thp: use compaction in kswapd for GFP_ATOMIC order > 0
[linux-2.6.git] / mm / vmscan.c
1 /*
2  *  linux/mm/vmscan.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  *
6  *  Swap reorganised 29.12.95, Stephen Tweedie.
7  *  kswapd added: 7.1.96  sct
8  *  Removed kswapd_ctl limits, and swap out as many pages as needed
9  *  to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10  *  Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11  *  Multiqueue VM started 5.8.00, Rik van Riel.
12  */
13
14 #include <linux/mm.h>
15 #include <linux/module.h>
16 #include <linux/gfp.h>
17 #include <linux/kernel_stat.h>
18 #include <linux/swap.h>
19 #include <linux/pagemap.h>
20 #include <linux/init.h>
21 #include <linux/highmem.h>
22 #include <linux/vmstat.h>
23 #include <linux/file.h>
24 #include <linux/writeback.h>
25 #include <linux/blkdev.h>
26 #include <linux/buffer_head.h>  /* for try_to_release_page(),
27                                         buffer_heads_over_limit */
28 #include <linux/mm_inline.h>
29 #include <linux/pagevec.h>
30 #include <linux/backing-dev.h>
31 #include <linux/rmap.h>
32 #include <linux/topology.h>
33 #include <linux/cpu.h>
34 #include <linux/cpuset.h>
35 #include <linux/compaction.h>
36 #include <linux/notifier.h>
37 #include <linux/rwsem.h>
38 #include <linux/delay.h>
39 #include <linux/kthread.h>
40 #include <linux/freezer.h>
41 #include <linux/memcontrol.h>
42 #include <linux/delayacct.h>
43 #include <linux/sysctl.h>
44 #include <linux/compaction.h>
45
46 #include <asm/tlbflush.h>
47 #include <asm/div64.h>
48
49 #include <linux/swapops.h>
50
51 #include "internal.h"
52
53 #define CREATE_TRACE_POINTS
54 #include <trace/events/vmscan.h>
55
56 /*
57  * reclaim_mode determines how the inactive list is shrunk
58  * RECLAIM_MODE_SINGLE: Reclaim only order-0 pages
59  * RECLAIM_MODE_ASYNC:  Do not block
60  * RECLAIM_MODE_SYNC:   Allow blocking e.g. call wait_on_page_writeback
61  * RECLAIM_MODE_LUMPYRECLAIM: For high-order allocations, take a reference
62  *                      page from the LRU and reclaim all pages within a
63  *                      naturally aligned range
64  * RECLAIM_MODE_COMPACTION: For high-order allocations, reclaim a number of
65  *                      order-0 pages and then compact the zone
66  */
67 typedef unsigned __bitwise__ reclaim_mode_t;
68 #define RECLAIM_MODE_SINGLE             ((__force reclaim_mode_t)0x01u)
69 #define RECLAIM_MODE_ASYNC              ((__force reclaim_mode_t)0x02u)
70 #define RECLAIM_MODE_SYNC               ((__force reclaim_mode_t)0x04u)
71 #define RECLAIM_MODE_LUMPYRECLAIM       ((__force reclaim_mode_t)0x08u)
72 #define RECLAIM_MODE_COMPACTION         ((__force reclaim_mode_t)0x10u)
73
74 struct scan_control {
75         /* Incremented by the number of inactive pages that were scanned */
76         unsigned long nr_scanned;
77
78         /* Number of pages freed so far during a call to shrink_zones() */
79         unsigned long nr_reclaimed;
80
81         /* How many pages shrink_list() should reclaim */
82         unsigned long nr_to_reclaim;
83
84         unsigned long hibernation_mode;
85
86         /* This context's GFP mask */
87         gfp_t gfp_mask;
88
89         int may_writepage;
90
91         /* Can mapped pages be reclaimed? */
92         int may_unmap;
93
94         /* Can pages be swapped as part of reclaim? */
95         int may_swap;
96
97         int swappiness;
98
99         int order;
100
101         /*
102          * Intend to reclaim enough continuous memory rather than reclaim
103          * enough amount of memory. i.e, mode for high order allocation.
104          */
105         reclaim_mode_t reclaim_mode;
106
107         /* Which cgroup do we reclaim from */
108         struct mem_cgroup *mem_cgroup;
109
110         /*
111          * Nodemask of nodes allowed by the caller. If NULL, all nodes
112          * are scanned.
113          */
114         nodemask_t      *nodemask;
115 };
116
117 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
118
119 #ifdef ARCH_HAS_PREFETCH
120 #define prefetch_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                         prefetch(&prev->_field);                        \
127                 }                                                       \
128         } while (0)
129 #else
130 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
131 #endif
132
133 #ifdef ARCH_HAS_PREFETCHW
134 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
135         do {                                                            \
136                 if ((_page)->lru.prev != _base) {                       \
137                         struct page *prev;                              \
138                                                                         \
139                         prev = lru_to_page(&(_page->lru));              \
140                         prefetchw(&prev->_field);                       \
141                 }                                                       \
142         } while (0)
143 #else
144 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
145 #endif
146
147 /*
148  * From 0 .. 100.  Higher means more swappy.
149  */
150 int vm_swappiness = 60;
151 long vm_total_pages;    /* The total number of pages which the VM controls */
152
153 static LIST_HEAD(shrinker_list);
154 static DECLARE_RWSEM(shrinker_rwsem);
155
156 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
157 #define scanning_global_lru(sc) (!(sc)->mem_cgroup)
158 #else
159 #define scanning_global_lru(sc) (1)
160 #endif
161
162 static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
163                                                   struct scan_control *sc)
164 {
165         if (!scanning_global_lru(sc))
166                 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
167
168         return &zone->reclaim_stat;
169 }
170
171 static unsigned long zone_nr_lru_pages(struct zone *zone,
172                                 struct scan_control *sc, enum lru_list lru)
173 {
174         if (!scanning_global_lru(sc))
175                 return mem_cgroup_zone_nr_pages(sc->mem_cgroup, zone, lru);
176
177         return zone_page_state(zone, NR_LRU_BASE + lru);
178 }
179
180
181 /*
182  * Add a shrinker callback to be called from the vm
183  */
184 void register_shrinker(struct shrinker *shrinker)
185 {
186         shrinker->nr = 0;
187         down_write(&shrinker_rwsem);
188         list_add_tail(&shrinker->list, &shrinker_list);
189         up_write(&shrinker_rwsem);
190 }
191 EXPORT_SYMBOL(register_shrinker);
192
193 /*
194  * Remove one
195  */
196 void unregister_shrinker(struct shrinker *shrinker)
197 {
198         down_write(&shrinker_rwsem);
199         list_del(&shrinker->list);
200         up_write(&shrinker_rwsem);
201 }
202 EXPORT_SYMBOL(unregister_shrinker);
203
204 #define SHRINK_BATCH 128
205 /*
206  * Call the shrink functions to age shrinkable caches
207  *
208  * Here we assume it costs one seek to replace a lru page and that it also
209  * takes a seek to recreate a cache object.  With this in mind we age equal
210  * percentages of the lru and ageable caches.  This should balance the seeks
211  * generated by these structures.
212  *
213  * If the vm encountered mapped pages on the LRU it increase the pressure on
214  * slab to avoid swapping.
215  *
216  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
217  *
218  * `lru_pages' represents the number of on-LRU pages in all the zones which
219  * are eligible for the caller's allocation attempt.  It is used for balancing
220  * slab reclaim versus page reclaim.
221  *
222  * Returns the number of slab objects which we shrunk.
223  */
224 unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
225                         unsigned long lru_pages)
226 {
227         struct shrinker *shrinker;
228         unsigned long ret = 0;
229
230         if (scanned == 0)
231                 scanned = SWAP_CLUSTER_MAX;
232
233         if (!down_read_trylock(&shrinker_rwsem))
234                 return 1;       /* Assume we'll be able to shrink next time */
235
236         list_for_each_entry(shrinker, &shrinker_list, list) {
237                 unsigned long long delta;
238                 unsigned long total_scan;
239                 unsigned long max_pass;
240
241                 max_pass = (*shrinker->shrink)(shrinker, 0, gfp_mask);
242                 delta = (4 * scanned) / shrinker->seeks;
243                 delta *= max_pass;
244                 do_div(delta, lru_pages + 1);
245                 shrinker->nr += delta;
246                 if (shrinker->nr < 0) {
247                         printk(KERN_ERR "shrink_slab: %pF negative objects to "
248                                "delete nr=%ld\n",
249                                shrinker->shrink, shrinker->nr);
250                         shrinker->nr = max_pass;
251                 }
252
253                 /*
254                  * Avoid risking looping forever due to too large nr value:
255                  * never try to free more than twice the estimate number of
256                  * freeable entries.
257                  */
258                 if (shrinker->nr > max_pass * 2)
259                         shrinker->nr = max_pass * 2;
260
261                 total_scan = shrinker->nr;
262                 shrinker->nr = 0;
263
264                 while (total_scan >= SHRINK_BATCH) {
265                         long this_scan = SHRINK_BATCH;
266                         int shrink_ret;
267                         int nr_before;
268
269                         nr_before = (*shrinker->shrink)(shrinker, 0, gfp_mask);
270                         shrink_ret = (*shrinker->shrink)(shrinker, this_scan,
271                                                                 gfp_mask);
272                         if (shrink_ret == -1)
273                                 break;
274                         if (shrink_ret < nr_before)
275                                 ret += nr_before - shrink_ret;
276                         count_vm_events(SLABS_SCANNED, this_scan);
277                         total_scan -= this_scan;
278
279                         cond_resched();
280                 }
281
282                 shrinker->nr += total_scan;
283         }
284         up_read(&shrinker_rwsem);
285         return ret;
286 }
287
288 static void set_reclaim_mode(int priority, struct scan_control *sc,
289                                    bool sync)
290 {
291         reclaim_mode_t syncmode = sync ? RECLAIM_MODE_SYNC : RECLAIM_MODE_ASYNC;
292
293         /*
294          * Initially assume we are entering either lumpy reclaim or
295          * reclaim/compaction.Depending on the order, we will either set the
296          * sync mode or just reclaim order-0 pages later.
297          */
298         if (COMPACTION_BUILD)
299                 sc->reclaim_mode = RECLAIM_MODE_COMPACTION;
300         else
301                 sc->reclaim_mode = RECLAIM_MODE_LUMPYRECLAIM;
302
303         /*
304          * Avoid using lumpy reclaim or reclaim/compaction if possible by
305          * restricting when its set to either costly allocations or when
306          * under memory pressure
307          */
308         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
309                 sc->reclaim_mode |= syncmode;
310         else if (sc->order && priority < DEF_PRIORITY - 2)
311                 sc->reclaim_mode |= syncmode;
312         else
313                 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
314 }
315
316 static void reset_reclaim_mode(struct scan_control *sc)
317 {
318         sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
319 }
320
321 static inline int is_page_cache_freeable(struct page *page)
322 {
323         /*
324          * A freeable page cache page is referenced only by the caller
325          * that isolated the page, the page cache radix tree and
326          * optional buffer heads at page->private.
327          */
328         return page_count(page) - page_has_private(page) == 2;
329 }
330
331 static int may_write_to_queue(struct backing_dev_info *bdi,
332                               struct scan_control *sc)
333 {
334         if (current->flags & PF_SWAPWRITE)
335                 return 1;
336         if (!bdi_write_congested(bdi))
337                 return 1;
338         if (bdi == current->backing_dev_info)
339                 return 1;
340
341         /* lumpy reclaim for hugepage often need a lot of write */
342         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
343                 return 1;
344         return 0;
345 }
346
347 /*
348  * We detected a synchronous write error writing a page out.  Probably
349  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
350  * fsync(), msync() or close().
351  *
352  * The tricky part is that after writepage we cannot touch the mapping: nothing
353  * prevents it from being freed up.  But we have a ref on the page and once
354  * that page is locked, the mapping is pinned.
355  *
356  * We're allowed to run sleeping lock_page() here because we know the caller has
357  * __GFP_FS.
358  */
359 static void handle_write_error(struct address_space *mapping,
360                                 struct page *page, int error)
361 {
362         lock_page_nosync(page);
363         if (page_mapping(page) == mapping)
364                 mapping_set_error(mapping, error);
365         unlock_page(page);
366 }
367
368 /* possible outcome of pageout() */
369 typedef enum {
370         /* failed to write page out, page is locked */
371         PAGE_KEEP,
372         /* move page to the active list, page is locked */
373         PAGE_ACTIVATE,
374         /* page has been sent to the disk successfully, page is unlocked */
375         PAGE_SUCCESS,
376         /* page is clean and locked */
377         PAGE_CLEAN,
378 } pageout_t;
379
380 /*
381  * pageout is called by shrink_page_list() for each dirty page.
382  * Calls ->writepage().
383  */
384 static pageout_t pageout(struct page *page, struct address_space *mapping,
385                          struct scan_control *sc)
386 {
387         /*
388          * If the page is dirty, only perform writeback if that write
389          * will be non-blocking.  To prevent this allocation from being
390          * stalled by pagecache activity.  But note that there may be
391          * stalls if we need to run get_block().  We could test
392          * PagePrivate for that.
393          *
394          * If this process is currently in __generic_file_aio_write() against
395          * this page's queue, we can perform writeback even if that
396          * will block.
397          *
398          * If the page is swapcache, write it back even if that would
399          * block, for some throttling. This happens by accident, because
400          * swap_backing_dev_info is bust: it doesn't reflect the
401          * congestion state of the swapdevs.  Easy to fix, if needed.
402          */
403         if (!is_page_cache_freeable(page))
404                 return PAGE_KEEP;
405         if (!mapping) {
406                 /*
407                  * Some data journaling orphaned pages can have
408                  * page->mapping == NULL while being dirty with clean buffers.
409                  */
410                 if (page_has_private(page)) {
411                         if (try_to_free_buffers(page)) {
412                                 ClearPageDirty(page);
413                                 printk("%s: orphaned page\n", __func__);
414                                 return PAGE_CLEAN;
415                         }
416                 }
417                 return PAGE_KEEP;
418         }
419         if (mapping->a_ops->writepage == NULL)
420                 return PAGE_ACTIVATE;
421         if (!may_write_to_queue(mapping->backing_dev_info, sc))
422                 return PAGE_KEEP;
423
424         if (clear_page_dirty_for_io(page)) {
425                 int res;
426                 struct writeback_control wbc = {
427                         .sync_mode = WB_SYNC_NONE,
428                         .nr_to_write = SWAP_CLUSTER_MAX,
429                         .range_start = 0,
430                         .range_end = LLONG_MAX,
431                         .for_reclaim = 1,
432                 };
433
434                 SetPageReclaim(page);
435                 res = mapping->a_ops->writepage(page, &wbc);
436                 if (res < 0)
437                         handle_write_error(mapping, page, res);
438                 if (res == AOP_WRITEPAGE_ACTIVATE) {
439                         ClearPageReclaim(page);
440                         return PAGE_ACTIVATE;
441                 }
442
443                 /*
444                  * Wait on writeback if requested to. This happens when
445                  * direct reclaiming a large contiguous area and the
446                  * first attempt to free a range of pages fails.
447                  */
448                 if (PageWriteback(page) &&
449                     (sc->reclaim_mode & RECLAIM_MODE_SYNC))
450                         wait_on_page_writeback(page);
451
452                 if (!PageWriteback(page)) {
453                         /* synchronous write or broken a_ops? */
454                         ClearPageReclaim(page);
455                 }
456                 trace_mm_vmscan_writepage(page,
457                         trace_reclaim_flags(page, sc->reclaim_mode));
458                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
459                 return PAGE_SUCCESS;
460         }
461
462         return PAGE_CLEAN;
463 }
464
465 /*
466  * Same as remove_mapping, but if the page is removed from the mapping, it
467  * gets returned with a refcount of 0.
468  */
469 static int __remove_mapping(struct address_space *mapping, struct page *page)
470 {
471         BUG_ON(!PageLocked(page));
472         BUG_ON(mapping != page_mapping(page));
473
474         spin_lock_irq(&mapping->tree_lock);
475         /*
476          * The non racy check for a busy page.
477          *
478          * Must be careful with the order of the tests. When someone has
479          * a ref to the page, it may be possible that they dirty it then
480          * drop the reference. So if PageDirty is tested before page_count
481          * here, then the following race may occur:
482          *
483          * get_user_pages(&page);
484          * [user mapping goes away]
485          * write_to(page);
486          *                              !PageDirty(page)    [good]
487          * SetPageDirty(page);
488          * put_page(page);
489          *                              !page_count(page)   [good, discard it]
490          *
491          * [oops, our write_to data is lost]
492          *
493          * Reversing the order of the tests ensures such a situation cannot
494          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
495          * load is not satisfied before that of page->_count.
496          *
497          * Note that if SetPageDirty is always performed via set_page_dirty,
498          * and thus under tree_lock, then this ordering is not required.
499          */
500         if (!page_freeze_refs(page, 2))
501                 goto cannot_free;
502         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
503         if (unlikely(PageDirty(page))) {
504                 page_unfreeze_refs(page, 2);
505                 goto cannot_free;
506         }
507
508         if (PageSwapCache(page)) {
509                 swp_entry_t swap = { .val = page_private(page) };
510                 __delete_from_swap_cache(page);
511                 spin_unlock_irq(&mapping->tree_lock);
512                 swapcache_free(swap, page);
513         } else {
514                 void (*freepage)(struct page *);
515
516                 freepage = mapping->a_ops->freepage;
517
518                 __remove_from_page_cache(page);
519                 spin_unlock_irq(&mapping->tree_lock);
520                 mem_cgroup_uncharge_cache_page(page);
521
522                 if (freepage != NULL)
523                         freepage(page);
524         }
525
526         return 1;
527
528 cannot_free:
529         spin_unlock_irq(&mapping->tree_lock);
530         return 0;
531 }
532
533 /*
534  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
535  * someone else has a ref on the page, abort and return 0.  If it was
536  * successfully detached, return 1.  Assumes the caller has a single ref on
537  * this page.
538  */
539 int remove_mapping(struct address_space *mapping, struct page *page)
540 {
541         if (__remove_mapping(mapping, page)) {
542                 /*
543                  * Unfreezing the refcount with 1 rather than 2 effectively
544                  * drops the pagecache ref for us without requiring another
545                  * atomic operation.
546                  */
547                 page_unfreeze_refs(page, 1);
548                 return 1;
549         }
550         return 0;
551 }
552
553 /**
554  * putback_lru_page - put previously isolated page onto appropriate LRU list
555  * @page: page to be put back to appropriate lru list
556  *
557  * Add previously isolated @page to appropriate LRU list.
558  * Page may still be unevictable for other reasons.
559  *
560  * lru_lock must not be held, interrupts must be enabled.
561  */
562 void putback_lru_page(struct page *page)
563 {
564         int lru;
565         int active = !!TestClearPageActive(page);
566         int was_unevictable = PageUnevictable(page);
567
568         VM_BUG_ON(PageLRU(page));
569
570 redo:
571         ClearPageUnevictable(page);
572
573         if (page_evictable(page, NULL)) {
574                 /*
575                  * For evictable pages, we can use the cache.
576                  * In event of a race, worst case is we end up with an
577                  * unevictable page on [in]active list.
578                  * We know how to handle that.
579                  */
580                 lru = active + page_lru_base_type(page);
581                 lru_cache_add_lru(page, lru);
582         } else {
583                 /*
584                  * Put unevictable pages directly on zone's unevictable
585                  * list.
586                  */
587                 lru = LRU_UNEVICTABLE;
588                 add_page_to_unevictable_list(page);
589                 /*
590                  * When racing with an mlock clearing (page is
591                  * unlocked), make sure that if the other thread does
592                  * not observe our setting of PG_lru and fails
593                  * isolation, we see PG_mlocked cleared below and move
594                  * the page back to the evictable list.
595                  *
596                  * The other side is TestClearPageMlocked().
597                  */
598                 smp_mb();
599         }
600
601         /*
602          * page's status can change while we move it among lru. If an evictable
603          * page is on unevictable list, it never be freed. To avoid that,
604          * check after we added it to the list, again.
605          */
606         if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
607                 if (!isolate_lru_page(page)) {
608                         put_page(page);
609                         goto redo;
610                 }
611                 /* This means someone else dropped this page from LRU
612                  * So, it will be freed or putback to LRU again. There is
613                  * nothing to do here.
614                  */
615         }
616
617         if (was_unevictable && lru != LRU_UNEVICTABLE)
618                 count_vm_event(UNEVICTABLE_PGRESCUED);
619         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
620                 count_vm_event(UNEVICTABLE_PGCULLED);
621
622         put_page(page);         /* drop ref from isolate */
623 }
624
625 enum page_references {
626         PAGEREF_RECLAIM,
627         PAGEREF_RECLAIM_CLEAN,
628         PAGEREF_KEEP,
629         PAGEREF_ACTIVATE,
630 };
631
632 static enum page_references page_check_references(struct page *page,
633                                                   struct scan_control *sc)
634 {
635         int referenced_ptes, referenced_page;
636         unsigned long vm_flags;
637
638         referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
639         referenced_page = TestClearPageReferenced(page);
640
641         /* Lumpy reclaim - ignore references */
642         if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
643                 return PAGEREF_RECLAIM;
644
645         /*
646          * Mlock lost the isolation race with us.  Let try_to_unmap()
647          * move the page to the unevictable list.
648          */
649         if (vm_flags & VM_LOCKED)
650                 return PAGEREF_RECLAIM;
651
652         if (referenced_ptes) {
653                 if (PageAnon(page))
654                         return PAGEREF_ACTIVATE;
655                 /*
656                  * All mapped pages start out with page table
657                  * references from the instantiating fault, so we need
658                  * to look twice if a mapped file page is used more
659                  * than once.
660                  *
661                  * Mark it and spare it for another trip around the
662                  * inactive list.  Another page table reference will
663                  * lead to its activation.
664                  *
665                  * Note: the mark is set for activated pages as well
666                  * so that recently deactivated but used pages are
667                  * quickly recovered.
668                  */
669                 SetPageReferenced(page);
670
671                 if (referenced_page)
672                         return PAGEREF_ACTIVATE;
673
674                 return PAGEREF_KEEP;
675         }
676
677         /* Reclaim if clean, defer dirty pages to writeback */
678         if (referenced_page && !PageSwapBacked(page))
679                 return PAGEREF_RECLAIM_CLEAN;
680
681         return PAGEREF_RECLAIM;
682 }
683
684 static noinline_for_stack void free_page_list(struct list_head *free_pages)
685 {
686         struct pagevec freed_pvec;
687         struct page *page, *tmp;
688
689         pagevec_init(&freed_pvec, 1);
690
691         list_for_each_entry_safe(page, tmp, free_pages, lru) {
692                 list_del(&page->lru);
693                 if (!pagevec_add(&freed_pvec, page)) {
694                         __pagevec_free(&freed_pvec);
695                         pagevec_reinit(&freed_pvec);
696                 }
697         }
698
699         pagevec_free(&freed_pvec);
700 }
701
702 /*
703  * shrink_page_list() returns the number of reclaimed pages
704  */
705 static unsigned long shrink_page_list(struct list_head *page_list,
706                                       struct zone *zone,
707                                       struct scan_control *sc)
708 {
709         LIST_HEAD(ret_pages);
710         LIST_HEAD(free_pages);
711         int pgactivate = 0;
712         unsigned long nr_dirty = 0;
713         unsigned long nr_congested = 0;
714         unsigned long nr_reclaimed = 0;
715
716         cond_resched();
717
718         while (!list_empty(page_list)) {
719                 enum page_references references;
720                 struct address_space *mapping;
721                 struct page *page;
722                 int may_enter_fs;
723
724                 cond_resched();
725
726                 page = lru_to_page(page_list);
727                 list_del(&page->lru);
728
729                 if (!trylock_page(page))
730                         goto keep;
731
732                 VM_BUG_ON(PageActive(page));
733                 VM_BUG_ON(page_zone(page) != zone);
734
735                 sc->nr_scanned++;
736
737                 if (unlikely(!page_evictable(page, NULL)))
738                         goto cull_mlocked;
739
740                 if (!sc->may_unmap && page_mapped(page))
741                         goto keep_locked;
742
743                 /* Double the slab pressure for mapped and swapcache pages */
744                 if (page_mapped(page) || PageSwapCache(page))
745                         sc->nr_scanned++;
746
747                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
748                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
749
750                 if (PageWriteback(page)) {
751                         /*
752                          * Synchronous reclaim is performed in two passes,
753                          * first an asynchronous pass over the list to
754                          * start parallel writeback, and a second synchronous
755                          * pass to wait for the IO to complete.  Wait here
756                          * for any page for which writeback has already
757                          * started.
758                          */
759                         if ((sc->reclaim_mode & RECLAIM_MODE_SYNC) &&
760                             may_enter_fs)
761                                 wait_on_page_writeback(page);
762                         else {
763                                 unlock_page(page);
764                                 goto keep_lumpy;
765                         }
766                 }
767
768                 references = page_check_references(page, sc);
769                 switch (references) {
770                 case PAGEREF_ACTIVATE:
771                         goto activate_locked;
772                 case PAGEREF_KEEP:
773                         goto keep_locked;
774                 case PAGEREF_RECLAIM:
775                 case PAGEREF_RECLAIM_CLEAN:
776                         ; /* try to reclaim the page below */
777                 }
778
779                 /*
780                  * Anonymous process memory has backing store?
781                  * Try to allocate it some swap space here.
782                  */
783                 if (PageAnon(page) && !PageSwapCache(page)) {
784                         if (!(sc->gfp_mask & __GFP_IO))
785                                 goto keep_locked;
786                         if (!add_to_swap(page))
787                                 goto activate_locked;
788                         may_enter_fs = 1;
789                 }
790
791                 mapping = page_mapping(page);
792
793                 /*
794                  * The page is mapped into the page tables of one or more
795                  * processes. Try to unmap it here.
796                  */
797                 if (page_mapped(page) && mapping) {
798                         switch (try_to_unmap(page, TTU_UNMAP)) {
799                         case SWAP_FAIL:
800                                 goto activate_locked;
801                         case SWAP_AGAIN:
802                                 goto keep_locked;
803                         case SWAP_MLOCK:
804                                 goto cull_mlocked;
805                         case SWAP_SUCCESS:
806                                 ; /* try to free the page below */
807                         }
808                 }
809
810                 if (PageDirty(page)) {
811                         nr_dirty++;
812
813                         if (references == PAGEREF_RECLAIM_CLEAN)
814                                 goto keep_locked;
815                         if (!may_enter_fs)
816                                 goto keep_locked;
817                         if (!sc->may_writepage)
818                                 goto keep_locked;
819
820                         /* Page is dirty, try to write it out here */
821                         switch (pageout(page, mapping, sc)) {
822                         case PAGE_KEEP:
823                                 nr_congested++;
824                                 goto keep_locked;
825                         case PAGE_ACTIVATE:
826                                 goto activate_locked;
827                         case PAGE_SUCCESS:
828                                 if (PageWriteback(page))
829                                         goto keep_lumpy;
830                                 if (PageDirty(page))
831                                         goto keep;
832
833                                 /*
834                                  * A synchronous write - probably a ramdisk.  Go
835                                  * ahead and try to reclaim the page.
836                                  */
837                                 if (!trylock_page(page))
838                                         goto keep;
839                                 if (PageDirty(page) || PageWriteback(page))
840                                         goto keep_locked;
841                                 mapping = page_mapping(page);
842                         case PAGE_CLEAN:
843                                 ; /* try to free the page below */
844                         }
845                 }
846
847                 /*
848                  * If the page has buffers, try to free the buffer mappings
849                  * associated with this page. If we succeed we try to free
850                  * the page as well.
851                  *
852                  * We do this even if the page is PageDirty().
853                  * try_to_release_page() does not perform I/O, but it is
854                  * possible for a page to have PageDirty set, but it is actually
855                  * clean (all its buffers are clean).  This happens if the
856                  * buffers were written out directly, with submit_bh(). ext3
857                  * will do this, as well as the blockdev mapping.
858                  * try_to_release_page() will discover that cleanness and will
859                  * drop the buffers and mark the page clean - it can be freed.
860                  *
861                  * Rarely, pages can have buffers and no ->mapping.  These are
862                  * the pages which were not successfully invalidated in
863                  * truncate_complete_page().  We try to drop those buffers here
864                  * and if that worked, and the page is no longer mapped into
865                  * process address space (page_count == 1) it can be freed.
866                  * Otherwise, leave the page on the LRU so it is swappable.
867                  */
868                 if (page_has_private(page)) {
869                         if (!try_to_release_page(page, sc->gfp_mask))
870                                 goto activate_locked;
871                         if (!mapping && page_count(page) == 1) {
872                                 unlock_page(page);
873                                 if (put_page_testzero(page))
874                                         goto free_it;
875                                 else {
876                                         /*
877                                          * rare race with speculative reference.
878                                          * the speculative reference will free
879                                          * this page shortly, so we may
880                                          * increment nr_reclaimed here (and
881                                          * leave it off the LRU).
882                                          */
883                                         nr_reclaimed++;
884                                         continue;
885                                 }
886                         }
887                 }
888
889                 if (!mapping || !__remove_mapping(mapping, page))
890                         goto keep_locked;
891
892                 /*
893                  * At this point, we have no other references and there is
894                  * no way to pick any more up (removed from LRU, removed
895                  * from pagecache). Can use non-atomic bitops now (and
896                  * we obviously don't have to worry about waking up a process
897                  * waiting on the page lock, because there are no references.
898                  */
899                 __clear_page_locked(page);
900 free_it:
901                 nr_reclaimed++;
902
903                 /*
904                  * Is there need to periodically free_page_list? It would
905                  * appear not as the counts should be low
906                  */
907                 list_add(&page->lru, &free_pages);
908                 continue;
909
910 cull_mlocked:
911                 if (PageSwapCache(page))
912                         try_to_free_swap(page);
913                 unlock_page(page);
914                 putback_lru_page(page);
915                 reset_reclaim_mode(sc);
916                 continue;
917
918 activate_locked:
919                 /* Not a candidate for swapping, so reclaim swap space. */
920                 if (PageSwapCache(page) && vm_swap_full())
921                         try_to_free_swap(page);
922                 VM_BUG_ON(PageActive(page));
923                 SetPageActive(page);
924                 pgactivate++;
925 keep_locked:
926                 unlock_page(page);
927 keep:
928                 reset_reclaim_mode(sc);
929 keep_lumpy:
930                 list_add(&page->lru, &ret_pages);
931                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
932         }
933
934         /*
935          * Tag a zone as congested if all the dirty pages encountered were
936          * backed by a congested BDI. In this case, reclaimers should just
937          * back off and wait for congestion to clear because further reclaim
938          * will encounter the same problem
939          */
940         if (nr_dirty == nr_congested && nr_dirty != 0)
941                 zone_set_flag(zone, ZONE_CONGESTED);
942
943         free_page_list(&free_pages);
944
945         list_splice(&ret_pages, page_list);
946         count_vm_events(PGACTIVATE, pgactivate);
947         return nr_reclaimed;
948 }
949
950 /*
951  * Attempt to remove the specified page from its LRU.  Only take this page
952  * if it is of the appropriate PageActive status.  Pages which are being
953  * freed elsewhere are also ignored.
954  *
955  * page:        page to consider
956  * mode:        one of the LRU isolation modes defined above
957  *
958  * returns 0 on success, -ve errno on failure.
959  */
960 int __isolate_lru_page(struct page *page, int mode, int file)
961 {
962         int ret = -EINVAL;
963
964         /* Only take pages on the LRU. */
965         if (!PageLRU(page))
966                 return ret;
967
968         /*
969          * When checking the active state, we need to be sure we are
970          * dealing with comparible boolean values.  Take the logical not
971          * of each.
972          */
973         if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
974                 return ret;
975
976         if (mode != ISOLATE_BOTH && page_is_file_cache(page) != file)
977                 return ret;
978
979         /*
980          * When this function is being called for lumpy reclaim, we
981          * initially look into all LRU pages, active, inactive and
982          * unevictable; only give shrink_page_list evictable pages.
983          */
984         if (PageUnevictable(page))
985                 return ret;
986
987         ret = -EBUSY;
988
989         if (likely(get_page_unless_zero(page))) {
990                 /*
991                  * Be careful not to clear PageLRU until after we're
992                  * sure the page is not being freed elsewhere -- the
993                  * page release code relies on it.
994                  */
995                 ClearPageLRU(page);
996                 ret = 0;
997         }
998
999         return ret;
1000 }
1001
1002 /*
1003  * zone->lru_lock is heavily contended.  Some of the functions that
1004  * shrink the lists perform better by taking out a batch of pages
1005  * and working on them outside the LRU lock.
1006  *
1007  * For pagecache intensive workloads, this function is the hottest
1008  * spot in the kernel (apart from copy_*_user functions).
1009  *
1010  * Appropriate locks must be held before calling this function.
1011  *
1012  * @nr_to_scan: The number of pages to look through on the list.
1013  * @src:        The LRU list to pull pages off.
1014  * @dst:        The temp list to put pages on to.
1015  * @scanned:    The number of pages that were scanned.
1016  * @order:      The caller's attempted allocation order
1017  * @mode:       One of the LRU isolation modes
1018  * @file:       True [1] if isolating file [!anon] pages
1019  *
1020  * returns how many pages were moved onto *@dst.
1021  */
1022 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1023                 struct list_head *src, struct list_head *dst,
1024                 unsigned long *scanned, int order, int mode, int file)
1025 {
1026         unsigned long nr_taken = 0;
1027         unsigned long nr_lumpy_taken = 0;
1028         unsigned long nr_lumpy_dirty = 0;
1029         unsigned long nr_lumpy_failed = 0;
1030         unsigned long scan;
1031
1032         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1033                 struct page *page;
1034                 unsigned long pfn;
1035                 unsigned long end_pfn;
1036                 unsigned long page_pfn;
1037                 int zone_id;
1038
1039                 page = lru_to_page(src);
1040                 prefetchw_prev_lru_page(page, src, flags);
1041
1042                 VM_BUG_ON(!PageLRU(page));
1043
1044                 switch (__isolate_lru_page(page, mode, file)) {
1045                 case 0:
1046                         list_move(&page->lru, dst);
1047                         mem_cgroup_del_lru(page);
1048                         nr_taken++;
1049                         break;
1050
1051                 case -EBUSY:
1052                         /* else it is being freed elsewhere */
1053                         list_move(&page->lru, src);
1054                         mem_cgroup_rotate_lru_list(page, page_lru(page));
1055                         continue;
1056
1057                 default:
1058                         BUG();
1059                 }
1060
1061                 if (!order)
1062                         continue;
1063
1064                 /*
1065                  * Attempt to take all pages in the order aligned region
1066                  * surrounding the tag page.  Only take those pages of
1067                  * the same active state as that tag page.  We may safely
1068                  * round the target page pfn down to the requested order
1069                  * as the mem_map is guarenteed valid out to MAX_ORDER,
1070                  * where that page is in a different zone we will detect
1071                  * it from its zone id and abort this block scan.
1072                  */
1073                 zone_id = page_zone_id(page);
1074                 page_pfn = page_to_pfn(page);
1075                 pfn = page_pfn & ~((1 << order) - 1);
1076                 end_pfn = pfn + (1 << order);
1077                 for (; pfn < end_pfn; pfn++) {
1078                         struct page *cursor_page;
1079
1080                         /* The target page is in the block, ignore it. */
1081                         if (unlikely(pfn == page_pfn))
1082                                 continue;
1083
1084                         /* Avoid holes within the zone. */
1085                         if (unlikely(!pfn_valid_within(pfn)))
1086                                 break;
1087
1088                         cursor_page = pfn_to_page(pfn);
1089
1090                         /* Check that we have not crossed a zone boundary. */
1091                         if (unlikely(page_zone_id(cursor_page) != zone_id))
1092                                 break;
1093
1094                         /*
1095                          * If we don't have enough swap space, reclaiming of
1096                          * anon page which don't already have a swap slot is
1097                          * pointless.
1098                          */
1099                         if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
1100                             !PageSwapCache(cursor_page))
1101                                 break;
1102
1103                         if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1104                                 list_move(&cursor_page->lru, dst);
1105                                 mem_cgroup_del_lru(cursor_page);
1106                                 nr_taken++;
1107                                 nr_lumpy_taken++;
1108                                 if (PageDirty(cursor_page))
1109                                         nr_lumpy_dirty++;
1110                                 scan++;
1111                         } else {
1112                                 /* the page is freed already. */
1113                                 if (!page_count(cursor_page))
1114                                         continue;
1115                                 break;
1116                         }
1117                 }
1118
1119                 /* If we break out of the loop above, lumpy reclaim failed */
1120                 if (pfn < end_pfn)
1121                         nr_lumpy_failed++;
1122         }
1123
1124         *scanned = scan;
1125
1126         trace_mm_vmscan_lru_isolate(order,
1127                         nr_to_scan, scan,
1128                         nr_taken,
1129                         nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1130                         mode);
1131         return nr_taken;
1132 }
1133
1134 static unsigned long isolate_pages_global(unsigned long nr,
1135                                         struct list_head *dst,
1136                                         unsigned long *scanned, int order,
1137                                         int mode, struct zone *z,
1138                                         int active, int file)
1139 {
1140         int lru = LRU_BASE;
1141         if (active)
1142                 lru += LRU_ACTIVE;
1143         if (file)
1144                 lru += LRU_FILE;
1145         return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1146                                                                 mode, file);
1147 }
1148
1149 /*
1150  * clear_active_flags() is a helper for shrink_active_list(), clearing
1151  * any active bits from the pages in the list.
1152  */
1153 static unsigned long clear_active_flags(struct list_head *page_list,
1154                                         unsigned int *count)
1155 {
1156         int nr_active = 0;
1157         int lru;
1158         struct page *page;
1159
1160         list_for_each_entry(page, page_list, lru) {
1161                 lru = page_lru_base_type(page);
1162                 if (PageActive(page)) {
1163                         lru += LRU_ACTIVE;
1164                         ClearPageActive(page);
1165                         nr_active++;
1166                 }
1167                 if (count)
1168                         count[lru]++;
1169         }
1170
1171         return nr_active;
1172 }
1173
1174 /**
1175  * isolate_lru_page - tries to isolate a page from its LRU list
1176  * @page: page to isolate from its LRU list
1177  *
1178  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1179  * vmstat statistic corresponding to whatever LRU list the page was on.
1180  *
1181  * Returns 0 if the page was removed from an LRU list.
1182  * Returns -EBUSY if the page was not on an LRU list.
1183  *
1184  * The returned page will have PageLRU() cleared.  If it was found on
1185  * the active list, it will have PageActive set.  If it was found on
1186  * the unevictable list, it will have the PageUnevictable bit set. That flag
1187  * may need to be cleared by the caller before letting the page go.
1188  *
1189  * The vmstat statistic corresponding to the list on which the page was
1190  * found will be decremented.
1191  *
1192  * Restrictions:
1193  * (1) Must be called with an elevated refcount on the page. This is a
1194  *     fundamentnal difference from isolate_lru_pages (which is called
1195  *     without a stable reference).
1196  * (2) the lru_lock must not be held.
1197  * (3) interrupts must be enabled.
1198  */
1199 int isolate_lru_page(struct page *page)
1200 {
1201         int ret = -EBUSY;
1202
1203         if (PageLRU(page)) {
1204                 struct zone *zone = page_zone(page);
1205
1206                 spin_lock_irq(&zone->lru_lock);
1207                 if (PageLRU(page) && get_page_unless_zero(page)) {
1208                         int lru = page_lru(page);
1209                         ret = 0;
1210                         ClearPageLRU(page);
1211
1212                         del_page_from_lru_list(zone, page, lru);
1213                 }
1214                 spin_unlock_irq(&zone->lru_lock);
1215         }
1216         return ret;
1217 }
1218
1219 /*
1220  * Are there way too many processes in the direct reclaim path already?
1221  */
1222 static int too_many_isolated(struct zone *zone, int file,
1223                 struct scan_control *sc)
1224 {
1225         unsigned long inactive, isolated;
1226
1227         if (current_is_kswapd())
1228                 return 0;
1229
1230         if (!scanning_global_lru(sc))
1231                 return 0;
1232
1233         if (file) {
1234                 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1235                 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1236         } else {
1237                 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1238                 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1239         }
1240
1241         return isolated > inactive;
1242 }
1243
1244 /*
1245  * TODO: Try merging with migrations version of putback_lru_pages
1246  */
1247 static noinline_for_stack void
1248 putback_lru_pages(struct zone *zone, struct scan_control *sc,
1249                                 unsigned long nr_anon, unsigned long nr_file,
1250                                 struct list_head *page_list)
1251 {
1252         struct page *page;
1253         struct pagevec pvec;
1254         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1255
1256         pagevec_init(&pvec, 1);
1257
1258         /*
1259          * Put back any unfreeable pages.
1260          */
1261         spin_lock(&zone->lru_lock);
1262         while (!list_empty(page_list)) {
1263                 int lru;
1264                 page = lru_to_page(page_list);
1265                 VM_BUG_ON(PageLRU(page));
1266                 list_del(&page->lru);
1267                 if (unlikely(!page_evictable(page, NULL))) {
1268                         spin_unlock_irq(&zone->lru_lock);
1269                         putback_lru_page(page);
1270                         spin_lock_irq(&zone->lru_lock);
1271                         continue;
1272                 }
1273                 SetPageLRU(page);
1274                 lru = page_lru(page);
1275                 add_page_to_lru_list(zone, page, lru);
1276                 if (is_active_lru(lru)) {
1277                         int file = is_file_lru(lru);
1278                         reclaim_stat->recent_rotated[file]++;
1279                 }
1280                 if (!pagevec_add(&pvec, page)) {
1281                         spin_unlock_irq(&zone->lru_lock);
1282                         __pagevec_release(&pvec);
1283                         spin_lock_irq(&zone->lru_lock);
1284                 }
1285         }
1286         __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1287         __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1288
1289         spin_unlock_irq(&zone->lru_lock);
1290         pagevec_release(&pvec);
1291 }
1292
1293 static noinline_for_stack void update_isolated_counts(struct zone *zone,
1294                                         struct scan_control *sc,
1295                                         unsigned long *nr_anon,
1296                                         unsigned long *nr_file,
1297                                         struct list_head *isolated_list)
1298 {
1299         unsigned long nr_active;
1300         unsigned int count[NR_LRU_LISTS] = { 0, };
1301         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1302
1303         nr_active = clear_active_flags(isolated_list, count);
1304         __count_vm_events(PGDEACTIVATE, nr_active);
1305
1306         __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1307                               -count[LRU_ACTIVE_FILE]);
1308         __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1309                               -count[LRU_INACTIVE_FILE]);
1310         __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1311                               -count[LRU_ACTIVE_ANON]);
1312         __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1313                               -count[LRU_INACTIVE_ANON]);
1314
1315         *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1316         *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1317         __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1318         __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1319
1320         reclaim_stat->recent_scanned[0] += *nr_anon;
1321         reclaim_stat->recent_scanned[1] += *nr_file;
1322 }
1323
1324 /*
1325  * Returns true if the caller should wait to clean dirty/writeback pages.
1326  *
1327  * If we are direct reclaiming for contiguous pages and we do not reclaim
1328  * everything in the list, try again and wait for writeback IO to complete.
1329  * This will stall high-order allocations noticeably. Only do that when really
1330  * need to free the pages under high memory pressure.
1331  */
1332 static inline bool should_reclaim_stall(unsigned long nr_taken,
1333                                         unsigned long nr_freed,
1334                                         int priority,
1335                                         struct scan_control *sc)
1336 {
1337         int lumpy_stall_priority;
1338
1339         /* kswapd should not stall on sync IO */
1340         if (current_is_kswapd())
1341                 return false;
1342
1343         /* Only stall on lumpy reclaim */
1344         if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
1345                 return false;
1346
1347         /* If we have relaimed everything on the isolated list, no stall */
1348         if (nr_freed == nr_taken)
1349                 return false;
1350
1351         /*
1352          * For high-order allocations, there are two stall thresholds.
1353          * High-cost allocations stall immediately where as lower
1354          * order allocations such as stacks require the scanning
1355          * priority to be much higher before stalling.
1356          */
1357         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1358                 lumpy_stall_priority = DEF_PRIORITY;
1359         else
1360                 lumpy_stall_priority = DEF_PRIORITY / 3;
1361
1362         return priority <= lumpy_stall_priority;
1363 }
1364
1365 /*
1366  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1367  * of reclaimed pages
1368  */
1369 static noinline_for_stack unsigned long
1370 shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1371                         struct scan_control *sc, int priority, int file)
1372 {
1373         LIST_HEAD(page_list);
1374         unsigned long nr_scanned;
1375         unsigned long nr_reclaimed = 0;
1376         unsigned long nr_taken;
1377         unsigned long nr_anon;
1378         unsigned long nr_file;
1379
1380         while (unlikely(too_many_isolated(zone, file, sc))) {
1381                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1382
1383                 /* We are about to die and free our memory. Return now. */
1384                 if (fatal_signal_pending(current))
1385                         return SWAP_CLUSTER_MAX;
1386         }
1387
1388         set_reclaim_mode(priority, sc, false);
1389         lru_add_drain();
1390         spin_lock_irq(&zone->lru_lock);
1391
1392         if (scanning_global_lru(sc)) {
1393                 nr_taken = isolate_pages_global(nr_to_scan,
1394                         &page_list, &nr_scanned, sc->order,
1395                         sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM ?
1396                                         ISOLATE_BOTH : ISOLATE_INACTIVE,
1397                         zone, 0, file);
1398                 zone->pages_scanned += nr_scanned;
1399                 if (current_is_kswapd())
1400                         __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1401                                                nr_scanned);
1402                 else
1403                         __count_zone_vm_events(PGSCAN_DIRECT, zone,
1404                                                nr_scanned);
1405         } else {
1406                 nr_taken = mem_cgroup_isolate_pages(nr_to_scan,
1407                         &page_list, &nr_scanned, sc->order,
1408                         sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM ?
1409                                         ISOLATE_BOTH : ISOLATE_INACTIVE,
1410                         zone, sc->mem_cgroup,
1411                         0, file);
1412                 /*
1413                  * mem_cgroup_isolate_pages() keeps track of
1414                  * scanned pages on its own.
1415                  */
1416         }
1417
1418         if (nr_taken == 0) {
1419                 spin_unlock_irq(&zone->lru_lock);
1420                 return 0;
1421         }
1422
1423         update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
1424
1425         spin_unlock_irq(&zone->lru_lock);
1426
1427         nr_reclaimed = shrink_page_list(&page_list, zone, sc);
1428
1429         /* Check if we should syncronously wait for writeback */
1430         if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1431                 set_reclaim_mode(priority, sc, true);
1432                 nr_reclaimed += shrink_page_list(&page_list, zone, sc);
1433         }
1434
1435         local_irq_disable();
1436         if (current_is_kswapd())
1437                 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1438         __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1439
1440         putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
1441
1442         trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1443                 zone_idx(zone),
1444                 nr_scanned, nr_reclaimed,
1445                 priority,
1446                 trace_shrink_flags(file, sc->reclaim_mode));
1447         return nr_reclaimed;
1448 }
1449
1450 /*
1451  * This moves pages from the active list to the inactive list.
1452  *
1453  * We move them the other way if the page is referenced by one or more
1454  * processes, from rmap.
1455  *
1456  * If the pages are mostly unmapped, the processing is fast and it is
1457  * appropriate to hold zone->lru_lock across the whole operation.  But if
1458  * the pages are mapped, the processing is slow (page_referenced()) so we
1459  * should drop zone->lru_lock around each page.  It's impossible to balance
1460  * this, so instead we remove the pages from the LRU while processing them.
1461  * It is safe to rely on PG_active against the non-LRU pages in here because
1462  * nobody will play with that bit on a non-LRU page.
1463  *
1464  * The downside is that we have to touch page->_count against each page.
1465  * But we had to alter page->flags anyway.
1466  */
1467
1468 static void move_active_pages_to_lru(struct zone *zone,
1469                                      struct list_head *list,
1470                                      enum lru_list lru)
1471 {
1472         unsigned long pgmoved = 0;
1473         struct pagevec pvec;
1474         struct page *page;
1475
1476         pagevec_init(&pvec, 1);
1477
1478         while (!list_empty(list)) {
1479                 page = lru_to_page(list);
1480
1481                 VM_BUG_ON(PageLRU(page));
1482                 SetPageLRU(page);
1483
1484                 list_move(&page->lru, &zone->lru[lru].list);
1485                 mem_cgroup_add_lru_list(page, lru);
1486                 pgmoved++;
1487
1488                 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1489                         spin_unlock_irq(&zone->lru_lock);
1490                         if (buffer_heads_over_limit)
1491                                 pagevec_strip(&pvec);
1492                         __pagevec_release(&pvec);
1493                         spin_lock_irq(&zone->lru_lock);
1494                 }
1495         }
1496         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1497         if (!is_active_lru(lru))
1498                 __count_vm_events(PGDEACTIVATE, pgmoved);
1499 }
1500
1501 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1502                         struct scan_control *sc, int priority, int file)
1503 {
1504         unsigned long nr_taken;
1505         unsigned long pgscanned;
1506         unsigned long vm_flags;
1507         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1508         LIST_HEAD(l_active);
1509         LIST_HEAD(l_inactive);
1510         struct page *page;
1511         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1512         unsigned long nr_rotated = 0;
1513
1514         lru_add_drain();
1515         spin_lock_irq(&zone->lru_lock);
1516         if (scanning_global_lru(sc)) {
1517                 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1518                                                 &pgscanned, sc->order,
1519                                                 ISOLATE_ACTIVE, zone,
1520                                                 1, file);
1521                 zone->pages_scanned += pgscanned;
1522         } else {
1523                 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1524                                                 &pgscanned, sc->order,
1525                                                 ISOLATE_ACTIVE, zone,
1526                                                 sc->mem_cgroup, 1, file);
1527                 /*
1528                  * mem_cgroup_isolate_pages() keeps track of
1529                  * scanned pages on its own.
1530                  */
1531         }
1532
1533         reclaim_stat->recent_scanned[file] += nr_taken;
1534
1535         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1536         if (file)
1537                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1538         else
1539                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1540         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1541         spin_unlock_irq(&zone->lru_lock);
1542
1543         while (!list_empty(&l_hold)) {
1544                 cond_resched();
1545                 page = lru_to_page(&l_hold);
1546                 list_del(&page->lru);
1547
1548                 if (unlikely(!page_evictable(page, NULL))) {
1549                         putback_lru_page(page);
1550                         continue;
1551                 }
1552
1553                 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1554                         nr_rotated++;
1555                         /*
1556                          * Identify referenced, file-backed active pages and
1557                          * give them one more trip around the active list. So
1558                          * that executable code get better chances to stay in
1559                          * memory under moderate memory pressure.  Anon pages
1560                          * are not likely to be evicted by use-once streaming
1561                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1562                          * so we ignore them here.
1563                          */
1564                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1565                                 list_add(&page->lru, &l_active);
1566                                 continue;
1567                         }
1568                 }
1569
1570                 ClearPageActive(page);  /* we are de-activating */
1571                 list_add(&page->lru, &l_inactive);
1572         }
1573
1574         /*
1575          * Move pages back to the lru list.
1576          */
1577         spin_lock_irq(&zone->lru_lock);
1578         /*
1579          * Count referenced pages from currently used mappings as rotated,
1580          * even though only some of them are actually re-activated.  This
1581          * helps balance scan pressure between file and anonymous pages in
1582          * get_scan_ratio.
1583          */
1584         reclaim_stat->recent_rotated[file] += nr_rotated;
1585
1586         move_active_pages_to_lru(zone, &l_active,
1587                                                 LRU_ACTIVE + file * LRU_FILE);
1588         move_active_pages_to_lru(zone, &l_inactive,
1589                                                 LRU_BASE   + file * LRU_FILE);
1590         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1591         spin_unlock_irq(&zone->lru_lock);
1592 }
1593
1594 #ifdef CONFIG_SWAP
1595 static int inactive_anon_is_low_global(struct zone *zone)
1596 {
1597         unsigned long active, inactive;
1598
1599         active = zone_page_state(zone, NR_ACTIVE_ANON);
1600         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1601
1602         if (inactive * zone->inactive_ratio < active)
1603                 return 1;
1604
1605         return 0;
1606 }
1607
1608 /**
1609  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1610  * @zone: zone to check
1611  * @sc:   scan control of this context
1612  *
1613  * Returns true if the zone does not have enough inactive anon pages,
1614  * meaning some active anon pages need to be deactivated.
1615  */
1616 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1617 {
1618         int low;
1619
1620         /*
1621          * If we don't have swap space, anonymous page deactivation
1622          * is pointless.
1623          */
1624         if (!total_swap_pages)
1625                 return 0;
1626
1627         if (scanning_global_lru(sc))
1628                 low = inactive_anon_is_low_global(zone);
1629         else
1630                 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1631         return low;
1632 }
1633 #else
1634 static inline int inactive_anon_is_low(struct zone *zone,
1635                                         struct scan_control *sc)
1636 {
1637         return 0;
1638 }
1639 #endif
1640
1641 static int inactive_file_is_low_global(struct zone *zone)
1642 {
1643         unsigned long active, inactive;
1644
1645         active = zone_page_state(zone, NR_ACTIVE_FILE);
1646         inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1647
1648         return (active > inactive);
1649 }
1650
1651 /**
1652  * inactive_file_is_low - check if file pages need to be deactivated
1653  * @zone: zone to check
1654  * @sc:   scan control of this context
1655  *
1656  * When the system is doing streaming IO, memory pressure here
1657  * ensures that active file pages get deactivated, until more
1658  * than half of the file pages are on the inactive list.
1659  *
1660  * Once we get to that situation, protect the system's working
1661  * set from being evicted by disabling active file page aging.
1662  *
1663  * This uses a different ratio than the anonymous pages, because
1664  * the page cache uses a use-once replacement algorithm.
1665  */
1666 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1667 {
1668         int low;
1669
1670         if (scanning_global_lru(sc))
1671                 low = inactive_file_is_low_global(zone);
1672         else
1673                 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1674         return low;
1675 }
1676
1677 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1678                                 int file)
1679 {
1680         if (file)
1681                 return inactive_file_is_low(zone, sc);
1682         else
1683                 return inactive_anon_is_low(zone, sc);
1684 }
1685
1686 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1687         struct zone *zone, struct scan_control *sc, int priority)
1688 {
1689         int file = is_file_lru(lru);
1690
1691         if (is_active_lru(lru)) {
1692                 if (inactive_list_is_low(zone, sc, file))
1693                     shrink_active_list(nr_to_scan, zone, sc, priority, file);
1694                 return 0;
1695         }
1696
1697         return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1698 }
1699
1700 /*
1701  * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1702  * until we collected @swap_cluster_max pages to scan.
1703  */
1704 static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
1705                                        unsigned long *nr_saved_scan)
1706 {
1707         unsigned long nr;
1708
1709         *nr_saved_scan += nr_to_scan;
1710         nr = *nr_saved_scan;
1711
1712         if (nr >= SWAP_CLUSTER_MAX)
1713                 *nr_saved_scan = 0;
1714         else
1715                 nr = 0;
1716
1717         return nr;
1718 }
1719
1720 /*
1721  * Determine how aggressively the anon and file LRU lists should be
1722  * scanned.  The relative value of each set of LRU lists is determined
1723  * by looking at the fraction of the pages scanned we did rotate back
1724  * onto the active list instead of evict.
1725  *
1726  * nr[0] = anon pages to scan; nr[1] = file pages to scan
1727  */
1728 static void get_scan_count(struct zone *zone, struct scan_control *sc,
1729                                         unsigned long *nr, int priority)
1730 {
1731         unsigned long anon, file, free;
1732         unsigned long anon_prio, file_prio;
1733         unsigned long ap, fp;
1734         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1735         u64 fraction[2], denominator;
1736         enum lru_list l;
1737         int noswap = 0;
1738
1739         /* If we have no swap space, do not bother scanning anon pages. */
1740         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1741                 noswap = 1;
1742                 fraction[0] = 0;
1743                 fraction[1] = 1;
1744                 denominator = 1;
1745                 goto out;
1746         }
1747
1748         anon  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1749                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1750         file  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1751                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1752
1753         if (scanning_global_lru(sc)) {
1754                 free  = zone_page_state(zone, NR_FREE_PAGES);
1755                 /* If we have very few page cache pages,
1756                    force-scan anon pages. */
1757                 if (unlikely(file + free <= high_wmark_pages(zone))) {
1758                         fraction[0] = 1;
1759                         fraction[1] = 0;
1760                         denominator = 1;
1761                         goto out;
1762                 }
1763         }
1764
1765         /*
1766          * With swappiness at 100, anonymous and file have the same priority.
1767          * This scanning priority is essentially the inverse of IO cost.
1768          */
1769         anon_prio = sc->swappiness;
1770         file_prio = 200 - sc->swappiness;
1771
1772         /*
1773          * OK, so we have swap space and a fair amount of page cache
1774          * pages.  We use the recently rotated / recently scanned
1775          * ratios to determine how valuable each cache is.
1776          *
1777          * Because workloads change over time (and to avoid overflow)
1778          * we keep these statistics as a floating average, which ends
1779          * up weighing recent references more than old ones.
1780          *
1781          * anon in [0], file in [1]
1782          */
1783         spin_lock_irq(&zone->lru_lock);
1784         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1785                 reclaim_stat->recent_scanned[0] /= 2;
1786                 reclaim_stat->recent_rotated[0] /= 2;
1787         }
1788
1789         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1790                 reclaim_stat->recent_scanned[1] /= 2;
1791                 reclaim_stat->recent_rotated[1] /= 2;
1792         }
1793
1794         /*
1795          * The amount of pressure on anon vs file pages is inversely
1796          * proportional to the fraction of recently scanned pages on
1797          * each list that were recently referenced and in active use.
1798          */
1799         ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1800         ap /= reclaim_stat->recent_rotated[0] + 1;
1801
1802         fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1803         fp /= reclaim_stat->recent_rotated[1] + 1;
1804         spin_unlock_irq(&zone->lru_lock);
1805
1806         fraction[0] = ap;
1807         fraction[1] = fp;
1808         denominator = ap + fp + 1;
1809 out:
1810         for_each_evictable_lru(l) {
1811                 int file = is_file_lru(l);
1812                 unsigned long scan;
1813
1814                 scan = zone_nr_lru_pages(zone, sc, l);
1815                 if (priority || noswap) {
1816                         scan >>= priority;
1817                         scan = div64_u64(scan * fraction[file], denominator);
1818                 }
1819                 nr[l] = nr_scan_try_batch(scan,
1820                                           &reclaim_stat->nr_saved_scan[l]);
1821         }
1822 }
1823
1824 /*
1825  * Reclaim/compaction depends on a number of pages being freed. To avoid
1826  * disruption to the system, a small number of order-0 pages continue to be
1827  * rotated and reclaimed in the normal fashion. However, by the time we get
1828  * back to the allocator and call try_to_compact_zone(), we ensure that
1829  * there are enough free pages for it to be likely successful
1830  */
1831 static inline bool should_continue_reclaim(struct zone *zone,
1832                                         unsigned long nr_reclaimed,
1833                                         unsigned long nr_scanned,
1834                                         struct scan_control *sc)
1835 {
1836         unsigned long pages_for_compaction;
1837         unsigned long inactive_lru_pages;
1838
1839         /* If not in reclaim/compaction mode, stop */
1840         if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
1841                 return false;
1842
1843         /*
1844          * If we failed to reclaim and have scanned the full list, stop.
1845          * NOTE: Checking just nr_reclaimed would exit reclaim/compaction far
1846          *       faster but obviously would be less likely to succeed
1847          *       allocation. If this is desirable, use GFP_REPEAT to decide
1848          *       if both reclaimed and scanned should be checked or just
1849          *       reclaimed
1850          */
1851         if (!nr_reclaimed && !nr_scanned)
1852                 return false;
1853
1854         /*
1855          * If we have not reclaimed enough pages for compaction and the
1856          * inactive lists are large enough, continue reclaiming
1857          */
1858         pages_for_compaction = (2UL << sc->order);
1859         inactive_lru_pages = zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON) +
1860                                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1861         if (sc->nr_reclaimed < pages_for_compaction &&
1862                         inactive_lru_pages > pages_for_compaction)
1863                 return true;
1864
1865         /* If compaction would go ahead or the allocation would succeed, stop */
1866         switch (compaction_suitable(zone, sc->order)) {
1867         case COMPACT_PARTIAL:
1868         case COMPACT_CONTINUE:
1869                 return false;
1870         default:
1871                 return true;
1872         }
1873 }
1874
1875 /*
1876  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1877  */
1878 static void shrink_zone(int priority, struct zone *zone,
1879                                 struct scan_control *sc)
1880 {
1881         unsigned long nr[NR_LRU_LISTS];
1882         unsigned long nr_to_scan;
1883         enum lru_list l;
1884         unsigned long nr_reclaimed;
1885         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1886         unsigned long nr_scanned = sc->nr_scanned;
1887
1888 restart:
1889         nr_reclaimed = 0;
1890         get_scan_count(zone, sc, nr, priority);
1891
1892         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1893                                         nr[LRU_INACTIVE_FILE]) {
1894                 for_each_evictable_lru(l) {
1895                         if (nr[l]) {
1896                                 nr_to_scan = min_t(unsigned long,
1897                                                    nr[l], SWAP_CLUSTER_MAX);
1898                                 nr[l] -= nr_to_scan;
1899
1900                                 nr_reclaimed += shrink_list(l, nr_to_scan,
1901                                                             zone, sc, priority);
1902                         }
1903                 }
1904                 /*
1905                  * On large memory systems, scan >> priority can become
1906                  * really large. This is fine for the starting priority;
1907                  * we want to put equal scanning pressure on each zone.
1908                  * However, if the VM has a harder time of freeing pages,
1909                  * with multiple processes reclaiming pages, the total
1910                  * freeing target can get unreasonably large.
1911                  */
1912                 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
1913                         break;
1914         }
1915         sc->nr_reclaimed += nr_reclaimed;
1916
1917         /*
1918          * Even if we did not try to evict anon pages at all, we want to
1919          * rebalance the anon lru active/inactive ratio.
1920          */
1921         if (inactive_anon_is_low(zone, sc))
1922                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1923
1924         /* reclaim/compaction might need reclaim to continue */
1925         if (should_continue_reclaim(zone, nr_reclaimed,
1926                                         sc->nr_scanned - nr_scanned, sc))
1927                 goto restart;
1928
1929         throttle_vm_writeout(sc->gfp_mask);
1930 }
1931
1932 /*
1933  * This is the direct reclaim path, for page-allocating processes.  We only
1934  * try to reclaim pages from zones which will satisfy the caller's allocation
1935  * request.
1936  *
1937  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1938  * Because:
1939  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1940  *    allocation or
1941  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1942  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1943  *    zone defense algorithm.
1944  *
1945  * If a zone is deemed to be full of pinned pages then just give it a light
1946  * scan then give up on it.
1947  */
1948 static void shrink_zones(int priority, struct zonelist *zonelist,
1949                                         struct scan_control *sc)
1950 {
1951         struct zoneref *z;
1952         struct zone *zone;
1953
1954         for_each_zone_zonelist_nodemask(zone, z, zonelist,
1955                                         gfp_zone(sc->gfp_mask), sc->nodemask) {
1956                 if (!populated_zone(zone))
1957                         continue;
1958                 /*
1959                  * Take care memory controller reclaiming has small influence
1960                  * to global LRU.
1961                  */
1962                 if (scanning_global_lru(sc)) {
1963                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1964                                 continue;
1965                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1966                                 continue;       /* Let kswapd poll it */
1967                 }
1968
1969                 shrink_zone(priority, zone, sc);
1970         }
1971 }
1972
1973 static bool zone_reclaimable(struct zone *zone)
1974 {
1975         return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
1976 }
1977
1978 /*
1979  * As hibernation is going on, kswapd is freezed so that it can't mark
1980  * the zone into all_unreclaimable. It can't handle OOM during hibernation.
1981  * So let's check zone's unreclaimable in direct reclaim as well as kswapd.
1982  */
1983 static bool all_unreclaimable(struct zonelist *zonelist,
1984                 struct scan_control *sc)
1985 {
1986         struct zoneref *z;
1987         struct zone *zone;
1988         bool all_unreclaimable = true;
1989
1990         for_each_zone_zonelist_nodemask(zone, z, zonelist,
1991                         gfp_zone(sc->gfp_mask), sc->nodemask) {
1992                 if (!populated_zone(zone))
1993                         continue;
1994                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1995                         continue;
1996                 if (zone_reclaimable(zone)) {
1997                         all_unreclaimable = false;
1998                         break;
1999                 }
2000         }
2001
2002         return all_unreclaimable;
2003 }
2004
2005 /*
2006  * This is the main entry point to direct page reclaim.
2007  *
2008  * If a full scan of the inactive list fails to free enough memory then we
2009  * are "out of memory" and something needs to be killed.
2010  *
2011  * If the caller is !__GFP_FS then the probability of a failure is reasonably
2012  * high - the zone may be full of dirty or under-writeback pages, which this
2013  * caller can't do much about.  We kick the writeback threads and take explicit
2014  * naps in the hope that some of these pages can be written.  But if the
2015  * allocating task holds filesystem locks which prevent writeout this might not
2016  * work, and the allocation attempt will fail.
2017  *
2018  * returns:     0, if no pages reclaimed
2019  *              else, the number of pages reclaimed
2020  */
2021 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2022                                         struct scan_control *sc)
2023 {
2024         int priority;
2025         unsigned long total_scanned = 0;
2026         struct reclaim_state *reclaim_state = current->reclaim_state;
2027         struct zoneref *z;
2028         struct zone *zone;
2029         unsigned long writeback_threshold;
2030
2031         get_mems_allowed();
2032         delayacct_freepages_start();
2033
2034         if (scanning_global_lru(sc))
2035                 count_vm_event(ALLOCSTALL);
2036
2037         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2038                 sc->nr_scanned = 0;
2039                 if (!priority)
2040                         disable_swap_token();
2041                 shrink_zones(priority, zonelist, sc);
2042                 /*
2043                  * Don't shrink slabs when reclaiming memory from
2044                  * over limit cgroups
2045                  */
2046                 if (scanning_global_lru(sc)) {
2047                         unsigned long lru_pages = 0;
2048                         for_each_zone_zonelist(zone, z, zonelist,
2049                                         gfp_zone(sc->gfp_mask)) {
2050                                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2051                                         continue;
2052
2053                                 lru_pages += zone_reclaimable_pages(zone);
2054                         }
2055
2056                         shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
2057                         if (reclaim_state) {
2058                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2059                                 reclaim_state->reclaimed_slab = 0;
2060                         }
2061                 }
2062                 total_scanned += sc->nr_scanned;
2063                 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2064                         goto out;
2065
2066                 /*
2067                  * Try to write back as many pages as we just scanned.  This
2068                  * tends to cause slow streaming writers to write data to the
2069                  * disk smoothly, at the dirtying rate, which is nice.   But
2070                  * that's undesirable in laptop mode, where we *want* lumpy
2071                  * writeout.  So in laptop mode, write out the whole world.
2072                  */
2073                 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2074                 if (total_scanned > writeback_threshold) {
2075                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
2076                         sc->may_writepage = 1;
2077                 }
2078
2079                 /* Take a nap, wait for some writeback to complete */
2080                 if (!sc->hibernation_mode && sc->nr_scanned &&
2081                     priority < DEF_PRIORITY - 2) {
2082                         struct zone *preferred_zone;
2083
2084                         first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2085                                                         NULL, &preferred_zone);
2086                         wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2087                 }
2088         }
2089
2090 out:
2091         delayacct_freepages_end();
2092         put_mems_allowed();
2093
2094         if (sc->nr_reclaimed)
2095                 return sc->nr_reclaimed;
2096
2097         /* top priority shrink_zones still had more to do? don't OOM, then */
2098         if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
2099                 return 1;
2100
2101         return 0;
2102 }
2103
2104 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2105                                 gfp_t gfp_mask, nodemask_t *nodemask)
2106 {
2107         unsigned long nr_reclaimed;
2108         struct scan_control sc = {
2109                 .gfp_mask = gfp_mask,
2110                 .may_writepage = !laptop_mode,
2111                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2112                 .may_unmap = 1,
2113                 .may_swap = 1,
2114                 .swappiness = vm_swappiness,
2115                 .order = order,
2116                 .mem_cgroup = NULL,
2117                 .nodemask = nodemask,
2118         };
2119
2120         trace_mm_vmscan_direct_reclaim_begin(order,
2121                                 sc.may_writepage,
2122                                 gfp_mask);
2123
2124         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2125
2126         trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2127
2128         return nr_reclaimed;
2129 }
2130
2131 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2132
2133 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2134                                                 gfp_t gfp_mask, bool noswap,
2135                                                 unsigned int swappiness,
2136                                                 struct zone *zone)
2137 {
2138         struct scan_control sc = {
2139                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2140                 .may_writepage = !laptop_mode,
2141                 .may_unmap = 1,
2142                 .may_swap = !noswap,
2143                 .swappiness = swappiness,
2144                 .order = 0,
2145                 .mem_cgroup = mem,
2146         };
2147         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2148                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2149
2150         trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2151                                                       sc.may_writepage,
2152                                                       sc.gfp_mask);
2153
2154         /*
2155          * NOTE: Although we can get the priority field, using it
2156          * here is not a good idea, since it limits the pages we can scan.
2157          * if we don't reclaim here, the shrink_zone from balance_pgdat
2158          * will pick up pages from other mem cgroup's as well. We hack
2159          * the priority and make it zero.
2160          */
2161         shrink_zone(0, zone, &sc);
2162
2163         trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2164
2165         return sc.nr_reclaimed;
2166 }
2167
2168 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
2169                                            gfp_t gfp_mask,
2170                                            bool noswap,
2171                                            unsigned int swappiness)
2172 {
2173         struct zonelist *zonelist;
2174         unsigned long nr_reclaimed;
2175         struct scan_control sc = {
2176                 .may_writepage = !laptop_mode,
2177                 .may_unmap = 1,
2178                 .may_swap = !noswap,
2179                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2180                 .swappiness = swappiness,
2181                 .order = 0,
2182                 .mem_cgroup = mem_cont,
2183                 .nodemask = NULL, /* we don't care the placement */
2184         };
2185
2186         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2187                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2188         zonelist = NODE_DATA(numa_node_id())->node_zonelists;
2189
2190         trace_mm_vmscan_memcg_reclaim_begin(0,
2191                                             sc.may_writepage,
2192                                             sc.gfp_mask);
2193
2194         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2195
2196         trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2197
2198         return nr_reclaimed;
2199 }
2200 #endif
2201
2202 /*
2203  * pgdat_balanced is used when checking if a node is balanced for high-order
2204  * allocations. Only zones that meet watermarks and are in a zone allowed
2205  * by the callers classzone_idx are added to balanced_pages. The total of
2206  * balanced pages must be at least 25% of the zones allowed by classzone_idx
2207  * for the node to be considered balanced. Forcing all zones to be balanced
2208  * for high orders can cause excessive reclaim when there are imbalanced zones.
2209  * The choice of 25% is due to
2210  *   o a 16M DMA zone that is balanced will not balance a zone on any
2211  *     reasonable sized machine
2212  *   o On all other machines, the top zone must be at least a reasonable
2213  *     precentage of the middle zones. For example, on 32-bit x86, highmem
2214  *     would need to be at least 256M for it to be balance a whole node.
2215  *     Similarly, on x86-64 the Normal zone would need to be at least 1G
2216  *     to balance a node on its own. These seemed like reasonable ratios.
2217  */
2218 static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2219                                                 int classzone_idx)
2220 {
2221         unsigned long present_pages = 0;
2222         int i;
2223
2224         for (i = 0; i <= classzone_idx; i++)
2225                 present_pages += pgdat->node_zones[i].present_pages;
2226
2227         return balanced_pages > (present_pages >> 2);
2228 }
2229
2230 /* is kswapd sleeping prematurely? */
2231 static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2232                                         int classzone_idx)
2233 {
2234         int i;
2235         unsigned long balanced = 0;
2236         bool all_zones_ok = true;
2237
2238         /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2239         if (remaining)
2240                 return true;
2241
2242         /* Check the watermark levels */
2243         for (i = 0; i < pgdat->nr_zones; i++) {
2244                 struct zone *zone = pgdat->node_zones + i;
2245
2246                 if (!populated_zone(zone))
2247                         continue;
2248
2249                 /*
2250                  * balance_pgdat() skips over all_unreclaimable after
2251                  * DEF_PRIORITY. Effectively, it considers them balanced so
2252                  * they must be considered balanced here as well if kswapd
2253                  * is to sleep
2254                  */
2255                 if (zone->all_unreclaimable) {
2256                         balanced += zone->present_pages;
2257                         continue;
2258                 }
2259
2260                 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2261                                                         classzone_idx, 0))
2262                         all_zones_ok = false;
2263                 else
2264                         balanced += zone->present_pages;
2265         }
2266
2267         /*
2268          * For high-order requests, the balanced zones must contain at least
2269          * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2270          * must be balanced
2271          */
2272         if (order)
2273                 return pgdat_balanced(pgdat, balanced, classzone_idx);
2274         else
2275                 return !all_zones_ok;
2276 }
2277
2278 /*
2279  * For kswapd, balance_pgdat() will work across all this node's zones until
2280  * they are all at high_wmark_pages(zone).
2281  *
2282  * Returns the final order kswapd was reclaiming at
2283  *
2284  * There is special handling here for zones which are full of pinned pages.
2285  * This can happen if the pages are all mlocked, or if they are all used by
2286  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
2287  * What we do is to detect the case where all pages in the zone have been
2288  * scanned twice and there has been zero successful reclaim.  Mark the zone as
2289  * dead and from now on, only perform a short scan.  Basically we're polling
2290  * the zone for when the problem goes away.
2291  *
2292  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
2293  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2294  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2295  * lower zones regardless of the number of free pages in the lower zones. This
2296  * interoperates with the page allocator fallback scheme to ensure that aging
2297  * of pages is balanced across the zones.
2298  */
2299 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2300                                                         int *classzone_idx)
2301 {
2302         int all_zones_ok;
2303         unsigned long balanced;
2304         int priority;
2305         int i;
2306         int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
2307         unsigned long total_scanned;
2308         struct reclaim_state *reclaim_state = current->reclaim_state;
2309         struct scan_control sc = {
2310                 .gfp_mask = GFP_KERNEL,
2311                 .may_unmap = 1,
2312                 .may_swap = 1,
2313                 /*
2314                  * kswapd doesn't want to be bailed out while reclaim. because
2315                  * we want to put equal scanning pressure on each zone.
2316                  */
2317                 .nr_to_reclaim = ULONG_MAX,
2318                 .swappiness = vm_swappiness,
2319                 .order = order,
2320                 .mem_cgroup = NULL,
2321         };
2322 loop_again:
2323         total_scanned = 0;
2324         sc.nr_reclaimed = 0;
2325         sc.may_writepage = !laptop_mode;
2326         count_vm_event(PAGEOUTRUN);
2327
2328         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2329                 unsigned long lru_pages = 0;
2330                 int has_under_min_watermark_zone = 0;
2331
2332                 /* The swap token gets in the way of swapout... */
2333                 if (!priority)
2334                         disable_swap_token();
2335
2336                 all_zones_ok = 1;
2337                 balanced = 0;
2338
2339                 /*
2340                  * Scan in the highmem->dma direction for the highest
2341                  * zone which needs scanning
2342                  */
2343                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2344                         struct zone *zone = pgdat->node_zones + i;
2345
2346                         if (!populated_zone(zone))
2347                                 continue;
2348
2349                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2350                                 continue;
2351
2352                         /*
2353                          * Do some background aging of the anon list, to give
2354                          * pages a chance to be referenced before reclaiming.
2355                          */
2356                         if (inactive_anon_is_low(zone, &sc))
2357                                 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2358                                                         &sc, priority, 0);
2359
2360                         if (!zone_watermark_ok_safe(zone, order,
2361                                         high_wmark_pages(zone), 0, 0)) {
2362                                 end_zone = i;
2363                                 *classzone_idx = i;
2364                                 break;
2365                         }
2366                 }
2367                 if (i < 0)
2368                         goto out;
2369
2370                 for (i = 0; i <= end_zone; i++) {
2371                         struct zone *zone = pgdat->node_zones + i;
2372
2373                         lru_pages += zone_reclaimable_pages(zone);
2374                 }
2375
2376                 /*
2377                  * Now scan the zone in the dma->highmem direction, stopping
2378                  * at the last zone which needs scanning.
2379                  *
2380                  * We do this because the page allocator works in the opposite
2381                  * direction.  This prevents the page allocator from allocating
2382                  * pages behind kswapd's direction of progress, which would
2383                  * cause too much scanning of the lower zones.
2384                  */
2385                 for (i = 0; i <= end_zone; i++) {
2386                         int compaction;
2387                         struct zone *zone = pgdat->node_zones + i;
2388                         int nr_slab;
2389
2390                         if (!populated_zone(zone))
2391                                 continue;
2392
2393                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2394                                 continue;
2395
2396                         sc.nr_scanned = 0;
2397
2398                         /*
2399                          * Call soft limit reclaim before calling shrink_zone.
2400                          * For now we ignore the return value
2401                          */
2402                         mem_cgroup_soft_limit_reclaim(zone, order, sc.gfp_mask);
2403
2404                         /*
2405                          * We put equal pressure on every zone, unless one
2406                          * zone has way too many pages free already.
2407                          */
2408                         if (!zone_watermark_ok_safe(zone, order,
2409                                         8*high_wmark_pages(zone), end_zone, 0))
2410                                 shrink_zone(priority, zone, &sc);
2411                         reclaim_state->reclaimed_slab = 0;
2412                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
2413                                                 lru_pages);
2414                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2415                         total_scanned += sc.nr_scanned;
2416
2417                         compaction = 0;
2418                         if (order &&
2419                             zone_watermark_ok(zone, 0,
2420                                                high_wmark_pages(zone),
2421                                               end_zone, 0) &&
2422                             !zone_watermark_ok(zone, order,
2423                                                high_wmark_pages(zone),
2424                                                end_zone, 0)) {
2425                                 compact_zone_order(zone,
2426                                                    order,
2427                                                    sc.gfp_mask, false,
2428                                                    COMPACT_MODE_KSWAPD);
2429                                 compaction = 1;
2430                         }
2431
2432                         if (zone->all_unreclaimable)
2433                                 continue;
2434                         if (!compaction && nr_slab == 0 &&
2435                             !zone_reclaimable(zone))
2436                                 zone->all_unreclaimable = 1;
2437                         /*
2438                          * If we've done a decent amount of scanning and
2439                          * the reclaim ratio is low, start doing writepage
2440                          * even in laptop mode
2441                          */
2442                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2443                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2444                                 sc.may_writepage = 1;
2445
2446                         if (!zone_watermark_ok_safe(zone, order,
2447                                         high_wmark_pages(zone), end_zone, 0)) {
2448                                 all_zones_ok = 0;
2449                                 /*
2450                                  * We are still under min water mark.  This
2451                                  * means that we have a GFP_ATOMIC allocation
2452                                  * failure risk. Hurry up!
2453                                  */
2454                                 if (!zone_watermark_ok_safe(zone, order,
2455                                             min_wmark_pages(zone), end_zone, 0))
2456                                         has_under_min_watermark_zone = 1;
2457                         } else {
2458                                 /*
2459                                  * If a zone reaches its high watermark,
2460                                  * consider it to be no longer congested. It's
2461                                  * possible there are dirty pages backed by
2462                                  * congested BDIs but as pressure is relieved,
2463                                  * spectulatively avoid congestion waits
2464                                  */
2465                                 zone_clear_flag(zone, ZONE_CONGESTED);
2466                                 if (i <= *classzone_idx)
2467                                         balanced += zone->present_pages;
2468                         }
2469
2470                 }
2471                 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2472                         break;          /* kswapd: all done */
2473                 /*
2474                  * OK, kswapd is getting into trouble.  Take a nap, then take
2475                  * another pass across the zones.
2476                  */
2477                 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2478                         if (has_under_min_watermark_zone)
2479                                 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2480                         else
2481                                 congestion_wait(BLK_RW_ASYNC, HZ/10);
2482                 }
2483
2484                 /*
2485                  * We do this so kswapd doesn't build up large priorities for
2486                  * example when it is freeing in parallel with allocators. It
2487                  * matches the direct reclaim path behaviour in terms of impact
2488                  * on zone->*_priority.
2489                  */
2490                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2491                         break;
2492         }
2493 out:
2494
2495         /*
2496          * order-0: All zones must meet high watermark for a balanced node
2497          * high-order: Balanced zones must make up at least 25% of the node
2498          *             for the node to be balanced
2499          */
2500         if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2501                 cond_resched();
2502
2503                 try_to_freeze();
2504
2505                 /*
2506                  * Fragmentation may mean that the system cannot be
2507                  * rebalanced for high-order allocations in all zones.
2508                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2509                  * it means the zones have been fully scanned and are still
2510                  * not balanced. For high-order allocations, there is
2511                  * little point trying all over again as kswapd may
2512                  * infinite loop.
2513                  *
2514                  * Instead, recheck all watermarks at order-0 as they
2515                  * are the most important. If watermarks are ok, kswapd will go
2516                  * back to sleep. High-order users can still perform direct
2517                  * reclaim if they wish.
2518                  */
2519                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2520                         order = sc.order = 0;
2521
2522                 goto loop_again;
2523         }
2524
2525         /*
2526          * If kswapd was reclaiming at a higher order, it has the option of
2527          * sleeping without all zones being balanced. Before it does, it must
2528          * ensure that the watermarks for order-0 on *all* zones are met and
2529          * that the congestion flags are cleared. The congestion flag must
2530          * be cleared as kswapd is the only mechanism that clears the flag
2531          * and it is potentially going to sleep here.
2532          */
2533         if (order) {
2534                 for (i = 0; i <= end_zone; i++) {
2535                         struct zone *zone = pgdat->node_zones + i;
2536
2537                         if (!populated_zone(zone))
2538                                 continue;
2539
2540                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2541                                 continue;
2542
2543                         /* Confirm the zone is balanced for order-0 */
2544                         if (!zone_watermark_ok(zone, 0,
2545                                         high_wmark_pages(zone), 0, 0)) {
2546                                 order = sc.order = 0;
2547                                 goto loop_again;
2548                         }
2549
2550                         /* If balanced, clear the congested flag */
2551                         zone_clear_flag(zone, ZONE_CONGESTED);
2552                 }
2553         }
2554
2555         /*
2556          * Return the order we were reclaiming at so sleeping_prematurely()
2557          * makes a decision on the order we were last reclaiming at. However,
2558          * if another caller entered the allocator slow path while kswapd
2559          * was awake, order will remain at the higher level
2560          */
2561         *classzone_idx = end_zone;
2562         return order;
2563 }
2564
2565 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2566 {
2567         long remaining = 0;
2568         DEFINE_WAIT(wait);
2569
2570         if (freezing(current) || kthread_should_stop())
2571                 return;
2572
2573         prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2574
2575         /* Try to sleep for a short interval */
2576         if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2577                 remaining = schedule_timeout(HZ/10);
2578                 finish_wait(&pgdat->kswapd_wait, &wait);
2579                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2580         }
2581
2582         /*
2583          * After a short sleep, check if it was a premature sleep. If not, then
2584          * go fully to sleep until explicitly woken up.
2585          */
2586         if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2587                 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2588
2589                 /*
2590                  * vmstat counters are not perfectly accurate and the estimated
2591                  * value for counters such as NR_FREE_PAGES can deviate from the
2592                  * true value by nr_online_cpus * threshold. To avoid the zone
2593                  * watermarks being breached while under pressure, we reduce the
2594                  * per-cpu vmstat threshold while kswapd is awake and restore
2595                  * them before going back to sleep.
2596                  */
2597                 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2598                 schedule();
2599                 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2600         } else {
2601                 if (remaining)
2602                         count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2603                 else
2604                         count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2605         }
2606         finish_wait(&pgdat->kswapd_wait, &wait);
2607 }
2608
2609 /*
2610  * The background pageout daemon, started as a kernel thread
2611  * from the init process.
2612  *
2613  * This basically trickles out pages so that we have _some_
2614  * free memory available even if there is no other activity
2615  * that frees anything up. This is needed for things like routing
2616  * etc, where we otherwise might have all activity going on in
2617  * asynchronous contexts that cannot page things out.
2618  *
2619  * If there are applications that are active memory-allocators
2620  * (most normal use), this basically shouldn't matter.
2621  */
2622 static int kswapd(void *p)
2623 {
2624         unsigned long order;
2625         int classzone_idx;
2626         pg_data_t *pgdat = (pg_data_t*)p;
2627         struct task_struct *tsk = current;
2628
2629         struct reclaim_state reclaim_state = {
2630                 .reclaimed_slab = 0,
2631         };
2632         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2633
2634         lockdep_set_current_reclaim_state(GFP_KERNEL);
2635
2636         if (!cpumask_empty(cpumask))
2637                 set_cpus_allowed_ptr(tsk, cpumask);
2638         current->reclaim_state = &reclaim_state;
2639
2640         /*
2641          * Tell the memory management that we're a "memory allocator",
2642          * and that if we need more memory we should get access to it
2643          * regardless (see "__alloc_pages()"). "kswapd" should
2644          * never get caught in the normal page freeing logic.
2645          *
2646          * (Kswapd normally doesn't need memory anyway, but sometimes
2647          * you need a small amount of memory in order to be able to
2648          * page out something else, and this flag essentially protects
2649          * us from recursively trying to free more memory as we're
2650          * trying to free the first piece of memory in the first place).
2651          */
2652         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2653         set_freezable();
2654
2655         order = 0;
2656         classzone_idx = MAX_NR_ZONES - 1;
2657         for ( ; ; ) {
2658                 unsigned long new_order;
2659                 int new_classzone_idx;
2660                 int ret;
2661
2662                 new_order = pgdat->kswapd_max_order;
2663                 new_classzone_idx = pgdat->classzone_idx;
2664                 pgdat->kswapd_max_order = 0;
2665                 pgdat->classzone_idx = MAX_NR_ZONES - 1;
2666                 if (order < new_order || classzone_idx > new_classzone_idx) {
2667                         /*
2668                          * Don't sleep if someone wants a larger 'order'
2669                          * allocation or has tigher zone constraints
2670                          */
2671                         order = new_order;
2672                         classzone_idx = new_classzone_idx;
2673                 } else {
2674                         kswapd_try_to_sleep(pgdat, order, classzone_idx);
2675                         order = pgdat->kswapd_max_order;
2676                         classzone_idx = pgdat->classzone_idx;
2677                         pgdat->kswapd_max_order = 0;
2678                         pgdat->classzone_idx = MAX_NR_ZONES - 1;
2679                 }
2680
2681                 ret = try_to_freeze();
2682                 if (kthread_should_stop())
2683                         break;
2684
2685                 /*
2686                  * We can speed up thawing tasks if we don't call balance_pgdat
2687                  * after returning from the refrigerator
2688                  */
2689                 if (!ret) {
2690                         trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2691                         order = balance_pgdat(pgdat, order, &classzone_idx);
2692                 }
2693         }
2694         return 0;
2695 }
2696
2697 /*
2698  * A zone is low on free memory, so wake its kswapd task to service it.
2699  */
2700 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
2701 {
2702         pg_data_t *pgdat;
2703
2704         if (!populated_zone(zone))
2705                 return;
2706
2707         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2708                 return;
2709         pgdat = zone->zone_pgdat;
2710         if (pgdat->kswapd_max_order < order) {
2711                 pgdat->kswapd_max_order = order;
2712                 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
2713         }
2714         if (!waitqueue_active(&pgdat->kswapd_wait))
2715                 return;
2716         if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
2717                 return;
2718
2719         trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
2720         wake_up_interruptible(&pgdat->kswapd_wait);
2721 }
2722
2723 /*
2724  * The reclaimable count would be mostly accurate.
2725  * The less reclaimable pages may be
2726  * - mlocked pages, which will be moved to unevictable list when encountered
2727  * - mapped pages, which may require several travels to be reclaimed
2728  * - dirty pages, which is not "instantly" reclaimable
2729  */
2730 unsigned long global_reclaimable_pages(void)
2731 {
2732         int nr;
2733
2734         nr = global_page_state(NR_ACTIVE_FILE) +
2735              global_page_state(NR_INACTIVE_FILE);
2736
2737         if (nr_swap_pages > 0)
2738                 nr += global_page_state(NR_ACTIVE_ANON) +
2739                       global_page_state(NR_INACTIVE_ANON);
2740
2741         return nr;
2742 }
2743
2744 unsigned long zone_reclaimable_pages(struct zone *zone)
2745 {
2746         int nr;
2747
2748         nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2749              zone_page_state(zone, NR_INACTIVE_FILE);
2750
2751         if (nr_swap_pages > 0)
2752                 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2753                       zone_page_state(zone, NR_INACTIVE_ANON);
2754
2755         return nr;
2756 }
2757
2758 #ifdef CONFIG_HIBERNATION
2759 /*
2760  * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2761  * freed pages.
2762  *
2763  * Rather than trying to age LRUs the aim is to preserve the overall
2764  * LRU order by reclaiming preferentially
2765  * inactive > active > active referenced > active mapped
2766  */
2767 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
2768 {
2769         struct reclaim_state reclaim_state;
2770         struct scan_control sc = {
2771                 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2772                 .may_swap = 1,
2773                 .may_unmap = 1,
2774                 .may_writepage = 1,
2775                 .nr_to_reclaim = nr_to_reclaim,
2776                 .hibernation_mode = 1,
2777                 .swappiness = vm_swappiness,
2778                 .order = 0,
2779         };
2780         struct zonelist * zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2781         struct task_struct *p = current;
2782         unsigned long nr_reclaimed;
2783
2784         p->flags |= PF_MEMALLOC;
2785         lockdep_set_current_reclaim_state(sc.gfp_mask);
2786         reclaim_state.reclaimed_slab = 0;
2787         p->reclaim_state = &reclaim_state;
2788
2789         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2790
2791         p->reclaim_state = NULL;
2792         lockdep_clear_current_reclaim_state();
2793         p->flags &= ~PF_MEMALLOC;
2794
2795         return nr_reclaimed;
2796 }
2797 #endif /* CONFIG_HIBERNATION */
2798
2799 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2800    not required for correctness.  So if the last cpu in a node goes
2801    away, we get changed to run anywhere: as the first one comes back,
2802    restore their cpu bindings. */
2803 static int __devinit cpu_callback(struct notifier_block *nfb,
2804                                   unsigned long action, void *hcpu)
2805 {
2806         int nid;
2807
2808         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2809                 for_each_node_state(nid, N_HIGH_MEMORY) {
2810                         pg_data_t *pgdat = NODE_DATA(nid);
2811                         const struct cpumask *mask;
2812
2813                         mask = cpumask_of_node(pgdat->node_id);
2814
2815                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2816                                 /* One of our CPUs online: restore mask */
2817                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2818                 }
2819         }
2820         return NOTIFY_OK;
2821 }
2822
2823 /*
2824  * This kswapd start function will be called by init and node-hot-add.
2825  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2826  */
2827 int kswapd_run(int nid)
2828 {
2829         pg_data_t *pgdat = NODE_DATA(nid);
2830         int ret = 0;
2831
2832         if (pgdat->kswapd)
2833                 return 0;
2834
2835         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2836         if (IS_ERR(pgdat->kswapd)) {
2837                 /* failure at boot is fatal */
2838                 BUG_ON(system_state == SYSTEM_BOOTING);
2839                 printk("Failed to start kswapd on node %d\n",nid);
2840                 ret = -1;
2841         }
2842         return ret;
2843 }
2844
2845 /*
2846  * Called by memory hotplug when all memory in a node is offlined.
2847  */
2848 void kswapd_stop(int nid)
2849 {
2850         struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
2851
2852         if (kswapd)
2853                 kthread_stop(kswapd);
2854 }
2855
2856 static int __init kswapd_init(void)
2857 {
2858         int nid;
2859
2860         swap_setup();
2861         for_each_node_state(nid, N_HIGH_MEMORY)
2862                 kswapd_run(nid);
2863         hotcpu_notifier(cpu_callback, 0);
2864         return 0;
2865 }
2866
2867 module_init(kswapd_init)
2868
2869 #ifdef CONFIG_NUMA
2870 /*
2871  * Zone reclaim mode
2872  *
2873  * If non-zero call zone_reclaim when the number of free pages falls below
2874  * the watermarks.
2875  */
2876 int zone_reclaim_mode __read_mostly;
2877
2878 #define RECLAIM_OFF 0
2879 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
2880 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
2881 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
2882
2883 /*
2884  * Priority for ZONE_RECLAIM. This determines the fraction of pages
2885  * of a node considered for each zone_reclaim. 4 scans 1/16th of
2886  * a zone.
2887  */
2888 #define ZONE_RECLAIM_PRIORITY 4
2889
2890 /*
2891  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2892  * occur.
2893  */
2894 int sysctl_min_unmapped_ratio = 1;
2895
2896 /*
2897  * If the number of slab pages in a zone grows beyond this percentage then
2898  * slab reclaim needs to occur.
2899  */
2900 int sysctl_min_slab_ratio = 5;
2901
2902 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
2903 {
2904         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
2905         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
2906                 zone_page_state(zone, NR_ACTIVE_FILE);
2907
2908         /*
2909          * It's possible for there to be more file mapped pages than
2910          * accounted for by the pages on the file LRU lists because
2911          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2912          */
2913         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
2914 }
2915
2916 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
2917 static long zone_pagecache_reclaimable(struct zone *zone)
2918 {
2919         long nr_pagecache_reclaimable;
2920         long delta = 0;
2921
2922         /*
2923          * If RECLAIM_SWAP is set, then all file pages are considered
2924          * potentially reclaimable. Otherwise, we have to worry about
2925          * pages like swapcache and zone_unmapped_file_pages() provides
2926          * a better estimate
2927          */
2928         if (zone_reclaim_mode & RECLAIM_SWAP)
2929                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
2930         else
2931                 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
2932
2933         /* If we can't clean pages, remove dirty pages from consideration */
2934         if (!(zone_reclaim_mode & RECLAIM_WRITE))
2935                 delta += zone_page_state(zone, NR_FILE_DIRTY);
2936
2937         /* Watch for any possible underflows due to delta */
2938         if (unlikely(delta > nr_pagecache_reclaimable))
2939                 delta = nr_pagecache_reclaimable;
2940
2941         return nr_pagecache_reclaimable - delta;
2942 }
2943
2944 /*
2945  * Try to free up some pages from this zone through reclaim.
2946  */
2947 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2948 {
2949         /* Minimum pages needed in order to stay on node */
2950         const unsigned long nr_pages = 1 << order;
2951         struct task_struct *p = current;
2952         struct reclaim_state reclaim_state;
2953         int priority;
2954         struct scan_control sc = {
2955                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2956                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2957                 .may_swap = 1,
2958                 .nr_to_reclaim = max_t(unsigned long, nr_pages,
2959                                        SWAP_CLUSTER_MAX),
2960                 .gfp_mask = gfp_mask,
2961                 .swappiness = vm_swappiness,
2962                 .order = order,
2963         };
2964         unsigned long nr_slab_pages0, nr_slab_pages1;
2965
2966         cond_resched();
2967         /*
2968          * We need to be able to allocate from the reserves for RECLAIM_SWAP
2969          * and we also need to be able to write out pages for RECLAIM_WRITE
2970          * and RECLAIM_SWAP.
2971          */
2972         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2973         lockdep_set_current_reclaim_state(gfp_mask);
2974         reclaim_state.reclaimed_slab = 0;
2975         p->reclaim_state = &reclaim_state;
2976
2977         if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
2978                 /*
2979                  * Free memory by calling shrink zone with increasing
2980                  * priorities until we have enough memory freed.
2981                  */
2982                 priority = ZONE_RECLAIM_PRIORITY;
2983                 do {
2984                         shrink_zone(priority, zone, &sc);
2985                         priority--;
2986                 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
2987         }
2988
2989         nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2990         if (nr_slab_pages0 > zone->min_slab_pages) {
2991                 /*
2992                  * shrink_slab() does not currently allow us to determine how
2993                  * many pages were freed in this zone. So we take the current
2994                  * number of slab pages and shake the slab until it is reduced
2995                  * by the same nr_pages that we used for reclaiming unmapped
2996                  * pages.
2997                  *
2998                  * Note that shrink_slab will free memory on all zones and may
2999                  * take a long time.
3000                  */
3001                 for (;;) {
3002                         unsigned long lru_pages = zone_reclaimable_pages(zone);
3003
3004                         /* No reclaimable slab or very low memory pressure */
3005                         if (!shrink_slab(sc.nr_scanned, gfp_mask, lru_pages))
3006                                 break;
3007
3008                         /* Freed enough memory */
3009                         nr_slab_pages1 = zone_page_state(zone,
3010                                                         NR_SLAB_RECLAIMABLE);
3011                         if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3012                                 break;
3013                 }
3014
3015                 /*
3016                  * Update nr_reclaimed by the number of slab pages we
3017                  * reclaimed from this zone.
3018                  */
3019                 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3020                 if (nr_slab_pages1 < nr_slab_pages0)
3021                         sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3022         }
3023
3024         p->reclaim_state = NULL;
3025         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3026         lockdep_clear_current_reclaim_state();
3027         return sc.nr_reclaimed >= nr_pages;
3028 }
3029
3030 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3031 {
3032         int node_id;
3033         int ret;
3034
3035         /*
3036          * Zone reclaim reclaims unmapped file backed pages and
3037          * slab pages if we are over the defined limits.
3038          *
3039          * A small portion of unmapped file backed pages is needed for
3040          * file I/O otherwise pages read by file I/O will be immediately
3041          * thrown out if the zone is overallocated. So we do not reclaim
3042          * if less than a specified percentage of the zone is used by
3043          * unmapped file backed pages.
3044          */
3045         if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3046             zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3047                 return ZONE_RECLAIM_FULL;
3048
3049         if (zone->all_unreclaimable)
3050                 return ZONE_RECLAIM_FULL;
3051
3052         /*
3053          * Do not scan if the allocation should not be delayed.
3054          */
3055         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3056                 return ZONE_RECLAIM_NOSCAN;
3057
3058         /*
3059          * Only run zone reclaim on the local zone or on zones that do not
3060          * have associated processors. This will favor the local processor
3061          * over remote processors and spread off node memory allocations
3062          * as wide as possible.
3063          */
3064         node_id = zone_to_nid(zone);
3065         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3066                 return ZONE_RECLAIM_NOSCAN;
3067
3068         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3069                 return ZONE_RECLAIM_NOSCAN;
3070
3071         ret = __zone_reclaim(zone, gfp_mask, order);
3072         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3073
3074         if (!ret)
3075                 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3076
3077         return ret;
3078 }
3079 #endif
3080
3081 /*
3082  * page_evictable - test whether a page is evictable
3083  * @page: the page to test
3084  * @vma: the VMA in which the page is or will be mapped, may be NULL
3085  *
3086  * Test whether page is evictable--i.e., should be placed on active/inactive
3087  * lists vs unevictable list.  The vma argument is !NULL when called from the
3088  * fault path to determine how to instantate a new page.
3089  *
3090  * Reasons page might not be evictable:
3091  * (1) page's mapping marked unevictable
3092  * (2) page is part of an mlocked VMA
3093  *
3094  */
3095 int page_evictable(struct page *page, struct vm_area_struct *vma)
3096 {
3097
3098         if (mapping_unevictable(page_mapping(page)))
3099                 return 0;
3100
3101         if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
3102                 return 0;
3103
3104         return 1;
3105 }
3106
3107 /**
3108  * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
3109  * @page: page to check evictability and move to appropriate lru list
3110  * @zone: zone page is in
3111  *
3112  * Checks a page for evictability and moves the page to the appropriate
3113  * zone lru list.
3114  *
3115  * Restrictions: zone->lru_lock must be held, page must be on LRU and must
3116  * have PageUnevictable set.
3117  */
3118 static void check_move_unevictable_page(struct page *page, struct zone *zone)
3119 {
3120         VM_BUG_ON(PageActive(page));
3121
3122 retry:
3123         ClearPageUnevictable(page);
3124         if (page_evictable(page, NULL)) {
3125                 enum lru_list l = page_lru_base_type(page);
3126
3127                 __dec_zone_state(zone, NR_UNEVICTABLE);
3128                 list_move(&page->lru, &zone->lru[l].list);
3129                 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
3130                 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
3131                 __count_vm_event(UNEVICTABLE_PGRESCUED);
3132         } else {
3133                 /*
3134                  * rotate unevictable list
3135                  */
3136                 SetPageUnevictable(page);
3137                 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
3138                 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
3139                 if (page_evictable(page, NULL))
3140                         goto retry;
3141         }
3142 }
3143
3144 /**
3145  * scan_mapping_unevictable_pages - scan an address space for evictable pages
3146  * @mapping: struct address_space to scan for evictable pages
3147  *
3148  * Scan all pages in mapping.  Check unevictable pages for
3149  * evictability and move them to the appropriate zone lru list.
3150  */
3151 void scan_mapping_unevictable_pages(struct address_space *mapping)
3152 {
3153         pgoff_t next = 0;
3154         pgoff_t end   = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
3155                          PAGE_CACHE_SHIFT;
3156         struct zone *zone;
3157         struct pagevec pvec;
3158
3159         if (mapping->nrpages == 0)
3160                 return;
3161
3162         pagevec_init(&pvec, 0);
3163         while (next < end &&
3164                 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
3165                 int i;
3166                 int pg_scanned = 0;
3167
3168                 zone = NULL;
3169
3170                 for (i = 0; i < pagevec_count(&pvec); i++) {
3171                         struct page *page = pvec.pages[i];
3172                         pgoff_t page_index = page->index;
3173                         struct zone *pagezone = page_zone(page);
3174
3175                         pg_scanned++;
3176                         if (page_index > next)
3177                                 next = page_index;
3178                         next++;
3179
3180                         if (pagezone != zone) {
3181                                 if (zone)
3182                                         spin_unlock_irq(&zone->lru_lock);
3183                                 zone = pagezone;
3184                                 spin_lock_irq(&zone->lru_lock);
3185                         }
3186
3187                         if (PageLRU(page) && PageUnevictable(page))
3188                                 check_move_unevictable_page(page, zone);
3189                 }
3190                 if (zone)
3191                         spin_unlock_irq(&zone->lru_lock);
3192                 pagevec_release(&pvec);
3193
3194                 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
3195         }
3196
3197 }
3198
3199 /**
3200  * scan_zone_unevictable_pages - check unevictable list for evictable pages
3201  * @zone - zone of which to scan the unevictable list
3202  *
3203  * Scan @zone's unevictable LRU lists to check for pages that have become
3204  * evictable.  Move those that have to @zone's inactive list where they
3205  * become candidates for reclaim, unless shrink_inactive_zone() decides
3206  * to reactivate them.  Pages that are still unevictable are rotated
3207  * back onto @zone's unevictable list.
3208  */
3209 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
3210 static void scan_zone_unevictable_pages(struct zone *zone)
3211 {
3212         struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
3213         unsigned long scan;
3214         unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
3215
3216         while (nr_to_scan > 0) {
3217                 unsigned long batch_size = min(nr_to_scan,
3218                                                 SCAN_UNEVICTABLE_BATCH_SIZE);
3219
3220                 spin_lock_irq(&zone->lru_lock);
3221                 for (scan = 0;  scan < batch_size; scan++) {
3222                         struct page *page = lru_to_page(l_unevictable);
3223
3224                         if (!trylock_page(page))
3225                                 continue;
3226
3227                         prefetchw_prev_lru_page(page, l_unevictable, flags);
3228
3229                         if (likely(PageLRU(page) && PageUnevictable(page)))
3230                                 check_move_unevictable_page(page, zone);
3231
3232                         unlock_page(page);
3233                 }
3234                 spin_unlock_irq(&zone->lru_lock);
3235
3236                 nr_to_scan -= batch_size;
3237         }
3238 }
3239
3240
3241 /**
3242  * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
3243  *
3244  * A really big hammer:  scan all zones' unevictable LRU lists to check for
3245  * pages that have become evictable.  Move those back to the zones'
3246  * inactive list where they become candidates for reclaim.
3247  * This occurs when, e.g., we have unswappable pages on the unevictable lists,
3248  * and we add swap to the system.  As such, it runs in the context of a task
3249  * that has possibly/probably made some previously unevictable pages
3250  * evictable.
3251  */
3252 static void scan_all_zones_unevictable_pages(void)
3253 {
3254         struct zone *zone;
3255
3256         for_each_zone(zone) {
3257                 scan_zone_unevictable_pages(zone);
3258         }
3259 }
3260
3261 /*
3262  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
3263  * all nodes' unevictable lists for evictable pages
3264  */
3265 unsigned long scan_unevictable_pages;
3266
3267 int scan_unevictable_handler(struct ctl_table *table, int write,
3268                            void __user *buffer,
3269                            size_t *length, loff_t *ppos)
3270 {
3271         proc_doulongvec_minmax(table, write, buffer, length, ppos);
3272
3273         if (write && *(unsigned long *)table->data)
3274                 scan_all_zones_unevictable_pages();
3275
3276         scan_unevictable_pages = 0;
3277         return 0;
3278 }
3279
3280 #ifdef CONFIG_NUMA
3281 /*
3282  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
3283  * a specified node's per zone unevictable lists for evictable pages.
3284  */
3285
3286 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
3287                                           struct sysdev_attribute *attr,
3288                                           char *buf)
3289 {
3290         return sprintf(buf, "0\n");     /* always zero; should fit... */
3291 }
3292
3293 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
3294                                            struct sysdev_attribute *attr,
3295                                         const char *buf, size_t count)
3296 {
3297         struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
3298         struct zone *zone;
3299         unsigned long res;
3300         unsigned long req = strict_strtoul(buf, 10, &res);
3301
3302         if (!req)
3303                 return 1;       /* zero is no-op */
3304
3305         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
3306                 if (!populated_zone(zone))
3307                         continue;
3308                 scan_zone_unevictable_pages(zone);
3309         }
3310         return 1;
3311 }
3312
3313
3314 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3315                         read_scan_unevictable_node,
3316                         write_scan_unevictable_node);
3317
3318 int scan_unevictable_register_node(struct node *node)
3319 {
3320         return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
3321 }
3322
3323 void scan_unevictable_unregister_node(struct node *node)
3324 {
3325         sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
3326 }
3327 #endif