleanness and will
860                  * drop the buffers and mark the page clean - it can be freed.
861                  *
862                  * Rarely, pages can have buffers and no ->mapping.  These are
863                  * the pages which were not successfully invalidated in
864                  * truncate_complete_page().  We try to drop those buffers here
865                  * and if that worked, and the page is no longer mapped into
866                  * process address space (page_count == 1) it can be freed.
867                  * Otherwise, leave the page on the LRU so it is swappable.
868                  */
869                 if (page_has_private(page)) {
870                         if (!try_to_release_page(page, sc->gfp_mask))
871                                 goto activate_locked;
872                         if (!mapping && page_count(page) == 1) {
873                                 unlock_page(page);
874                                 if (put_page_testzero(page))
875                                         goto free_it;
876                                 else {
877                                         /*
878                                          * rare race with speculative reference.
879                                          * the speculative reference will free
880                                          * this page shortly, so we may
881                                          * increment nr_reclaimed here (and
882                                          * leave it off the LRU).
883                                          */
884                                         nr_reclaimed++;
885                                         continue;
886                                 }
887                         }
888                 }
889
890                 if (!mapping || !__remove_mapping(mapping, page))
891                         goto keep_locked;
892
893                 /*
894                  * At this point, we have no other references and there is
895                  * no way to pick any more up (removed from LRU, removed
896                  * from pagecache). Can use non-atomic bitops now (and
897                  * we obviously don't have to worry about waking up a process
898                  * waiting on the page lock, because there are no references.
899                  */
900                 __clear_page_locked(page);
901 free_it:
902                 nr_reclaimed++;
903
904                 /*
905                  * Is there need to periodically free_page_list? It would
906                  * appear not as the counts should be low
907                  */
908                 list_add(&page->lru, &free_pages);
909                 continue;
910
911 cull_mlocked:
912                 if (PageSwapCache(page))
913                         try_to_free_swap(page);
914                 unlock_page(page);
915                 putback_lru_page(page);
916                 reset_reclaim_mode(sc);
917                 continue;
918
919 activate_locked:
920                 /* Not a candidate for swapping, so reclaim swap space. */
921                 if (PageSwapCache(page) && vm_swap_full())
922                         try_to_free_swap(page);
923                 VM_BUG_ON(PageActive(page));
924                 SetPageActive(page);
925                 pgactivate++;
926 keep_locked:
927                 unlock_page(page);
928 keep:
929                 reset_reclaim_mode(sc);
930 keep_lumpy:
931                 list_add(&page->lru, &ret_pages);
932                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
933         }
934
935         /*
936          * Tag a zone as congested if all the dirty pages encountered were
937          * backed by a congested BDI. In this case, reclaimers should just
938          * back off and wait for congestion to clear because further reclaim
939          * will encounter the same problem
940          */
941         if (nr_dirty && nr_dirty == nr_congested && scanning_global_lru(sc))
942                 zone_set_flag(zone, ZONE_CONGESTED);
943
944         free_page_list(&free_pages);
945
946         list_splice(&ret_pages, page_list);
947         count_vm_events(PGACTIVATE, pgactivate);
948         return nr_reclaimed;
949 }
950
951 /*
952  * Attempt to remove the specified page from its LRU.  Only take this page
953  * if it is of the appropriate PageActive status.  Pages which are being
954  * freed elsewhere are also ignored.
955  *
956  * page:        page to consider
957  * mode:        one of the LRU isolation modes defined above
958  *
959  * returns 0 on success, -ve errno on failure.
960  */
961 int __isolate_lru_page(struct page *page, int mode, int file)
962 {
963         int ret = -EINVAL;
964
965         /* Only take pages on the LRU. */
966         if (!PageLRU(page))
967                 return ret;
968
969         /*
970          * When checking the active state, we need to be sure we are
971          * dealing with comparible boolean values.  Take the logical not
972          * of each.
973          */
974         if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
975                 return ret;
976
977         if (mode != ISOLATE_BOTH && page_is_file_cache(page) != file)
978                 return ret;
979
980         /*
981          * When this function is being called for lumpy reclaim, we
982          * initially look into all LRU pages, active, inactive and
983          * unevictable; only give shrink_page_list evictable pages.
984          */
985         if (PageUnevictable(page))
986                 return ret;
987
988         ret = -EBUSY;
989
990         if (likely(get_page_unless_zero(page))) {
991                 /*
992                  * Be careful not to clear PageLRU until after we're
993                  * sure the page is not being freed elsewhere -- the
994                  * page release code relies on it.
995                  */
996                 ClearPageLRU(page);
997                 ret = 0;
998         }
999
1000         return ret;
1001 }
1002
1003 /*
1004  * zone->lru_lock is heavily contended.  Some of the functions that
1005  * shrink the lists perform better by taking out a batch of pages
1006  * and working on them outside the LRU lock.
1007  *
1008  * For pagecache intensive workloads, this function is the hottest
1009  * spot in the kernel (apart from copy_*_user functions).
1010  *
1011  * Appropriate locks must be held before calling this function.
1012  *
1013  * @nr_to_scan: The number of pages to look through on the list.
1014  * @src:        The LRU list to pull pages off.
1015  * @dst:        The temp list to put pages on to.
1016  * @scanned:    The number of pages that were scanned.
1017  * @order:      The caller's attempted allocation order
1018  * @mode:       One of the LRU isolation modes
1019  * @file:       True [1] if isolating file [!anon] pages
1020  *
1021  * returns how many pages were moved onto *@dst.
1022  */
1023 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1024                 struct list_head *src, struct list_head *dst,
1025                 unsigned long *scanned, int order, int mode, int file)
1026 {
1027         unsigned long nr_taken = 0;
1028         unsigned long nr_lumpy_taken = 0;
1029         unsigned long nr_lumpy_dirty = 0;
1030         unsigned long nr_lumpy_failed = 0;
1031         unsigned long scan;
1032
1033         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1034                 struct page *page;
1035                 unsigned long pfn;
1036                 unsigned long end_pfn;
1037                 unsigned long page_pfn;
1038                 int zone_id;
1039
1040                 page = lru_to_page(src);
1041                 prefetchw_prev_lru_page(page, src, flags);
1042
1043                 VM_BUG_ON(!PageLRU(page));
1044
1045                 switch (__isolate_lru_page(page, mode, file)) {
1046                 case 0:
1047                         list_move(&page->lru, dst);
1048                         mem_cgroup_del_lru(page);
1049                         nr_taken += hpage_nr_pages(page);
1050                         break;
1051
1052                 case -EBUSY:
1053                         /* else it is being freed elsewhere */
1054                         list_move(&page->lru, src);
1055                         mem_cgroup_rotate_lru_list(page, page_lru(page));
1056                         continue;
1057
1058                 default:
1059                         BUG();
1060                 }
1061
1062                 if (!order)
1063                         continue;
1064
1065                 /*
1066                  * Attempt to take all pages in the order aligned region
1067                  * surrounding the tag page.  Only take those pages of
1068                  * the same active state as that tag page.  We may safely
1069                  * round the target page pfn down to the requested order
1070                  * as the mem_map is guaranteed valid out to MAX_ORDER,
1071                  * where that page is in a different zone we will detect
1072                  * it from its zone id and abort this block scan.
1073                  */
1074                 zone_id = page_zone_id(page);
1075                 page_pfn = page_to_pfn(page);
1076                 pfn = page_pfn & ~((1 << order) - 1);
1077                 end_pfn = pfn + (1 << order);
1078                 for (; pfn < end_pfn; pfn++) {
1079                         struct page *cursor_page;
1080
1081                         /* The target page is in the block, ignore it. */
1082                         if (unlikely(pfn == page_pfn))
1083                                 continue;
1084
1085                         /* Avoid holes within the zone. */
1086                         if (unlikely(!pfn_valid_within(pfn)))
1087                                 break;
1088
1089                         cursor_page = pfn_to_page(pfn);
1090
1091                         /* Check that we have not crossed a zone boundary. */
1092                         if (unlikely(page_zone_id(cursor_page) != zone_id))
1093                                 break;
1094
1095                         /*
1096                          * If we don't have enough swap space, reclaiming of
1097                          * anon page which don't already have a swap slot is
1098                          * pointless.
1099                          */
1100                         if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
1101                             !PageSwapCache(cursor_page))
1102                                 break;
1103
1104                         if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1105                                 list_move(&cursor_page->lru, dst);
1106                                 mem_cgroup_del_lru(cursor_page);
1107                                 nr_taken += hpage_nr_pages(page);
1108                                 nr_lumpy_taken++;
1109                                 if (PageDirty(cursor_page))
1110                                         nr_lumpy_dirty++;
1111                                 scan++;
1112                         } else {
1113                                 /* the page is freed already. */
1114                                 if (!page_count(cursor_page))
1115                                         continue;
1116                                 break;
1117                         }
1118                 }
1119
1120                 /* If we break out of the loop above, lumpy reclaim failed */
1121                 if (pfn < end_pfn)
1122                         nr_lumpy_failed++;
1123         }
1124
1125         *scanned = scan;
1126
1127         trace_mm_vmscan_lru_isolate(order,
1128                         nr_to_scan, scan,
1129                         nr_taken,
1130                         nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1131                         mode);
1132         return nr_taken;
1133 }
1134
1135 static unsigned long isolate_pages_global(unsigned long nr,
1136                                         struct list_head *dst,
1137                                         unsigned long *scanned, int order,
1138                                         int mode, struct zone *z,
1139                                         int active, int file)
1140 {
1141         int lru = LRU_BASE;
1142         if (active)
1143                 lru += LRU_ACTIVE;
1144         if (file)
1145                 lru += LRU_FILE;
1146         return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1147                                                                 mode, file);
1148 }
1149
1150 /*
1151  * clear_active_flags() is a helper for shrink_active_list(), clearing
1152  * any active bits from the pages in the list.
1153  */
1154 static unsigned long clear_active_flags(struct list_head *page_list,
1155                                         unsigned int *count)
1156 {
1157         int nr_active = 0;
1158         int lru;
1159         struct page *page;
1160
1161         list_for_each_entry(page, page_list, lru) {
1162                 int numpages = hpage_nr_pages(page);
1163                 lru = page_lru_base_type(page);
1164                 if (PageActive(page)) {
1165                         lru += LRU_ACTIVE;
1166                         ClearPageActive(page);
1167                         nr_active += numpages;
1168                 }
1169                 if (count)
1170                         count[lru] += numpages;
1171         }
1172
1173         return nr_active;
1174 }
1175
1176 /**
1177  * isolate_lru_page - tries to isolate a page from its LRU list
1178  * @page: page to isolate from its LRU list
1179  *
1180  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1181  * vmstat statistic corresponding to whatever LRU list the page was on.
1182  *
1183  * Returns 0 if the page was removed from an LRU list.
1184  * Returns -EBUSY if the page was not on an LRU list.
1185  *
1186  * The returned page will have PageLRU() cleared.  If it was found on
1187  * the active list, it will have PageActive set.  If it was found on
1188  * the unevictable list, it will have the PageUnevictable bit set. That flag
1189  * may need to be cleared by the caller before letting the page go.
1190  *
1191  * The vmstat statistic corresponding to the list on which the page was
1192  * found will be decremented.
1193  *
1194  * Restrictions:
1195  * (1) Must be called with an elevated refcount on the page. This is a
1196  *     fundamentnal difference from isolate_lru_pages (which is called
1197  *     without a stable reference).
1198  * (2) the lru_lock must not be held.
1199  * (3) interrupts must be enabled.
1200  */
1201 int isolate_lru_page(struct page *page)
1202 {
1203         int ret = -EBUSY;
1204
1205         if (PageLRU(page)) {
1206                 struct zone *zone = page_zone(page);
1207
1208                 spin_lock_irq(&zone->lru_lock);
1209                 if (PageLRU(page) && get_page_unless_zero(page)) {
1210                         int lru = page_lru(page);
1211                         ret = 0;
1212                         ClearPageLRU(page);
1213
1214                         del_page_from_lru_list(zone, page, lru);
1215                 }
1216                 spin_unlock_irq(&zone->lru_lock);
1217         }
1218         return ret;
1219 }
1220
1221 /*
1222  * Are there way too many processes in the direct reclaim path already?
1223  */
1224 static int too_many_isolated(struct zone *zone, int file,
1225                 struct scan_control *sc)
1226 {
1227         unsigned long inactive, isolated;
1228
1229         if (current_is_kswapd())
1230                 return 0;
1231
1232         if (!scanning_global_lru(sc))
1233                 return 0;
1234
1235         if (file) {
1236                 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1237                 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1238         } else {
1239                 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1240                 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1241         }
1242
1243         return isolated > inactive;
1244 }
1245
1246 /*
1247  * TODO: Try merging with migrations version of putback_lru_pages
1248  */
1249 static noinline_for_stack void
1250 putback_lru_pages(struct zone *zone, struct scan_control *sc,
1251                                 unsigned long nr_anon, unsigned long nr_file,
1252                                 struct list_head *page_list)
1253 {
1254         struct page *page;
1255         struct pagevec pvec;
1256         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1257
1258         pagevec_init(&pvec, 1);
1259
1260         /*
1261          * Put back any unfreeable pages.
1262          */
1263         spin_lock(&zone->lru_lock);
1264         while (!list_empty(page_list)) {
1265                 int lru;
1266                 page = lru_to_page(page_list);
1267                 VM_BUG_ON(PageLRU(page));
1268                 list_del(&page->lru);
1269                 if (unlikely(!page_evictable(page, NULL))) {
1270                         spin_unlock_irq(&zone->lru_lock);
1271                         putback_lru_page(page);
1272                         spin_lock_irq(&zone->lru_lock);
1273                         continue;
1274                 }
1275                 SetPageLRU(page);
1276                 lru = page_lru(page);
1277                 add_page_to_lru_list(zone, page, lru);
1278                 if (is_active_lru(lru)) {
1279                         int file = is_file_lru(lru);
1280                         int numpages = hpage_nr_pages(page);
1281                         reclaim_stat->recent_rotated[file] += numpages;
1282                 }
1283                 if (!pagevec_add(&pvec, page)) {
1284                         spin_unlock_irq(&zone->lru_lock);
1285                         __pagevec_release(&pvec);
1286                         spin_lock_irq(&zone->lru_lock);
1287                 }
1288         }
1289         __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1290         __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1291
1292         spin_unlock_irq(&zone->lru_lock);
1293         pagevec_release(&pvec);
1294 }
1295
1296 static noinline_for_stack void update_isolated_counts(struct zone *zone,
1297                                         struct scan_control *sc,
1298                                         unsigned long *nr_anon,
1299                                         unsigned long *nr_file,
1300                                         struct list_head *isolated_list)
1301 {
1302         unsigned long nr_active;
1303         unsigned int count[NR_LRU_LISTS] = { 0, };
1304         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1305
1306         nr_active = clear_active_flags(isolated_list, count);
1307         __count_vm_events(PGDEACTIVATE, nr_active);
1308
1309         __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1310                               -count[LRU_ACTIVE_FILE]);
1311         __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1312                               -count[LRU_INACTIVE_FILE]);
1313         __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1314                               -count[LRU_ACTIVE_ANON]);
1315         __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1316                               -count[LRU_INACTIVE_ANON]);
1317
1318         *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1319         *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1320         __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1321         __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1322
1323         reclaim_stat->recent_scanned[0] += *nr_anon;
1324         reclaim_stat->recent_scanned[1] += *nr_file;
1325 }
1326
1327 /*
1328  * Returns true if the caller should wait to clean dirty/writeback pages.
1329  *
1330  * If we are direct reclaiming for contiguous pages and we do not reclaim
1331  * everything in the list, try again and wait for writeback IO to complete.
1332  * This will stall high-order allocations noticeably. Only do that when really
1333  * need to free the pages under high memory pressure.
1334  */
1335 static inline bool should_reclaim_stall(unsigned long nr_taken,
1336                                         unsigned long nr_freed,
1337                                         int priority,
1338                                         struct scan_control *sc)
1339 {
1340         int lumpy_stall_priority;
1341
1342         /* kswapd should not stall on sync IO */
1343         if (current_is_kswapd())
1344                 return false;
1345
1346         /* Only stall on lumpy reclaim */
1347         if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
1348                 return false;
1349
1350         /* If we have relaimed everything on the isolated list, no stall */
1351         if (nr_freed == nr_taken)
1352                 return false;
1353
1354         /*
1355          * For high-order allocations, there are two stall thresholds.
1356          * High-cost allocations stall immediately where as lower
1357          * order allocations such as stacks require the scanning
1358          * priority to be much higher before stalling.
1359          */
1360         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1361                 lumpy_stall_priority = DEF_PRIORITY;
1362         else
1363                 lumpy_stall_priority = DEF_PRIORITY / 3;
1364
1365         return priority <= lumpy_stall_priority;
1366 }
1367
1368 /*
1369  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1370  * of reclaimed pages
1371  */
1372 static noinline_for_stack unsigned long
1373 shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1374                         struct scan_control *sc, int priority, int file)
1375 {
1376         LIST_HEAD(page_list);
1377         unsigned long nr_scanned;
1378         unsigned long nr_reclaimed = 0;
1379         unsigned long nr_taken;
1380         unsigned long nr_anon;
1381         unsigned long nr_file;
1382
1383         while (unlikely(too_many_isolated(zone, file, sc))) {
1384                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1385
1386                 /* We are about to die and free our memory. Return now. */
1387                 if (fatal_signal_pending(current))
1388                         return SWAP_CLUSTER_MAX;
1389         }
1390
1391         set_reclaim_mode(priority, sc, false);
1392         lru_add_drain();
1393         spin_lock_irq(&zone->lru_lock);
1394
1395         if (scanning_global_lru(sc)) {
1396                 nr_taken = isolate_pages_global(nr_to_scan,
1397                         &page_list, &nr_scanned, sc->order,
1398                         sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM ?
1399                                         ISOLATE_BOTH : ISOLATE_INACTIVE,
1400                         zone, 0, file);
1401                 zone->pages_scanned += nr_scanned;
1402                 if (current_is_kswapd())
1403                         __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1404                                                nr_scanned);
1405                 else
1406                         __count_zone_vm_events(PGSCAN_DIRECT, zone,
1407                                                nr_scanned);
1408         } else {
1409                 nr_taken = mem_cgroup_isolate_pages(nr_to_scan,
1410                         &page_list, &nr_scanned, sc->order,
1411                         sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM ?
1412                                         ISOLATE_BOTH : ISOLATE_INACTIVE,
1413                         zone, sc->mem_cgroup,
1414                         0, file);
1415                 /*
1416                  * mem_cgroup_isolate_pages() keeps track of
1417                  * scanned pages on its own.
1418                  */
1419         }
1420
1421         if (nr_taken == 0) {
1422                 spin_unlock_irq(&zone->lru_lock);
1423                 return 0;
1424         }
1425
1426         update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
1427
1428         spin_unlock_irq(&zone->lru_lock);
1429
1430         nr_reclaimed = shrink_page_list(&page_list, zone, sc);
1431
1432         /* Check if we should syncronously wait for writeback */
1433         if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1434                 set_reclaim_mode(priority, sc, true);
1435                 nr_reclaimed += shrink_page_list(&page_list, zone, sc);
1436         }
1437
1438         local_irq_disable();
1439         if (current_is_kswapd())
1440                 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1441         __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1442
1443         putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
1444
1445         trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1446                 zone_idx(zone),
1447                 nr_scanned, nr_reclaimed,
1448                 priority,
1449                 trace_shrink_flags(file, sc->reclaim_mode));
1450         return nr_reclaimed;
1451 }
1452
1453 /*
1454  * This moves pages from the active list to the inactive list.
1455  *
1456  * We move them the other way if the page is referenced by one or more
1457  * processes, from rmap.
1458  *
1459  * If the pages are mostly unmapped, the processing is fast and it is
1460  * appropriate to hold zone->lru_lock across the whole operation.  But if
1461  * the pages are mapped, the processing is slow (page_referenced()) so we
1462  * should drop zone->lru_lock around each page.  It's impossible to balance
1463  * this, so instead we remove the pages from the LRU while processing them.
1464  * It is safe to rely on PG_active against the non-LRU pages in here because
1465  * nobody will play with that bit on a non-LRU page.
1466  *
1467  * The downside is that we have to touch page->_count against each page.
1468  * But we had to alter page->flags anyway.
1469  */
1470
1471 static void move_active_pages_to_lru(struct zone *zone,
1472                                      struct list_head *list,
1473                                      enum lru_list lru)
1474 {
1475         unsigned long pgmoved = 0;
1476         struct pagevec pvec;
1477         struct page *page;
1478
1479         pagevec_init(&pvec, 1);
1480
1481         while (!list_empty(list)) {
1482                 page = lru_to_page(list);
1483
1484                 VM_BUG_ON(PageLRU(page));
1485                 SetPageLRU(page);
1486
1487                 list_move(&page->lru, &zone->lru[lru].list);
1488                 mem_cgroup_add_lru_list(page, lru);
1489                 pgmoved += hpage_nr_pages(page);
1490
1491                 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1492                         spin_unlock_irq(&zone->lru_lock);
1493                         if (buffer_heads_over_limit)
1494                                 pagevec_strip(&pvec);
1495                         __pagevec_release(&pvec);
1496                         spin_lock_irq(&zone->lru_lock);
1497                 }
1498         }
1499         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1500         if (!is_active_lru(lru))
1501                 __count_vm_events(PGDEACTIVATE, pgmoved);
1502 }
1503
1504 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1505                         struct scan_control *sc, int priority, int file)
1506 {
1507         unsigned long nr_taken;
1508         unsigned long pgscanned;
1509         unsigned long vm_flags;
1510         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1511         LIST_HEAD(l_active);
1512         LIST_HEAD(l_inactive);
1513         struct page *page;
1514         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1515         unsigned long nr_rotated = 0;
1516
1517         lru_add_drain();
1518         spin_lock_irq(&zone->lru_lock);
1519         if (scanning_global_lru(sc)) {
1520                 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1521                                                 &pgscanned, sc->order,
1522                                                 ISOLATE_ACTIVE, zone,
1523                                                 1, file);
1524                 zone->pages_scanned += pgscanned;
1525         } else {
1526                 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1527                                                 &pgscanned, sc->order,
1528                                                 ISOLATE_ACTIVE, zone,
1529                                                 sc->mem_cgroup, 1, file);
1530                 /*
1531                  * mem_cgroup_isolate_pages() keeps track of
1532                  * scanned pages on its own.
1533                  */
1534         }
1535
1536         reclaim_stat->recent_scanned[file] += nr_taken;
1537
1538         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1539         if (file)
1540                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1541         else
1542                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1543         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1544         spin_unlock_irq(&zone->lru_lock);
1545
1546         while (!list_empty(&l_hold)) {
1547                 cond_resched();
1548                 page = lru_to_page(&l_hold);
1549                 list_del(&page->lru);
1550
1551                 if (unlikely(!page_evictable(page, NULL))) {
1552                         putback_lru_page(page);
1553                         continue;
1554                 }
1555
1556                 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1557                         nr_rotated += hpage_nr_pages(page);
1558                         /*
1559                          * Identify referenced, file-backed active pages and
1560                          * give them one more trip around the active list. So
1561                          * that executable code get better chances to stay in
1562                          * memory under moderate memory pressure.  Anon pages
1563                          * are not likely to be evicted by use-once streaming
1564                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1565                          * so we ignore them here.
1566                          */
1567                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1568                                 list_add(&page->lru, &l_active);
1569                                 continue;
1570                         }
1571                 }
1572
1573                 ClearPageActive(page);  /* we are de-activating */
1574                 list_add(&page->lru, &l_inactive);
1575         }
1576
1577         /*
1578          * Move pages back to the lru list.
1579          */
1580         spin_lock_irq(&zone->lru_lock);
1581         /*
1582          * Count referenced pages from currently used mappings as rotated,
1583          * even though only some of them are actually re-activated.  This
1584          * helps balance scan pressure between file and anonymous pages in
1585          * get_scan_ratio.
1586          */
1587         reclaim_stat->recent_rotated[file] += nr_rotated;
1588
1589         move_active_pages_to_lru(zone, &l_active,
1590                                                 LRU_ACTIVE + file * LRU_FILE);
1591         move_active_pages_to_lru(zone, &l_inactive,
1592                                                 LRU_BASE   + file * LRU_FILE);
1593         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1594         spin_unlock_irq(&zone->lru_lock);
1595 }
1596
1597 #ifdef CONFIG_SWAP
1598 static int inactive_anon_is_low_global(struct zone *zone)
1599 {
1600         unsigned long active, inactive;
1601
1602         active = zone_page_state(zone, NR_ACTIVE_ANON);
1603         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1604
1605         if (inactive * zone->inactive_ratio < active)
1606                 return 1;
1607
1608         return 0;
1609 }
1610
1611 /**
1612  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1613  * @zone: zone to check
1614  * @sc:   scan control of this context
1615  *
1616  * Returns true if the zone does not have enough inactive anon pages,
1617  * meaning some active anon pages need to be deactivated.
1618  */
1619 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1620 {
1621         int low;
1622
1623         /*
1624          * If we don't have swap space, anonymous page deactivation
1625          * is pointless.
1626          */
1627         if (!total_swap_pages)
1628                 return 0;
1629
1630         if (scanning_global_lru(sc))
1631                 low = inactive_anon_is_low_global(zone);
1632         else
1633                 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1634         return low;
1635 }
1636 #else
1637 static inline int inactive_anon_is_low(struct zone *zone,
1638                                         struct scan_control *sc)
1639 {
1640         return 0;
1641 }
1642 #endif
1643
1644 static int inactive_file_is_low_global(struct zone *zone)
1645 {
1646         unsigned long active, inactive;
1647
1648         active = zone_page_state(zone, NR_ACTIVE_FILE);
1649         inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1650
1651         return (active > inactive);
1652 }
1653
1654 /**
1655  * inactive_file_is_low - check if file pages need to be deactivated
1656  * @zone: zone to check
1657  * @sc:   scan control of this context
1658  *
1659  * When the system is doing streaming IO, memory pressure here
1660  * ensures that active file pages get deactivated, until more
1661  * than half of the file pages are on the inactive list.
1662  *
1663  * Once we get to that situation, protect the system's working
1664  * set from being evicted by disabling active file page aging.
1665  *
1666  * This uses a different ratio than the anonymous pages, because
1667  * the page cache uses a use-once replacement algorithm.
1668  */
1669 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1670 {
1671         int low;
1672
1673         if (scanning_global_lru(sc))
1674                 low = inactive_file_is_low_global(zone);
1675         else
1676                 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1677         return low;
1678 }
1679
1680 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1681                                 int file)
1682 {
1683         if (file)
1684                 return inactive_file_is_low(zone, sc);
1685         else
1686                 return inactive_anon_is_low(zone, sc);
1687 }
1688
1689 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1690         struct zone *zone, struct scan_control *sc, int priority)
1691 {
1692         int file = is_file_lru(lru);
1693
1694         if (is_active_lru(lru)) {
1695                 if (inactive_list_is_low(zone, sc, file))
1696                     shrink_active_list(nr_to_scan, zone, sc, priority, file);
1697                 return 0;
1698         }
1699
1700         return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1701 }
1702
1703 /*
1704  * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1705  * until we collected @swap_cluster_max pages to scan.
1706  */
1707 static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
1708                                        unsigned long *nr_saved_scan)
1709 {
1710         unsigned long nr;
1711
1712         *nr_saved_scan += nr_to_scan;
1713         nr = *nr_saved_scan;
1714
1715         if (nr >= SWAP_CLUSTER_MAX)
1716                 *nr_saved_scan = 0;
1717         else
1718                 nr = 0;
1719
1720         return nr;
1721 }
1722
1723 /*
1724  * Determine how aggressively the anon and file LRU lists should be
1725  * scanned.  The relative value of each set of LRU lists is determined
1726  * by looking at the fraction of the pages scanned we did rotate back
1727  * onto the active list instead of evict.
1728  *
1729  * nr[0] = anon pages to scan; nr[1] = file pages to scan
1730  */
1731 static void get_scan_count(struct zone *zone, struct scan_control *sc,
1732                                         unsigned long *nr, int priority)
1733 {
1734         unsigned long anon, file, free;
1735         unsigned long anon_prio, file_prio;
1736         unsigned long ap, fp;
1737         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1738         u64 fraction[2], denominator;
1739         enum lru_list l;
1740         int noswap = 0;
1741
1742         /* If we have no swap space, do not bother scanning anon pages. */
1743         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1744                 noswap = 1;
1745                 fraction[0] = 0;
1746                 fraction[1] = 1;
1747                 denominator = 1;
1748                 goto out;
1749         }
1750
1751         anon  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1752                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1753         file  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1754                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1755
1756         if (scanning_global_lru(sc)) {
1757                 free  = zone_page_state(zone, NR_FREE_PAGES);
1758                 /* If we have very few page cache pages,
1759                    force-scan anon pages. */
1760                 if (unlikely(file + free <= high_wmark_pages(zone))) {
1761                         fraction[0] = 1;
1762                         fraction[1] = 0;
1763                         denominator = 1;
1764                         goto out;
1765                 }
1766         }
1767
1768         /*
1769          * With swappiness at 100, anonymous and file have the same priority.
1770          * This scanning priority is essentially the inverse of IO cost.
1771          */
1772         anon_prio = sc->swappiness;
1773         file_prio = 200 - sc->swappiness;
1774
1775         /*
1776          * OK, so we have swap space and a fair amount of page cache
1777          * pages.  We use the recently rotated / recently scanned
1778          * ratios to determine how valuable each cache is.
1779          *
1780          * Because workloads change over time (and to avoid overflow)
1781          * we keep these statistics as a floating average, which ends
1782          * up weighing recent references more than old ones.
1783          *
1784          * anon in [0], file in [1]
1785          */
1786         spin_lock_irq(&zone->lru_lock);
1787         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1788                 reclaim_stat->recent_scanned[0] /= 2;
1789                 reclaim_stat->recent_rotated[0] /= 2;
1790         }
1791
1792         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1793                 reclaim_stat->recent_scanned[1] /= 2;
1794                 reclaim_stat->recent_rotated[1] /= 2;
1795         }
1796
1797         /*
1798          * The amount of pressure on anon vs file pages is inversely
1799          * proportional to the fraction of recently scanned pages on
1800          * each list that were recently referenced and in active use.
1801          */
1802         ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1803         ap /= reclaim_stat->recent_rotated[0] + 1;
1804
1805         fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1806         fp /= reclaim_stat->recent_rotated[1] + 1;
1807         spin_unlock_irq(&zone->lru_lock);
1808
1809         fraction[0] = ap;
1810         fraction[1] = fp;
1811         denominator = ap + fp + 1;
1812 out:
1813         for_each_evictable_lru(l) {
1814                 int file = is_file_lru(l);
1815                 unsigned long scan;
1816
1817                 scan = zone_nr_lru_pages(zone, sc, l);
1818                 if (priority || noswap) {
1819                         scan >>= priority;
1820                         scan = div64_u64(scan * fraction[file], denominator);
1821                 }
1822                 nr[l] = nr_scan_try_batch(scan,
1823                                           &reclaim_stat->nr_saved_scan[l]);
1824         }
1825 }
1826
1827 /*
1828  * Reclaim/compaction depends on a number of pages being freed. To avoid
1829  * disruption to the system, a small number of order-0 pages continue to be
1830  * rotated and reclaimed in the normal fashion. However, by the time we get
1831  * back to the allocator and call try_to_compact_zone(), we ensure that
1832  * there are enough free pages for it to be likely successful
1833  */
1834 static inline bool should_continue_reclaim(struct zone *zone,
1835                                         unsigned long nr_reclaimed,
1836                                         unsigned long nr_scanned,
1837                                         struct scan_control *sc)
1838 {
1839         unsigned long pages_for_compaction;
1840         unsigned long inactive_lru_pages;
1841
1842         /* If not in reclaim/compaction mode, stop */
1843         if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
1844                 return false;
1845
1846         /* Consider stopping depending on scan and reclaim activity */
1847         if (sc->gfp_mask & __GFP_REPEAT) {
1848                 /*
1849                  * For __GFP_REPEAT allocations, stop reclaiming if the
1850                  * full LRU list has been scanned and we are still failing
1851                  * to reclaim pages. This full LRU scan is potentially
1852                  * expensive but a __GFP_REPEAT caller really wants to succeed
1853                  */
1854                 if (!nr_reclaimed && !nr_scanned)
1855                         return false;
1856         } else {
1857                 /*
1858                  * For non-__GFP_REPEAT allocations which can presumably
1859                  * fail without consequence, stop if we failed to reclaim
1860                  * any pages from the last SWAP_CLUSTER_MAX number of
1861                  * pages that were scanned. This will return to the
1862                  * caller faster at the risk reclaim/compaction and
1863                  * the resulting allocation attempt fails
1864                  */
1865                 if (!nr_reclaimed)
1866                         return false;
1867         }
1868
1869         /*
1870          * If we have not reclaimed enough pages for compaction and the
1871          * inactive lists are large enough, continue reclaiming
1872          */
1873         pages_for_compaction = (2UL << sc->order);
1874         inactive_lru_pages = zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON) +
1875                                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1876         if (sc->nr_reclaimed < pages_for_compaction &&
1877                         inactive_lru_pages > pages_for_compaction)
1878                 return true;
1879
1880         /* If compaction would go ahead or the allocation would succeed, stop */
1881         switch (compaction_suitable(zone, sc->order)) {
1882         case COMPACT_PARTIAL:
1883         case COMPACT_CONTINUE:
1884                 return false;
1885         default:
1886                 return true;
1887         }
1888 }
1889
1890 /*
1891  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1892  */
1893 static void shrink_zone(int priority, struct zone *zone,
1894                                 struct scan_control *sc)
1895 {
1896         unsigned long nr[NR_LRU_LISTS];
1897         unsigned long nr_to_scan;
1898         enum lru_list l;
1899         unsigned long nr_reclaimed, nr_scanned;
1900         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1901
1902 restart:
1903         nr_reclaimed = 0;
1904         nr_scanned = sc->nr_scanned;
1905         get_scan_count(zone, sc, nr, priority);
1906
1907         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1908                                         nr[LRU_INACTIVE_FILE]) {
1909                 for_each_evictable_lru(l) {
1910                         if (nr[l]) {
1911                                 nr_to_scan = min_t(unsigned long,
1912                                                    nr[l], SWAP_CLUSTER_MAX);
1913                                 nr[l] -= nr_to_scan;
1914
1915                                 nr_reclaimed += shrink_list(l, nr_to_scan,
1916                                                             zone, sc, priority);
1917                         }
1918                 }
1919                 /*
1920                  * On large memory systems, scan >> priority can become
1921                  * really large. This is fine for the starting priority;
1922                  * we want to put equal scanning pressure on each zone.
1923                  * However, if the VM has a harder time of freeing pages,
1924                  * with multiple processes reclaiming pages, the total
1925                  * freeing target can get unreasonably large.
1926                  */
1927                 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
1928                         break;
1929         }
1930         sc->nr_reclaimed += nr_reclaimed;
1931
1932         /*
1933          * Even if we did not try to evict anon pages at all, we want to
1934          * rebalance the anon lru active/inactive ratio.
1935          */
1936         if (inactive_anon_is_low(zone, sc))
1937                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1938
1939         /* reclaim/compaction might need reclaim to continue */
1940         if (should_continue_reclaim(zone, nr_reclaimed,
1941                                         sc->nr_scanned - nr_scanned, sc))
1942                 goto restart;
1943
1944         throttle_vm_writeout(sc->gfp_mask);
1945 }
1946
1947 /*
1948  * This is the direct reclaim path, for page-allocating processes.  We only
1949  * try to reclaim pages from zones which will satisfy the caller's allocation
1950  * request.
1951  *
1952  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1953  * Because:
1954  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1955  *    allocation or
1956  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1957  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1958  *    zone defense algorithm.
1959  *
1960  * If a zone is deemed to be full of pinned pages then just give it a light
1961  * scan then give up on it.
1962  */
1963 static void shrink_zones(int priority, struct zonelist *zonelist,
1964                                         struct scan_control *sc)
1965 {
1966         struct zoneref *z;
1967         struct zone *zone;
1968
1969         for_each_zone_zonelist_nodemask(zone, z, zonelist,
1970                                         gfp_zone(sc->gfp_mask), sc->nodemask) {
1971                 if (!populated_zone(zone))
1972                         continue;
1973                 /*
1974                  * Take care memory controller reclaiming has small influence
1975                  * to global LRU.
1976                  */
1977                 if (scanning_global_lru(sc)) {
1978                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1979                                 continue;
1980                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1981                                 continue;       /* Let kswapd poll it */
1982                 }
1983
1984                 shrink_zone(priority, zone, sc);
1985         }
1986 }
1987
1988 static bool zone_reclaimable(struct zone *zone)
1989 {
1990         return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
1991 }
1992
1993 /* All zones in zonelist are unreclaimable? */
1994 static bool all_unreclaimable(struct zonelist *zonelist,
1995                 struct scan_control *sc)
1996 {
1997         struct zoneref *z;
1998         struct zone *zone;
1999
2000         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2001                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2002                 if (!populated_zone(zone))
2003                         continue;
2004                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2005                         continue;
2006                 if (!zone->all_unreclaimable)
2007                         return false;
2008         }
2009
2010         return true;
2011 }
2012
2013 /*
2014  * This is the main entry point to direct page reclaim.
2015  *
2016  * If a full scan of the inactive list fails to free enough memory then we
2017  * are "out of memory" and something needs to be killed.
2018  *
2019  * If the caller is !__GFP_FS then the probability of a failure is reasonably
2020  * high - the zone may be full of dirty or under-writeback pages, which this
2021  * caller can't do much about.  We kick the writeback threads and take explicit
2022  * naps in the hope that some of these pages can be written.  But if the
2023  * allocating task holds filesystem locks which prevent writeout this might not
2024  * work, and the allocation attempt will fail.
2025  *
2026  * returns:     0, if no pages reclaimed
2027  *              else, the number of pages reclaimed
2028  */
2029 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2030                                         struct scan_control *sc)
2031 {
2032         int priority;
2033         unsigned long total_scanned = 0;
2034         struct reclaim_state *reclaim_state = current->reclaim_state;
2035         struct zoneref *z;
2036         struct zone *zone;
2037         unsigned long writeback_threshold;
2038
2039         get_mems_allowed();
2040         delayacct_freepages_start();
2041
2042         if (scanning_global_lru(sc))
2043                 count_vm_event(ALLOCSTALL);
2044
2045         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2046                 sc->nr_scanned = 0;
2047                 if (!priority)
2048                         disable_swap_token();
2049                 shrink_zones(priority, zonelist, sc);
2050                 /*
2051                  * Don't shrink slabs when reclaiming memory from
2052                  * over limit cgroups
2053                  */
2054                 if (scanning_global_lru(sc)) {
2055                         unsigned long lru_pages = 0;
2056                         for_each_zone_zonelist(zone, z, zonelist,
2057                                         gfp_zone(sc->gfp_mask)) {
2058                                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2059                                         continue;
2060
2061                                 lru_pages += zone_reclaimable_pages(zone);
2062                         }
2063
2064                         shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
2065                         if (reclaim_state) {
2066                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2067                                 reclaim_state->reclaimed_slab = 0;
2068                         }
2069                 }
2070                 total_scanned += sc->nr_scanned;
2071                 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2072                         goto out;
2073
2074                 /*
2075                  * Try to write back as many pages as we just scanned.  This
2076                  * tends to cause slow streaming writers to write data to the
2077                  * disk smoothly, at the dirtying rate, which is nice.   But
2078                  * that's undesirable in laptop mode, where we *want* lumpy
2079                  * writeout.  So in laptop mode, write out the whole world.
2080                  */
2081                 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2082                 if (total_scanned > writeback_threshold) {
2083                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
2084                         sc->may_writepage = 1;
2085                 }
2086
2087                 /* Take a nap, wait for some writeback to complete */
2088                 if (!sc->hibernation_mode && sc->nr_scanned &&
2089                     priority < DEF_PRIORITY - 2) {
2090                         struct zone *preferred_zone;
2091
2092                         first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2093                                                 &cpuset_current_mems_allowed,
2094                                                 &preferred_zone);
2095                         wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2096                 }
2097         }
2098
2099 out:
2100         delayacct_freepages_end();
2101         put_mems_allowed();
2102
2103         if (sc->nr_reclaimed)
2104                 return sc->nr_reclaimed;
2105
2106         /*
2107          * As hibernation is going on, kswapd is freezed so that it can't mark
2108          * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2109          * check.
2110          */
2111         if (oom_killer_disabled)
2112                 return 0;
2113
2114         /* top priority shrink_zones still had more to do? don't OOM, then */
2115         if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
2116                 return 1;
2117
2118         return 0;
2119 }
2120
2121 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2122                                 gfp_t gfp_mask, nodemask_t *nodemask)
2123 {
2124         unsigned long nr_reclaimed;
2125         struct scan_control sc = {
2126                 .gfp_mask = gfp_mask,
2127                 .may_writepage = !laptop_mode,
2128                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2129                 .may_unmap = 1,
2130                 .may_swap = 1,
2131                 .swappiness = vm_swappiness,
2132                 .order = order,
2133                 .mem_cgroup = NULL,
2134                 .nodemask = nodemask,
2135         };
2136
2137         trace_mm_vmscan_direct_reclaim_begin(order,
2138                                 sc.may_writepage,
2139                                 gfp_mask);
2140
2141         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2142
2143         trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2144
2145         return nr_reclaimed;
2146 }
2147
2148 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2149
2150 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2151                                                 gfp_t gfp_mask, bool noswap,
2152                                                 unsigned int swappiness,
2153                                                 struct zone *zone)
2154 {
2155         struct scan_control sc = {
2156                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2157                 .may_writepage = !laptop_mode,
2158                 .may_unmap = 1,
2159                 .may_swap = !noswap,
2160                 .swappiness = swappiness,
2161                 .order = 0,
2162                 .mem_cgroup = mem,
2163         };
2164         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2165                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2166
2167         trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2168                                                       sc.may_writepage,
2169                                                       sc.gfp_mask);
2170
2171         /*
2172          * NOTE: Although we can get the priority field, using it
2173          * here is not a good idea, since it limits the pages we can scan.
2174          * if we don't reclaim here, the shrink_zone from balance_pgdat
2175          * will pick up pages from other mem cgroup's as well. We hack
2176          * the priority and make it zero.
2177          */
2178         shrink_zone(0, zone, &sc);
2179
2180         trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2181
2182         return sc.nr_reclaimed;
2183 }
2184
2185 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
2186                                            gfp_t gfp_mask,
2187                                            bool noswap,
2188                                            unsigned int swappiness)
2189 {
2190         struct zonelist *zonelist;
2191         unsigned long nr_reclaimed;
2192         struct scan_control sc = {
2193                 .may_writepage = !laptop_mode,
2194                 .may_unmap = 1,
2195                 .may_swap = !noswap,
2196                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2197                 .swappiness = swappiness,
2198                 .order = 0,
2199                 .mem_cgroup = mem_cont,
2200                 .nodemask = NULL, /* we don't care the placement */
2201         };
2202
2203         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2204                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2205         zonelist = NODE_DATA(numa_node_id())->node_zonelists;
2206
2207         trace_mm_vmscan_memcg_reclaim_begin(0,
2208                                             sc.may_writepage,
2209                                             sc.gfp_mask);
2210
2211         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2212
2213         trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2214
2215         return nr_reclaimed;
2216 }
2217 #endif
2218
2219 /*
2220  * pgdat_balanced is used when checking if a node is balanced for high-order
2221  * allocations. Only zones that meet watermarks and are in a zone allowed
2222  * by the callers classzone_idx are added to balanced_pages. The total of
2223  * balanced pages must be at least 25% of the zones allowed by classzone_idx
2224  * for the node to be considered balanced. Forcing all zones to be balanced
2225  * for high orders can cause excessive reclaim when there are imbalanced zones.
2226  * The choice of 25% is due to
2227  *   o a 16M DMA zone that is balanced will not balance a zone on any
2228  *     reasonable sized machine
2229  *   o On all other machines, the top zone must be at least a reasonable
2230  *     percentage of the middle zones. For example, on 32-bit x86, highmem
2231  *     would need to be at least 256M for it to be balance a whole node.
2232  *     Similarly, on x86-64 the Normal zone would need to be at least 1G
2233  *     to balance a node on its own. These seemed like reasonable ratios.
2234  */
2235 static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2236                                                 int classzone_idx)
2237 {
2238         unsigned long present_pages = 0;
2239         int i;
2240
2241         for (i = 0; i <= classzone_idx; i++)
2242                 present_pages += pgdat->node_zones[i].present_pages;
2243
2244         return balanced_pages > (present_pages >> 2);
2245 }
2246
2247 /* is kswapd sleeping prematurely? */
2248 static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2249                                         int classzone_idx)
2250 {
2251         int i;
2252         unsigned long balanced = 0;
2253         bool all_zones_ok = true;
2254
2255         /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2256         if (remaining)
2257                 return true;
2258
2259         /* Check the watermark levels */
2260         for (i = 0; i < pgdat->nr_zones; i++) {
2261                 struct zone *zone = pgdat->node_zones + i;
2262
2263                 if (!populated_zone(zone))
2264                         continue;
2265
2266                 /*
2267                  * balance_pgdat() skips over all_unreclaimable after
2268                  * DEF_PRIORITY. Effectively, it considers them balanced so
2269                  * they must be considered balanced here as well if kswapd
2270                  * is to sleep
2271                  */
2272                 if (zone->all_unreclaimable) {
2273                         balanced += zone->present_pages;
2274                         continue;
2275                 }
2276
2277                 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2278                                                         classzone_idx, 0))
2279                         all_zones_ok = false;
2280                 else
2281                         balanced += zone->present_pages;
2282         }
2283
2284         /*
2285          * For high-order requests, the balanced zones must contain at least
2286          * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2287          * must be balanced
2288          */
2289         if (order)
2290                 return pgdat_balanced(pgdat, balanced, classzone_idx);
2291         else
2292                 return !all_zones_ok;
2293 }
2294
2295 /*
2296  * For kswapd, balance_pgdat() will work across all this node's zones until
2297  * they are all at high_wmark_pages(zone).
2298  *
2299  * Returns the final order kswapd was reclaiming at
2300  *
2301  * There is special handling here for zones which are full of pinned pages.
2302  * This can happen if the pages are all mlocked, or if they are all used by
2303  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
2304  * What we do is to detect the case where all pages in the zone have been
2305  * scanned twice and there has been zero successful reclaim.  Mark the zone as
2306  * dead and from now on, only perform a short scan.  Basically we're polling
2307  * the zone for when the problem goes away.
2308  *
2309  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
2310  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2311  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2312  * lower zones regardless of the number of free pages in the lower zones. This
2313  * interoperates with the page allocator fallback scheme to ensure that aging
2314  * of pages is balanced across the zones.
2315  */
2316 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2317                                                         int *classzone_idx)
2318 {
2319         int all_zones_ok;
2320         unsigned long balanced;
2321         int priority;
2322         int i;
2323         int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
2324         unsigned long total_scanned;
2325         struct reclaim_state *reclaim_state = current->reclaim_state;
2326         struct scan_control sc = {
2327                 .gfp_mask = GFP_KERNEL,
2328                 .may_unmap = 1,
2329                 .may_swap = 1,
2330                 /*
2331                  * kswapd doesn't want to be bailed out while reclaim. because
2332                  * we want to put equal scanning pressure on each zone.
2333                  */
2334                 .nr_to_reclaim = ULONG_MAX,
2335                 .swappiness = vm_swappiness,
2336                 .order = order,
2337                 .mem_cgroup = NULL,
2338         };
2339 loop_again:
2340         total_scanned = 0;
2341         sc.nr_reclaimed = 0;
2342         sc.may_writepage = !laptop_mode;
2343         count_vm_event(PAGEOUTRUN);
2344
2345         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2346                 unsigned long lru_pages = 0;
2347                 int has_under_min_watermark_zone = 0;
2348
2349                 /* The swap token gets in the way of swapout... */
2350                 if (!priority)
2351                         disable_swap_token();
2352
2353                 all_zones_ok = 1;
2354                 balanced = 0;
2355
2356                 /*
2357                  * Scan in the highmem->dma direction for the highest
2358                  * zone which needs scanning
2359                  */
2360                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2361                         struct zone *zone = pgdat->node_zones + i;
2362
2363                         if (!populated_zone(zone))
2364                                 continue;
2365
2366                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2367                                 continue;
2368
2369                         /*
2370                          * Do some background aging of the anon list, to give
2371                          * pages a chance to be referenced before reclaiming.
2372                          */
2373                         if (inactive_anon_is_low(zone, &sc))
2374                                 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2375                                                         &sc, priority, 0);
2376
2377                         if (!zone_watermark_ok_safe(zone, order,
2378                                         high_wmark_pages(zone), 0, 0)) {
2379                                 end_zone = i;
2380                                 *classzone_idx = i;
2381                                 break;
2382                         }
2383                 }
2384                 if (i < 0)
2385                         goto out;
2386
2387                 for (i = 0; i <= end_zone; i++) {
2388                         struct zone *zone = pgdat->node_zones + i;
2389
2390                         lru_pages += zone_reclaimable_pages(zone);
2391                 }
2392
2393                 /*
2394                  * Now scan the zone in the dma->highmem direction, stopping
2395                  * at the last zone which needs scanning.
2396                  *
2397                  * We do this because the page allocator works in the opposite
2398                  * direction.  This prevents the page allocator from allocating
2399                  * pages behind kswapd's direction of progress, which would
2400                  * cause too much scanning of the lower zones.
2401                  */
2402                 for (i = 0; i <= end_zone; i++) {
2403                         struct zone *zone = pgdat->node_zones + i;
2404                         int nr_slab;
2405                         unsigned long balance_gap;
2406
2407                         if (!populated_zone(zone))
2408                                 continue;
2409
2410                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2411                                 continue;
2412
2413                         sc.nr_scanned = 0;
2414
2415                         /*
2416                          * Call soft limit reclaim before calling shrink_zone.
2417                          * For now we ignore the return value
2418                          */
2419                         mem_cgroup_soft_limit_reclaim(zone, order, sc.gfp_mask);
2420
2421                         /*
2422                          * We put equal pressure on every zone, unless
2423                          * one zone has way too many pages free
2424                          * already. The "too many pages" is defined
2425                          * as the high wmark plus a "gap" where the
2426                          * gap is either the low watermark or 1%
2427                          * of the zone, whichever is smaller.
2428                          */
2429                         balance_gap = min(low_wmark_pages(zone),
2430                                 (zone->present_pages +
2431                                         KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2432                                 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2433                         if (!zone_watermark_ok_safe(zone, order,
2434                                         high_wmark_pages(zone) + balance_gap,
2435                                         end_zone, 0))
2436                                 shrink_zone(priority, zone, &sc);
2437                         reclaim_state->reclaimed_slab = 0;
2438                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
2439                                                 lru_pages);
2440                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2441                         total_scanned += sc.nr_scanned;
2442
2443                         if (zone->all_unreclaimable)
2444                                 continue;
2445                         if (nr_slab == 0 &&
2446                             !zone_reclaimable(zone))
2447                                 zone->all_unreclaimable = 1;
2448                         /*
2449                          * If we've done a decent amount of scanning and
2450                          * the reclaim ratio is low, start doing writepage
2451                          * even in laptop mode
2452                          */
2453                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2454                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2455                                 sc.may_writepage = 1;
2456
2457                         if (!zone_watermark_ok_safe(zone, order,
2458                                         high_wmark_pages(zone), end_zone, 0)) {
2459                                 all_zones_ok = 0;
2460                                 /*
2461                                  * We are still under min water mark.  This
2462                                  * means that we have a GFP_ATOMIC allocation
2463                                  * failure risk. Hurry up!
2464                                  */
2465                                 if (!zone_watermark_ok_safe(zone, order,
2466                                             min_wmark_pages(zone), end_zone, 0))
2467                                         has_under_min_watermark_zone = 1;
2468                         } else {
2469                                 /*
2470                                  * If a zone reaches its high watermark,
2471                                  * consider it to be no longer congested. It's
2472                                  * possible there are dirty pages backed by
2473                                  * congested BDIs but as pressure is relieved,
2474                                  * spectulatively avoid congestion waits
2475                                  */
2476                                 zone_clear_flag(zone, ZONE_CONGESTED);
2477                                 if (i <= *classzone_idx)
2478                                         balanced += zone->present_pages;
2479                         }
2480
2481                 }
2482                 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2483                         break;          /* kswapd: all done */
2484                 /*
2485                  * OK, kswapd is getting into trouble.  Take a nap, then take
2486                  * another pass across the zones.
2487                  */
2488                 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2489                         if (has_under_min_watermark_zone)
2490                                 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2491                         else
2492                                 congestion_wait(BLK_RW_ASYNC, HZ/10);
2493                 }
2494
2495                 /*
2496                  * We do this so kswapd doesn't build up large priorities for
2497                  * example when it is freeing in parallel with allocators. It
2498                  * matches the direct reclaim path behaviour in terms of impact
2499                  * on zone->*_priority.
2500                  */
2501                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2502                         break;
2503         }
2504 out:
2505
2506         /*
2507          * order-0: All zones must meet high watermark for a balanced node
2508          * high-order: Balanced zones must make up at least 25% of the node
2509          *             for the node to be balanced
2510          */
2511         if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2512                 cond_resched();
2513
2514                 try_to_freeze();
2515
2516                 /*
2517                  * Fragmentation may mean that the system cannot be
2518                  * rebalanced for high-order allocations in all zones.
2519                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2520                  * it means the zones have been fully scanned and are still
2521                  * not balanced. For high-order allocations, there is
2522                  * little point trying all over again as kswapd may
2523                  * infinite loop.
2524                  *
2525                  * Instead, recheck all watermarks at order-0 as they
2526                  * are the most important. If watermarks are ok, kswapd will go
2527                  * back to sleep. High-order users can still perform direct
2528                  * reclaim if they wish.
2529                  */
2530                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2531                         order = sc.order = 0;
2532
2533                 goto loop_again;
2534         }
2535
2536         /*
2537          * If kswapd was reclaiming at a higher order, it has the option of
2538          * sleeping without all zones being balanced. Before it does, it must
2539          * ensure that the watermarks for order-0 on *all* zones are met and
2540          * that the congestion flags are cleared. The congestion flag must
2541          * be cleared as kswapd is the only mechanism that clears the flag
2542          * and it is potentially going to sleep here.
2543          */
2544         if (order) {
2545                 for (i = 0; i <= end_zone; i++) {
2546                         struct zone *zone = pgdat->node_zones + i;
2547
2548                         if (!populated_zone(zone))
2549                                 continue;
2550
2551                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2552                                 continue;
2553
2554                         /* Confirm the zone is balanced for order-0 */
2555                         if (!zone_watermark_ok(zone, 0,
2556                                         high_wmark_pages(zone), 0, 0)) {
2557                                 order = sc.order = 0;
2558                                 goto loop_again;
2559                         }
2560
2561                         /* If balanced, clear the congested flag */
2562                         zone_clear_flag(zone, ZONE_CONGESTED);
2563                 }
2564         }
2565
2566         /*
2567          * Return the order we were reclaiming at so sleeping_prematurely()
2568          * makes a decision on the order we were last reclaiming at. However,
2569          * if another caller entered the allocator slow path while kswapd
2570          * was awake, order will remain at the higher level
2571          */
2572         *classzone_idx = end_zone;
2573         return order;
2574 }
2575
2576 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2577 {
2578         long remaining = 0;
2579         DEFINE_WAIT(wait);
2580
2581         if (freezing(current) || kthread_should_stop())
2582                 return;
2583
2584         prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2585
2586         /* Try to sleep for a short interval */
2587         if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2588                 remaining = schedule_timeout(HZ/10);
2589                 finish_wait(&pgdat->kswapd_wait, &wait);
2590                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2591         }
2592
2593         /*
2594          * After a short sleep, check if it was a premature sleep. If not, then
2595          * go fully to sleep until explicitly woken up.
2596          */
2597         if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2598                 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2599
2600                 /*
2601                  * vmstat counters are not perfectly accurate and the estimated
2602                  * value for counters such as NR_FREE_PAGES can deviate from the
2603                  * true value by nr_online_cpus * threshold. To avoid the zone
2604                  * watermarks being breached while under pressure, we reduce the
2605                  * per-cpu vmstat threshold while kswapd is awake and restore
2606                  * them before going back to sleep.
2607                  */
2608                 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2609                 schedule();
2610                 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2611         } else {
2612                 if (remaining)
2613                         count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2614                 else
2615                         count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2616         }
2617         finish_wait(&pgdat->kswapd_wait, &wait);
2618 }
2619
2620 /*
2621  * The background pageout daemon, started as a kernel thread
2622  * from the init process.
2623  *
2624  * This basically trickles out pages so that we have _some_
2625  * free memory available even if there is no other activity
2626  * that frees anything up. This is needed for things like routing
2627  * etc, where we otherwise might have all activity going on in
2628  * asynchronous contexts that cannot page things out.
2629  *
2630  * If there are applications that are active memory-allocators
2631  * (most normal use), this basically shouldn't matter.
2632  */
2633 static int kswapd(void *p)
2634 {
2635         unsigned long order;
2636         int classzone_idx;
2637         pg_data_t *pgdat = (pg_data_t*)p;
2638         struct task_struct *tsk = current;
2639
2640         struct reclaim_state reclaim_state = {
2641                 .reclaimed_slab = 0,
2642         };
2643         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2644
2645         lockdep_set_current_reclaim_state(GFP_KERNEL);
2646
2647         if (!cpumask_empty(cpumask))
2648                 set_cpus_allowed_ptr(tsk, cpumask);
2649         current->reclaim_state = &reclaim_state;
2650
2651         /*
2652          * Tell the memory management that we're a "memory allocator",
2653          * and that if we need more memory we should get access to it
2654          * regardless (see "__alloc_pages()"). "kswapd" should
2655          * never get caught in the normal page freeing logic.
2656          *
2657          * (Kswapd normally doesn't need memory anyway, but sometimes
2658          * you need a small amount of memory in order to be able to
2659          * page out something else, and this flag essentially protects
2660          * us from recursively trying to free more memory as we're
2661          * trying to free the first piece of memory in the first place).
2662          */
2663         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2664         set_freezable();
2665
2666         order = 0;
2667         classzone_idx = MAX_NR_ZONES - 1;
2668         for ( ; ; ) {
2669                 unsigned long new_order;
2670                 int new_classzone_idx;
2671                 int ret;
2672
2673                 new_order = pgdat->kswapd_max_order;
2674                 new_classzone_idx = pgdat->classzone_idx;
2675                 pgdat->kswapd_max_order = 0;
2676                 pgdat->classzone_idx = MAX_NR_ZONES - 1;
2677                 if (order < new_order || classzone_idx > new_classzone_idx) {
2678                         /*
2679                          * Don't sleep if someone wants a larger 'order'
2680                          * allocation or has tigher zone constraints
2681                          */
2682                         order = new_order;
2683                         classzone_idx = new_classzone_idx;
2684                 } else {
2685                         kswapd_try_to_sleep(pgdat, order, classzone_idx);
2686                         order = pgdat->kswapd_max_order;
2687                         classzone_idx = pgdat->classzone_idx;
2688                         pgdat->kswapd_max_order = 0;
2689                         pgdat->classzone_idx = MAX_NR_ZONES - 1;
2690                 }
2691
2692                 ret = try_to_freeze();
2693                 if (kthread_should_stop())
2694                         break;
2695
2696                 /*
2697                  * We can speed up thawing tasks if we don't call balance_pgdat
2698                  * after returning from the refrigerator
2699                  */
2700                 if (!ret) {
2701                         trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2702                         order = balance_pgdat(pgdat, order, &classzone_idx);
2703                 }
2704         }
2705         return 0;
2706 }
2707
2708 /*
2709  * A zone is low on free memory, so wake its kswapd task to service it.
2710  */
2711 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
2712 {
2713         pg_data_t *pgdat;
2714
2715         if (!populated_zone(zone))
2716                 return;
2717
2718         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2719                 return;
2720         pgdat = zone->zone_pgdat;
2721         if (pgdat->kswapd_max_order < order) {
2722                 pgdat->kswapd_max_order = order;
2723                 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
2724         }
2725         if (!waitqueue_active(&pgdat->kswapd_wait))
2726                 return;
2727         if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
2728                 return;
2729
2730         trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
2731         wake_up_interruptible(&pgdat->kswapd_wait);
2732 }
2733
2734 /*
2735  * The reclaimable count would be mostly accurate.
2736  * The less reclaimable pages may be
2737  * - mlocked pages, which will be moved to unevictable list when encountered
2738  * - mapped pages, which may require several travels to be reclaimed
2739  * - dirty pages, which is not "instantly" reclaimable
2740  */
2741 unsigned long global_reclaimable_pages(void)
2742 {
2743         int nr;
2744
2745         nr = global_page_state(NR_ACTIVE_FILE) +
2746              global_page_state(NR_INACTIVE_FILE);
2747
2748         if (nr_swap_pages > 0)
2749                 nr += global_page_state(NR_ACTIVE_ANON) +
2750                       global_page_state(NR_INACTIVE_ANON);
2751
2752         return nr;
2753 }
2754
2755 unsigned long zone_reclaimable_pages(struct zone *zone)
2756 {
2757         int nr;
2758
2759         nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2760              zone_page_state(zone, NR_INACTIVE_FILE);
2761
2762         if (nr_swap_pages > 0)
2763                 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2764                       zone_page_state(zone, NR_INACTIVE_ANON);
2765
2766         return nr;
2767 }
2768
2769 #ifdef CONFIG_HIBERNATION
2770 /*
2771  * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2772  * freed pages.
2773  *
2774  * Rather than trying to age LRUs the aim is to preserve the overall
2775  * LRU order by reclaiming preferentially
2776  * inactive > active > active referenced > active mapped
2777  */
2778 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
2779 {
2780         struct reclaim_state reclaim_state;
2781         struct scan_control sc = {
2782                 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2783                 .may_swap = 1,
2784                 .may_unmap = 1,
2785                 .may_writepage = 1,
2786                 .nr_to_reclaim = nr_to_reclaim,
2787                 .hibernation_mode = 1,
2788                 .swappiness = vm_swappiness,
2789                 .order = 0,
2790         };
2791         struct zonelist * zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2792         struct task_struct *p = current;
2793         unsigned long nr_reclaimed;
2794
2795         p->flags |= PF_MEMALLOC;
2796         lockdep_set_current_reclaim_state(sc.gfp_mask);
2797         reclaim_state.reclaimed_slab = 0;
2798         p->reclaim_state = &reclaim_state;
2799
2800         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2801
2802         p->reclaim_state = NULL;
2803         lockdep_clear_current_reclaim_state();
2804         p->flags &= ~PF_MEMALLOC;
2805
2806         return nr_reclaimed;
2807 }
2808 #endif /* CONFIG_HIBERNATION */
2809
2810 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2811    not required for correctness.  So if the last cpu in a node goes
2812    away, we get changed to run anywhere: as the first one comes back,
2813    restore their cpu bindings. */
2814 static int __devinit cpu_callback(struct notifier_block *nfb,
2815                                   unsigned long action, void *hcpu)
2816 {
2817         int nid;
2818
2819         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2820                 for_each_node_state(nid, N_HIGH_MEMORY) {
2821                         pg_data_t *pgdat = NODE_DATA(nid);
2822                         const struct cpumask *mask;
2823
2824                         mask = cpumask_of_node(pgdat->node_id);
2825
2826                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2827                                 /* One of our CPUs online: restore mask */
2828                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2829                 }
2830         }
2831         return NOTIFY_OK;
2832 }
2833
2834 /*
2835  * This kswapd start function will be called by init and node-hot-add.
2836  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2837  */
2838 int kswapd_run(int nid)
2839 {
2840         pg_data_t *pgdat = NODE_DATA(nid);
2841         int ret = 0;
2842
2843         if (pgdat->kswapd)
2844                 return 0;
2845
2846         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2847         if (IS_ERR(pgdat->kswapd)) {
2848                 /* failure at boot is fatal */
2849                 BUG_ON(system_state == SYSTEM_BOOTING);
2850                 printk("Failed to start kswapd on node %d\n",nid);
2851                 ret = -1;
2852         }
2853         return ret;
2854 }
2855
2856 /*
2857  * Called by memory hotplug when all memory in a node is offlined.
2858  */
2859 void kswapd_stop(int nid)
2860 {
2861         struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
2862
2863         if (kswapd)
2864                 kthread_stop(kswapd);
2865 }
2866
2867 static int __init kswapd_init(void)
2868 {
2869         int nid;
2870
2871         swap_setup();
2872         for_each_node_state(nid, N_HIGH_MEMORY)
2873                 kswapd_run(nid);
2874         hotcpu_notifier(cpu_callback, 0);
2875         return 0;
2876 }
2877
2878 module_init(kswapd_init)
2879
2880 #ifdef CONFIG_NUMA
2881 /*
2882  * Zone reclaim mode
2883  *
2884  * If non-zero call zone_reclaim when the number of free pages falls below
2885  * the watermarks.
2886  */
2887 int zone_reclaim_mode __read_mostly;
2888
2889 #define RECLAIM_OFF 0
2890 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
2891 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
2892 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
2893
2894 /*
2895  * Priority for ZONE_RECLAIM. This determines the fraction of pages
2896  * of a node considered for each zone_reclaim. 4 scans 1/16th of
2897  * a zone.
2898  */
2899 #define ZONE_RECLAIM_PRIORITY 4
2900
2901 /*
2902  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2903  * occur.
2904  */
2905 int sysctl_min_unmapped_ratio = 1;
2906
2907 /*
2908  * If the number of slab pages in a zone grows beyond this percentage then
2909  * slab reclaim needs to occur.
2910  */
2911 int sysctl_min_slab_ratio = 5;
2912
2913 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
2914 {
2915         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
2916         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
2917                 zone_page_state(zone, NR_ACTIVE_FILE);
2918
2919         /*
2920          * It's possible for there to be more file mapped pages than
2921          * accounted for by the pages on the file LRU lists because
2922          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2923          */
2924         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
2925 }
2926
2927 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
2928 static long zone_pagecache_reclaimable(struct zone *zone)
2929 {
2930         long nr_pagecache_reclaimable;
2931         long delta = 0;
2932
2933         /*
2934          * If RECLAIM_SWAP is set, then all file pages are considered
2935          * potentially reclaimable. Otherwise, we have to worry about
2936          * pages like swapcache and zone_unmapped_file_pages() provides
2937          * a better estimate
2938          */
2939         if (zone_reclaim_mode & RECLAIM_SWAP)
2940                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
2941         else
2942                 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
2943
2944         /* If we can't clean pages, remove dirty pages from consideration */
2945         if (!(zone_reclaim_mode & RECLAIM_WRITE))
2946                 delta += zone_page_state(zone, NR_FILE_DIRTY);
2947
2948         /* Watch for any possible underflows due to delta */
2949         if (unlikely(delta > nr_pagecache_reclaimable))
2950                 delta = nr_pagecache_reclaimable;
2951
2952         return nr_pagecache_reclaimable - delta;
2953 }
2954
2955 /*
2956  * Try to free up some pages from this zone through reclaim.
2957  */
2958 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2959 {
2960         /* Minimum pages needed in order to stay on node */
2961         const unsigned long nr_pages = 1 << order;
2962         struct task_struct *p = current;
2963         struct reclaim_state reclaim_state;
2964         int priority;
2965         struct scan_control sc = {
2966                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2967                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2968                 .may_swap = 1,
2969                 .nr_to_reclaim = max_t(unsigned long, nr_pages,
2970                                        SWAP_CLUSTER_MAX),
2971                 .gfp_mask = gfp_mask,
2972                 .swappiness = vm_swappiness,
2973                 .order = order,
2974         };
2975         unsigned long nr_slab_pages0, nr_slab_pages1;
2976
2977         cond_resched();
2978         /*
2979          * We need to be able to allocate from the reserves for RECLAIM_SWAP
2980          * and we also need to be able to write out pages for RECLAIM_WRITE
2981          * and RECLAIM_SWAP.
2982          */
2983         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2984         lockdep_set_current_reclaim_state(gfp_mask);
2985         reclaim_state.reclaimed_slab = 0;
2986         p->reclaim_state = &reclaim_state;
2987
2988         if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
2989                 /*
2990                  * Free memory by calling shrink zone with increasing
2991                  * priorities until we have enough memory freed.
2992                  */
2993                 priority = ZONE_RECLAIM_PRIORITY;
2994                 do {
2995                         shrink_zone(priority, zone, &sc);
2996                         priority--;
2997                 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
2998         }
2999
3000         nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3001         if (nr_slab_pages0 > zone->min_slab_pages) {
3002                 /*
3003                  * shrink_slab() does not currently allow us to determine how
3004                  * many pages were freed in this zone. So we take the current
3005                  * number of slab pages and shake the slab until it is reduced
3006                  * by the same nr_pages that we used for reclaiming unmapped
3007                  * pages.
3008                  *
3009                  * Note that shrink_slab will free memory on all zones and may
3010                  * take a long time.
3011                  */
3012                 for (;;) {
3013                         unsigned long lru_pages = zone_reclaimable_pages(zone);
3014
3015                         /* No reclaimable slab or very low memory pressure */
3016                         if (!shrink_slab(sc.nr_scanned, gfp_mask, lru_pages))
3017                                 break;
3018
3019                         /* Freed enough memory */
3020                         nr_slab_pages1 = zone_page_state(zone,
3021                                                         NR_SLAB_RECLAIMABLE);
3022                         if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3023                                 break;
3024                 }
3025
3026                 /*
3027                  * Update nr_reclaimed by the number of slab pages we
3028                  * reclaimed from this zone.
3029                  */
3030                 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3031                 if (nr_slab_pages1 < nr_slab_pages0)
3032                         sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3033         }
3034
3035         p->reclaim_state = NULL;
3036         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3037         lockdep_clear_current_reclaim_state();
3038         return sc.nr_reclaimed >= nr_pages;
3039 }
3040
3041 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3042 {
3043         int node_id;
3044         int ret;
3045
3046         /*
3047          * Zone reclaim reclaims unmapped file backed pages and
3048          * slab pages if we are over the defined limits.
3049          *
3050          * A small portion of unmapped file backed pages is needed for
3051          * file I/O otherwise pages read by file I/O will be immediately
3052          * thrown out if the zone is overallocated. So we do not reclaim
3053          * if less than a specified percentage of the zone is used by
3054          * unmapped file backed pages.
3055          */
3056         if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3057             zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3058                 return ZONE_RECLAIM_FULL;
3059
3060         if (zone->all_unreclaimable)
3061                 return ZONE_RECLAIM_FULL;
3062
3063         /*
3064          * Do not scan if the allocation should not be delayed.
3065          */
3066         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3067                 return ZONE_RECLAIM_NOSCAN;
3068
3069         /*
3070          * Only run zone reclaim on the local zone or on zones that do not
3071          * have associated processors. This will favor the local processor
3072          * over remote processors and spread off node memory allocations
3073          * as wide as possible.
3074          */
3075         node_id = zone_to_nid(zone);
3076         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3077                 return ZONE_RECLAIM_NOSCAN;
3078
3079         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3080                 return ZONE_RECLAIM_NOSCAN;
3081
3082         ret = __zone_reclaim(zone, gfp_mask, order);
3083         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3084
3085         if (!ret)
3086                 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3087
3088         return ret;
3089 }
3090 #endif
3091
3092 /*
3093  * page_evictable - test whether a page is evictable
3094  * @page: the page to test
3095  * @vma: the VMA in which the page is or will be mapped, may be NULL
3096  *
3097  * Test whether page is evictable--i.e., should be placed on active/inactive
3098  * lists vs unevictable list.  The vma argument is !NULL when called from the
3099  * fault path to determine how to instantate a new page.
3100  *
3101  * Reasons page might not be evictable:
3102  * (1) page's mapping marked unevictable
3103  * (2) page is part of an mlocked VMA
3104  *
3105  */
3106 int page_evictable(struct page *page, struct vm_area_struct *vma)
3107 {
3108
3109         if (mapping_unevictable(page_mapping(page)))
3110                 return 0;
3111
3112         if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
3113                 return 0;
3114
3115         return 1;
3116 }
3117
3118 /**
3119  * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
3120  * @page: page to check evictability and move to appropriate lru list
3121  * @zone: zone page is in
3122  *
3123  * Checks a page for evictability and moves the page to the appropriate
3124  * zone lru list.
3125  *
3126  * Restrictions: zone->lru_lock must be held, page must be on LRU and must
3127  * have PageUnevictable set.
3128  */
3129 static void check_move_unevictable_page(struct page *page, struct zone *zone)
3130 {
3131         VM_BUG_ON(PageActive(page));
3132
3133 retry:
3134         ClearPageUnevictable(page);
3135         if (page_evictable(page, NULL)) {
3136                 enum lru_list l = page_lru_base_type(page);
3137
3138                 __dec_zone_state(zone, NR_UNEVICTABLE);
3139                 list_move(&page->lru, &zone->lru[l].list);
3140                 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
3141                 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
3142                 __count_vm_event(UNEVICTABLE_PGRESCUED);
3143         } else {
3144                 /*
3145                  * rotate unevictable list
3146                  */
3147                 SetPageUnevictable(page);
3148                 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
3149                 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
3150                 if (page_evictable(page, NULL))
3151                         goto retry;
3152         }
3153 }
3154
3155 /**
3156  * scan_mapping_unevictable_pages - scan an address space for evictable pages
3157  * @mapping: struct address_space to scan for evictable pages
3158  *
3159  * Scan all pages in mapping.  Check unevictable pages for
3160  * evictability and move them to the appropriate zone lru list.
3161  */
3162 void scan_mapping_unevictable_pages(struct address_space *mapping)
3163 {
3164         pgoff_t next = 0;
3165         pgoff_t end   = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
3166                          PAGE_CACHE_SHIFT;
3167         struct zone *zone;
3168         struct pagevec pvec;
3169
3170         if (mapping->nrpages == 0)
3171                 return;
3172
3173         pagevec_init(&pvec, 0);
3174         while (next < end &&
3175                 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
3176                 int i;
3177                 int pg_scanned = 0;
3178
3179                 zone = NULL;
3180
3181                 for (i = 0; i < pagevec_count(&pvec); i++) {
3182                         struct page *page = pvec.pages[i];
3183                         pgoff_t page_index = page->index;
3184                         struct zone *pagezone = page_zone(page);
3185
3186                         pg_scanned++;
3187                         if (page_index > next)
3188                                 next = page_index;
3189                         next++;
3190
3191                         if (pagezone != zone) {
3192                                 if (zone)
3193                                         spin_unlock_irq(&zone->lru_lock);
3194                                 zone = pagezone;
3195                                 spin_lock_irq(&zone->lru_lock);
3196                         }
3197
3198                         if (PageLRU(page) && PageUnevictable(page))
3199                                 check_move_unevictable_page(page, zone);
3200                 }
3201                 if (zone)
3202                         spin_unlock_irq(&zone->lru_lock);
3203                 pagevec_release(&pvec);
3204
3205                 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
3206         }
3207
3208 }
3209
3210 /**
3211  * scan_zone_unevictable_pages - check unevictable list for evictable pages
3212  * @zone - zone of which to scan the unevictable list
3213  *
3214  * Scan @zone's unevictable LRU lists to check for pages that have become
3215  * evictable.  Move those that have to @zone's inactive list where they
3216  * become candidates for reclaim, unless shrink_inactive_zone() decides
3217  * to reactivate them.  Pages that are still unevictable are rotated
3218  * back onto @zone's unevictable list.
3219  */
3220 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
3221 static void scan_zone_unevictable_pages(struct zone *zone)
3222 {
3223         struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
3224         unsigned long scan;
3225         unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
3226
3227         while (nr_to_scan > 0) {
3228                 unsigned long batch_size = min(nr_to_scan,
3229                                                 SCAN_UNEVICTABLE_BATCH_SIZE);
3230
3231                 spin_lock_irq(&zone->lru_lock);
3232                 for (scan = 0;  scan < batch_size; scan++) {
3233                         struct page *page = lru_to_page(l_unevictable);
3234
3235                         if (!trylock_page(page))
3236                                 continue;
3237
3238                         prefetchw_prev_lru_page(page, l_unevictable, flags);
3239
3240                         if (likely(PageLRU(page) && PageUnevictable(page)))
3241                                 check_move_unevictable_page(page, zone);
3242
3243                         unlock_page(page);
3244                 }
3245                 spin_unlock_irq(&zone->lru_lock);
3246
3247                 nr_to_scan -= batch_size;
3248         }
3249 }
3250
3251
3252 /**
3253  * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
3254  *
3255  * A really big hammer:  scan all zones' unevictable LRU lists to check for
3256  * pages that have become evictable.  Move those back to the zones'
3257  * inactive list where they become candidates for reclaim.
3258  * This occurs when, e.g., we have unswappable pages on the unevictable lists,
3259  * and we add swap to the system.  As such, it runs in the context of a task
3260  * that has possibly/probably made some previously unevictable pages
3261  * evictable.
3262  */
3263 static void scan_all_zones_unevictable_pages(void)
3264 {
3265         struct zone *zone;
3266
3267         for_each_zone(zone) {
3268                 scan_zone_unevictable_pages(zone);
3269         }
3270 }
3271
3272 /*
3273  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
3274  * all nodes' unevictable lists for evictable pages
3275  */
3276 unsigned long scan_unevictable_pages;
3277
3278 int scan_unevictable_handler(struct ctl_table *table, int write,
3279                            void __user *buffer,
3280                            size_t *length, loff_t *ppos)
3281 {
3282         proc_doulongvec_minmax(table, write, buffer, length, ppos);
3283
3284         if (write && *(unsigned long *)table->data)
3285                 scan_all_zones_unevictable_pages();
3286
3287         scan_unevictable_pages = 0;
3288         return 0;
3289 }
3290
3291 #ifdef CONFIG_NUMA
3292 /*
3293  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
3294  * a specified node's per zone unevictable lists for evictable pages.
3295  */
3296
3297 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
3298                                           struct sysdev_attribute *attr,
3299                                           char *buf)
3300 {
3301         return sprintf(buf, "0\n");     /* always zero; should fit... */
3302 }
3303
3304 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
3305                                            struct sysdev_attribute *attr,
3306                                         const char *buf, size_t count)
3307 {
3308         struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
3309         struct zone *zone;
3310         unsigned long res;
3311         unsigned long req = strict_strtoul(buf, 10, &res);
3312
3313         if (!req)
3314                 return 1;       /* zero is no-op */
3315
3316         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
3317                 if (!populated_zone(zone))
3318                         continue;
3319                 scan_zone_unevictable_pages(zone);
3320         }
3321         return 1;
3322 }
3323
3324
3325 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3326                         read_scan_unevictable_node,
3327                         write_scan_unevictable_node);
3328
3329 int scan_unevictable_register_node(struct node *node)
3330 {
3331         return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
3332 }
3333
3334 void scan_unevictable_unregister_node(struct node *node)
3335 {
3336         sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
3337 }
3338 #endif