writeback: do not sleep on the congestion queue if there are no congested BDIs or...
[linux-2.6.git] / mm / vmscan.c
1 /*
2  *  linux/mm/vmscan.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  *
6  *  Swap reorganised 29.12.95, Stephen Tweedie.
7  *  kswapd added: 7.1.96  sct
8  *  Removed kswapd_ctl limits, and swap out as many pages as needed
9  *  to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10  *  Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11  *  Multiqueue VM started 5.8.00, Rik van Riel.
12  */
13
14 #include <linux/mm.h>
15 #include <linux/module.h>
16 #include <linux/gfp.h>
17 #include <linux/kernel_stat.h>
18 #include <linux/swap.h>
19 #include <linux/pagemap.h>
20 #include <linux/init.h>
21 #include <linux/highmem.h>
22 #include <linux/vmstat.h>
23 #include <linux/file.h>
24 #include <linux/writeback.h>
25 #include <linux/blkdev.h>
26 #include <linux/buffer_head.h>  /* for try_to_release_page(),
27                                         buffer_heads_over_limit */
28 #include <linux/mm_inline.h>
29 #include <linux/pagevec.h>
30 #include <linux/backing-dev.h>
31 #include <linux/rmap.h>
32 #include <linux/topology.h>
33 #include <linux/cpu.h>
34 #include <linux/cpuset.h>
35 #include <linux/notifier.h>
36 #include <linux/rwsem.h>
37 #include <linux/delay.h>
38 #include <linux/kthread.h>
39 #include <linux/freezer.h>
40 #include <linux/memcontrol.h>
41 #include <linux/delayacct.h>
42 #include <linux/sysctl.h>
43
44 #include <asm/tlbflush.h>
45 #include <asm/div64.h>
46
47 #include <linux/swapops.h>
48
49 #include "internal.h"
50
51 #define CREATE_TRACE_POINTS
52 #include <trace/events/vmscan.h>
53
54 enum lumpy_mode {
55         LUMPY_MODE_NONE,
56         LUMPY_MODE_ASYNC,
57         LUMPY_MODE_SYNC,
58 };
59
60 struct scan_control {
61         /* Incremented by the number of inactive pages that were scanned */
62         unsigned long nr_scanned;
63
64         /* Number of pages freed so far during a call to shrink_zones() */
65         unsigned long nr_reclaimed;
66
67         /* How many pages shrink_list() should reclaim */
68         unsigned long nr_to_reclaim;
69
70         unsigned long hibernation_mode;
71
72         /* This context's GFP mask */
73         gfp_t gfp_mask;
74
75         int may_writepage;
76
77         /* Can mapped pages be reclaimed? */
78         int may_unmap;
79
80         /* Can pages be swapped as part of reclaim? */
81         int may_swap;
82
83         int swappiness;
84
85         int order;
86
87         /*
88          * Intend to reclaim enough continuous memory rather than reclaim
89          * enough amount of memory. i.e, mode for high order allocation.
90          */
91         enum lumpy_mode lumpy_reclaim_mode;
92
93         /* Which cgroup do we reclaim from */
94         struct mem_cgroup *mem_cgroup;
95
96         /*
97          * Nodemask of nodes allowed by the caller. If NULL, all nodes
98          * are scanned.
99          */
100         nodemask_t      *nodemask;
101 };
102
103 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
104
105 #ifdef ARCH_HAS_PREFETCH
106 #define prefetch_prev_lru_page(_page, _base, _field)                    \
107         do {                                                            \
108                 if ((_page)->lru.prev != _base) {                       \
109                         struct page *prev;                              \
110                                                                         \
111                         prev = lru_to_page(&(_page->lru));              \
112                         prefetch(&prev->_field);                        \
113                 }                                                       \
114         } while (0)
115 #else
116 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
117 #endif
118
119 #ifdef ARCH_HAS_PREFETCHW
120 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
121         do {                                                            \
122                 if ((_page)->lru.prev != _base) {                       \
123                         struct page *prev;                              \
124                                                                         \
125                         prev = lru_to_page(&(_page->lru));              \
126                         prefetchw(&prev->_field);                       \
127                 }                                                       \
128         } while (0)
129 #else
130 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
131 #endif
132
133 /*
134  * From 0 .. 100.  Higher means more swappy.
135  */
136 int vm_swappiness = 60;
137 long vm_total_pages;    /* The total number of pages which the VM controls */
138
139 static LIST_HEAD(shrinker_list);
140 static DECLARE_RWSEM(shrinker_rwsem);
141
142 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
143 #define scanning_global_lru(sc) (!(sc)->mem_cgroup)
144 #else
145 #define scanning_global_lru(sc) (1)
146 #endif
147
148 static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
149                                                   struct scan_control *sc)
150 {
151         if (!scanning_global_lru(sc))
152                 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
153
154         return &zone->reclaim_stat;
155 }
156
157 static unsigned long zone_nr_lru_pages(struct zone *zone,
158                                 struct scan_control *sc, enum lru_list lru)
159 {
160         if (!scanning_global_lru(sc))
161                 return mem_cgroup_zone_nr_pages(sc->mem_cgroup, zone, lru);
162
163         return zone_page_state(zone, NR_LRU_BASE + lru);
164 }
165
166
167 /*
168  * Add a shrinker callback to be called from the vm
169  */
170 void register_shrinker(struct shrinker *shrinker)
171 {
172         shrinker->nr = 0;
173         down_write(&shrinker_rwsem);
174         list_add_tail(&shrinker->list, &shrinker_list);
175         up_write(&shrinker_rwsem);
176 }
177 EXPORT_SYMBOL(register_shrinker);
178
179 /*
180  * Remove one
181  */
182 void unregister_shrinker(struct shrinker *shrinker)
183 {
184         down_write(&shrinker_rwsem);
185         list_del(&shrinker->list);
186         up_write(&shrinker_rwsem);
187 }
188 EXPORT_SYMBOL(unregister_shrinker);
189
190 #define SHRINK_BATCH 128
191 /*
192  * Call the shrink functions to age shrinkable caches
193  *
194  * Here we assume it costs one seek to replace a lru page and that it also
195  * takes a seek to recreate a cache object.  With this in mind we age equal
196  * percentages of the lru and ageable caches.  This should balance the seeks
197  * generated by these structures.
198  *
199  * If the vm encountered mapped pages on the LRU it increase the pressure on
200  * slab to avoid swapping.
201  *
202  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
203  *
204  * `lru_pages' represents the number of on-LRU pages in all the zones which
205  * are eligible for the caller's allocation attempt.  It is used for balancing
206  * slab reclaim versus page reclaim.
207  *
208  * Returns the number of slab objects which we shrunk.
209  */
210 unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
211                         unsigned long lru_pages)
212 {
213         struct shrinker *shrinker;
214         unsigned long ret = 0;
215
216         if (scanned == 0)
217                 scanned = SWAP_CLUSTER_MAX;
218
219         if (!down_read_trylock(&shrinker_rwsem))
220                 return 1;       /* Assume we'll be able to shrink next time */
221
222         list_for_each_entry(shrinker, &shrinker_list, list) {
223                 unsigned long long delta;
224                 unsigned long total_scan;
225                 unsigned long max_pass;
226
227                 max_pass = (*shrinker->shrink)(shrinker, 0, gfp_mask);
228                 delta = (4 * scanned) / shrinker->seeks;
229                 delta *= max_pass;
230                 do_div(delta, lru_pages + 1);
231                 shrinker->nr += delta;
232                 if (shrinker->nr < 0) {
233                         printk(KERN_ERR "shrink_slab: %pF negative objects to "
234                                "delete nr=%ld\n",
235                                shrinker->shrink, shrinker->nr);
236                         shrinker->nr = max_pass;
237                 }
238
239                 /*
240                  * Avoid risking looping forever due to too large nr value:
241                  * never try to free more than twice the estimate number of
242                  * freeable entries.
243                  */
244                 if (shrinker->nr > max_pass * 2)
245                         shrinker->nr = max_pass * 2;
246
247                 total_scan = shrinker->nr;
248                 shrinker->nr = 0;
249
250                 while (total_scan >= SHRINK_BATCH) {
251                         long this_scan = SHRINK_BATCH;
252                         int shrink_ret;
253                         int nr_before;
254
255                         nr_before = (*shrinker->shrink)(shrinker, 0, gfp_mask);
256                         shrink_ret = (*shrinker->shrink)(shrinker, this_scan,
257                                                                 gfp_mask);
258                         if (shrink_ret == -1)
259                                 break;
260                         if (shrink_ret < nr_before)
261                                 ret += nr_before - shrink_ret;
262                         count_vm_events(SLABS_SCANNED, this_scan);
263                         total_scan -= this_scan;
264
265                         cond_resched();
266                 }
267
268                 shrinker->nr += total_scan;
269         }
270         up_read(&shrinker_rwsem);
271         return ret;
272 }
273
274 static void set_lumpy_reclaim_mode(int priority, struct scan_control *sc,
275                                    bool sync)
276 {
277         enum lumpy_mode mode = sync ? LUMPY_MODE_SYNC : LUMPY_MODE_ASYNC;
278
279         /*
280          * Some reclaim have alredy been failed. No worth to try synchronous
281          * lumpy reclaim.
282          */
283         if (sync && sc->lumpy_reclaim_mode == LUMPY_MODE_NONE)
284                 return;
285
286         /*
287          * If we need a large contiguous chunk of memory, or have
288          * trouble getting a small set of contiguous pages, we
289          * will reclaim both active and inactive pages.
290          */
291         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
292                 sc->lumpy_reclaim_mode = mode;
293         else if (sc->order && priority < DEF_PRIORITY - 2)
294                 sc->lumpy_reclaim_mode = mode;
295         else
296                 sc->lumpy_reclaim_mode = LUMPY_MODE_NONE;
297 }
298
299 static void disable_lumpy_reclaim_mode(struct scan_control *sc)
300 {
301         sc->lumpy_reclaim_mode = LUMPY_MODE_NONE;
302 }
303
304 static inline int is_page_cache_freeable(struct page *page)
305 {
306         /*
307          * A freeable page cache page is referenced only by the caller
308          * that isolated the page, the page cache radix tree and
309          * optional buffer heads at page->private.
310          */
311         return page_count(page) - page_has_private(page) == 2;
312 }
313
314 static int may_write_to_queue(struct backing_dev_info *bdi,
315                               struct scan_control *sc)
316 {
317         if (current->flags & PF_SWAPWRITE)
318                 return 1;
319         if (!bdi_write_congested(bdi))
320                 return 1;
321         if (bdi == current->backing_dev_info)
322                 return 1;
323
324         /* lumpy reclaim for hugepage often need a lot of write */
325         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
326                 return 1;
327         return 0;
328 }
329
330 /*
331  * We detected a synchronous write error writing a page out.  Probably
332  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
333  * fsync(), msync() or close().
334  *
335  * The tricky part is that after writepage we cannot touch the mapping: nothing
336  * prevents it from being freed up.  But we have a ref on the page and once
337  * that page is locked, the mapping is pinned.
338  *
339  * We're allowed to run sleeping lock_page() here because we know the caller has
340  * __GFP_FS.
341  */
342 static void handle_write_error(struct address_space *mapping,
343                                 struct page *page, int error)
344 {
345         lock_page_nosync(page);
346         if (page_mapping(page) == mapping)
347                 mapping_set_error(mapping, error);
348         unlock_page(page);
349 }
350
351 /* possible outcome of pageout() */
352 typedef enum {
353         /* failed to write page out, page is locked */
354         PAGE_KEEP,
355         /* move page to the active list, page is locked */
356         PAGE_ACTIVATE,
357         /* page has been sent to the disk successfully, page is unlocked */
358         PAGE_SUCCESS,
359         /* page is clean and locked */
360         PAGE_CLEAN,
361 } pageout_t;
362
363 /*
364  * pageout is called by shrink_page_list() for each dirty page.
365  * Calls ->writepage().
366  */
367 static pageout_t pageout(struct page *page, struct address_space *mapping,
368                          struct scan_control *sc)
369 {
370         /*
371          * If the page is dirty, only perform writeback if that write
372          * will be non-blocking.  To prevent this allocation from being
373          * stalled by pagecache activity.  But note that there may be
374          * stalls if we need to run get_block().  We could test
375          * PagePrivate for that.
376          *
377          * If this process is currently in __generic_file_aio_write() against
378          * this page's queue, we can perform writeback even if that
379          * will block.
380          *
381          * If the page is swapcache, write it back even if that would
382          * block, for some throttling. This happens by accident, because
383          * swap_backing_dev_info is bust: it doesn't reflect the
384          * congestion state of the swapdevs.  Easy to fix, if needed.
385          */
386         if (!is_page_cache_freeable(page))
387                 return PAGE_KEEP;
388         if (!mapping) {
389                 /*
390                  * Some data journaling orphaned pages can have
391                  * page->mapping == NULL while being dirty with clean buffers.
392                  */
393                 if (page_has_private(page)) {
394                         if (try_to_free_buffers(page)) {
395                                 ClearPageDirty(page);
396                                 printk("%s: orphaned page\n", __func__);
397                                 return PAGE_CLEAN;
398                         }
399                 }
400                 return PAGE_KEEP;
401         }
402         if (mapping->a_ops->writepage == NULL)
403                 return PAGE_ACTIVATE;
404         if (!may_write_to_queue(mapping->backing_dev_info, sc))
405                 return PAGE_KEEP;
406
407         if (clear_page_dirty_for_io(page)) {
408                 int res;
409                 struct writeback_control wbc = {
410                         .sync_mode = WB_SYNC_NONE,
411                         .nr_to_write = SWAP_CLUSTER_MAX,
412                         .range_start = 0,
413                         .range_end = LLONG_MAX,
414                         .for_reclaim = 1,
415                 };
416
417                 SetPageReclaim(page);
418                 res = mapping->a_ops->writepage(page, &wbc);
419                 if (res < 0)
420                         handle_write_error(mapping, page, res);
421                 if (res == AOP_WRITEPAGE_ACTIVATE) {
422                         ClearPageReclaim(page);
423                         return PAGE_ACTIVATE;
424                 }
425
426                 /*
427                  * Wait on writeback if requested to. This happens when
428                  * direct reclaiming a large contiguous area and the
429                  * first attempt to free a range of pages fails.
430                  */
431                 if (PageWriteback(page) &&
432                     sc->lumpy_reclaim_mode == LUMPY_MODE_SYNC)
433                         wait_on_page_writeback(page);
434
435                 if (!PageWriteback(page)) {
436                         /* synchronous write or broken a_ops? */
437                         ClearPageReclaim(page);
438                 }
439                 trace_mm_vmscan_writepage(page,
440                         trace_reclaim_flags(page, sc->lumpy_reclaim_mode));
441                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
442                 return PAGE_SUCCESS;
443         }
444
445         return PAGE_CLEAN;
446 }
447
448 /*
449  * Same as remove_mapping, but if the page is removed from the mapping, it
450  * gets returned with a refcount of 0.
451  */
452 static int __remove_mapping(struct address_space *mapping, struct page *page)
453 {
454         BUG_ON(!PageLocked(page));
455         BUG_ON(mapping != page_mapping(page));
456
457         spin_lock_irq(&mapping->tree_lock);
458         /*
459          * The non racy check for a busy page.
460          *
461          * Must be careful with the order of the tests. When someone has
462          * a ref to the page, it may be possible that they dirty it then
463          * drop the reference. So if PageDirty is tested before page_count
464          * here, then the following race may occur:
465          *
466          * get_user_pages(&page);
467          * [user mapping goes away]
468          * write_to(page);
469          *                              !PageDirty(page)    [good]
470          * SetPageDirty(page);
471          * put_page(page);
472          *                              !page_count(page)   [good, discard it]
473          *
474          * [oops, our write_to data is lost]
475          *
476          * Reversing the order of the tests ensures such a situation cannot
477          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
478          * load is not satisfied before that of page->_count.
479          *
480          * Note that if SetPageDirty is always performed via set_page_dirty,
481          * and thus under tree_lock, then this ordering is not required.
482          */
483         if (!page_freeze_refs(page, 2))
484                 goto cannot_free;
485         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
486         if (unlikely(PageDirty(page))) {
487                 page_unfreeze_refs(page, 2);
488                 goto cannot_free;
489         }
490
491         if (PageSwapCache(page)) {
492                 swp_entry_t swap = { .val = page_private(page) };
493                 __delete_from_swap_cache(page);
494                 spin_unlock_irq(&mapping->tree_lock);
495                 swapcache_free(swap, page);
496         } else {
497                 __remove_from_page_cache(page);
498                 spin_unlock_irq(&mapping->tree_lock);
499                 mem_cgroup_uncharge_cache_page(page);
500         }
501
502         return 1;
503
504 cannot_free:
505         spin_unlock_irq(&mapping->tree_lock);
506         return 0;
507 }
508
509 /*
510  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
511  * someone else has a ref on the page, abort and return 0.  If it was
512  * successfully detached, return 1.  Assumes the caller has a single ref on
513  * this page.
514  */
515 int remove_mapping(struct address_space *mapping, struct page *page)
516 {
517         if (__remove_mapping(mapping, page)) {
518                 /*
519                  * Unfreezing the refcount with 1 rather than 2 effectively
520                  * drops the pagecache ref for us without requiring another
521                  * atomic operation.
522                  */
523                 page_unfreeze_refs(page, 1);
524                 return 1;
525         }
526         return 0;
527 }
528
529 /**
530  * putback_lru_page - put previously isolated page onto appropriate LRU list
531  * @page: page to be put back to appropriate lru list
532  *
533  * Add previously isolated @page to appropriate LRU list.
534  * Page may still be unevictable for other reasons.
535  *
536  * lru_lock must not be held, interrupts must be enabled.
537  */
538 void putback_lru_page(struct page *page)
539 {
540         int lru;
541         int active = !!TestClearPageActive(page);
542         int was_unevictable = PageUnevictable(page);
543
544         VM_BUG_ON(PageLRU(page));
545
546 redo:
547         ClearPageUnevictable(page);
548
549         if (page_evictable(page, NULL)) {
550                 /*
551                  * For evictable pages, we can use the cache.
552                  * In event of a race, worst case is we end up with an
553                  * unevictable page on [in]active list.
554                  * We know how to handle that.
555                  */
556                 lru = active + page_lru_base_type(page);
557                 lru_cache_add_lru(page, lru);
558         } else {
559                 /*
560                  * Put unevictable pages directly on zone's unevictable
561                  * list.
562                  */
563                 lru = LRU_UNEVICTABLE;
564                 add_page_to_unevictable_list(page);
565                 /*
566                  * When racing with an mlock clearing (page is
567                  * unlocked), make sure that if the other thread does
568                  * not observe our setting of PG_lru and fails
569                  * isolation, we see PG_mlocked cleared below and move
570                  * the page back to the evictable list.
571                  *
572                  * The other side is TestClearPageMlocked().
573                  */
574                 smp_mb();
575         }
576
577         /*
578          * page's status can change while we move it among lru. If an evictable
579          * page is on unevictable list, it never be freed. To avoid that,
580          * check after we added it to the list, again.
581          */
582         if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
583                 if (!isolate_lru_page(page)) {
584                         put_page(page);
585                         goto redo;
586                 }
587                 /* This means someone else dropped this page from LRU
588                  * So, it will be freed or putback to LRU again. There is
589                  * nothing to do here.
590                  */
591         }
592
593         if (was_unevictable && lru != LRU_UNEVICTABLE)
594                 count_vm_event(UNEVICTABLE_PGRESCUED);
595         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
596                 count_vm_event(UNEVICTABLE_PGCULLED);
597
598         put_page(page);         /* drop ref from isolate */
599 }
600
601 enum page_references {
602         PAGEREF_RECLAIM,
603         PAGEREF_RECLAIM_CLEAN,
604         PAGEREF_KEEP,
605         PAGEREF_ACTIVATE,
606 };
607
608 static enum page_references page_check_references(struct page *page,
609                                                   struct scan_control *sc)
610 {
611         int referenced_ptes, referenced_page;
612         unsigned long vm_flags;
613
614         referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
615         referenced_page = TestClearPageReferenced(page);
616
617         /* Lumpy reclaim - ignore references */
618         if (sc->lumpy_reclaim_mode != LUMPY_MODE_NONE)
619                 return PAGEREF_RECLAIM;
620
621         /*
622          * Mlock lost the isolation race with us.  Let try_to_unmap()
623          * move the page to the unevictable list.
624          */
625         if (vm_flags & VM_LOCKED)
626                 return PAGEREF_RECLAIM;
627
628         if (referenced_ptes) {
629                 if (PageAnon(page))
630                         return PAGEREF_ACTIVATE;
631                 /*
632                  * All mapped pages start out with page table
633                  * references from the instantiating fault, so we need
634                  * to look twice if a mapped file page is used more
635                  * than once.
636                  *
637                  * Mark it and spare it for another trip around the
638                  * inactive list.  Another page table reference will
639                  * lead to its activation.
640                  *
641                  * Note: the mark is set for activated pages as well
642                  * so that recently deactivated but used pages are
643                  * quickly recovered.
644                  */
645                 SetPageReferenced(page);
646
647                 if (referenced_page)
648                         return PAGEREF_ACTIVATE;
649
650                 return PAGEREF_KEEP;
651         }
652
653         /* Reclaim if clean, defer dirty pages to writeback */
654         if (referenced_page)
655                 return PAGEREF_RECLAIM_CLEAN;
656
657         return PAGEREF_RECLAIM;
658 }
659
660 static noinline_for_stack void free_page_list(struct list_head *free_pages)
661 {
662         struct pagevec freed_pvec;
663         struct page *page, *tmp;
664
665         pagevec_init(&freed_pvec, 1);
666
667         list_for_each_entry_safe(page, tmp, free_pages, lru) {
668                 list_del(&page->lru);
669                 if (!pagevec_add(&freed_pvec, page)) {
670                         __pagevec_free(&freed_pvec);
671                         pagevec_reinit(&freed_pvec);
672                 }
673         }
674
675         pagevec_free(&freed_pvec);
676 }
677
678 /*
679  * shrink_page_list() returns the number of reclaimed pages
680  */
681 static unsigned long shrink_page_list(struct list_head *page_list,
682                                       struct zone *zone,
683                                       struct scan_control *sc)
684 {
685         LIST_HEAD(ret_pages);
686         LIST_HEAD(free_pages);
687         int pgactivate = 0;
688         unsigned long nr_dirty = 0;
689         unsigned long nr_congested = 0;
690         unsigned long nr_reclaimed = 0;
691
692         cond_resched();
693
694         while (!list_empty(page_list)) {
695                 enum page_references references;
696                 struct address_space *mapping;
697                 struct page *page;
698                 int may_enter_fs;
699
700                 cond_resched();
701
702                 page = lru_to_page(page_list);
703                 list_del(&page->lru);
704
705                 if (!trylock_page(page))
706                         goto keep;
707
708                 VM_BUG_ON(PageActive(page));
709                 VM_BUG_ON(page_zone(page) != zone);
710
711                 sc->nr_scanned++;
712
713                 if (unlikely(!page_evictable(page, NULL)))
714                         goto cull_mlocked;
715
716                 if (!sc->may_unmap && page_mapped(page))
717                         goto keep_locked;
718
719                 /* Double the slab pressure for mapped and swapcache pages */
720                 if (page_mapped(page) || PageSwapCache(page))
721                         sc->nr_scanned++;
722
723                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
724                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
725
726                 if (PageWriteback(page)) {
727                         /*
728                          * Synchronous reclaim is performed in two passes,
729                          * first an asynchronous pass over the list to
730                          * start parallel writeback, and a second synchronous
731                          * pass to wait for the IO to complete.  Wait here
732                          * for any page for which writeback has already
733                          * started.
734                          */
735                         if (sc->lumpy_reclaim_mode == LUMPY_MODE_SYNC &&
736                             may_enter_fs)
737                                 wait_on_page_writeback(page);
738                         else {
739                                 unlock_page(page);
740                                 goto keep_lumpy;
741                         }
742                 }
743
744                 references = page_check_references(page, sc);
745                 switch (references) {
746                 case PAGEREF_ACTIVATE:
747                         goto activate_locked;
748                 case PAGEREF_KEEP:
749                         goto keep_locked;
750                 case PAGEREF_RECLAIM:
751                 case PAGEREF_RECLAIM_CLEAN:
752                         ; /* try to reclaim the page below */
753                 }
754
755                 /*
756                  * Anonymous process memory has backing store?
757                  * Try to allocate it some swap space here.
758                  */
759                 if (PageAnon(page) && !PageSwapCache(page)) {
760                         if (!(sc->gfp_mask & __GFP_IO))
761                                 goto keep_locked;
762                         if (!add_to_swap(page))
763                                 goto activate_locked;
764                         may_enter_fs = 1;
765                 }
766
767                 mapping = page_mapping(page);
768
769                 /*
770                  * The page is mapped into the page tables of one or more
771                  * processes. Try to unmap it here.
772                  */
773                 if (page_mapped(page) && mapping) {
774                         switch (try_to_unmap(page, TTU_UNMAP)) {
775                         case SWAP_FAIL:
776                                 goto activate_locked;
777                         case SWAP_AGAIN:
778                                 goto keep_locked;
779                         case SWAP_MLOCK:
780                                 goto cull_mlocked;
781                         case SWAP_SUCCESS:
782                                 ; /* try to free the page below */
783                         }
784                 }
785
786                 if (PageDirty(page)) {
787                         nr_dirty++;
788
789                         if (references == PAGEREF_RECLAIM_CLEAN)
790                                 goto keep_locked;
791                         if (!may_enter_fs)
792                                 goto keep_locked;
793                         if (!sc->may_writepage)
794                                 goto keep_locked;
795
796                         /* Page is dirty, try to write it out here */
797                         switch (pageout(page, mapping, sc)) {
798                         case PAGE_KEEP:
799                                 nr_congested++;
800                                 goto keep_locked;
801                         case PAGE_ACTIVATE:
802                                 goto activate_locked;
803                         case PAGE_SUCCESS:
804                                 if (PageWriteback(page))
805                                         goto keep_lumpy;
806                                 if (PageDirty(page))
807                                         goto keep;
808
809                                 /*
810                                  * A synchronous write - probably a ramdisk.  Go
811                                  * ahead and try to reclaim the page.
812                                  */
813                                 if (!trylock_page(page))
814                                         goto keep;
815                                 if (PageDirty(page) || PageWriteback(page))
816                                         goto keep_locked;
817                                 mapping = page_mapping(page);
818                         case PAGE_CLEAN:
819                                 ; /* try to free the page below */
820                         }
821                 }
822
823                 /*
824                  * If the page has buffers, try to free the buffer mappings
825                  * associated with this page. If we succeed we try to free
826                  * the page as well.
827                  *
828                  * We do this even if the page is PageDirty().
829                  * try_to_release_page() does not perform I/O, but it is
830                  * possible for a page to have PageDirty set, but it is actually
831                  * clean (all its buffers are clean).  This happens if the
832                  * buffers were written out directly, with submit_bh(). ext3
833                  * will do this, as well as the blockdev mapping.
834                  * try_to_release_page() will discover that cleanness and will
835                  * drop the buffers and mark the page clean - it can be freed.
836                  *
837                  * Rarely, pages can have buffers and no ->mapping.  These are
838                  * the pages which were not successfully invalidated in
839                  * truncate_complete_page().  We try to drop those buffers here
840                  * and if that worked, and the page is no longer mapped into
841                  * process address space (page_count == 1) it can be freed.
842                  * Otherwise, leave the page on the LRU so it is swappable.
843                  */
844                 if (page_has_private(page)) {
845                         if (!try_to_release_page(page, sc->gfp_mask))
846                                 goto activate_locked;
847                         if (!mapping && page_count(page) == 1) {
848                                 unlock_page(page);
849                                 if (put_page_testzero(page))
850                                         goto free_it;
851                                 else {
852                                         /*
853                                          * rare race with speculative reference.
854                                          * the speculative reference will free
855                                          * this page shortly, so we may
856                                          * increment nr_reclaimed here (and
857                                          * leave it off the LRU).
858                                          */
859                                         nr_reclaimed++;
860                                         continue;
861                                 }
862                         }
863                 }
864
865                 if (!mapping || !__remove_mapping(mapping, page))
866                         goto keep_locked;
867
868                 /*
869                  * At this point, we have no other references and there is
870                  * no way to pick any more up (removed from LRU, removed
871                  * from pagecache). Can use non-atomic bitops now (and
872                  * we obviously don't have to worry about waking up a process
873                  * waiting on the page lock, because there are no references.
874                  */
875                 __clear_page_locked(page);
876 free_it:
877                 nr_reclaimed++;
878
879                 /*
880                  * Is there need to periodically free_page_list? It would
881                  * appear not as the counts should be low
882                  */
883                 list_add(&page->lru, &free_pages);
884                 continue;
885
886 cull_mlocked:
887                 if (PageSwapCache(page))
888                         try_to_free_swap(page);
889                 unlock_page(page);
890                 putback_lru_page(page);
891                 disable_lumpy_reclaim_mode(sc);
892                 continue;
893
894 activate_locked:
895                 /* Not a candidate for swapping, so reclaim swap space. */
896                 if (PageSwapCache(page) && vm_swap_full())
897                         try_to_free_swap(page);
898                 VM_BUG_ON(PageActive(page));
899                 SetPageActive(page);
900                 pgactivate++;
901 keep_locked:
902                 unlock_page(page);
903 keep:
904                 disable_lumpy_reclaim_mode(sc);
905 keep_lumpy:
906                 list_add(&page->lru, &ret_pages);
907                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
908         }
909
910         /*
911          * Tag a zone as congested if all the dirty pages encountered were
912          * backed by a congested BDI. In this case, reclaimers should just
913          * back off and wait for congestion to clear because further reclaim
914          * will encounter the same problem
915          */
916         if (nr_dirty == nr_congested)
917                 zone_set_flag(zone, ZONE_CONGESTED);
918
919         free_page_list(&free_pages);
920
921         list_splice(&ret_pages, page_list);
922         count_vm_events(PGACTIVATE, pgactivate);
923         return nr_reclaimed;
924 }
925
926 /*
927  * Attempt to remove the specified page from its LRU.  Only take this page
928  * if it is of the appropriate PageActive status.  Pages which are being
929  * freed elsewhere are also ignored.
930  *
931  * page:        page to consider
932  * mode:        one of the LRU isolation modes defined above
933  *
934  * returns 0 on success, -ve errno on failure.
935  */
936 int __isolate_lru_page(struct page *page, int mode, int file)
937 {
938         int ret = -EINVAL;
939
940         /* Only take pages on the LRU. */
941         if (!PageLRU(page))
942                 return ret;
943
944         /*
945          * When checking the active state, we need to be sure we are
946          * dealing with comparible boolean values.  Take the logical not
947          * of each.
948          */
949         if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
950                 return ret;
951
952         if (mode != ISOLATE_BOTH && page_is_file_cache(page) != file)
953                 return ret;
954
955         /*
956          * When this function is being called for lumpy reclaim, we
957          * initially look into all LRU pages, active, inactive and
958          * unevictable; only give shrink_page_list evictable pages.
959          */
960         if (PageUnevictable(page))
961                 return ret;
962
963         ret = -EBUSY;
964
965         if (likely(get_page_unless_zero(page))) {
966                 /*
967                  * Be careful not to clear PageLRU until after we're
968                  * sure the page is not being freed elsewhere -- the
969                  * page release code relies on it.
970                  */
971                 ClearPageLRU(page);
972                 ret = 0;
973         }
974
975         return ret;
976 }
977
978 /*
979  * zone->lru_lock is heavily contended.  Some of the functions that
980  * shrink the lists perform better by taking out a batch of pages
981  * and working on them outside the LRU lock.
982  *
983  * For pagecache intensive workloads, this function is the hottest
984  * spot in the kernel (apart from copy_*_user functions).
985  *
986  * Appropriate locks must be held before calling this function.
987  *
988  * @nr_to_scan: The number of pages to look through on the list.
989  * @src:        The LRU list to pull pages off.
990  * @dst:        The temp list to put pages on to.
991  * @scanned:    The number of pages that were scanned.
992  * @order:      The caller's attempted allocation order
993  * @mode:       One of the LRU isolation modes
994  * @file:       True [1] if isolating file [!anon] pages
995  *
996  * returns how many pages were moved onto *@dst.
997  */
998 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
999                 struct list_head *src, struct list_head *dst,
1000                 unsigned long *scanned, int order, int mode, int file)
1001 {
1002         unsigned long nr_taken = 0;
1003         unsigned long nr_lumpy_taken = 0;
1004         unsigned long nr_lumpy_dirty = 0;
1005         unsigned long nr_lumpy_failed = 0;
1006         unsigned long scan;
1007
1008         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1009                 struct page *page;
1010                 unsigned long pfn;
1011                 unsigned long end_pfn;
1012                 unsigned long page_pfn;
1013                 int zone_id;
1014
1015                 page = lru_to_page(src);
1016                 prefetchw_prev_lru_page(page, src, flags);
1017
1018                 VM_BUG_ON(!PageLRU(page));
1019
1020                 switch (__isolate_lru_page(page, mode, file)) {
1021                 case 0:
1022                         list_move(&page->lru, dst);
1023                         mem_cgroup_del_lru(page);
1024                         nr_taken++;
1025                         break;
1026
1027                 case -EBUSY:
1028                         /* else it is being freed elsewhere */
1029                         list_move(&page->lru, src);
1030                         mem_cgroup_rotate_lru_list(page, page_lru(page));
1031                         continue;
1032
1033                 default:
1034                         BUG();
1035                 }
1036
1037                 if (!order)
1038                         continue;
1039
1040                 /*
1041                  * Attempt to take all pages in the order aligned region
1042                  * surrounding the tag page.  Only take those pages of
1043                  * the same active state as that tag page.  We may safely
1044                  * round the target page pfn down to the requested order
1045                  * as the mem_map is guarenteed valid out to MAX_ORDER,
1046                  * where that page is in a different zone we will detect
1047                  * it from its zone id and abort this block scan.
1048                  */
1049                 zone_id = page_zone_id(page);
1050                 page_pfn = page_to_pfn(page);
1051                 pfn = page_pfn & ~((1 << order) - 1);
1052                 end_pfn = pfn + (1 << order);
1053                 for (; pfn < end_pfn; pfn++) {
1054                         struct page *cursor_page;
1055
1056                         /* The target page is in the block, ignore it. */
1057                         if (unlikely(pfn == page_pfn))
1058                                 continue;
1059
1060                         /* Avoid holes within the zone. */
1061                         if (unlikely(!pfn_valid_within(pfn)))
1062                                 break;
1063
1064                         cursor_page = pfn_to_page(pfn);
1065
1066                         /* Check that we have not crossed a zone boundary. */
1067                         if (unlikely(page_zone_id(cursor_page) != zone_id))
1068                                 break;
1069
1070                         /*
1071                          * If we don't have enough swap space, reclaiming of
1072                          * anon page which don't already have a swap slot is
1073                          * pointless.
1074                          */
1075                         if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
1076                             !PageSwapCache(cursor_page))
1077                                 break;
1078
1079                         if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1080                                 list_move(&cursor_page->lru, dst);
1081                                 mem_cgroup_del_lru(cursor_page);
1082                                 nr_taken++;
1083                                 nr_lumpy_taken++;
1084                                 if (PageDirty(cursor_page))
1085                                         nr_lumpy_dirty++;
1086                                 scan++;
1087                         } else {
1088                                 /* the page is freed already. */
1089                                 if (!page_count(cursor_page))
1090                                         continue;
1091                                 break;
1092                         }
1093                 }
1094
1095                 /* If we break out of the loop above, lumpy reclaim failed */
1096                 if (pfn < end_pfn)
1097                         nr_lumpy_failed++;
1098         }
1099
1100         *scanned = scan;
1101
1102         trace_mm_vmscan_lru_isolate(order,
1103                         nr_to_scan, scan,
1104                         nr_taken,
1105                         nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1106                         mode);
1107         return nr_taken;
1108 }
1109
1110 static unsigned long isolate_pages_global(unsigned long nr,
1111                                         struct list_head *dst,
1112                                         unsigned long *scanned, int order,
1113                                         int mode, struct zone *z,
1114                                         int active, int file)
1115 {
1116         int lru = LRU_BASE;
1117         if (active)
1118                 lru += LRU_ACTIVE;
1119         if (file)
1120                 lru += LRU_FILE;
1121         return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1122                                                                 mode, file);
1123 }
1124
1125 /*
1126  * clear_active_flags() is a helper for shrink_active_list(), clearing
1127  * any active bits from the pages in the list.
1128  */
1129 static unsigned long clear_active_flags(struct list_head *page_list,
1130                                         unsigned int *count)
1131 {
1132         int nr_active = 0;
1133         int lru;
1134         struct page *page;
1135
1136         list_for_each_entry(page, page_list, lru) {
1137                 lru = page_lru_base_type(page);
1138                 if (PageActive(page)) {
1139                         lru += LRU_ACTIVE;
1140                         ClearPageActive(page);
1141                         nr_active++;
1142                 }
1143                 if (count)
1144                         count[lru]++;
1145         }
1146
1147         return nr_active;
1148 }
1149
1150 /**
1151  * isolate_lru_page - tries to isolate a page from its LRU list
1152  * @page: page to isolate from its LRU list
1153  *
1154  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1155  * vmstat statistic corresponding to whatever LRU list the page was on.
1156  *
1157  * Returns 0 if the page was removed from an LRU list.
1158  * Returns -EBUSY if the page was not on an LRU list.
1159  *
1160  * The returned page will have PageLRU() cleared.  If it was found on
1161  * the active list, it will have PageActive set.  If it was found on
1162  * the unevictable list, it will have the PageUnevictable bit set. That flag
1163  * may need to be cleared by the caller before letting the page go.
1164  *
1165  * The vmstat statistic corresponding to the list on which the page was
1166  * found will be decremented.
1167  *
1168  * Restrictions:
1169  * (1) Must be called with an elevated refcount on the page. This is a
1170  *     fundamentnal difference from isolate_lru_pages (which is called
1171  *     without a stable reference).
1172  * (2) the lru_lock must not be held.
1173  * (3) interrupts must be enabled.
1174  */
1175 int isolate_lru_page(struct page *page)
1176 {
1177         int ret = -EBUSY;
1178
1179         if (PageLRU(page)) {
1180                 struct zone *zone = page_zone(page);
1181
1182                 spin_lock_irq(&zone->lru_lock);
1183                 if (PageLRU(page) && get_page_unless_zero(page)) {
1184                         int lru = page_lru(page);
1185                         ret = 0;
1186                         ClearPageLRU(page);
1187
1188                         del_page_from_lru_list(zone, page, lru);
1189                 }
1190                 spin_unlock_irq(&zone->lru_lock);
1191         }
1192         return ret;
1193 }
1194
1195 /*
1196  * Are there way too many processes in the direct reclaim path already?
1197  */
1198 static int too_many_isolated(struct zone *zone, int file,
1199                 struct scan_control *sc)
1200 {
1201         unsigned long inactive, isolated;
1202
1203         if (current_is_kswapd())
1204                 return 0;
1205
1206         if (!scanning_global_lru(sc))
1207                 return 0;
1208
1209         if (file) {
1210                 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1211                 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1212         } else {
1213                 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1214                 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1215         }
1216
1217         return isolated > inactive;
1218 }
1219
1220 /*
1221  * TODO: Try merging with migrations version of putback_lru_pages
1222  */
1223 static noinline_for_stack void
1224 putback_lru_pages(struct zone *zone, struct scan_control *sc,
1225                                 unsigned long nr_anon, unsigned long nr_file,
1226                                 struct list_head *page_list)
1227 {
1228         struct page *page;
1229         struct pagevec pvec;
1230         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1231
1232         pagevec_init(&pvec, 1);
1233
1234         /*
1235          * Put back any unfreeable pages.
1236          */
1237         spin_lock(&zone->lru_lock);
1238         while (!list_empty(page_list)) {
1239                 int lru;
1240                 page = lru_to_page(page_list);
1241                 VM_BUG_ON(PageLRU(page));
1242                 list_del(&page->lru);
1243                 if (unlikely(!page_evictable(page, NULL))) {
1244                         spin_unlock_irq(&zone->lru_lock);
1245                         putback_lru_page(page);
1246                         spin_lock_irq(&zone->lru_lock);
1247                         continue;
1248                 }
1249                 SetPageLRU(page);
1250                 lru = page_lru(page);
1251                 add_page_to_lru_list(zone, page, lru);
1252                 if (is_active_lru(lru)) {
1253                         int file = is_file_lru(lru);
1254                         reclaim_stat->recent_rotated[file]++;
1255                 }
1256                 if (!pagevec_add(&pvec, page)) {
1257                         spin_unlock_irq(&zone->lru_lock);
1258                         __pagevec_release(&pvec);
1259                         spin_lock_irq(&zone->lru_lock);
1260                 }
1261         }
1262         __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1263         __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1264
1265         spin_unlock_irq(&zone->lru_lock);
1266         pagevec_release(&pvec);
1267 }
1268
1269 static noinline_for_stack void update_isolated_counts(struct zone *zone,
1270                                         struct scan_control *sc,
1271                                         unsigned long *nr_anon,
1272                                         unsigned long *nr_file,
1273                                         struct list_head *isolated_list)
1274 {
1275         unsigned long nr_active;
1276         unsigned int count[NR_LRU_LISTS] = { 0, };
1277         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1278
1279         nr_active = clear_active_flags(isolated_list, count);
1280         __count_vm_events(PGDEACTIVATE, nr_active);
1281
1282         __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1283                               -count[LRU_ACTIVE_FILE]);
1284         __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1285                               -count[LRU_INACTIVE_FILE]);
1286         __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1287                               -count[LRU_ACTIVE_ANON]);
1288         __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1289                               -count[LRU_INACTIVE_ANON]);
1290
1291         *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1292         *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1293         __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1294         __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1295
1296         reclaim_stat->recent_scanned[0] += *nr_anon;
1297         reclaim_stat->recent_scanned[1] += *nr_file;
1298 }
1299
1300 /*
1301  * Returns true if the caller should wait to clean dirty/writeback pages.
1302  *
1303  * If we are direct reclaiming for contiguous pages and we do not reclaim
1304  * everything in the list, try again and wait for writeback IO to complete.
1305  * This will stall high-order allocations noticeably. Only do that when really
1306  * need to free the pages under high memory pressure.
1307  */
1308 static inline bool should_reclaim_stall(unsigned long nr_taken,
1309                                         unsigned long nr_freed,
1310                                         int priority,
1311                                         struct scan_control *sc)
1312 {
1313         int lumpy_stall_priority;
1314
1315         /* kswapd should not stall on sync IO */
1316         if (current_is_kswapd())
1317                 return false;
1318
1319         /* Only stall on lumpy reclaim */
1320         if (sc->lumpy_reclaim_mode == LUMPY_MODE_NONE)
1321                 return false;
1322
1323         /* If we have relaimed everything on the isolated list, no stall */
1324         if (nr_freed == nr_taken)
1325                 return false;
1326
1327         /*
1328          * For high-order allocations, there are two stall thresholds.
1329          * High-cost allocations stall immediately where as lower
1330          * order allocations such as stacks require the scanning
1331          * priority to be much higher before stalling.
1332          */
1333         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1334                 lumpy_stall_priority = DEF_PRIORITY;
1335         else
1336                 lumpy_stall_priority = DEF_PRIORITY / 3;
1337
1338         return priority <= lumpy_stall_priority;
1339 }
1340
1341 /*
1342  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1343  * of reclaimed pages
1344  */
1345 static noinline_for_stack unsigned long
1346 shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1347                         struct scan_control *sc, int priority, int file)
1348 {
1349         LIST_HEAD(page_list);
1350         unsigned long nr_scanned;
1351         unsigned long nr_reclaimed = 0;
1352         unsigned long nr_taken;
1353         unsigned long nr_anon;
1354         unsigned long nr_file;
1355
1356         while (unlikely(too_many_isolated(zone, file, sc))) {
1357                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1358
1359                 /* We are about to die and free our memory. Return now. */
1360                 if (fatal_signal_pending(current))
1361                         return SWAP_CLUSTER_MAX;
1362         }
1363
1364         set_lumpy_reclaim_mode(priority, sc, false);
1365         lru_add_drain();
1366         spin_lock_irq(&zone->lru_lock);
1367
1368         if (scanning_global_lru(sc)) {
1369                 nr_taken = isolate_pages_global(nr_to_scan,
1370                         &page_list, &nr_scanned, sc->order,
1371                         sc->lumpy_reclaim_mode == LUMPY_MODE_NONE ?
1372                                         ISOLATE_INACTIVE : ISOLATE_BOTH,
1373                         zone, 0, file);
1374                 zone->pages_scanned += nr_scanned;
1375                 if (current_is_kswapd())
1376                         __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1377                                                nr_scanned);
1378                 else
1379                         __count_zone_vm_events(PGSCAN_DIRECT, zone,
1380                                                nr_scanned);
1381         } else {
1382                 nr_taken = mem_cgroup_isolate_pages(nr_to_scan,
1383                         &page_list, &nr_scanned, sc->order,
1384                         sc->lumpy_reclaim_mode == LUMPY_MODE_NONE ?
1385                                         ISOLATE_INACTIVE : ISOLATE_BOTH,
1386                         zone, sc->mem_cgroup,
1387                         0, file);
1388                 /*
1389                  * mem_cgroup_isolate_pages() keeps track of
1390                  * scanned pages on its own.
1391                  */
1392         }
1393
1394         if (nr_taken == 0) {
1395                 spin_unlock_irq(&zone->lru_lock);
1396                 return 0;
1397         }
1398
1399         update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
1400
1401         spin_unlock_irq(&zone->lru_lock);
1402
1403         nr_reclaimed = shrink_page_list(&page_list, zone, sc);
1404
1405         /* Check if we should syncronously wait for writeback */
1406         if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1407                 set_lumpy_reclaim_mode(priority, sc, true);
1408                 nr_reclaimed += shrink_page_list(&page_list, zone, sc);
1409         }
1410
1411         local_irq_disable();
1412         if (current_is_kswapd())
1413                 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1414         __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1415
1416         putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
1417
1418         trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1419                 zone_idx(zone),
1420                 nr_scanned, nr_reclaimed,
1421                 priority,
1422                 trace_shrink_flags(file, sc->lumpy_reclaim_mode));
1423         return nr_reclaimed;
1424 }
1425
1426 /*
1427  * This moves pages from the active list to the inactive list.
1428  *
1429  * We move them the other way if the page is referenced by one or more
1430  * processes, from rmap.
1431  *
1432  * If the pages are mostly unmapped, the processing is fast and it is
1433  * appropriate to hold zone->lru_lock across the whole operation.  But if
1434  * the pages are mapped, the processing is slow (page_referenced()) so we
1435  * should drop zone->lru_lock around each page.  It's impossible to balance
1436  * this, so instead we remove the pages from the LRU while processing them.
1437  * It is safe to rely on PG_active against the non-LRU pages in here because
1438  * nobody will play with that bit on a non-LRU page.
1439  *
1440  * The downside is that we have to touch page->_count against each page.
1441  * But we had to alter page->flags anyway.
1442  */
1443
1444 static void move_active_pages_to_lru(struct zone *zone,
1445                                      struct list_head *list,
1446                                      enum lru_list lru)
1447 {
1448         unsigned long pgmoved = 0;
1449         struct pagevec pvec;
1450         struct page *page;
1451
1452         pagevec_init(&pvec, 1);
1453
1454         while (!list_empty(list)) {
1455                 page = lru_to_page(list);
1456
1457                 VM_BUG_ON(PageLRU(page));
1458                 SetPageLRU(page);
1459
1460                 list_move(&page->lru, &zone->lru[lru].list);
1461                 mem_cgroup_add_lru_list(page, lru);
1462                 pgmoved++;
1463
1464                 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1465                         spin_unlock_irq(&zone->lru_lock);
1466                         if (buffer_heads_over_limit)
1467                                 pagevec_strip(&pvec);
1468                         __pagevec_release(&pvec);
1469                         spin_lock_irq(&zone->lru_lock);
1470                 }
1471         }
1472         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1473         if (!is_active_lru(lru))
1474                 __count_vm_events(PGDEACTIVATE, pgmoved);
1475 }
1476
1477 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1478                         struct scan_control *sc, int priority, int file)
1479 {
1480         unsigned long nr_taken;
1481         unsigned long pgscanned;
1482         unsigned long vm_flags;
1483         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1484         LIST_HEAD(l_active);
1485         LIST_HEAD(l_inactive);
1486         struct page *page;
1487         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1488         unsigned long nr_rotated = 0;
1489
1490         lru_add_drain();
1491         spin_lock_irq(&zone->lru_lock);
1492         if (scanning_global_lru(sc)) {
1493                 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1494                                                 &pgscanned, sc->order,
1495                                                 ISOLATE_ACTIVE, zone,
1496                                                 1, file);
1497                 zone->pages_scanned += pgscanned;
1498         } else {
1499                 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1500                                                 &pgscanned, sc->order,
1501                                                 ISOLATE_ACTIVE, zone,
1502                                                 sc->mem_cgroup, 1, file);
1503                 /*
1504                  * mem_cgroup_isolate_pages() keeps track of
1505                  * scanned pages on its own.
1506                  */
1507         }
1508
1509         reclaim_stat->recent_scanned[file] += nr_taken;
1510
1511         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1512         if (file)
1513                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1514         else
1515                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1516         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1517         spin_unlock_irq(&zone->lru_lock);
1518
1519         while (!list_empty(&l_hold)) {
1520                 cond_resched();
1521                 page = lru_to_page(&l_hold);
1522                 list_del(&page->lru);
1523
1524                 if (unlikely(!page_evictable(page, NULL))) {
1525                         putback_lru_page(page);
1526                         continue;
1527                 }
1528
1529                 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1530                         nr_rotated++;
1531                         /*
1532                          * Identify referenced, file-backed active pages and
1533                          * give them one more trip around the active list. So
1534                          * that executable code get better chances to stay in
1535                          * memory under moderate memory pressure.  Anon pages
1536                          * are not likely to be evicted by use-once streaming
1537                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1538                          * so we ignore them here.
1539                          */
1540                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1541                                 list_add(&page->lru, &l_active);
1542                                 continue;
1543                         }
1544                 }
1545
1546                 ClearPageActive(page);  /* we are de-activating */
1547                 list_add(&page->lru, &l_inactive);
1548         }
1549
1550         /*
1551          * Move pages back to the lru list.
1552          */
1553         spin_lock_irq(&zone->lru_lock);
1554         /*
1555          * Count referenced pages from currently used mappings as rotated,
1556          * even though only some of them are actually re-activated.  This
1557          * helps balance scan pressure between file and anonymous pages in
1558          * get_scan_ratio.
1559          */
1560         reclaim_stat->recent_rotated[file] += nr_rotated;
1561
1562         move_active_pages_to_lru(zone, &l_active,
1563                                                 LRU_ACTIVE + file * LRU_FILE);
1564         move_active_pages_to_lru(zone, &l_inactive,
1565                                                 LRU_BASE   + file * LRU_FILE);
1566         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1567         spin_unlock_irq(&zone->lru_lock);
1568 }
1569
1570 #ifdef CONFIG_SWAP
1571 static int inactive_anon_is_low_global(struct zone *zone)
1572 {
1573         unsigned long active, inactive;
1574
1575         active = zone_page_state(zone, NR_ACTIVE_ANON);
1576         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1577
1578         if (inactive * zone->inactive_ratio < active)
1579                 return 1;
1580
1581         return 0;
1582 }
1583
1584 /**
1585  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1586  * @zone: zone to check
1587  * @sc:   scan control of this context
1588  *
1589  * Returns true if the zone does not have enough inactive anon pages,
1590  * meaning some active anon pages need to be deactivated.
1591  */
1592 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1593 {
1594         int low;
1595
1596         /*
1597          * If we don't have swap space, anonymous page deactivation
1598          * is pointless.
1599          */
1600         if (!total_swap_pages)
1601                 return 0;
1602
1603         if (scanning_global_lru(sc))
1604                 low = inactive_anon_is_low_global(zone);
1605         else
1606                 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1607         return low;
1608 }
1609 #else
1610 static inline int inactive_anon_is_low(struct zone *zone,
1611                                         struct scan_control *sc)
1612 {
1613         return 0;
1614 }
1615 #endif
1616
1617 static int inactive_file_is_low_global(struct zone *zone)
1618 {
1619         unsigned long active, inactive;
1620
1621         active = zone_page_state(zone, NR_ACTIVE_FILE);
1622         inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1623
1624         return (active > inactive);
1625 }
1626
1627 /**
1628  * inactive_file_is_low - check if file pages need to be deactivated
1629  * @zone: zone to check
1630  * @sc:   scan control of this context
1631  *
1632  * When the system is doing streaming IO, memory pressure here
1633  * ensures that active file pages get deactivated, until more
1634  * than half of the file pages are on the inactive list.
1635  *
1636  * Once we get to that situation, protect the system's working
1637  * set from being evicted by disabling active file page aging.
1638  *
1639  * This uses a different ratio than the anonymous pages, because
1640  * the page cache uses a use-once replacement algorithm.
1641  */
1642 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1643 {
1644         int low;
1645
1646         if (scanning_global_lru(sc))
1647                 low = inactive_file_is_low_global(zone);
1648         else
1649                 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1650         return low;
1651 }
1652
1653 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1654                                 int file)
1655 {
1656         if (file)
1657                 return inactive_file_is_low(zone, sc);
1658         else
1659                 return inactive_anon_is_low(zone, sc);
1660 }
1661
1662 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1663         struct zone *zone, struct scan_control *sc, int priority)
1664 {
1665         int file = is_file_lru(lru);
1666
1667         if (is_active_lru(lru)) {
1668                 if (inactive_list_is_low(zone, sc, file))
1669                     shrink_active_list(nr_to_scan, zone, sc, priority, file);
1670                 return 0;
1671         }
1672
1673         return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1674 }
1675
1676 /*
1677  * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1678  * until we collected @swap_cluster_max pages to scan.
1679  */
1680 static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
1681                                        unsigned long *nr_saved_scan)
1682 {
1683         unsigned long nr;
1684
1685         *nr_saved_scan += nr_to_scan;
1686         nr = *nr_saved_scan;
1687
1688         if (nr >= SWAP_CLUSTER_MAX)
1689                 *nr_saved_scan = 0;
1690         else
1691                 nr = 0;
1692
1693         return nr;
1694 }
1695
1696 /*
1697  * Determine how aggressively the anon and file LRU lists should be
1698  * scanned.  The relative value of each set of LRU lists is determined
1699  * by looking at the fraction of the pages scanned we did rotate back
1700  * onto the active list instead of evict.
1701  *
1702  * nr[0] = anon pages to scan; nr[1] = file pages to scan
1703  */
1704 static void get_scan_count(struct zone *zone, struct scan_control *sc,
1705                                         unsigned long *nr, int priority)
1706 {
1707         unsigned long anon, file, free;
1708         unsigned long anon_prio, file_prio;
1709         unsigned long ap, fp;
1710         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1711         u64 fraction[2], denominator;
1712         enum lru_list l;
1713         int noswap = 0;
1714
1715         /* If we have no swap space, do not bother scanning anon pages. */
1716         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1717                 noswap = 1;
1718                 fraction[0] = 0;
1719                 fraction[1] = 1;
1720                 denominator = 1;
1721                 goto out;
1722         }
1723
1724         anon  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1725                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1726         file  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1727                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1728
1729         if (scanning_global_lru(sc)) {
1730                 free  = zone_page_state(zone, NR_FREE_PAGES);
1731                 /* If we have very few page cache pages,
1732                    force-scan anon pages. */
1733                 if (unlikely(file + free <= high_wmark_pages(zone))) {
1734                         fraction[0] = 1;
1735                         fraction[1] = 0;
1736                         denominator = 1;
1737                         goto out;
1738                 }
1739         }
1740
1741         /*
1742          * With swappiness at 100, anonymous and file have the same priority.
1743          * This scanning priority is essentially the inverse of IO cost.
1744          */
1745         anon_prio = sc->swappiness;
1746         file_prio = 200 - sc->swappiness;
1747
1748         /*
1749          * OK, so we have swap space and a fair amount of page cache
1750          * pages.  We use the recently rotated / recently scanned
1751          * ratios to determine how valuable each cache is.
1752          *
1753          * Because workloads change over time (and to avoid overflow)
1754          * we keep these statistics as a floating average, which ends
1755          * up weighing recent references more than old ones.
1756          *
1757          * anon in [0], file in [1]
1758          */
1759         spin_lock_irq(&zone->lru_lock);
1760         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1761                 reclaim_stat->recent_scanned[0] /= 2;
1762                 reclaim_stat->recent_rotated[0] /= 2;
1763         }
1764
1765         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1766                 reclaim_stat->recent_scanned[1] /= 2;
1767                 reclaim_stat->recent_rotated[1] /= 2;
1768         }
1769
1770         /*
1771          * The amount of pressure on anon vs file pages is inversely
1772          * proportional to the fraction of recently scanned pages on
1773          * each list that were recently referenced and in active use.
1774          */
1775         ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1776         ap /= reclaim_stat->recent_rotated[0] + 1;
1777
1778         fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1779         fp /= reclaim_stat->recent_rotated[1] + 1;
1780         spin_unlock_irq(&zone->lru_lock);
1781
1782         fraction[0] = ap;
1783         fraction[1] = fp;
1784         denominator = ap + fp + 1;
1785 out:
1786         for_each_evictable_lru(l) {
1787                 int file = is_file_lru(l);
1788                 unsigned long scan;
1789
1790                 scan = zone_nr_lru_pages(zone, sc, l);
1791                 if (priority || noswap) {
1792                         scan >>= priority;
1793                         scan = div64_u64(scan * fraction[file], denominator);
1794                 }
1795                 nr[l] = nr_scan_try_batch(scan,
1796                                           &reclaim_stat->nr_saved_scan[l]);
1797         }
1798 }
1799
1800 /*
1801  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1802  */
1803 static void shrink_zone(int priority, struct zone *zone,
1804                                 struct scan_control *sc)
1805 {
1806         unsigned long nr[NR_LRU_LISTS];
1807         unsigned long nr_to_scan;
1808         enum lru_list l;
1809         unsigned long nr_reclaimed = sc->nr_reclaimed;
1810         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1811
1812         get_scan_count(zone, sc, nr, priority);
1813
1814         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1815                                         nr[LRU_INACTIVE_FILE]) {
1816                 for_each_evictable_lru(l) {
1817                         if (nr[l]) {
1818                                 nr_to_scan = min_t(unsigned long,
1819                                                    nr[l], SWAP_CLUSTER_MAX);
1820                                 nr[l] -= nr_to_scan;
1821
1822                                 nr_reclaimed += shrink_list(l, nr_to_scan,
1823                                                             zone, sc, priority);
1824                         }
1825                 }
1826                 /*
1827                  * On large memory systems, scan >> priority can become
1828                  * really large. This is fine for the starting priority;
1829                  * we want to put equal scanning pressure on each zone.
1830                  * However, if the VM has a harder time of freeing pages,
1831                  * with multiple processes reclaiming pages, the total
1832                  * freeing target can get unreasonably large.
1833                  */
1834                 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
1835                         break;
1836         }
1837
1838         sc->nr_reclaimed = nr_reclaimed;
1839
1840         /*
1841          * Even if we did not try to evict anon pages at all, we want to
1842          * rebalance the anon lru active/inactive ratio.
1843          */
1844         if (inactive_anon_is_low(zone, sc))
1845                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1846
1847         throttle_vm_writeout(sc->gfp_mask);
1848 }
1849
1850 /*
1851  * This is the direct reclaim path, for page-allocating processes.  We only
1852  * try to reclaim pages from zones which will satisfy the caller's allocation
1853  * request.
1854  *
1855  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1856  * Because:
1857  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1858  *    allocation or
1859  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1860  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1861  *    zone defense algorithm.
1862  *
1863  * If a zone is deemed to be full of pinned pages then just give it a light
1864  * scan then give up on it.
1865  */
1866 static void shrink_zones(int priority, struct zonelist *zonelist,
1867                                         struct scan_control *sc)
1868 {
1869         struct zoneref *z;
1870         struct zone *zone;
1871
1872         for_each_zone_zonelist_nodemask(zone, z, zonelist,
1873                                         gfp_zone(sc->gfp_mask), sc->nodemask) {
1874                 if (!populated_zone(zone))
1875                         continue;
1876                 /*
1877                  * Take care memory controller reclaiming has small influence
1878                  * to global LRU.
1879                  */
1880                 if (scanning_global_lru(sc)) {
1881                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1882                                 continue;
1883                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1884                                 continue;       /* Let kswapd poll it */
1885                 }
1886
1887                 shrink_zone(priority, zone, sc);
1888         }
1889 }
1890
1891 static bool zone_reclaimable(struct zone *zone)
1892 {
1893         return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
1894 }
1895
1896 /*
1897  * As hibernation is going on, kswapd is freezed so that it can't mark
1898  * the zone into all_unreclaimable. It can't handle OOM during hibernation.
1899  * So let's check zone's unreclaimable in direct reclaim as well as kswapd.
1900  */
1901 static bool all_unreclaimable(struct zonelist *zonelist,
1902                 struct scan_control *sc)
1903 {
1904         struct zoneref *z;
1905         struct zone *zone;
1906         bool all_unreclaimable = true;
1907
1908         for_each_zone_zonelist_nodemask(zone, z, zonelist,
1909                         gfp_zone(sc->gfp_mask), sc->nodemask) {
1910                 if (!populated_zone(zone))
1911                         continue;
1912                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1913                         continue;
1914                 if (zone_reclaimable(zone)) {
1915                         all_unreclaimable = false;
1916                         break;
1917                 }
1918         }
1919
1920         return all_unreclaimable;
1921 }
1922
1923 /*
1924  * This is the main entry point to direct page reclaim.
1925  *
1926  * If a full scan of the inactive list fails to free enough memory then we
1927  * are "out of memory" and something needs to be killed.
1928  *
1929  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1930  * high - the zone may be full of dirty or under-writeback pages, which this
1931  * caller can't do much about.  We kick the writeback threads and take explicit
1932  * naps in the hope that some of these pages can be written.  But if the
1933  * allocating task holds filesystem locks which prevent writeout this might not
1934  * work, and the allocation attempt will fail.
1935  *
1936  * returns:     0, if no pages reclaimed
1937  *              else, the number of pages reclaimed
1938  */
1939 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
1940                                         struct scan_control *sc)
1941 {
1942         int priority;
1943         unsigned long total_scanned = 0;
1944         struct reclaim_state *reclaim_state = current->reclaim_state;
1945         struct zoneref *z;
1946         struct zone *zone;
1947         unsigned long writeback_threshold;
1948
1949         get_mems_allowed();
1950         delayacct_freepages_start();
1951
1952         if (scanning_global_lru(sc))
1953                 count_vm_event(ALLOCSTALL);
1954
1955         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1956                 sc->nr_scanned = 0;
1957                 if (!priority)
1958                         disable_swap_token();
1959                 shrink_zones(priority, zonelist, sc);
1960                 /*
1961                  * Don't shrink slabs when reclaiming memory from
1962                  * over limit cgroups
1963                  */
1964                 if (scanning_global_lru(sc)) {
1965                         unsigned long lru_pages = 0;
1966                         for_each_zone_zonelist(zone, z, zonelist,
1967                                         gfp_zone(sc->gfp_mask)) {
1968                                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1969                                         continue;
1970
1971                                 lru_pages += zone_reclaimable_pages(zone);
1972                         }
1973
1974                         shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
1975                         if (reclaim_state) {
1976                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
1977                                 reclaim_state->reclaimed_slab = 0;
1978                         }
1979                 }
1980                 total_scanned += sc->nr_scanned;
1981                 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
1982                         goto out;
1983
1984                 /*
1985                  * Try to write back as many pages as we just scanned.  This
1986                  * tends to cause slow streaming writers to write data to the
1987                  * disk smoothly, at the dirtying rate, which is nice.   But
1988                  * that's undesirable in laptop mode, where we *want* lumpy
1989                  * writeout.  So in laptop mode, write out the whole world.
1990                  */
1991                 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
1992                 if (total_scanned > writeback_threshold) {
1993                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
1994                         sc->may_writepage = 1;
1995                 }
1996
1997                 /* Take a nap, wait for some writeback to complete */
1998                 if (!sc->hibernation_mode && sc->nr_scanned &&
1999                     priority < DEF_PRIORITY - 2) {
2000                         struct zone *preferred_zone;
2001
2002                         first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2003                                                         NULL, &preferred_zone);
2004                         wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2005                 }
2006         }
2007
2008 out:
2009         delayacct_freepages_end();
2010         put_mems_allowed();
2011
2012         if (sc->nr_reclaimed)
2013                 return sc->nr_reclaimed;
2014
2015         /* top priority shrink_zones still had more to do? don't OOM, then */
2016         if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
2017                 return 1;
2018
2019         return 0;
2020 }
2021
2022 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2023                                 gfp_t gfp_mask, nodemask_t *nodemask)
2024 {
2025         unsigned long nr_reclaimed;
2026         struct scan_control sc = {
2027                 .gfp_mask = gfp_mask,
2028                 .may_writepage = !laptop_mode,
2029                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2030                 .may_unmap = 1,
2031                 .may_swap = 1,
2032                 .swappiness = vm_swappiness,
2033                 .order = order,
2034                 .mem_cgroup = NULL,
2035                 .nodemask = nodemask,
2036         };
2037
2038         trace_mm_vmscan_direct_reclaim_begin(order,
2039                                 sc.may_writepage,
2040                                 gfp_mask);
2041
2042         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2043
2044         trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2045
2046         return nr_reclaimed;
2047 }
2048
2049 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2050
2051 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2052                                                 gfp_t gfp_mask, bool noswap,
2053                                                 unsigned int swappiness,
2054                                                 struct zone *zone)
2055 {
2056         struct scan_control sc = {
2057                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2058                 .may_writepage = !laptop_mode,
2059                 .may_unmap = 1,
2060                 .may_swap = !noswap,
2061                 .swappiness = swappiness,
2062                 .order = 0,
2063                 .mem_cgroup = mem,
2064         };
2065         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2066                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2067
2068         trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2069                                                       sc.may_writepage,
2070                                                       sc.gfp_mask);
2071
2072         /*
2073          * NOTE: Although we can get the priority field, using it
2074          * here is not a good idea, since it limits the pages we can scan.
2075          * if we don't reclaim here, the shrink_zone from balance_pgdat
2076          * will pick up pages from other mem cgroup's as well. We hack
2077          * the priority and make it zero.
2078          */
2079         shrink_zone(0, zone, &sc);
2080
2081         trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2082
2083         return sc.nr_reclaimed;
2084 }
2085
2086 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
2087                                            gfp_t gfp_mask,
2088                                            bool noswap,
2089                                            unsigned int swappiness)
2090 {
2091         struct zonelist *zonelist;
2092         unsigned long nr_reclaimed;
2093         struct scan_control sc = {
2094                 .may_writepage = !laptop_mode,
2095                 .may_unmap = 1,
2096                 .may_swap = !noswap,
2097                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2098                 .swappiness = swappiness,
2099                 .order = 0,
2100                 .mem_cgroup = mem_cont,
2101                 .nodemask = NULL, /* we don't care the placement */
2102         };
2103
2104         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2105                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2106         zonelist = NODE_DATA(numa_node_id())->node_zonelists;
2107
2108         trace_mm_vmscan_memcg_reclaim_begin(0,
2109                                             sc.may_writepage,
2110                                             sc.gfp_mask);
2111
2112         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2113
2114         trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2115
2116         return nr_reclaimed;
2117 }
2118 #endif
2119
2120 /* is kswapd sleeping prematurely? */
2121 static int sleeping_prematurely(pg_data_t *pgdat, int order, long remaining)
2122 {
2123         int i;
2124
2125         /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2126         if (remaining)
2127                 return 1;
2128
2129         /* If after HZ/10, a zone is below the high mark, it's premature */
2130         for (i = 0; i < pgdat->nr_zones; i++) {
2131                 struct zone *zone = pgdat->node_zones + i;
2132
2133                 if (!populated_zone(zone))
2134                         continue;
2135
2136                 if (zone->all_unreclaimable)
2137                         continue;
2138
2139                 if (!zone_watermark_ok(zone, order, high_wmark_pages(zone),
2140                                                                 0, 0))
2141                         return 1;
2142         }
2143
2144         return 0;
2145 }
2146
2147 /*
2148  * For kswapd, balance_pgdat() will work across all this node's zones until
2149  * they are all at high_wmark_pages(zone).
2150  *
2151  * Returns the number of pages which were actually freed.
2152  *
2153  * There is special handling here for zones which are full of pinned pages.
2154  * This can happen if the pages are all mlocked, or if they are all used by
2155  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
2156  * What we do is to detect the case where all pages in the zone have been
2157  * scanned twice and there has been zero successful reclaim.  Mark the zone as
2158  * dead and from now on, only perform a short scan.  Basically we're polling
2159  * the zone for when the problem goes away.
2160  *
2161  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
2162  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2163  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2164  * lower zones regardless of the number of free pages in the lower zones. This
2165  * interoperates with the page allocator fallback scheme to ensure that aging
2166  * of pages is balanced across the zones.
2167  */
2168 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
2169 {
2170         int all_zones_ok;
2171         int priority;
2172         int i;
2173         unsigned long total_scanned;
2174         struct reclaim_state *reclaim_state = current->reclaim_state;
2175         struct scan_control sc = {
2176                 .gfp_mask = GFP_KERNEL,
2177                 .may_unmap = 1,
2178                 .may_swap = 1,
2179                 /*
2180                  * kswapd doesn't want to be bailed out while reclaim. because
2181                  * we want to put equal scanning pressure on each zone.
2182                  */
2183                 .nr_to_reclaim = ULONG_MAX,
2184                 .swappiness = vm_swappiness,
2185                 .order = order,
2186                 .mem_cgroup = NULL,
2187         };
2188 loop_again:
2189         total_scanned = 0;
2190         sc.nr_reclaimed = 0;
2191         sc.may_writepage = !laptop_mode;
2192         count_vm_event(PAGEOUTRUN);
2193
2194         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2195                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
2196                 unsigned long lru_pages = 0;
2197                 int has_under_min_watermark_zone = 0;
2198
2199                 /* The swap token gets in the way of swapout... */
2200                 if (!priority)
2201                         disable_swap_token();
2202
2203                 all_zones_ok = 1;
2204
2205                 /*
2206                  * Scan in the highmem->dma direction for the highest
2207                  * zone which needs scanning
2208                  */
2209                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2210                         struct zone *zone = pgdat->node_zones + i;
2211
2212                         if (!populated_zone(zone))
2213                                 continue;
2214
2215                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2216                                 continue;
2217
2218                         /*
2219                          * Do some background aging of the anon list, to give
2220                          * pages a chance to be referenced before reclaiming.
2221                          */
2222                         if (inactive_anon_is_low(zone, &sc))
2223                                 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2224                                                         &sc, priority, 0);
2225
2226                         if (!zone_watermark_ok(zone, order,
2227                                         high_wmark_pages(zone), 0, 0)) {
2228                                 end_zone = i;
2229                                 break;
2230                         }
2231                 }
2232                 if (i < 0)
2233                         goto out;
2234
2235                 for (i = 0; i <= end_zone; i++) {
2236                         struct zone *zone = pgdat->node_zones + i;
2237
2238                         lru_pages += zone_reclaimable_pages(zone);
2239                 }
2240
2241                 /*
2242                  * Now scan the zone in the dma->highmem direction, stopping
2243                  * at the last zone which needs scanning.
2244                  *
2245                  * We do this because the page allocator works in the opposite
2246                  * direction.  This prevents the page allocator from allocating
2247                  * pages behind kswapd's direction of progress, which would
2248                  * cause too much scanning of the lower zones.
2249                  */
2250                 for (i = 0; i <= end_zone; i++) {
2251                         struct zone *zone = pgdat->node_zones + i;
2252                         int nr_slab;
2253
2254                         if (!populated_zone(zone))
2255                                 continue;
2256
2257                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2258                                 continue;
2259
2260                         sc.nr_scanned = 0;
2261
2262                         /*
2263                          * Call soft limit reclaim before calling shrink_zone.
2264                          * For now we ignore the return value
2265                          */
2266                         mem_cgroup_soft_limit_reclaim(zone, order, sc.gfp_mask);
2267
2268                         /*
2269                          * We put equal pressure on every zone, unless one
2270                          * zone has way too many pages free already.
2271                          */
2272                         if (!zone_watermark_ok(zone, order,
2273                                         8*high_wmark_pages(zone), end_zone, 0))
2274                                 shrink_zone(priority, zone, &sc);
2275                         reclaim_state->reclaimed_slab = 0;
2276                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
2277                                                 lru_pages);
2278                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2279                         total_scanned += sc.nr_scanned;
2280                         if (zone->all_unreclaimable)
2281                                 continue;
2282                         if (nr_slab == 0 && !zone_reclaimable(zone))
2283                                 zone->all_unreclaimable = 1;
2284                         /*
2285                          * If we've done a decent amount of scanning and
2286                          * the reclaim ratio is low, start doing writepage
2287                          * even in laptop mode
2288                          */
2289                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2290                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2291                                 sc.may_writepage = 1;
2292
2293                         if (!zone_watermark_ok(zone, order,
2294                                         high_wmark_pages(zone), end_zone, 0)) {
2295                                 all_zones_ok = 0;
2296                                 /*
2297                                  * We are still under min water mark.  This
2298                                  * means that we have a GFP_ATOMIC allocation
2299                                  * failure risk. Hurry up!
2300                                  */
2301                                 if (!zone_watermark_ok(zone, order,
2302                                             min_wmark_pages(zone), end_zone, 0))
2303                                         has_under_min_watermark_zone = 1;
2304                         } else {
2305                                 /*
2306                                  * If a zone reaches its high watermark,
2307                                  * consider it to be no longer congested. It's
2308                                  * possible there are dirty pages backed by
2309                                  * congested BDIs but as pressure is relieved,
2310                                  * spectulatively avoid congestion waits
2311                                  */
2312                                 zone_clear_flag(zone, ZONE_CONGESTED);
2313                         }
2314
2315                 }
2316                 if (all_zones_ok)
2317                         break;          /* kswapd: all done */
2318                 /*
2319                  * OK, kswapd is getting into trouble.  Take a nap, then take
2320                  * another pass across the zones.
2321                  */
2322                 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2323                         if (has_under_min_watermark_zone)
2324                                 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2325                         else
2326                                 congestion_wait(BLK_RW_ASYNC, HZ/10);
2327                 }
2328
2329                 /*
2330                  * We do this so kswapd doesn't build up large priorities for
2331                  * example when it is freeing in parallel with allocators. It
2332                  * matches the direct reclaim path behaviour in terms of impact
2333                  * on zone->*_priority.
2334                  */
2335                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2336                         break;
2337         }
2338 out:
2339         if (!all_zones_ok) {
2340                 cond_resched();
2341
2342                 try_to_freeze();
2343
2344                 /*
2345                  * Fragmentation may mean that the system cannot be
2346                  * rebalanced for high-order allocations in all zones.
2347                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2348                  * it means the zones have been fully scanned and are still
2349                  * not balanced. For high-order allocations, there is
2350                  * little point trying all over again as kswapd may
2351                  * infinite loop.
2352                  *
2353                  * Instead, recheck all watermarks at order-0 as they
2354                  * are the most important. If watermarks are ok, kswapd will go
2355                  * back to sleep. High-order users can still perform direct
2356                  * reclaim if they wish.
2357                  */
2358                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2359                         order = sc.order = 0;
2360
2361                 goto loop_again;
2362         }
2363
2364         return sc.nr_reclaimed;
2365 }
2366
2367 /*
2368  * The background pageout daemon, started as a kernel thread
2369  * from the init process.
2370  *
2371  * This basically trickles out pages so that we have _some_
2372  * free memory available even if there is no other activity
2373  * that frees anything up. This is needed for things like routing
2374  * etc, where we otherwise might have all activity going on in
2375  * asynchronous contexts that cannot page things out.
2376  *
2377  * If there are applications that are active memory-allocators
2378  * (most normal use), this basically shouldn't matter.
2379  */
2380 static int kswapd(void *p)
2381 {
2382         unsigned long order;
2383         pg_data_t *pgdat = (pg_data_t*)p;
2384         struct task_struct *tsk = current;
2385         DEFINE_WAIT(wait);
2386         struct reclaim_state reclaim_state = {
2387                 .reclaimed_slab = 0,
2388         };
2389         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2390
2391         lockdep_set_current_reclaim_state(GFP_KERNEL);
2392
2393         if (!cpumask_empty(cpumask))
2394                 set_cpus_allowed_ptr(tsk, cpumask);
2395         current->reclaim_state = &reclaim_state;
2396
2397         /*
2398          * Tell the memory management that we're a "memory allocator",
2399          * and that if we need more memory we should get access to it
2400          * regardless (see "__alloc_pages()"). "kswapd" should
2401          * never get caught in the normal page freeing logic.
2402          *
2403          * (Kswapd normally doesn't need memory anyway, but sometimes
2404          * you need a small amount of memory in order to be able to
2405          * page out something else, and this flag essentially protects
2406          * us from recursively trying to free more memory as we're
2407          * trying to free the first piece of memory in the first place).
2408          */
2409         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2410         set_freezable();
2411
2412         order = 0;
2413         for ( ; ; ) {
2414                 unsigned long new_order;
2415                 int ret;
2416
2417                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2418                 new_order = pgdat->kswapd_max_order;
2419                 pgdat->kswapd_max_order = 0;
2420                 if (order < new_order) {
2421                         /*
2422                          * Don't sleep if someone wants a larger 'order'
2423                          * allocation
2424                          */
2425                         order = new_order;
2426                 } else {
2427                         if (!freezing(current) && !kthread_should_stop()) {
2428                                 long remaining = 0;
2429
2430                                 /* Try to sleep for a short interval */
2431                                 if (!sleeping_prematurely(pgdat, order, remaining)) {
2432                                         remaining = schedule_timeout(HZ/10);
2433                                         finish_wait(&pgdat->kswapd_wait, &wait);
2434                                         prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2435                                 }
2436
2437                                 /*
2438                                  * After a short sleep, check if it was a
2439                                  * premature sleep. If not, then go fully
2440                                  * to sleep until explicitly woken up
2441                                  */
2442                                 if (!sleeping_prematurely(pgdat, order, remaining)) {
2443                                         trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2444                                         schedule();
2445                                 } else {
2446                                         if (remaining)
2447                                                 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2448                                         else
2449                                                 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2450                                 }
2451                         }
2452
2453                         order = pgdat->kswapd_max_order;
2454                 }
2455                 finish_wait(&pgdat->kswapd_wait, &wait);
2456
2457                 ret = try_to_freeze();
2458                 if (kthread_should_stop())
2459                         break;
2460
2461                 /*
2462                  * We can speed up thawing tasks if we don't call balance_pgdat
2463                  * after returning from the refrigerator
2464                  */
2465                 if (!ret) {
2466                         trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2467                         balance_pgdat(pgdat, order);
2468                 }
2469         }
2470         return 0;
2471 }
2472
2473 /*
2474  * A zone is low on free memory, so wake its kswapd task to service it.
2475  */
2476 void wakeup_kswapd(struct zone *zone, int order)
2477 {
2478         pg_data_t *pgdat;
2479
2480         if (!populated_zone(zone))
2481                 return;
2482
2483         pgdat = zone->zone_pgdat;
2484         if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
2485                 return;
2486         if (pgdat->kswapd_max_order < order)
2487                 pgdat->kswapd_max_order = order;
2488         trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
2489         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2490                 return;
2491         if (!waitqueue_active(&pgdat->kswapd_wait))
2492                 return;
2493         wake_up_interruptible(&pgdat->kswapd_wait);
2494 }
2495
2496 /*
2497  * The reclaimable count would be mostly accurate.
2498  * The less reclaimable pages may be
2499  * - mlocked pages, which will be moved to unevictable list when encountered
2500  * - mapped pages, which may require several travels to be reclaimed
2501  * - dirty pages, which is not "instantly" reclaimable
2502  */
2503 unsigned long global_reclaimable_pages(void)
2504 {
2505         int nr;
2506
2507         nr = global_page_state(NR_ACTIVE_FILE) +
2508              global_page_state(NR_INACTIVE_FILE);
2509
2510         if (nr_swap_pages > 0)
2511                 nr += global_page_state(NR_ACTIVE_ANON) +
2512                       global_page_state(NR_INACTIVE_ANON);
2513
2514         return nr;
2515 }
2516
2517 unsigned long zone_reclaimable_pages(struct zone *zone)
2518 {
2519         int nr;
2520
2521         nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2522              zone_page_state(zone, NR_INACTIVE_FILE);
2523
2524         if (nr_swap_pages > 0)
2525                 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2526                       zone_page_state(zone, NR_INACTIVE_ANON);
2527
2528         return nr;
2529 }
2530
2531 #ifdef CONFIG_HIBERNATION
2532 /*
2533  * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2534  * freed pages.
2535  *
2536  * Rather than trying to age LRUs the aim is to preserve the overall
2537  * LRU order by reclaiming preferentially
2538  * inactive > active > active referenced > active mapped
2539  */
2540 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
2541 {
2542         struct reclaim_state reclaim_state;
2543         struct scan_control sc = {
2544                 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2545                 .may_swap = 1,
2546                 .may_unmap = 1,
2547                 .may_writepage = 1,
2548                 .nr_to_reclaim = nr_to_reclaim,
2549                 .hibernation_mode = 1,
2550                 .swappiness = vm_swappiness,
2551                 .order = 0,
2552         };
2553         struct zonelist * zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2554         struct task_struct *p = current;
2555         unsigned long nr_reclaimed;
2556
2557         p->flags |= PF_MEMALLOC;
2558         lockdep_set_current_reclaim_state(sc.gfp_mask);
2559         reclaim_state.reclaimed_slab = 0;
2560         p->reclaim_state = &reclaim_state;
2561
2562         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2563
2564         p->reclaim_state = NULL;
2565         lockdep_clear_current_reclaim_state();
2566         p->flags &= ~PF_MEMALLOC;
2567
2568         return nr_reclaimed;
2569 }
2570 #endif /* CONFIG_HIBERNATION */
2571
2572 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2573    not required for correctness.  So if the last cpu in a node goes
2574    away, we get changed to run anywhere: as the first one comes back,
2575    restore their cpu bindings. */
2576 static int __devinit cpu_callback(struct notifier_block *nfb,
2577                                   unsigned long action, void *hcpu)
2578 {
2579         int nid;
2580
2581         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2582                 for_each_node_state(nid, N_HIGH_MEMORY) {
2583                         pg_data_t *pgdat = NODE_DATA(nid);
2584                         const struct cpumask *mask;
2585
2586                         mask = cpumask_of_node(pgdat->node_id);
2587
2588                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2589                                 /* One of our CPUs online: restore mask */
2590                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2591                 }
2592         }
2593         return NOTIFY_OK;
2594 }
2595
2596 /*
2597  * This kswapd start function will be called by init and node-hot-add.
2598  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2599  */
2600 int kswapd_run(int nid)
2601 {
2602         pg_data_t *pgdat = NODE_DATA(nid);
2603         int ret = 0;
2604
2605         if (pgdat->kswapd)
2606                 return 0;
2607
2608         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2609         if (IS_ERR(pgdat->kswapd)) {
2610                 /* failure at boot is fatal */
2611                 BUG_ON(system_state == SYSTEM_BOOTING);
2612                 printk("Failed to start kswapd on node %d\n",nid);
2613                 ret = -1;
2614         }
2615         return ret;
2616 }
2617
2618 /*
2619  * Called by memory hotplug when all memory in a node is offlined.
2620  */
2621 void kswapd_stop(int nid)
2622 {
2623         struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
2624
2625         if (kswapd)
2626                 kthread_stop(kswapd);
2627 }
2628
2629 static int __init kswapd_init(void)
2630 {
2631         int nid;
2632
2633         swap_setup();
2634         for_each_node_state(nid, N_HIGH_MEMORY)
2635                 kswapd_run(nid);
2636         hotcpu_notifier(cpu_callback, 0);
2637         return 0;
2638 }
2639
2640 module_init(kswapd_init)
2641
2642 #ifdef CONFIG_NUMA
2643 /*
2644  * Zone reclaim mode
2645  *
2646  * If non-zero call zone_reclaim when the number of free pages falls below
2647  * the watermarks.
2648  */
2649 int zone_reclaim_mode __read_mostly;
2650
2651 #define RECLAIM_OFF 0
2652 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
2653 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
2654 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
2655
2656 /*
2657  * Priority for ZONE_RECLAIM. This determines the fraction of pages
2658  * of a node considered for each zone_reclaim. 4 scans 1/16th of
2659  * a zone.
2660  */
2661 #define ZONE_RECLAIM_PRIORITY 4
2662
2663 /*
2664  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2665  * occur.
2666  */
2667 int sysctl_min_unmapped_ratio = 1;
2668
2669 /*
2670  * If the number of slab pages in a zone grows beyond this percentage then
2671  * slab reclaim needs to occur.
2672  */
2673 int sysctl_min_slab_ratio = 5;
2674
2675 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
2676 {
2677         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
2678         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
2679                 zone_page_state(zone, NR_ACTIVE_FILE);
2680
2681         /*
2682          * It's possible for there to be more file mapped pages than
2683          * accounted for by the pages on the file LRU lists because
2684          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2685          */
2686         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
2687 }
2688
2689 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
2690 static long zone_pagecache_reclaimable(struct zone *zone)
2691 {
2692         long nr_pagecache_reclaimable;
2693         long delta = 0;
2694
2695         /*
2696          * If RECLAIM_SWAP is set, then all file pages are considered
2697          * potentially reclaimable. Otherwise, we have to worry about
2698          * pages like swapcache and zone_unmapped_file_pages() provides
2699          * a better estimate
2700          */
2701         if (zone_reclaim_mode & RECLAIM_SWAP)
2702                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
2703         else
2704                 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
2705
2706         /* If we can't clean pages, remove dirty pages from consideration */
2707         if (!(zone_reclaim_mode & RECLAIM_WRITE))
2708                 delta += zone_page_state(zone, NR_FILE_DIRTY);
2709
2710         /* Watch for any possible underflows due to delta */
2711         if (unlikely(delta > nr_pagecache_reclaimable))
2712                 delta = nr_pagecache_reclaimable;
2713
2714         return nr_pagecache_reclaimable - delta;
2715 }
2716
2717 /*
2718  * Try to free up some pages from this zone through reclaim.
2719  */
2720 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2721 {
2722         /* Minimum pages needed in order to stay on node */
2723         const unsigned long nr_pages = 1 << order;
2724         struct task_struct *p = current;
2725         struct reclaim_state reclaim_state;
2726         int priority;
2727         struct scan_control sc = {
2728                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2729                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2730                 .may_swap = 1,
2731                 .nr_to_reclaim = max_t(unsigned long, nr_pages,
2732                                        SWAP_CLUSTER_MAX),
2733                 .gfp_mask = gfp_mask,
2734                 .swappiness = vm_swappiness,
2735                 .order = order,
2736         };
2737         unsigned long nr_slab_pages0, nr_slab_pages1;
2738
2739         cond_resched();
2740         /*
2741          * We need to be able to allocate from the reserves for RECLAIM_SWAP
2742          * and we also need to be able to write out pages for RECLAIM_WRITE
2743          * and RECLAIM_SWAP.
2744          */
2745         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2746         lockdep_set_current_reclaim_state(gfp_mask);
2747         reclaim_state.reclaimed_slab = 0;
2748         p->reclaim_state = &reclaim_state;
2749
2750         if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
2751                 /*
2752                  * Free memory by calling shrink zone with increasing
2753                  * priorities until we have enough memory freed.
2754                  */
2755                 priority = ZONE_RECLAIM_PRIORITY;
2756                 do {
2757                         shrink_zone(priority, zone, &sc);
2758                         priority--;
2759                 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
2760         }
2761
2762         nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2763         if (nr_slab_pages0 > zone->min_slab_pages) {
2764                 /*
2765                  * shrink_slab() does not currently allow us to determine how
2766                  * many pages were freed in this zone. So we take the current
2767                  * number of slab pages and shake the slab until it is reduced
2768                  * by the same nr_pages that we used for reclaiming unmapped
2769                  * pages.
2770                  *
2771                  * Note that shrink_slab will free memory on all zones and may
2772                  * take a long time.
2773                  */
2774                 for (;;) {
2775                         unsigned long lru_pages = zone_reclaimable_pages(zone);
2776
2777                         /* No reclaimable slab or very low memory pressure */
2778                         if (!shrink_slab(sc.nr_scanned, gfp_mask, lru_pages))
2779                                 break;
2780
2781                         /* Freed enough memory */
2782                         nr_slab_pages1 = zone_page_state(zone,
2783                                                         NR_SLAB_RECLAIMABLE);
2784                         if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
2785                                 break;
2786                 }
2787
2788                 /*
2789                  * Update nr_reclaimed by the number of slab pages we
2790                  * reclaimed from this zone.
2791                  */
2792                 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2793                 if (nr_slab_pages1 < nr_slab_pages0)
2794                         sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
2795         }
2796
2797         p->reclaim_state = NULL;
2798         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
2799         lockdep_clear_current_reclaim_state();
2800         return sc.nr_reclaimed >= nr_pages;
2801 }
2802
2803 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2804 {
2805         int node_id;
2806         int ret;
2807
2808         /*
2809          * Zone reclaim reclaims unmapped file backed pages and
2810          * slab pages if we are over the defined limits.
2811          *
2812          * A small portion of unmapped file backed pages is needed for
2813          * file I/O otherwise pages read by file I/O will be immediately
2814          * thrown out if the zone is overallocated. So we do not reclaim
2815          * if less than a specified percentage of the zone is used by
2816          * unmapped file backed pages.
2817          */
2818         if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
2819             zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
2820                 return ZONE_RECLAIM_FULL;
2821
2822         if (zone->all_unreclaimable)
2823                 return ZONE_RECLAIM_FULL;
2824
2825         /*
2826          * Do not scan if the allocation should not be delayed.
2827          */
2828         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
2829                 return ZONE_RECLAIM_NOSCAN;
2830
2831         /*
2832          * Only run zone reclaim on the local zone or on zones that do not
2833          * have associated processors. This will favor the local processor
2834          * over remote processors and spread off node memory allocations
2835          * as wide as possible.
2836          */
2837         node_id = zone_to_nid(zone);
2838         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
2839                 return ZONE_RECLAIM_NOSCAN;
2840
2841         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2842                 return ZONE_RECLAIM_NOSCAN;
2843
2844         ret = __zone_reclaim(zone, gfp_mask, order);
2845         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2846
2847         if (!ret)
2848                 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
2849
2850         return ret;
2851 }
2852 #endif
2853
2854 /*
2855  * page_evictable - test whether a page is evictable
2856  * @page: the page to test
2857  * @vma: the VMA in which the page is or will be mapped, may be NULL
2858  *
2859  * Test whether page is evictable--i.e., should be placed on active/inactive
2860  * lists vs unevictable list.  The vma argument is !NULL when called from the
2861  * fault path to determine how to instantate a new page.
2862  *
2863  * Reasons page might not be evictable:
2864  * (1) page's mapping marked unevictable
2865  * (2) page is part of an mlocked VMA
2866  *
2867  */
2868 int page_evictable(struct page *page, struct vm_area_struct *vma)
2869 {
2870
2871         if (mapping_unevictable(page_mapping(page)))
2872                 return 0;
2873
2874         if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2875                 return 0;
2876
2877         return 1;
2878 }
2879
2880 /**
2881  * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2882  * @page: page to check evictability and move to appropriate lru list
2883  * @zone: zone page is in
2884  *
2885  * Checks a page for evictability and moves the page to the appropriate
2886  * zone lru list.
2887  *
2888  * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2889  * have PageUnevictable set.
2890  */
2891 static void check_move_unevictable_page(struct page *page, struct zone *zone)
2892 {
2893         VM_BUG_ON(PageActive(page));
2894
2895 retry:
2896         ClearPageUnevictable(page);
2897         if (page_evictable(page, NULL)) {
2898                 enum lru_list l = page_lru_base_type(page);
2899
2900                 __dec_zone_state(zone, NR_UNEVICTABLE);
2901                 list_move(&page->lru, &zone->lru[l].list);
2902                 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
2903                 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
2904                 __count_vm_event(UNEVICTABLE_PGRESCUED);
2905         } else {
2906                 /*
2907                  * rotate unevictable list
2908                  */
2909                 SetPageUnevictable(page);
2910                 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
2911                 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
2912                 if (page_evictable(page, NULL))
2913                         goto retry;
2914         }
2915 }
2916
2917 /**
2918  * scan_mapping_unevictable_pages - scan an address space for evictable pages
2919  * @mapping: struct address_space to scan for evictable pages
2920  *
2921  * Scan all pages in mapping.  Check unevictable pages for
2922  * evictability and move them to the appropriate zone lru list.
2923  */
2924 void scan_mapping_unevictable_pages(struct address_space *mapping)
2925 {
2926         pgoff_t next = 0;
2927         pgoff_t end   = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
2928                          PAGE_CACHE_SHIFT;
2929         struct zone *zone;
2930         struct pagevec pvec;
2931
2932         if (mapping->nrpages == 0)
2933                 return;
2934
2935         pagevec_init(&pvec, 0);
2936         while (next < end &&
2937                 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
2938                 int i;
2939                 int pg_scanned = 0;
2940
2941                 zone = NULL;
2942
2943                 for (i = 0; i < pagevec_count(&pvec); i++) {
2944                         struct page *page = pvec.pages[i];
2945                         pgoff_t page_index = page->index;
2946                         struct zone *pagezone = page_zone(page);
2947
2948                         pg_scanned++;
2949                         if (page_index > next)
2950                                 next = page_index;
2951                         next++;
2952
2953                         if (pagezone != zone) {
2954                                 if (zone)
2955                                         spin_unlock_irq(&zone->lru_lock);
2956                                 zone = pagezone;
2957                                 spin_lock_irq(&zone->lru_lock);
2958                         }
2959
2960                         if (PageLRU(page) && PageUnevictable(page))
2961                                 check_move_unevictable_page(page, zone);
2962                 }
2963                 if (zone)
2964                         spin_unlock_irq(&zone->lru_lock);
2965                 pagevec_release(&pvec);
2966
2967                 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
2968         }
2969
2970 }
2971
2972 /**
2973  * scan_zone_unevictable_pages - check unevictable list for evictable pages
2974  * @zone - zone of which to scan the unevictable list
2975  *
2976  * Scan @zone's unevictable LRU lists to check for pages that have become
2977  * evictable.  Move those that have to @zone's inactive list where they
2978  * become candidates for reclaim, unless shrink_inactive_zone() decides
2979  * to reactivate them.  Pages that are still unevictable are rotated
2980  * back onto @zone's unevictable list.
2981  */
2982 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
2983 static void scan_zone_unevictable_pages(struct zone *zone)
2984 {
2985         struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
2986         unsigned long scan;
2987         unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
2988
2989         while (nr_to_scan > 0) {
2990                 unsigned long batch_size = min(nr_to_scan,
2991                                                 SCAN_UNEVICTABLE_BATCH_SIZE);
2992
2993                 spin_lock_irq(&zone->lru_lock);
2994                 for (scan = 0;  scan < batch_size; scan++) {
2995                         struct page *page = lru_to_page(l_unevictable);
2996
2997                         if (!trylock_page(page))
2998                                 continue;
2999
3000                         prefetchw_prev_lru_page(page, l_unevictable, flags);
3001
3002                         if (likely(PageLRU(page) && PageUnevictable(page)))
3003                                 check_move_unevictable_page(page, zone);
3004
3005                         unlock_page(page);
3006                 }
3007                 spin_unlock_irq(&zone->lru_lock);
3008
3009                 nr_to_scan -= batch_size;
3010         }
3011 }
3012
3013
3014 /**
3015  * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
3016  *
3017  * A really big hammer:  scan all zones' unevictable LRU lists to check for
3018  * pages that have become evictable.  Move those back to the zones'
3019  * inactive list where they become candidates for reclaim.
3020  * This occurs when, e.g., we have unswappable pages on the unevictable lists,
3021  * and we add swap to the system.  As such, it runs in the context of a task
3022  * that has possibly/probably made some previously unevictable pages
3023  * evictable.
3024  */
3025 static void scan_all_zones_unevictable_pages(void)
3026 {
3027         struct zone *zone;
3028
3029         for_each_zone(zone) {
3030                 scan_zone_unevictable_pages(zone);
3031         }
3032 }
3033
3034 /*
3035  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
3036  * all nodes' unevictable lists for evictable pages
3037  */
3038 unsigned long scan_unevictable_pages;
3039
3040 int scan_unevictable_handler(struct ctl_table *table, int write,
3041                            void __user *buffer,
3042                            size_t *length, loff_t *ppos)
3043 {
3044         proc_doulongvec_minmax(table, write, buffer, length, ppos);
3045
3046         if (write && *(unsigned long *)table->data)
3047                 scan_all_zones_unevictable_pages();
3048
3049         scan_unevictable_pages = 0;
3050         return 0;
3051 }
3052
3053 #ifdef CONFIG_NUMA
3054 /*
3055  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
3056  * a specified node's per zone unevictable lists for evictable pages.
3057  */
3058
3059 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
3060                                           struct sysdev_attribute *attr,
3061                                           char *buf)
3062 {
3063         return sprintf(buf, "0\n");     /* always zero; should fit... */
3064 }
3065
3066 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
3067                                            struct sysdev_attribute *attr,
3068                                         const char *buf, size_t count)
3069 {
3070         struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
3071         struct zone *zone;
3072         unsigned long res;
3073         unsigned long req = strict_strtoul(buf, 10, &res);
3074
3075         if (!req)
3076                 return 1;       /* zero is no-op */
3077
3078         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
3079                 if (!populated_zone(zone))
3080                         continue;
3081                 scan_zone_unevictable_pages(zone);
3082         }
3083         return 1;
3084 }
3085
3086
3087 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3088                         read_scan_unevictable_node,
3089                         write_scan_unevictable_node);
3090
3091 int scan_unevictable_register_node(struct node *node)
3092 {
3093         return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
3094 }
3095
3096 void scan_unevictable_unregister_node(struct node *node)
3097 {
3098         sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
3099 }
3100 #endif