memcg: reclaim memory from nodes in round-robin order
[linux-2.6.git] / mm / vmscan.c
1 /*
2  *  linux/mm/vmscan.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  *
6  *  Swap reorganised 29.12.95, Stephen Tweedie.
7  *  kswapd added: 7.1.96  sct
8  *  Removed kswapd_ctl limits, and swap out as many pages as needed
9  *  to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10  *  Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11  *  Multiqueue VM started 5.8.00, Rik van Riel.
12  */
13
14 #include <linux/mm.h>
15 #include <linux/module.h>
16 #include <linux/gfp.h>
17 #include <linux/kernel_stat.h>
18 #include <linux/swap.h>
19 #include <linux/pagemap.h>
20 #include <linux/init.h>
21 #include <linux/highmem.h>
22 #include <linux/vmstat.h>
23 #include <linux/file.h>
24 #include <linux/writeback.h>
25 #include <linux/blkdev.h>
26 #include <linux/buffer_head.h>  /* for try_to_release_page(),
27                                         buffer_heads_over_limit */
28 #include <linux/mm_inline.h>
29 #include <linux/pagevec.h>
30 #include <linux/backing-dev.h>
31 #include <linux/rmap.h>
32 #include <linux/topology.h>
33 #include <linux/cpu.h>
34 #include <linux/cpuset.h>
35 #include <linux/compaction.h>
36 #include <linux/notifier.h>
37 #include <linux/rwsem.h>
38 #include <linux/delay.h>
39 #include <linux/kthread.h>
40 #include <linux/freezer.h>
41 #include <linux/memcontrol.h>
42 #include <linux/delayacct.h>
43 #include <linux/sysctl.h>
44 #include <linux/oom.h>
45 #include <linux/prefetch.h>
46
47 #include <asm/tlbflush.h>
48 #include <asm/div64.h>
49
50 #include <linux/swapops.h>
51
52 #include "internal.h"
53
54 #define CREATE_TRACE_POINTS
55 #include <trace/events/vmscan.h>
56
57 /*
58  * reclaim_mode determines how the inactive list is shrunk
59  * RECLAIM_MODE_SINGLE: Reclaim only order-0 pages
60  * RECLAIM_MODE_ASYNC:  Do not block
61  * RECLAIM_MODE_SYNC:   Allow blocking e.g. call wait_on_page_writeback
62  * RECLAIM_MODE_LUMPYRECLAIM: For high-order allocations, take a reference
63  *                      page from the LRU and reclaim all pages within a
64  *                      naturally aligned range
65  * RECLAIM_MODE_COMPACTION: For high-order allocations, reclaim a number of
66  *                      order-0 pages and then compact the zone
67  */
68 typedef unsigned __bitwise__ reclaim_mode_t;
69 #define RECLAIM_MODE_SINGLE             ((__force reclaim_mode_t)0x01u)
70 #define RECLAIM_MODE_ASYNC              ((__force reclaim_mode_t)0x02u)
71 #define RECLAIM_MODE_SYNC               ((__force reclaim_mode_t)0x04u)
72 #define RECLAIM_MODE_LUMPYRECLAIM       ((__force reclaim_mode_t)0x08u)
73 #define RECLAIM_MODE_COMPACTION         ((__force reclaim_mode_t)0x10u)
74
75 struct scan_control {
76         /* Incremented by the number of inactive pages that were scanned */
77         unsigned long nr_scanned;
78
79         /* Number of pages freed so far during a call to shrink_zones() */
80         unsigned long nr_reclaimed;
81
82         /* How many pages shrink_list() should reclaim */
83         unsigned long nr_to_reclaim;
84
85         unsigned long hibernation_mode;
86
87         /* This context's GFP mask */
88         gfp_t gfp_mask;
89
90         int may_writepage;
91
92         /* Can mapped pages be reclaimed? */
93         int may_unmap;
94
95         /* Can pages be swapped as part of reclaim? */
96         int may_swap;
97
98         int swappiness;
99
100         int order;
101
102         /*
103          * Intend to reclaim enough continuous memory rather than reclaim
104          * enough amount of memory. i.e, mode for high order allocation.
105          */
106         reclaim_mode_t reclaim_mode;
107
108         /* Which cgroup do we reclaim from */
109         struct mem_cgroup *mem_cgroup;
110
111         /*
112          * Nodemask of nodes allowed by the caller. If NULL, all nodes
113          * are scanned.
114          */
115         nodemask_t      *nodemask;
116 };
117
118 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
119
120 #ifdef ARCH_HAS_PREFETCH
121 #define prefetch_prev_lru_page(_page, _base, _field)                    \
122         do {                                                            \
123                 if ((_page)->lru.prev != _base) {                       \
124                         struct page *prev;                              \
125                                                                         \
126                         prev = lru_to_page(&(_page->lru));              \
127                         prefetch(&prev->_field);                        \
128                 }                                                       \
129         } while (0)
130 #else
131 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
132 #endif
133
134 #ifdef ARCH_HAS_PREFETCHW
135 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
136         do {                                                            \
137                 if ((_page)->lru.prev != _base) {                       \
138                         struct page *prev;                              \
139                                                                         \
140                         prev = lru_to_page(&(_page->lru));              \
141                         prefetchw(&prev->_field);                       \
142                 }                                                       \
143         } while (0)
144 #else
145 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
146 #endif
147
148 /*
149  * From 0 .. 100.  Higher means more swappy.
150  */
151 int vm_swappiness = 60;
152 long vm_total_pages;    /* The total number of pages which the VM controls */
153
154 static LIST_HEAD(shrinker_list);
155 static DECLARE_RWSEM(shrinker_rwsem);
156
157 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
158 #define scanning_global_lru(sc) (!(sc)->mem_cgroup)
159 #else
160 #define scanning_global_lru(sc) (1)
161 #endif
162
163 static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
164                                                   struct scan_control *sc)
165 {
166         if (!scanning_global_lru(sc))
167                 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
168
169         return &zone->reclaim_stat;
170 }
171
172 static unsigned long zone_nr_lru_pages(struct zone *zone,
173                                 struct scan_control *sc, enum lru_list lru)
174 {
175         if (!scanning_global_lru(sc))
176                 return mem_cgroup_zone_nr_pages(sc->mem_cgroup, zone, lru);
177
178         return zone_page_state(zone, NR_LRU_BASE + lru);
179 }
180
181
182 /*
183  * Add a shrinker callback to be called from the vm
184  */
185 void register_shrinker(struct shrinker *shrinker)
186 {
187         shrinker->nr = 0;
188         down_write(&shrinker_rwsem);
189         list_add_tail(&shrinker->list, &shrinker_list);
190         up_write(&shrinker_rwsem);
191 }
192 EXPORT_SYMBOL(register_shrinker);
193
194 /*
195  * Remove one
196  */
197 void unregister_shrinker(struct shrinker *shrinker)
198 {
199         down_write(&shrinker_rwsem);
200         list_del(&shrinker->list);
201         up_write(&shrinker_rwsem);
202 }
203 EXPORT_SYMBOL(unregister_shrinker);
204
205 static inline int do_shrinker_shrink(struct shrinker *shrinker,
206                                      struct shrink_control *sc,
207                                      unsigned long nr_to_scan)
208 {
209         sc->nr_to_scan = nr_to_scan;
210         return (*shrinker->shrink)(shrinker, sc);
211 }
212
213 #define SHRINK_BATCH 128
214 /*
215  * Call the shrink functions to age shrinkable caches
216  *
217  * Here we assume it costs one seek to replace a lru page and that it also
218  * takes a seek to recreate a cache object.  With this in mind we age equal
219  * percentages of the lru and ageable caches.  This should balance the seeks
220  * generated by these structures.
221  *
222  * If the vm encountered mapped pages on the LRU it increase the pressure on
223  * slab to avoid swapping.
224  *
225  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
226  *
227  * `lru_pages' represents the number of on-LRU pages in all the zones which
228  * are eligible for the caller's allocation attempt.  It is used for balancing
229  * slab reclaim versus page reclaim.
230  *
231  * Returns the number of slab objects which we shrunk.
232  */
233 unsigned long shrink_slab(struct shrink_control *shrink,
234                           unsigned long nr_pages_scanned,
235                           unsigned long lru_pages)
236 {
237         struct shrinker *shrinker;
238         unsigned long ret = 0;
239
240         if (nr_pages_scanned == 0)
241                 nr_pages_scanned = SWAP_CLUSTER_MAX;
242
243         if (!down_read_trylock(&shrinker_rwsem)) {
244                 /* Assume we'll be able to shrink next time */
245                 ret = 1;
246                 goto out;
247         }
248
249         list_for_each_entry(shrinker, &shrinker_list, list) {
250                 unsigned long long delta;
251                 unsigned long total_scan;
252                 unsigned long max_pass;
253
254                 max_pass = do_shrinker_shrink(shrinker, shrink, 0);
255                 delta = (4 * nr_pages_scanned) / shrinker->seeks;
256                 delta *= max_pass;
257                 do_div(delta, lru_pages + 1);
258                 shrinker->nr += delta;
259                 if (shrinker->nr < 0) {
260                         printk(KERN_ERR "shrink_slab: %pF negative objects to "
261                                "delete nr=%ld\n",
262                                shrinker->shrink, shrinker->nr);
263                         shrinker->nr = max_pass;
264                 }
265
266                 /*
267                  * Avoid risking looping forever due to too large nr value:
268                  * never try to free more than twice the estimate number of
269                  * freeable entries.
270                  */
271                 if (shrinker->nr > max_pass * 2)
272                         shrinker->nr = max_pass * 2;
273
274                 total_scan = shrinker->nr;
275                 shrinker->nr = 0;
276
277                 while (total_scan >= SHRINK_BATCH) {
278                         long this_scan = SHRINK_BATCH;
279                         int shrink_ret;
280                         int nr_before;
281
282                         nr_before = do_shrinker_shrink(shrinker, shrink, 0);
283                         shrink_ret = do_shrinker_shrink(shrinker, shrink,
284                                                         this_scan);
285                         if (shrink_ret == -1)
286                                 break;
287                         if (shrink_ret < nr_before)
288                                 ret += nr_before - shrink_ret;
289                         count_vm_events(SLABS_SCANNED, this_scan);
290                         total_scan -= this_scan;
291
292                         cond_resched();
293                 }
294
295                 shrinker->nr += total_scan;
296         }
297         up_read(&shrinker_rwsem);
298 out:
299         cond_resched();
300         return ret;
301 }
302
303 static void set_reclaim_mode(int priority, struct scan_control *sc,
304                                    bool sync)
305 {
306         reclaim_mode_t syncmode = sync ? RECLAIM_MODE_SYNC : RECLAIM_MODE_ASYNC;
307
308         /*
309          * Initially assume we are entering either lumpy reclaim or
310          * reclaim/compaction.Depending on the order, we will either set the
311          * sync mode or just reclaim order-0 pages later.
312          */
313         if (COMPACTION_BUILD)
314                 sc->reclaim_mode = RECLAIM_MODE_COMPACTION;
315         else
316                 sc->reclaim_mode = RECLAIM_MODE_LUMPYRECLAIM;
317
318         /*
319          * Avoid using lumpy reclaim or reclaim/compaction if possible by
320          * restricting when its set to either costly allocations or when
321          * under memory pressure
322          */
323         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
324                 sc->reclaim_mode |= syncmode;
325         else if (sc->order && priority < DEF_PRIORITY - 2)
326                 sc->reclaim_mode |= syncmode;
327         else
328                 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
329 }
330
331 static void reset_reclaim_mode(struct scan_control *sc)
332 {
333         sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
334 }
335
336 static inline int is_page_cache_freeable(struct page *page)
337 {
338         /*
339          * A freeable page cache page is referenced only by the caller
340          * that isolated the page, the page cache radix tree and
341          * optional buffer heads at page->private.
342          */
343         return page_count(page) - page_has_private(page) == 2;
344 }
345
346 static int may_write_to_queue(struct backing_dev_info *bdi,
347                               struct scan_control *sc)
348 {
349         if (current->flags & PF_SWAPWRITE)
350                 return 1;
351         if (!bdi_write_congested(bdi))
352                 return 1;
353         if (bdi == current->backing_dev_info)
354                 return 1;
355
356         /* lumpy reclaim for hugepage often need a lot of write */
357         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
358                 return 1;
359         return 0;
360 }
361
362 /*
363  * We detected a synchronous write error writing a page out.  Probably
364  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
365  * fsync(), msync() or close().
366  *
367  * The tricky part is that after writepage we cannot touch the mapping: nothing
368  * prevents it from being freed up.  But we have a ref on the page and once
369  * that page is locked, the mapping is pinned.
370  *
371  * We're allowed to run sleeping lock_page() here because we know the caller has
372  * __GFP_FS.
373  */
374 static void handle_write_error(struct address_space *mapping,
375                                 struct page *page, int error)
376 {
377         lock_page(page);
378         if (page_mapping(page) == mapping)
379                 mapping_set_error(mapping, error);
380         unlock_page(page);
381 }
382
383 /* possible outcome of pageout() */
384 typedef enum {
385         /* failed to write page out, page is locked */
386         PAGE_KEEP,
387         /* move page to the active list, page is locked */
388         PAGE_ACTIVATE,
389         /* page has been sent to the disk successfully, page is unlocked */
390         PAGE_SUCCESS,
391         /* page is clean and locked */
392         PAGE_CLEAN,
393 } pageout_t;
394
395 /*
396  * pageout is called by shrink_page_list() for each dirty page.
397  * Calls ->writepage().
398  */
399 static pageout_t pageout(struct page *page, struct address_space *mapping,
400                          struct scan_control *sc)
401 {
402         /*
403          * If the page is dirty, only perform writeback if that write
404          * will be non-blocking.  To prevent this allocation from being
405          * stalled by pagecache activity.  But note that there may be
406          * stalls if we need to run get_block().  We could test
407          * PagePrivate for that.
408          *
409          * If this process is currently in __generic_file_aio_write() against
410          * this page's queue, we can perform writeback even if that
411          * will block.
412          *
413          * If the page is swapcache, write it back even if that would
414          * block, for some throttling. This happens by accident, because
415          * swap_backing_dev_info is bust: it doesn't reflect the
416          * congestion state of the swapdevs.  Easy to fix, if needed.
417          */
418         if (!is_page_cache_freeable(page))
419                 return PAGE_KEEP;
420         if (!mapping) {
421                 /*
422                  * Some data journaling orphaned pages can have
423                  * page->mapping == NULL while being dirty with clean buffers.
424                  */
425                 if (page_has_private(page)) {
426                         if (try_to_free_buffers(page)) {
427                                 ClearPageDirty(page);
428                                 printk("%s: orphaned page\n", __func__);
429                                 return PAGE_CLEAN;
430                         }
431                 }
432                 return PAGE_KEEP;
433         }
434         if (mapping->a_ops->writepage == NULL)
435                 return PAGE_ACTIVATE;
436         if (!may_write_to_queue(mapping->backing_dev_info, sc))
437                 return PAGE_KEEP;
438
439         if (clear_page_dirty_for_io(page)) {
440                 int res;
441                 struct writeback_control wbc = {
442                         .sync_mode = WB_SYNC_NONE,
443                         .nr_to_write = SWAP_CLUSTER_MAX,
444                         .range_start = 0,
445                         .range_end = LLONG_MAX,
446                         .for_reclaim = 1,
447                 };
448
449                 SetPageReclaim(page);
450                 res = mapping->a_ops->writepage(page, &wbc);
451                 if (res < 0)
452                         handle_write_error(mapping, page, res);
453                 if (res == AOP_WRITEPAGE_ACTIVATE) {
454                         ClearPageReclaim(page);
455                         return PAGE_ACTIVATE;
456                 }
457
458                 /*
459                  * Wait on writeback if requested to. This happens when
460                  * direct reclaiming a large contiguous area and the
461                  * first attempt to free a range of pages fails.
462                  */
463                 if (PageWriteback(page) &&
464                     (sc->reclaim_mode & RECLAIM_MODE_SYNC))
465                         wait_on_page_writeback(page);
466
467                 if (!PageWriteback(page)) {
468                         /* synchronous write or broken a_ops? */
469                         ClearPageReclaim(page);
470                 }
471                 trace_mm_vmscan_writepage(page,
472                         trace_reclaim_flags(page, sc->reclaim_mode));
473                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
474                 return PAGE_SUCCESS;
475         }
476
477         return PAGE_CLEAN;
478 }
479
480 /*
481  * Same as remove_mapping, but if the page is removed from the mapping, it
482  * gets returned with a refcount of 0.
483  */
484 static int __remove_mapping(struct address_space *mapping, struct page *page)
485 {
486         BUG_ON(!PageLocked(page));
487         BUG_ON(mapping != page_mapping(page));
488
489         spin_lock_irq(&mapping->tree_lock);
490         /*
491          * The non racy check for a busy page.
492          *
493          * Must be careful with the order of the tests. When someone has
494          * a ref to the page, it may be possible that they dirty it then
495          * drop the reference. So if PageDirty is tested before page_count
496          * here, then the following race may occur:
497          *
498          * get_user_pages(&page);
499          * [user mapping goes away]
500          * write_to(page);
501          *                              !PageDirty(page)    [good]
502          * SetPageDirty(page);
503          * put_page(page);
504          *                              !page_count(page)   [good, discard it]
505          *
506          * [oops, our write_to data is lost]
507          *
508          * Reversing the order of the tests ensures such a situation cannot
509          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
510          * load is not satisfied before that of page->_count.
511          *
512          * Note that if SetPageDirty is always performed via set_page_dirty,
513          * and thus under tree_lock, then this ordering is not required.
514          */
515         if (!page_freeze_refs(page, 2))
516                 goto cannot_free;
517         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
518         if (unlikely(PageDirty(page))) {
519                 page_unfreeze_refs(page, 2);
520                 goto cannot_free;
521         }
522
523         if (PageSwapCache(page)) {
524                 swp_entry_t swap = { .val = page_private(page) };
525                 __delete_from_swap_cache(page);
526                 spin_unlock_irq(&mapping->tree_lock);
527                 swapcache_free(swap, page);
528         } else {
529                 void (*freepage)(struct page *);
530
531                 freepage = mapping->a_ops->freepage;
532
533                 __delete_from_page_cache(page);
534                 spin_unlock_irq(&mapping->tree_lock);
535                 mem_cgroup_uncharge_cache_page(page);
536
537                 if (freepage != NULL)
538                         freepage(page);
539         }
540
541         return 1;
542
543 cannot_free:
544         spin_unlock_irq(&mapping->tree_lock);
545         return 0;
546 }
547
548 /*
549  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
550  * someone else has a ref on the page, abort and return 0.  If it was
551  * successfully detached, return 1.  Assumes the caller has a single ref on
552  * this page.
553  */
554 int remove_mapping(struct address_space *mapping, struct page *page)
555 {
556         if (__remove_mapping(mapping, page)) {
557                 /*
558                  * Unfreezing the refcount with 1 rather than 2 effectively
559                  * drops the pagecache ref for us without requiring another
560                  * atomic operation.
561                  */
562                 page_unfreeze_refs(page, 1);
563                 return 1;
564         }
565         return 0;
566 }
567
568 /**
569  * putback_lru_page - put previously isolated page onto appropriate LRU list
570  * @page: page to be put back to appropriate lru list
571  *
572  * Add previously isolated @page to appropriate LRU list.
573  * Page may still be unevictable for other reasons.
574  *
575  * lru_lock must not be held, interrupts must be enabled.
576  */
577 void putback_lru_page(struct page *page)
578 {
579         int lru;
580         int active = !!TestClearPageActive(page);
581         int was_unevictable = PageUnevictable(page);
582
583         VM_BUG_ON(PageLRU(page));
584
585 redo:
586         ClearPageUnevictable(page);
587
588         if (page_evictable(page, NULL)) {
589                 /*
590                  * For evictable pages, we can use the cache.
591                  * In event of a race, worst case is we end up with an
592                  * unevictable page on [in]active list.
593                  * We know how to handle that.
594                  */
595                 lru = active + page_lru_base_type(page);
596                 lru_cache_add_lru(page, lru);
597         } else {
598                 /*
599                  * Put unevictable pages directly on zone's unevictable
600                  * list.
601                  */
602                 lru = LRU_UNEVICTABLE;
603                 add_page_to_unevictable_list(page);
604                 /*
605                  * When racing with an mlock clearing (page is
606                  * unlocked), make sure that if the other thread does
607                  * not observe our setting of PG_lru and fails
608                  * isolation, we see PG_mlocked cleared below and move
609                  * the page back to the evictable list.
610                  *
611                  * The other side is TestClearPageMlocked().
612                  */
613                 smp_mb();
614         }
615
616         /*
617          * page's status can change while we move it among lru. If an evictable
618          * page is on unevictable list, it never be freed. To avoid that,
619          * check after we added it to the list, again.
620          */
621         if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
622                 if (!isolate_lru_page(page)) {
623                         put_page(page);
624                         goto redo;
625                 }
626                 /* This means someone else dropped this page from LRU
627                  * So, it will be freed or putback to LRU again. There is
628                  * nothing to do here.
629                  */
630         }
631
632         if (was_unevictable && lru != LRU_UNEVICTABLE)
633                 count_vm_event(UNEVICTABLE_PGRESCUED);
634         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
635                 count_vm_event(UNEVICTABLE_PGCULLED);
636
637         put_page(page);         /* drop ref from isolate */
638 }
639
640 enum page_references {
641         PAGEREF_RECLAIM,
642         PAGEREF_RECLAIM_CLEAN,
643         PAGEREF_KEEP,
644         PAGEREF_ACTIVATE,
645 };
646
647 static enum page_references page_check_references(struct page *page,
648                                                   struct scan_control *sc)
649 {
650         int referenced_ptes, referenced_page;
651         unsigned long vm_flags;
652
653         referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
654         referenced_page = TestClearPageReferenced(page);
655
656         /* Lumpy reclaim - ignore references */
657         if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
658                 return PAGEREF_RECLAIM;
659
660         /*
661          * Mlock lost the isolation race with us.  Let try_to_unmap()
662          * move the page to the unevictable list.
663          */
664         if (vm_flags & VM_LOCKED)
665                 return PAGEREF_RECLAIM;
666
667         if (referenced_ptes) {
668                 if (PageAnon(page))
669                         return PAGEREF_ACTIVATE;
670                 /*
671                  * All mapped pages start out with page table
672                  * references from the instantiating fault, so we need
673                  * to look twice if a mapped file page is used more
674                  * than once.
675                  *
676                  * Mark it and spare it for another trip around the
677                  * inactive list.  Another page table reference will
678                  * lead to its activation.
679                  *
680                  * Note: the mark is set for activated pages as well
681                  * so that recently deactivated but used pages are
682                  * quickly recovered.
683                  */
684                 SetPageReferenced(page);
685
686                 if (referenced_page)
687                         return PAGEREF_ACTIVATE;
688
689                 return PAGEREF_KEEP;
690         }
691
692         /* Reclaim if clean, defer dirty pages to writeback */
693         if (referenced_page && !PageSwapBacked(page))
694                 return PAGEREF_RECLAIM_CLEAN;
695
696         return PAGEREF_RECLAIM;
697 }
698
699 static noinline_for_stack void free_page_list(struct list_head *free_pages)
700 {
701         struct pagevec freed_pvec;
702         struct page *page, *tmp;
703
704         pagevec_init(&freed_pvec, 1);
705
706         list_for_each_entry_safe(page, tmp, free_pages, lru) {
707                 list_del(&page->lru);
708                 if (!pagevec_add(&freed_pvec, page)) {
709                         __pagevec_free(&freed_pvec);
710                         pagevec_reinit(&freed_pvec);
711                 }
712         }
713
714         pagevec_free(&freed_pvec);
715 }
716
717 /*
718  * shrink_page_list() returns the number of reclaimed pages
719  */
720 static unsigned long shrink_page_list(struct list_head *page_list,
721                                       struct zone *zone,
722                                       struct scan_control *sc)
723 {
724         LIST_HEAD(ret_pages);
725         LIST_HEAD(free_pages);
726         int pgactivate = 0;
727         unsigned long nr_dirty = 0;
728         unsigned long nr_congested = 0;
729         unsigned long nr_reclaimed = 0;
730
731         cond_resched();
732
733         while (!list_empty(page_list)) {
734                 enum page_references references;
735                 struct address_space *mapping;
736                 struct page *page;
737                 int may_enter_fs;
738
739                 cond_resched();
740
741                 page = lru_to_page(page_list);
742                 list_del(&page->lru);
743
744                 if (!trylock_page(page))
745                         goto keep;
746
747                 VM_BUG_ON(PageActive(page));
748                 VM_BUG_ON(page_zone(page) != zone);
749
750                 sc->nr_scanned++;
751
752                 if (unlikely(!page_evictable(page, NULL)))
753                         goto cull_mlocked;
754
755                 if (!sc->may_unmap && page_mapped(page))
756                         goto keep_locked;
757
758                 /* Double the slab pressure for mapped and swapcache pages */
759                 if (page_mapped(page) || PageSwapCache(page))
760                         sc->nr_scanned++;
761
762                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
763                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
764
765                 if (PageWriteback(page)) {
766                         /*
767                          * Synchronous reclaim is performed in two passes,
768                          * first an asynchronous pass over the list to
769                          * start parallel writeback, and a second synchronous
770                          * pass to wait for the IO to complete.  Wait here
771                          * for any page for which writeback has already
772                          * started.
773                          */
774                         if ((sc->reclaim_mode & RECLAIM_MODE_SYNC) &&
775                             may_enter_fs)
776                                 wait_on_page_writeback(page);
777                         else {
778                                 unlock_page(page);
779                                 goto keep_lumpy;
780                         }
781                 }
782
783                 references = page_check_references(page, sc);
784                 switch (references) {
785                 case PAGEREF_ACTIVATE:
786                         goto activate_locked;
787                 case PAGEREF_KEEP:
788                         goto keep_locked;
789                 case PAGEREF_RECLAIM:
790                 case PAGEREF_RECLAIM_CLEAN:
791                         ; /* try to reclaim the page below */
792                 }
793
794                 /*
795                  * Anonymous process memory has backing store?
796                  * Try to allocate it some swap space here.
797                  */
798                 if (PageAnon(page) && !PageSwapCache(page)) {
799                         if (!(sc->gfp_mask & __GFP_IO))
800                                 goto keep_locked;
801                         if (!add_to_swap(page))
802                                 goto activate_locked;
803                         may_enter_fs = 1;
804                 }
805
806                 mapping = page_mapping(page);
807
808                 /*
809                  * The page is mapped into the page tables of one or more
810                  * processes. Try to unmap it here.
811                  */
812                 if (page_mapped(page) && mapping) {
813                         switch (try_to_unmap(page, TTU_UNMAP)) {
814                         case SWAP_FAIL:
815                                 goto activate_locked;
816                         case SWAP_AGAIN:
817                                 goto keep_locked;
818                         case SWAP_MLOCK:
819                                 goto cull_mlocked;
820                         case SWAP_SUCCESS:
821                                 ; /* try to free the page below */
822                         }
823                 }
824
825                 if (PageDirty(page)) {
826                         nr_dirty++;
827
828                         if (references == PAGEREF_RECLAIM_CLEAN)
829                                 goto keep_locked;
830                         if (!may_enter_fs)
831                                 goto keep_locked;
832                         if (!sc->may_writepage)
833                                 goto keep_locked;
834
835                         /* Page is dirty, try to write it out here */
836                         switch (pageout(page, mapping, sc)) {
837                         case PAGE_KEEP:
838                                 nr_congested++;
839                                 goto keep_locked;
840                         case PAGE_ACTIVATE:
841                                 goto activate_locked;
842                         case PAGE_SUCCESS:
843                                 if (PageWriteback(page))
844                                         goto keep_lumpy;
845                                 if (PageDirty(page))
846                                         goto keep;
847
848                                 /*
849                                  * A synchronous write - probably a ramdisk.  Go
850                                  * ahead and try to reclaim the page.
851                                  */
852                                 if (!trylock_page(page))
853                                         goto keep;
854                                 if (PageDirty(page) || PageWriteback(page))
855                                         goto keep_locked;
856                                 mapping = page_mapping(page);
857                         case PAGE_CLEAN:
858                                 ; /* try to free the page below */
859                         }
860                 }
861
862                 /*
863                  * If the page has buffers, try to free the buffer mappings
864                  * associated with this page. If we succeed we try to free
865                  * the page as well.
866                  *
867                  * We do this even if the page is PageDirty().
868                  * try_to_release_page() does not perform I/O, but it is
869                  * possible for a page to have PageDirty set, but it is actually
870                  * clean (all its buffers are clean).  This happens if the
871                  * buffers were written out directly, with submit_bh(). ext3
872                  * will do this, as well as the blockdev mapping.
873                  * try_to_release_page() will discover that cleanness and will
874                  * drop the buffers and mark the page clean - it can be freed.
875                  *
876                  * Rarely, pages can have buffers and no ->mapping.  These are
877                  * the pages which were not successfully invalidated in
878                  * truncate_complete_page().  We try to drop those buffers here
879                  * and if that worked, and the page is no longer mapped into
880                  * process address space (page_count == 1) it can be freed.
881                  * Otherwise, leave the page on the LRU so it is swappable.
882                  */
883                 if (page_has_private(page)) {
884                         if (!try_to_release_page(page, sc->gfp_mask))
885                                 goto activate_locked;
886                         if (!mapping && page_count(page) == 1) {
887                                 unlock_page(page);
888                                 if (put_page_testzero(page))
889                                         goto free_it;
890                                 else {
891                                         /*
892                                          * rare race with speculative reference.
893                                          * the speculative reference will free
894                                          * this page shortly, so we may
895                                          * increment nr_reclaimed here (and
896                                          * leave it off the LRU).
897                                          */
898                                         nr_reclaimed++;
899                                         continue;
900                                 }
901                         }
902                 }
903
904                 if (!mapping || !__remove_mapping(mapping, page))
905                         goto keep_locked;
906
907                 /*
908                  * At this point, we have no other references and there is
909                  * no way to pick any more up (removed from LRU, removed
910                  * from pagecache). Can use non-atomic bitops now (and
911                  * we obviously don't have to worry about waking up a process
912                  * waiting on the page lock, because there are no references.
913                  */
914                 __clear_page_locked(page);
915 free_it:
916                 nr_reclaimed++;
917
918                 /*
919                  * Is there need to periodically free_page_list? It would
920                  * appear not as the counts should be low
921                  */
922                 list_add(&page->lru, &free_pages);
923                 continue;
924
925 cull_mlocked:
926                 if (PageSwapCache(page))
927                         try_to_free_swap(page);
928                 unlock_page(page);
929                 putback_lru_page(page);
930                 reset_reclaim_mode(sc);
931                 continue;
932
933 activate_locked:
934                 /* Not a candidate for swapping, so reclaim swap space. */
935                 if (PageSwapCache(page) && vm_swap_full())
936                         try_to_free_swap(page);
937                 VM_BUG_ON(PageActive(page));
938                 SetPageActive(page);
939                 pgactivate++;
940 keep_locked:
941                 unlock_page(page);
942 keep:
943                 reset_reclaim_mode(sc);
944 keep_lumpy:
945                 list_add(&page->lru, &ret_pages);
946                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
947         }
948
949         /*
950          * Tag a zone as congested if all the dirty pages encountered were
951          * backed by a congested BDI. In this case, reclaimers should just
952          * back off and wait for congestion to clear because further reclaim
953          * will encounter the same problem
954          */
955         if (nr_dirty && nr_dirty == nr_congested && scanning_global_lru(sc))
956                 zone_set_flag(zone, ZONE_CONGESTED);
957
958         free_page_list(&free_pages);
959
960         list_splice(&ret_pages, page_list);
961         count_vm_events(PGACTIVATE, pgactivate);
962         return nr_reclaimed;
963 }
964
965 /*
966  * Attempt to remove the specified page from its LRU.  Only take this page
967  * if it is of the appropriate PageActive status.  Pages which are being
968  * freed elsewhere are also ignored.
969  *
970  * page:        page to consider
971  * mode:        one of the LRU isolation modes defined above
972  *
973  * returns 0 on success, -ve errno on failure.
974  */
975 int __isolate_lru_page(struct page *page, int mode, int file)
976 {
977         int ret = -EINVAL;
978
979         /* Only take pages on the LRU. */
980         if (!PageLRU(page))
981                 return ret;
982
983         /*
984          * When checking the active state, we need to be sure we are
985          * dealing with comparible boolean values.  Take the logical not
986          * of each.
987          */
988         if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
989                 return ret;
990
991         if (mode != ISOLATE_BOTH && page_is_file_cache(page) != file)
992                 return ret;
993
994         /*
995          * When this function is being called for lumpy reclaim, we
996          * initially look into all LRU pages, active, inactive and
997          * unevictable; only give shrink_page_list evictable pages.
998          */
999         if (PageUnevictable(page))
1000                 return ret;
1001
1002         ret = -EBUSY;
1003
1004         if (likely(get_page_unless_zero(page))) {
1005                 /*
1006                  * Be careful not to clear PageLRU until after we're
1007                  * sure the page is not being freed elsewhere -- the
1008                  * page release code relies on it.
1009                  */
1010                 ClearPageLRU(page);
1011                 ret = 0;
1012         }
1013
1014         return ret;
1015 }
1016
1017 /*
1018  * zone->lru_lock is heavily contended.  Some of the functions that
1019  * shrink the lists perform better by taking out a batch of pages
1020  * and working on them outside the LRU lock.
1021  *
1022  * For pagecache intensive workloads, this function is the hottest
1023  * spot in the kernel (apart from copy_*_user functions).
1024  *
1025  * Appropriate locks must be held before calling this function.
1026  *
1027  * @nr_to_scan: The number of pages to look through on the list.
1028  * @src:        The LRU list to pull pages off.
1029  * @dst:        The temp list to put pages on to.
1030  * @scanned:    The number of pages that were scanned.
1031  * @order:      The caller's attempted allocation order
1032  * @mode:       One of the LRU isolation modes
1033  * @file:       True [1] if isolating file [!anon] pages
1034  *
1035  * returns how many pages were moved onto *@dst.
1036  */
1037 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1038                 struct list_head *src, struct list_head *dst,
1039                 unsigned long *scanned, int order, int mode, int file)
1040 {
1041         unsigned long nr_taken = 0;
1042         unsigned long nr_lumpy_taken = 0;
1043         unsigned long nr_lumpy_dirty = 0;
1044         unsigned long nr_lumpy_failed = 0;
1045         unsigned long scan;
1046
1047         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1048                 struct page *page;
1049                 unsigned long pfn;
1050                 unsigned long end_pfn;
1051                 unsigned long page_pfn;
1052                 int zone_id;
1053
1054                 page = lru_to_page(src);
1055                 prefetchw_prev_lru_page(page, src, flags);
1056
1057                 VM_BUG_ON(!PageLRU(page));
1058
1059                 switch (__isolate_lru_page(page, mode, file)) {
1060                 case 0:
1061                         list_move(&page->lru, dst);
1062                         mem_cgroup_del_lru(page);
1063                         nr_taken += hpage_nr_pages(page);
1064                         break;
1065
1066                 case -EBUSY:
1067                         /* else it is being freed elsewhere */
1068                         list_move(&page->lru, src);
1069                         mem_cgroup_rotate_lru_list(page, page_lru(page));
1070                         continue;
1071
1072                 default:
1073                         BUG();
1074                 }
1075
1076                 if (!order)
1077                         continue;
1078
1079                 /*
1080                  * Attempt to take all pages in the order aligned region
1081                  * surrounding the tag page.  Only take those pages of
1082                  * the same active state as that tag page.  We may safely
1083                  * round the target page pfn down to the requested order
1084                  * as the mem_map is guaranteed valid out to MAX_ORDER,
1085                  * where that page is in a different zone we will detect
1086                  * it from its zone id and abort this block scan.
1087                  */
1088                 zone_id = page_zone_id(page);
1089                 page_pfn = page_to_pfn(page);
1090                 pfn = page_pfn & ~((1 << order) - 1);
1091                 end_pfn = pfn + (1 << order);
1092                 for (; pfn < end_pfn; pfn++) {
1093                         struct page *cursor_page;
1094
1095                         /* The target page is in the block, ignore it. */
1096                         if (unlikely(pfn == page_pfn))
1097                                 continue;
1098
1099                         /* Avoid holes within the zone. */
1100                         if (unlikely(!pfn_valid_within(pfn)))
1101                                 break;
1102
1103                         cursor_page = pfn_to_page(pfn);
1104
1105                         /* Check that we have not crossed a zone boundary. */
1106                         if (unlikely(page_zone_id(cursor_page) != zone_id))
1107                                 break;
1108
1109                         /*
1110                          * If we don't have enough swap space, reclaiming of
1111                          * anon page which don't already have a swap slot is
1112                          * pointless.
1113                          */
1114                         if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
1115                             !PageSwapCache(cursor_page))
1116                                 break;
1117
1118                         if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1119                                 list_move(&cursor_page->lru, dst);
1120                                 mem_cgroup_del_lru(cursor_page);
1121                                 nr_taken += hpage_nr_pages(page);
1122                                 nr_lumpy_taken++;
1123                                 if (PageDirty(cursor_page))
1124                                         nr_lumpy_dirty++;
1125                                 scan++;
1126                         } else {
1127                                 /* the page is freed already. */
1128                                 if (!page_count(cursor_page))
1129                                         continue;
1130                                 break;
1131                         }
1132                 }
1133
1134                 /* If we break out of the loop above, lumpy reclaim failed */
1135                 if (pfn < end_pfn)
1136                         nr_lumpy_failed++;
1137         }
1138
1139         *scanned = scan;
1140
1141         trace_mm_vmscan_lru_isolate(order,
1142                         nr_to_scan, scan,
1143                         nr_taken,
1144                         nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1145                         mode);
1146         return nr_taken;
1147 }
1148
1149 static unsigned long isolate_pages_global(unsigned long nr,
1150                                         struct list_head *dst,
1151                                         unsigned long *scanned, int order,
1152                                         int mode, struct zone *z,
1153                                         int active, int file)
1154 {
1155         int lru = LRU_BASE;
1156         if (active)
1157                 lru += LRU_ACTIVE;
1158         if (file)
1159                 lru += LRU_FILE;
1160         return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1161                                                                 mode, file);
1162 }
1163
1164 /*
1165  * clear_active_flags() is a helper for shrink_active_list(), clearing
1166  * any active bits from the pages in the list.
1167  */
1168 static unsigned long clear_active_flags(struct list_head *page_list,
1169                                         unsigned int *count)
1170 {
1171         int nr_active = 0;
1172         int lru;
1173         struct page *page;
1174
1175         list_for_each_entry(page, page_list, lru) {
1176                 int numpages = hpage_nr_pages(page);
1177                 lru = page_lru_base_type(page);
1178                 if (PageActive(page)) {
1179                         lru += LRU_ACTIVE;
1180                         ClearPageActive(page);
1181                         nr_active += numpages;
1182                 }
1183                 if (count)
1184                         count[lru] += numpages;
1185         }
1186
1187         return nr_active;
1188 }
1189
1190 /**
1191  * isolate_lru_page - tries to isolate a page from its LRU list
1192  * @page: page to isolate from its LRU list
1193  *
1194  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1195  * vmstat statistic corresponding to whatever LRU list the page was on.
1196  *
1197  * Returns 0 if the page was removed from an LRU list.
1198  * Returns -EBUSY if the page was not on an LRU list.
1199  *
1200  * The returned page will have PageLRU() cleared.  If it was found on
1201  * the active list, it will have PageActive set.  If it was found on
1202  * the unevictable list, it will have the PageUnevictable bit set. That flag
1203  * may need to be cleared by the caller before letting the page go.
1204  *
1205  * The vmstat statistic corresponding to the list on which the page was
1206  * found will be decremented.
1207  *
1208  * Restrictions:
1209  * (1) Must be called with an elevated refcount on the page. This is a
1210  *     fundamentnal difference from isolate_lru_pages (which is called
1211  *     without a stable reference).
1212  * (2) the lru_lock must not be held.
1213  * (3) interrupts must be enabled.
1214  */
1215 int isolate_lru_page(struct page *page)
1216 {
1217         int ret = -EBUSY;
1218
1219         VM_BUG_ON(!page_count(page));
1220
1221         if (PageLRU(page)) {
1222                 struct zone *zone = page_zone(page);
1223
1224                 spin_lock_irq(&zone->lru_lock);
1225                 if (PageLRU(page)) {
1226                         int lru = page_lru(page);
1227                         ret = 0;
1228                         get_page(page);
1229                         ClearPageLRU(page);
1230
1231                         del_page_from_lru_list(zone, page, lru);
1232                 }
1233                 spin_unlock_irq(&zone->lru_lock);
1234         }
1235         return ret;
1236 }
1237
1238 /*
1239  * Are there way too many processes in the direct reclaim path already?
1240  */
1241 static int too_many_isolated(struct zone *zone, int file,
1242                 struct scan_control *sc)
1243 {
1244         unsigned long inactive, isolated;
1245
1246         if (current_is_kswapd())
1247                 return 0;
1248
1249         if (!scanning_global_lru(sc))
1250                 return 0;
1251
1252         if (file) {
1253                 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1254                 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1255         } else {
1256                 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1257                 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1258         }
1259
1260         return isolated > inactive;
1261 }
1262
1263 /*
1264  * TODO: Try merging with migrations version of putback_lru_pages
1265  */
1266 static noinline_for_stack void
1267 putback_lru_pages(struct zone *zone, struct scan_control *sc,
1268                                 unsigned long nr_anon, unsigned long nr_file,
1269                                 struct list_head *page_list)
1270 {
1271         struct page *page;
1272         struct pagevec pvec;
1273         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1274
1275         pagevec_init(&pvec, 1);
1276
1277         /*
1278          * Put back any unfreeable pages.
1279          */
1280         spin_lock(&zone->lru_lock);
1281         while (!list_empty(page_list)) {
1282                 int lru;
1283                 page = lru_to_page(page_list);
1284                 VM_BUG_ON(PageLRU(page));
1285                 list_del(&page->lru);
1286                 if (unlikely(!page_evictable(page, NULL))) {
1287                         spin_unlock_irq(&zone->lru_lock);
1288                         putback_lru_page(page);
1289                         spin_lock_irq(&zone->lru_lock);
1290                         continue;
1291                 }
1292                 SetPageLRU(page);
1293                 lru = page_lru(page);
1294                 add_page_to_lru_list(zone, page, lru);
1295                 if (is_active_lru(lru)) {
1296                         int file = is_file_lru(lru);
1297                         int numpages = hpage_nr_pages(page);
1298                         reclaim_stat->recent_rotated[file] += numpages;
1299                 }
1300                 if (!pagevec_add(&pvec, page)) {
1301                         spin_unlock_irq(&zone->lru_lock);
1302                         __pagevec_release(&pvec);
1303                         spin_lock_irq(&zone->lru_lock);
1304                 }
1305         }
1306         __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1307         __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1308
1309         spin_unlock_irq(&zone->lru_lock);
1310         pagevec_release(&pvec);
1311 }
1312
1313 static noinline_for_stack void update_isolated_counts(struct zone *zone,
1314                                         struct scan_control *sc,
1315                                         unsigned long *nr_anon,
1316                                         unsigned long *nr_file,
1317                                         struct list_head *isolated_list)
1318 {
1319         unsigned long nr_active;
1320         unsigned int count[NR_LRU_LISTS] = { 0, };
1321         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1322
1323         nr_active = clear_active_flags(isolated_list, count);
1324         __count_vm_events(PGDEACTIVATE, nr_active);
1325
1326         __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1327                               -count[LRU_ACTIVE_FILE]);
1328         __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1329                               -count[LRU_INACTIVE_FILE]);
1330         __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1331                               -count[LRU_ACTIVE_ANON]);
1332         __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1333                               -count[LRU_INACTIVE_ANON]);
1334
1335         *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1336         *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1337         __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1338         __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1339
1340         reclaim_stat->recent_scanned[0] += *nr_anon;
1341         reclaim_stat->recent_scanned[1] += *nr_file;
1342 }
1343
1344 /*
1345  * Returns true if the caller should wait to clean dirty/writeback pages.
1346  *
1347  * If we are direct reclaiming for contiguous pages and we do not reclaim
1348  * everything in the list, try again and wait for writeback IO to complete.
1349  * This will stall high-order allocations noticeably. Only do that when really
1350  * need to free the pages under high memory pressure.
1351  */
1352 static inline bool should_reclaim_stall(unsigned long nr_taken,
1353                                         unsigned long nr_freed,
1354                                         int priority,
1355                                         struct scan_control *sc)
1356 {
1357         int lumpy_stall_priority;
1358
1359         /* kswapd should not stall on sync IO */
1360         if (current_is_kswapd())
1361                 return false;
1362
1363         /* Only stall on lumpy reclaim */
1364         if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
1365                 return false;
1366
1367         /* If we have relaimed everything on the isolated list, no stall */
1368         if (nr_freed == nr_taken)
1369                 return false;
1370
1371         /*
1372          * For high-order allocations, there are two stall thresholds.
1373          * High-cost allocations stall immediately where as lower
1374          * order allocations such as stacks require the scanning
1375          * priority to be much higher before stalling.
1376          */
1377         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1378                 lumpy_stall_priority = DEF_PRIORITY;
1379         else
1380                 lumpy_stall_priority = DEF_PRIORITY / 3;
1381
1382         return priority <= lumpy_stall_priority;
1383 }
1384
1385 /*
1386  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1387  * of reclaimed pages
1388  */
1389 static noinline_for_stack unsigned long
1390 shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1391                         struct scan_control *sc, int priority, int file)
1392 {
1393         LIST_HEAD(page_list);
1394         unsigned long nr_scanned;
1395         unsigned long nr_reclaimed = 0;
1396         unsigned long nr_taken;
1397         unsigned long nr_anon;
1398         unsigned long nr_file;
1399
1400         while (unlikely(too_many_isolated(zone, file, sc))) {
1401                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1402
1403                 /* We are about to die and free our memory. Return now. */
1404                 if (fatal_signal_pending(current))
1405                         return SWAP_CLUSTER_MAX;
1406         }
1407
1408         set_reclaim_mode(priority, sc, false);
1409         lru_add_drain();
1410         spin_lock_irq(&zone->lru_lock);
1411
1412         if (scanning_global_lru(sc)) {
1413                 nr_taken = isolate_pages_global(nr_to_scan,
1414                         &page_list, &nr_scanned, sc->order,
1415                         sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM ?
1416                                         ISOLATE_BOTH : ISOLATE_INACTIVE,
1417                         zone, 0, file);
1418                 zone->pages_scanned += nr_scanned;
1419                 if (current_is_kswapd())
1420                         __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1421                                                nr_scanned);
1422                 else
1423                         __count_zone_vm_events(PGSCAN_DIRECT, zone,
1424                                                nr_scanned);
1425         } else {
1426                 nr_taken = mem_cgroup_isolate_pages(nr_to_scan,
1427                         &page_list, &nr_scanned, sc->order,
1428                         sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM ?
1429                                         ISOLATE_BOTH : ISOLATE_INACTIVE,
1430                         zone, sc->mem_cgroup,
1431                         0, file);
1432                 /*
1433                  * mem_cgroup_isolate_pages() keeps track of
1434                  * scanned pages on its own.
1435                  */
1436         }
1437
1438         if (nr_taken == 0) {
1439                 spin_unlock_irq(&zone->lru_lock);
1440                 return 0;
1441         }
1442
1443         update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
1444
1445         spin_unlock_irq(&zone->lru_lock);
1446
1447         nr_reclaimed = shrink_page_list(&page_list, zone, sc);
1448
1449         /* Check if we should syncronously wait for writeback */
1450         if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1451                 set_reclaim_mode(priority, sc, true);
1452                 nr_reclaimed += shrink_page_list(&page_list, zone, sc);
1453         }
1454
1455         local_irq_disable();
1456         if (current_is_kswapd())
1457                 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1458         __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1459
1460         putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
1461
1462         trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1463                 zone_idx(zone),
1464                 nr_scanned, nr_reclaimed,
1465                 priority,
1466                 trace_shrink_flags(file, sc->reclaim_mode));
1467         return nr_reclaimed;
1468 }
1469
1470 /*
1471  * This moves pages from the active list to the inactive list.
1472  *
1473  * We move them the other way if the page is referenced by one or more
1474  * processes, from rmap.
1475  *
1476  * If the pages are mostly unmapped, the processing is fast and it is
1477  * appropriate to hold zone->lru_lock across the whole operation.  But if
1478  * the pages are mapped, the processing is slow (page_referenced()) so we
1479  * should drop zone->lru_lock around each page.  It's impossible to balance
1480  * this, so instead we remove the pages from the LRU while processing them.
1481  * It is safe to rely on PG_active against the non-LRU pages in here because
1482  * nobody will play with that bit on a non-LRU page.
1483  *
1484  * The downside is that we have to touch page->_count against each page.
1485  * But we had to alter page->flags anyway.
1486  */
1487
1488 static void move_active_pages_to_lru(struct zone *zone,
1489                                      struct list_head *list,
1490                                      enum lru_list lru)
1491 {
1492         unsigned long pgmoved = 0;
1493         struct pagevec pvec;
1494         struct page *page;
1495
1496         pagevec_init(&pvec, 1);
1497
1498         while (!list_empty(list)) {
1499                 page = lru_to_page(list);
1500
1501                 VM_BUG_ON(PageLRU(page));
1502                 SetPageLRU(page);
1503
1504                 list_move(&page->lru, &zone->lru[lru].list);
1505                 mem_cgroup_add_lru_list(page, lru);
1506                 pgmoved += hpage_nr_pages(page);
1507
1508                 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1509                         spin_unlock_irq(&zone->lru_lock);
1510                         if (buffer_heads_over_limit)
1511                                 pagevec_strip(&pvec);
1512                         __pagevec_release(&pvec);
1513                         spin_lock_irq(&zone->lru_lock);
1514                 }
1515         }
1516         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1517         if (!is_active_lru(lru))
1518                 __count_vm_events(PGDEACTIVATE, pgmoved);
1519 }
1520
1521 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1522                         struct scan_control *sc, int priority, int file)
1523 {
1524         unsigned long nr_taken;
1525         unsigned long pgscanned;
1526         unsigned long vm_flags;
1527         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1528         LIST_HEAD(l_active);
1529         LIST_HEAD(l_inactive);
1530         struct page *page;
1531         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1532         unsigned long nr_rotated = 0;
1533
1534         lru_add_drain();
1535         spin_lock_irq(&zone->lru_lock);
1536         if (scanning_global_lru(sc)) {
1537                 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1538                                                 &pgscanned, sc->order,
1539                                                 ISOLATE_ACTIVE, zone,
1540                                                 1, file);
1541                 zone->pages_scanned += pgscanned;
1542         } else {
1543                 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1544                                                 &pgscanned, sc->order,
1545                                                 ISOLATE_ACTIVE, zone,
1546                                                 sc->mem_cgroup, 1, file);
1547                 /*
1548                  * mem_cgroup_isolate_pages() keeps track of
1549                  * scanned pages on its own.
1550                  */
1551         }
1552
1553         reclaim_stat->recent_scanned[file] += nr_taken;
1554
1555         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1556         if (file)
1557                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1558         else
1559                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1560         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1561         spin_unlock_irq(&zone->lru_lock);
1562
1563         while (!list_empty(&l_hold)) {
1564                 cond_resched();
1565                 page = lru_to_page(&l_hold);
1566                 list_del(&page->lru);
1567
1568                 if (unlikely(!page_evictable(page, NULL))) {
1569                         putback_lru_page(page);
1570                         continue;
1571                 }
1572
1573                 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1574                         nr_rotated += hpage_nr_pages(page);
1575                         /*
1576                          * Identify referenced, file-backed active pages and
1577                          * give them one more trip around the active list. So
1578                          * that executable code get better chances to stay in
1579                          * memory under moderate memory pressure.  Anon pages
1580                          * are not likely to be evicted by use-once streaming
1581                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1582                          * so we ignore them here.
1583                          */
1584                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1585                                 list_add(&page->lru, &l_active);
1586                                 continue;
1587                         }
1588                 }
1589
1590                 ClearPageActive(page);  /* we are de-activating */
1591                 list_add(&page->lru, &l_inactive);
1592         }
1593
1594         /*
1595          * Move pages back to the lru list.
1596          */
1597         spin_lock_irq(&zone->lru_lock);
1598         /*
1599          * Count referenced pages from currently used mappings as rotated,
1600          * even though only some of them are actually re-activated.  This
1601          * helps balance scan pressure between file and anonymous pages in
1602          * get_scan_ratio.
1603          */
1604         reclaim_stat->recent_rotated[file] += nr_rotated;
1605
1606         move_active_pages_to_lru(zone, &l_active,
1607                                                 LRU_ACTIVE + file * LRU_FILE);
1608         move_active_pages_to_lru(zone, &l_inactive,
1609                                                 LRU_BASE   + file * LRU_FILE);
1610         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1611         spin_unlock_irq(&zone->lru_lock);
1612 }
1613
1614 #ifdef CONFIG_SWAP
1615 static int inactive_anon_is_low_global(struct zone *zone)
1616 {
1617         unsigned long active, inactive;
1618
1619         active = zone_page_state(zone, NR_ACTIVE_ANON);
1620         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1621
1622         if (inactive * zone->inactive_ratio < active)
1623                 return 1;
1624
1625         return 0;
1626 }
1627
1628 /**
1629  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1630  * @zone: zone to check
1631  * @sc:   scan control of this context
1632  *
1633  * Returns true if the zone does not have enough inactive anon pages,
1634  * meaning some active anon pages need to be deactivated.
1635  */
1636 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1637 {
1638         int low;
1639
1640         /*
1641          * If we don't have swap space, anonymous page deactivation
1642          * is pointless.
1643          */
1644         if (!total_swap_pages)
1645                 return 0;
1646
1647         if (scanning_global_lru(sc))
1648                 low = inactive_anon_is_low_global(zone);
1649         else
1650                 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1651         return low;
1652 }
1653 #else
1654 static inline int inactive_anon_is_low(struct zone *zone,
1655                                         struct scan_control *sc)
1656 {
1657         return 0;
1658 }
1659 #endif
1660
1661 static int inactive_file_is_low_global(struct zone *zone)
1662 {
1663         unsigned long active, inactive;
1664
1665         active = zone_page_state(zone, NR_ACTIVE_FILE);
1666         inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1667
1668         return (active > inactive);
1669 }
1670
1671 /**
1672  * inactive_file_is_low - check if file pages need to be deactivated
1673  * @zone: zone to check
1674  * @sc:   scan control of this context
1675  *
1676  * When the system is doing streaming IO, memory pressure here
1677  * ensures that active file pages get deactivated, until more
1678  * than half of the file pages are on the inactive list.
1679  *
1680  * Once we get to that situation, protect the system's working
1681  * set from being evicted by disabling active file page aging.
1682  *
1683  * This uses a different ratio than the anonymous pages, because
1684  * the page cache uses a use-once replacement algorithm.
1685  */
1686 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1687 {
1688         int low;
1689
1690         if (scanning_global_lru(sc))
1691                 low = inactive_file_is_low_global(zone);
1692         else
1693                 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1694         return low;
1695 }
1696
1697 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1698                                 int file)
1699 {
1700         if (file)
1701                 return inactive_file_is_low(zone, sc);
1702         else
1703                 return inactive_anon_is_low(zone, sc);
1704 }
1705
1706 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1707         struct zone *zone, struct scan_control *sc, int priority)
1708 {
1709         int file = is_file_lru(lru);
1710
1711         if (is_active_lru(lru)) {
1712                 if (inactive_list_is_low(zone, sc, file))
1713                     shrink_active_list(nr_to_scan, zone, sc, priority, file);
1714                 return 0;
1715         }
1716
1717         return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1718 }
1719
1720 /*
1721  * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1722  * until we collected @swap_cluster_max pages to scan.
1723  */
1724 static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
1725                                        unsigned long *nr_saved_scan)
1726 {
1727         unsigned long nr;
1728
1729         *nr_saved_scan += nr_to_scan;
1730         nr = *nr_saved_scan;
1731
1732         if (nr >= SWAP_CLUSTER_MAX)
1733                 *nr_saved_scan = 0;
1734         else
1735                 nr = 0;
1736
1737         return nr;
1738 }
1739
1740 /*
1741  * Determine how aggressively the anon and file LRU lists should be
1742  * scanned.  The relative value of each set of LRU lists is determined
1743  * by looking at the fraction of the pages scanned we did rotate back
1744  * onto the active list instead of evict.
1745  *
1746  * nr[0] = anon pages to scan; nr[1] = file pages to scan
1747  */
1748 static void get_scan_count(struct zone *zone, struct scan_control *sc,
1749                                         unsigned long *nr, int priority)
1750 {
1751         unsigned long anon, file, free;
1752         unsigned long anon_prio, file_prio;
1753         unsigned long ap, fp;
1754         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1755         u64 fraction[2], denominator;
1756         enum lru_list l;
1757         int noswap = 0;
1758
1759         /* If we have no swap space, do not bother scanning anon pages. */
1760         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1761                 noswap = 1;
1762                 fraction[0] = 0;
1763                 fraction[1] = 1;
1764                 denominator = 1;
1765                 goto out;
1766         }
1767
1768         anon  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1769                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1770         file  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1771                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1772
1773         if (scanning_global_lru(sc)) {
1774                 free  = zone_page_state(zone, NR_FREE_PAGES);
1775                 /* If we have very few page cache pages,
1776                    force-scan anon pages. */
1777                 if (unlikely(file + free <= high_wmark_pages(zone))) {
1778                         fraction[0] = 1;
1779                         fraction[1] = 0;
1780                         denominator = 1;
1781                         goto out;
1782                 }
1783         }
1784
1785         /*
1786          * With swappiness at 100, anonymous and file have the same priority.
1787          * This scanning priority is essentially the inverse of IO cost.
1788          */
1789         anon_prio = sc->swappiness;
1790         file_prio = 200 - sc->swappiness;
1791
1792         /*
1793          * OK, so we have swap space and a fair amount of page cache
1794          * pages.  We use the recently rotated / recently scanned
1795          * ratios to determine how valuable each cache is.
1796          *
1797          * Because workloads change over time (and to avoid overflow)
1798          * we keep these statistics as a floating average, which ends
1799          * up weighing recent references more than old ones.
1800          *
1801          * anon in [0], file in [1]
1802          */
1803         spin_lock_irq(&zone->lru_lock);
1804         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1805                 reclaim_stat->recent_scanned[0] /= 2;
1806                 reclaim_stat->recent_rotated[0] /= 2;
1807         }
1808
1809         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1810                 reclaim_stat->recent_scanned[1] /= 2;
1811                 reclaim_stat->recent_rotated[1] /= 2;
1812         }
1813
1814         /*
1815          * The amount of pressure on anon vs file pages is inversely
1816          * proportional to the fraction of recently scanned pages on
1817          * each list that were recently referenced and in active use.
1818          */
1819         ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1820         ap /= reclaim_stat->recent_rotated[0] + 1;
1821
1822         fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1823         fp /= reclaim_stat->recent_rotated[1] + 1;
1824         spin_unlock_irq(&zone->lru_lock);
1825
1826         fraction[0] = ap;
1827         fraction[1] = fp;
1828         denominator = ap + fp + 1;
1829 out:
1830         for_each_evictable_lru(l) {
1831                 int file = is_file_lru(l);
1832                 unsigned long scan;
1833
1834                 scan = zone_nr_lru_pages(zone, sc, l);
1835                 if (priority || noswap) {
1836                         scan >>= priority;
1837                         scan = div64_u64(scan * fraction[file], denominator);
1838                 }
1839                 nr[l] = nr_scan_try_batch(scan,
1840                                           &reclaim_stat->nr_saved_scan[l]);
1841         }
1842 }
1843
1844 /*
1845  * Reclaim/compaction depends on a number of pages being freed. To avoid
1846  * disruption to the system, a small number of order-0 pages continue to be
1847  * rotated and reclaimed in the normal fashion. However, by the time we get
1848  * back to the allocator and call try_to_compact_zone(), we ensure that
1849  * there are enough free pages for it to be likely successful
1850  */
1851 static inline bool should_continue_reclaim(struct zone *zone,
1852                                         unsigned long nr_reclaimed,
1853                                         unsigned long nr_scanned,
1854                                         struct scan_control *sc)
1855 {
1856         unsigned long pages_for_compaction;
1857         unsigned long inactive_lru_pages;
1858
1859         /* If not in reclaim/compaction mode, stop */
1860         if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
1861                 return false;
1862
1863         /* Consider stopping depending on scan and reclaim activity */
1864         if (sc->gfp_mask & __GFP_REPEAT) {
1865                 /*
1866                  * For __GFP_REPEAT allocations, stop reclaiming if the
1867                  * full LRU list has been scanned and we are still failing
1868                  * to reclaim pages. This full LRU scan is potentially
1869                  * expensive but a __GFP_REPEAT caller really wants to succeed
1870                  */
1871                 if (!nr_reclaimed && !nr_scanned)
1872                         return false;
1873         } else {
1874                 /*
1875                  * For non-__GFP_REPEAT allocations which can presumably
1876                  * fail without consequence, stop if we failed to reclaim
1877                  * any pages from the last SWAP_CLUSTER_MAX number of
1878                  * pages that were scanned. This will return to the
1879                  * caller faster at the risk reclaim/compaction and
1880                  * the resulting allocation attempt fails
1881                  */
1882                 if (!nr_reclaimed)
1883                         return false;
1884         }
1885
1886         /*
1887          * If we have not reclaimed enough pages for compaction and the
1888          * inactive lists are large enough, continue reclaiming
1889          */
1890         pages_for_compaction = (2UL << sc->order);
1891         inactive_lru_pages = zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON) +
1892                                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1893         if (sc->nr_reclaimed < pages_for_compaction &&
1894                         inactive_lru_pages > pages_for_compaction)
1895                 return true;
1896
1897         /* If compaction would go ahead or the allocation would succeed, stop */
1898         switch (compaction_suitable(zone, sc->order)) {
1899         case COMPACT_PARTIAL:
1900         case COMPACT_CONTINUE:
1901                 return false;
1902         default:
1903                 return true;
1904         }
1905 }
1906
1907 /*
1908  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1909  */
1910 static void shrink_zone(int priority, struct zone *zone,
1911                                 struct scan_control *sc)
1912 {
1913         unsigned long nr[NR_LRU_LISTS];
1914         unsigned long nr_to_scan;
1915         enum lru_list l;
1916         unsigned long nr_reclaimed, nr_scanned;
1917         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1918
1919 restart:
1920         nr_reclaimed = 0;
1921         nr_scanned = sc->nr_scanned;
1922         get_scan_count(zone, sc, nr, priority);
1923
1924         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1925                                         nr[LRU_INACTIVE_FILE]) {
1926                 for_each_evictable_lru(l) {
1927                         if (nr[l]) {
1928                                 nr_to_scan = min_t(unsigned long,
1929                                                    nr[l], SWAP_CLUSTER_MAX);
1930                                 nr[l] -= nr_to_scan;
1931
1932                                 nr_reclaimed += shrink_list(l, nr_to_scan,
1933                                                             zone, sc, priority);
1934                         }
1935                 }
1936                 /*
1937                  * On large memory systems, scan >> priority can become
1938                  * really large. This is fine for the starting priority;
1939                  * we want to put equal scanning pressure on each zone.
1940                  * However, if the VM has a harder time of freeing pages,
1941                  * with multiple processes reclaiming pages, the total
1942                  * freeing target can get unreasonably large.
1943                  */
1944                 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
1945                         break;
1946         }
1947         sc->nr_reclaimed += nr_reclaimed;
1948
1949         /*
1950          * Even if we did not try to evict anon pages at all, we want to
1951          * rebalance the anon lru active/inactive ratio.
1952          */
1953         if (inactive_anon_is_low(zone, sc))
1954                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1955
1956         /* reclaim/compaction might need reclaim to continue */
1957         if (should_continue_reclaim(zone, nr_reclaimed,
1958                                         sc->nr_scanned - nr_scanned, sc))
1959                 goto restart;
1960
1961         throttle_vm_writeout(sc->gfp_mask);
1962 }
1963
1964 /*
1965  * This is the direct reclaim path, for page-allocating processes.  We only
1966  * try to reclaim pages from zones which will satisfy the caller's allocation
1967  * request.
1968  *
1969  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1970  * Because:
1971  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1972  *    allocation or
1973  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1974  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1975  *    zone defense algorithm.
1976  *
1977  * If a zone is deemed to be full of pinned pages then just give it a light
1978  * scan then give up on it.
1979  */
1980 static unsigned long shrink_zones(int priority, struct zonelist *zonelist,
1981                                         struct scan_control *sc)
1982 {
1983         struct zoneref *z;
1984         struct zone *zone;
1985         unsigned long nr_soft_reclaimed;
1986         unsigned long nr_soft_scanned;
1987         unsigned long total_scanned = 0;
1988
1989         for_each_zone_zonelist_nodemask(zone, z, zonelist,
1990                                         gfp_zone(sc->gfp_mask), sc->nodemask) {
1991                 if (!populated_zone(zone))
1992                         continue;
1993                 /*
1994                  * Take care memory controller reclaiming has small influence
1995                  * to global LRU.
1996                  */
1997                 if (scanning_global_lru(sc)) {
1998                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1999                                 continue;
2000                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2001                                 continue;       /* Let kswapd poll it */
2002                 }
2003
2004                 nr_soft_scanned = 0;
2005                 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2006                                                         sc->order, sc->gfp_mask,
2007                                                         &nr_soft_scanned);
2008                 sc->nr_reclaimed += nr_soft_reclaimed;
2009                 total_scanned += nr_soft_scanned;
2010
2011                 shrink_zone(priority, zone, sc);
2012         }
2013
2014         return total_scanned;
2015 }
2016
2017 static bool zone_reclaimable(struct zone *zone)
2018 {
2019         return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2020 }
2021
2022 /* All zones in zonelist are unreclaimable? */
2023 static bool all_unreclaimable(struct zonelist *zonelist,
2024                 struct scan_control *sc)
2025 {
2026         struct zoneref *z;
2027         struct zone *zone;
2028
2029         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2030                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2031                 if (!populated_zone(zone))
2032                         continue;
2033                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2034                         continue;
2035                 if (!zone->all_unreclaimable)
2036                         return false;
2037         }
2038
2039         return true;
2040 }
2041
2042 /*
2043  * This is the main entry point to direct page reclaim.
2044  *
2045  * If a full scan of the inactive list fails to free enough memory then we
2046  * are "out of memory" and something needs to be killed.
2047  *
2048  * If the caller is !__GFP_FS then the probability of a failure is reasonably
2049  * high - the zone may be full of dirty or under-writeback pages, which this
2050  * caller can't do much about.  We kick the writeback threads and take explicit
2051  * naps in the hope that some of these pages can be written.  But if the
2052  * allocating task holds filesystem locks which prevent writeout this might not
2053  * work, and the allocation attempt will fail.
2054  *
2055  * returns:     0, if no pages reclaimed
2056  *              else, the number of pages reclaimed
2057  */
2058 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2059                                         struct scan_control *sc,
2060                                         struct shrink_control *shrink)
2061 {
2062         int priority;
2063         unsigned long total_scanned = 0;
2064         struct reclaim_state *reclaim_state = current->reclaim_state;
2065         struct zoneref *z;
2066         struct zone *zone;
2067         unsigned long writeback_threshold;
2068
2069         get_mems_allowed();
2070         delayacct_freepages_start();
2071
2072         if (scanning_global_lru(sc))
2073                 count_vm_event(ALLOCSTALL);
2074
2075         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2076                 sc->nr_scanned = 0;
2077                 if (!priority)
2078                         disable_swap_token();
2079                 total_scanned += shrink_zones(priority, zonelist, sc);
2080                 /*
2081                  * Don't shrink slabs when reclaiming memory from
2082                  * over limit cgroups
2083                  */
2084                 if (scanning_global_lru(sc)) {
2085                         unsigned long lru_pages = 0;
2086                         for_each_zone_zonelist(zone, z, zonelist,
2087                                         gfp_zone(sc->gfp_mask)) {
2088                                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2089                                         continue;
2090
2091                                 lru_pages += zone_reclaimable_pages(zone);
2092                         }
2093
2094                         shrink_slab(shrink, sc->nr_scanned, lru_pages);
2095                         if (reclaim_state) {
2096                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2097                                 reclaim_state->reclaimed_slab = 0;
2098                         }
2099                 }
2100                 total_scanned += sc->nr_scanned;
2101                 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2102                         goto out;
2103
2104                 /*
2105                  * Try to write back as many pages as we just scanned.  This
2106                  * tends to cause slow streaming writers to write data to the
2107                  * disk smoothly, at the dirtying rate, which is nice.   But
2108                  * that's undesirable in laptop mode, where we *want* lumpy
2109                  * writeout.  So in laptop mode, write out the whole world.
2110                  */
2111                 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2112                 if (total_scanned > writeback_threshold) {
2113                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
2114                         sc->may_writepage = 1;
2115                 }
2116
2117                 /* Take a nap, wait for some writeback to complete */
2118                 if (!sc->hibernation_mode && sc->nr_scanned &&
2119                     priority < DEF_PRIORITY - 2) {
2120                         struct zone *preferred_zone;
2121
2122                         first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2123                                                 &cpuset_current_mems_allowed,
2124                                                 &preferred_zone);
2125                         wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2126                 }
2127         }
2128
2129 out:
2130         delayacct_freepages_end();
2131         put_mems_allowed();
2132
2133         if (sc->nr_reclaimed)
2134                 return sc->nr_reclaimed;
2135
2136         /*
2137          * As hibernation is going on, kswapd is freezed so that it can't mark
2138          * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2139          * check.
2140          */
2141         if (oom_killer_disabled)
2142                 return 0;
2143
2144         /* top priority shrink_zones still had more to do? don't OOM, then */
2145         if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
2146                 return 1;
2147
2148         return 0;
2149 }
2150
2151 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2152                                 gfp_t gfp_mask, nodemask_t *nodemask)
2153 {
2154         unsigned long nr_reclaimed;
2155         struct scan_control sc = {
2156                 .gfp_mask = gfp_mask,
2157                 .may_writepage = !laptop_mode,
2158                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2159                 .may_unmap = 1,
2160                 .may_swap = 1,
2161                 .swappiness = vm_swappiness,
2162                 .order = order,
2163                 .mem_cgroup = NULL,
2164                 .nodemask = nodemask,
2165         };
2166         struct shrink_control shrink = {
2167                 .gfp_mask = sc.gfp_mask,
2168         };
2169
2170         trace_mm_vmscan_direct_reclaim_begin(order,
2171                                 sc.may_writepage,
2172                                 gfp_mask);
2173
2174         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2175
2176         trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2177
2178         return nr_reclaimed;
2179 }
2180
2181 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2182
2183 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2184                                                 gfp_t gfp_mask, bool noswap,
2185                                                 unsigned int swappiness,
2186                                                 struct zone *zone,
2187                                                 unsigned long *nr_scanned)
2188 {
2189         struct scan_control sc = {
2190                 .nr_scanned = 0,
2191                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2192                 .may_writepage = !laptop_mode,
2193                 .may_unmap = 1,
2194                 .may_swap = !noswap,
2195                 .swappiness = swappiness,
2196                 .order = 0,
2197                 .mem_cgroup = mem,
2198         };
2199
2200         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2201                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2202
2203         trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2204                                                       sc.may_writepage,
2205                                                       sc.gfp_mask);
2206
2207         /*
2208          * NOTE: Although we can get the priority field, using it
2209          * here is not a good idea, since it limits the pages we can scan.
2210          * if we don't reclaim here, the shrink_zone from balance_pgdat
2211          * will pick up pages from other mem cgroup's as well. We hack
2212          * the priority and make it zero.
2213          */
2214         shrink_zone(0, zone, &sc);
2215
2216         trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2217
2218         *nr_scanned = sc.nr_scanned;
2219         return sc.nr_reclaimed;
2220 }
2221
2222 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
2223                                            gfp_t gfp_mask,
2224                                            bool noswap,
2225                                            unsigned int swappiness)
2226 {
2227         struct zonelist *zonelist;
2228         unsigned long nr_reclaimed;
2229         int nid;
2230         struct scan_control sc = {
2231                 .may_writepage = !laptop_mode,
2232                 .may_unmap = 1,
2233                 .may_swap = !noswap,
2234                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2235                 .swappiness = swappiness,
2236                 .order = 0,
2237                 .mem_cgroup = mem_cont,
2238                 .nodemask = NULL, /* we don't care the placement */
2239                 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2240                                 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2241         };
2242         struct shrink_control shrink = {
2243                 .gfp_mask = sc.gfp_mask,
2244         };
2245
2246         /*
2247          * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2248          * take care of from where we get pages. So the node where we start the
2249          * scan does not need to be the current node.
2250          */
2251         nid = mem_cgroup_select_victim_node(mem_cont);
2252
2253         zonelist = NODE_DATA(nid)->node_zonelists;
2254
2255         trace_mm_vmscan_memcg_reclaim_begin(0,
2256                                             sc.may_writepage,
2257                                             sc.gfp_mask);
2258
2259         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2260
2261         trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2262
2263         return nr_reclaimed;
2264 }
2265 #endif
2266
2267 /*
2268  * pgdat_balanced is used when checking if a node is balanced for high-order
2269  * allocations. Only zones that meet watermarks and are in a zone allowed
2270  * by the callers classzone_idx are added to balanced_pages. The total of
2271  * balanced pages must be at least 25% of the zones allowed by classzone_idx
2272  * for the node to be considered balanced. Forcing all zones to be balanced
2273  * for high orders can cause excessive reclaim when there are imbalanced zones.
2274  * The choice of 25% is due to
2275  *   o a 16M DMA zone that is balanced will not balance a zone on any
2276  *     reasonable sized machine
2277  *   o On all other machines, the top zone must be at least a reasonable
2278  *     percentage of the middle zones. For example, on 32-bit x86, highmem
2279  *     would need to be at least 256M for it to be balance a whole node.
2280  *     Similarly, on x86-64 the Normal zone would need to be at least 1G
2281  *     to balance a node on its own. These seemed like reasonable ratios.
2282  */
2283 static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2284                                                 int classzone_idx)
2285 {
2286         unsigned long present_pages = 0;
2287         int i;
2288
2289         for (i = 0; i <= classzone_idx; i++)
2290                 present_pages += pgdat->node_zones[i].present_pages;
2291
2292         return balanced_pages > (present_pages >> 2);
2293 }
2294
2295 /* is kswapd sleeping prematurely? */
2296 static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2297                                         int classzone_idx)
2298 {
2299         int i;
2300         unsigned long balanced = 0;
2301         bool all_zones_ok = true;
2302
2303         /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2304         if (remaining)
2305                 return true;
2306
2307         /* Check the watermark levels */
2308         for (i = 0; i < pgdat->nr_zones; i++) {
2309                 struct zone *zone = pgdat->node_zones + i;
2310
2311                 if (!populated_zone(zone))
2312                         continue;
2313
2314                 /*
2315                  * balance_pgdat() skips over all_unreclaimable after
2316                  * DEF_PRIORITY. Effectively, it considers them balanced so
2317                  * they must be considered balanced here as well if kswapd
2318                  * is to sleep
2319                  */
2320                 if (zone->all_unreclaimable) {
2321                         balanced += zone->present_pages;
2322                         continue;
2323                 }
2324
2325                 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2326                                                         classzone_idx, 0))
2327                         all_zones_ok = false;
2328                 else
2329                         balanced += zone->present_pages;
2330         }
2331
2332         /*
2333          * For high-order requests, the balanced zones must contain at least
2334          * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2335          * must be balanced
2336          */
2337         if (order)
2338                 return !pgdat_balanced(pgdat, balanced, classzone_idx);
2339         else
2340                 return !all_zones_ok;
2341 }
2342
2343 /*
2344  * For kswapd, balance_pgdat() will work across all this node's zones until
2345  * they are all at high_wmark_pages(zone).
2346  *
2347  * Returns the final order kswapd was reclaiming at
2348  *
2349  * There is special handling here for zones which are full of pinned pages.
2350  * This can happen if the pages are all mlocked, or if they are all used by
2351  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
2352  * What we do is to detect the case where all pages in the zone have been
2353  * scanned twice and there has been zero successful reclaim.  Mark the zone as
2354  * dead and from now on, only perform a short scan.  Basically we're polling
2355  * the zone for when the problem goes away.
2356  *
2357  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
2358  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2359  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2360  * lower zones regardless of the number of free pages in the lower zones. This
2361  * interoperates with the page allocator fallback scheme to ensure that aging
2362  * of pages is balanced across the zones.
2363  */
2364 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2365                                                         int *classzone_idx)
2366 {
2367         int all_zones_ok;
2368         unsigned long balanced;
2369         int priority;
2370         int i;
2371         int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
2372         unsigned long total_scanned;
2373         struct reclaim_state *reclaim_state = current->reclaim_state;
2374         unsigned long nr_soft_reclaimed;
2375         unsigned long nr_soft_scanned;
2376         struct scan_control sc = {
2377                 .gfp_mask = GFP_KERNEL,
2378                 .may_unmap = 1,
2379                 .may_swap = 1,
2380                 /*
2381                  * kswapd doesn't want to be bailed out while reclaim. because
2382                  * we want to put equal scanning pressure on each zone.
2383                  */
2384                 .nr_to_reclaim = ULONG_MAX,
2385                 .swappiness = vm_swappiness,
2386                 .order = order,
2387                 .mem_cgroup = NULL,
2388         };
2389         struct shrink_control shrink = {
2390                 .gfp_mask = sc.gfp_mask,
2391         };
2392 loop_again:
2393         total_scanned = 0;
2394         sc.nr_reclaimed = 0;
2395         sc.may_writepage = !laptop_mode;
2396         count_vm_event(PAGEOUTRUN);
2397
2398         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2399                 unsigned long lru_pages = 0;
2400                 int has_under_min_watermark_zone = 0;
2401
2402                 /* The swap token gets in the way of swapout... */
2403                 if (!priority)
2404                         disable_swap_token();
2405
2406                 all_zones_ok = 1;
2407                 balanced = 0;
2408
2409                 /*
2410                  * Scan in the highmem->dma direction for the highest
2411                  * zone which needs scanning
2412                  */
2413                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2414                         struct zone *zone = pgdat->node_zones + i;
2415
2416                         if (!populated_zone(zone))
2417                                 continue;
2418
2419                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2420                                 continue;
2421
2422                         /*
2423                          * Do some background aging of the anon list, to give
2424                          * pages a chance to be referenced before reclaiming.
2425                          */
2426                         if (inactive_anon_is_low(zone, &sc))
2427                                 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2428                                                         &sc, priority, 0);
2429
2430                         if (!zone_watermark_ok_safe(zone, order,
2431                                         high_wmark_pages(zone), 0, 0)) {
2432                                 end_zone = i;
2433                                 *classzone_idx = i;
2434                                 break;
2435                         }
2436                 }
2437                 if (i < 0)
2438                         goto out;
2439
2440                 for (i = 0; i <= end_zone; i++) {
2441                         struct zone *zone = pgdat->node_zones + i;
2442
2443                         lru_pages += zone_reclaimable_pages(zone);
2444                 }
2445
2446                 /*
2447                  * Now scan the zone in the dma->highmem direction, stopping
2448                  * at the last zone which needs scanning.
2449                  *
2450                  * We do this because the page allocator works in the opposite
2451                  * direction.  This prevents the page allocator from allocating
2452                  * pages behind kswapd's direction of progress, which would
2453                  * cause too much scanning of the lower zones.
2454                  */
2455                 for (i = 0; i <= end_zone; i++) {
2456                         struct zone *zone = pgdat->node_zones + i;
2457                         int nr_slab;
2458                         unsigned long balance_gap;
2459
2460                         if (!populated_zone(zone))
2461                                 continue;
2462
2463                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2464                                 continue;
2465
2466                         sc.nr_scanned = 0;
2467
2468                         nr_soft_scanned = 0;
2469                         /*
2470                          * Call soft limit reclaim before calling shrink_zone.
2471                          */
2472                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2473                                                         order, sc.gfp_mask,
2474                                                         &nr_soft_scanned);
2475                         sc.nr_reclaimed += nr_soft_reclaimed;
2476                         total_scanned += nr_soft_scanned;
2477
2478                         /*
2479                          * We put equal pressure on every zone, unless
2480                          * one zone has way too many pages free
2481                          * already. The "too many pages" is defined
2482                          * as the high wmark plus a "gap" where the
2483                          * gap is either the low watermark or 1%
2484                          * of the zone, whichever is smaller.
2485                          */
2486                         balance_gap = min(low_wmark_pages(zone),
2487                                 (zone->present_pages +
2488                                         KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2489                                 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2490                         if (!zone_watermark_ok_safe(zone, order,
2491                                         high_wmark_pages(zone) + balance_gap,
2492                                         end_zone, 0))
2493                                 shrink_zone(priority, zone, &sc);
2494                         reclaim_state->reclaimed_slab = 0;
2495                         nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2496                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2497                         total_scanned += sc.nr_scanned;
2498
2499                         if (zone->all_unreclaimable)
2500                                 continue;
2501                         if (nr_slab == 0 &&
2502                             !zone_reclaimable(zone))
2503                                 zone->all_unreclaimable = 1;
2504                         /*
2505                          * If we've done a decent amount of scanning and
2506                          * the reclaim ratio is low, start doing writepage
2507                          * even in laptop mode
2508                          */
2509                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2510                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2511                                 sc.may_writepage = 1;
2512
2513                         if (!zone_watermark_ok_safe(zone, order,
2514                                         high_wmark_pages(zone), end_zone, 0)) {
2515                                 all_zones_ok = 0;
2516                                 /*
2517                                  * We are still under min water mark.  This
2518                                  * means that we have a GFP_ATOMIC allocation
2519                                  * failure risk. Hurry up!
2520                                  */
2521                                 if (!zone_watermark_ok_safe(zone, order,
2522                                             min_wmark_pages(zone), end_zone, 0))
2523                                         has_under_min_watermark_zone = 1;
2524                         } else {
2525                                 /*
2526                                  * If a zone reaches its high watermark,
2527                                  * consider it to be no longer congested. It's
2528                                  * possible there are dirty pages backed by
2529                                  * congested BDIs but as pressure is relieved,
2530                                  * spectulatively avoid congestion waits
2531                                  */
2532                                 zone_clear_flag(zone, ZONE_CONGESTED);
2533                                 if (i <= *classzone_idx)
2534                                         balanced += zone->present_pages;
2535                         }
2536
2537                 }
2538                 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2539                         break;          /* kswapd: all done */
2540                 /*
2541                  * OK, kswapd is getting into trouble.  Take a nap, then take
2542                  * another pass across the zones.
2543                  */
2544                 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2545                         if (has_under_min_watermark_zone)
2546                                 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2547                         else
2548                                 congestion_wait(BLK_RW_ASYNC, HZ/10);
2549                 }
2550
2551                 /*
2552                  * We do this so kswapd doesn't build up large priorities for
2553                  * example when it is freeing in parallel with allocators. It
2554                  * matches the direct reclaim path behaviour in terms of impact
2555                  * on zone->*_priority.
2556                  */
2557                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2558                         break;
2559         }
2560 out:
2561
2562         /*
2563          * order-0: All zones must meet high watermark for a balanced node
2564          * high-order: Balanced zones must make up at least 25% of the node
2565          *             for the node to be balanced
2566          */
2567         if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2568                 cond_resched();
2569
2570                 try_to_freeze();
2571
2572                 /*
2573                  * Fragmentation may mean that the system cannot be
2574                  * rebalanced for high-order allocations in all zones.
2575                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2576                  * it means the zones have been fully scanned and are still
2577                  * not balanced. For high-order allocations, there is
2578                  * little point trying all over again as kswapd may
2579                  * infinite loop.
2580                  *
2581                  * Instead, recheck all watermarks at order-0 as they
2582                  * are the most important. If watermarks are ok, kswapd will go
2583                  * back to sleep. High-order users can still perform direct
2584                  * reclaim if they wish.
2585                  */
2586                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2587                         order = sc.order = 0;
2588
2589                 goto loop_again;
2590         }
2591
2592         /*
2593          * If kswapd was reclaiming at a higher order, it has the option of
2594          * sleeping without all zones being balanced. Before it does, it must
2595          * ensure that the watermarks for order-0 on *all* zones are met and
2596          * that the congestion flags are cleared. The congestion flag must
2597          * be cleared as kswapd is the only mechanism that clears the flag
2598          * and it is potentially going to sleep here.
2599          */
2600         if (order) {
2601                 for (i = 0; i <= end_zone; i++) {
2602                         struct zone *zone = pgdat->node_zones + i;
2603
2604                         if (!populated_zone(zone))
2605                                 continue;
2606
2607                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2608                                 continue;
2609
2610                         /* Confirm the zone is balanced for order-0 */
2611                         if (!zone_watermark_ok(zone, 0,
2612                                         high_wmark_pages(zone), 0, 0)) {
2613                                 order = sc.order = 0;
2614                                 goto loop_again;
2615                         }
2616
2617                         /* If balanced, clear the congested flag */
2618                         zone_clear_flag(zone, ZONE_CONGESTED);
2619                 }
2620         }
2621
2622         /*
2623          * Return the order we were reclaiming at so sleeping_prematurely()
2624          * makes a decision on the order we were last reclaiming at. However,
2625          * if another caller entered the allocator slow path while kswapd
2626          * was awake, order will remain at the higher level
2627          */
2628         *classzone_idx = end_zone;
2629         return order;
2630 }
2631
2632 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2633 {
2634         long remaining = 0;
2635         DEFINE_WAIT(wait);
2636
2637         if (freezing(current) || kthread_should_stop())
2638                 return;
2639
2640         prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2641
2642         /* Try to sleep for a short interval */
2643         if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2644                 remaining = schedule_timeout(HZ/10);
2645                 finish_wait(&pgdat->kswapd_wait, &wait);
2646                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2647         }
2648
2649         /*
2650          * After a short sleep, check if it was a premature sleep. If not, then
2651          * go fully to sleep until explicitly woken up.
2652          */
2653         if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2654                 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2655
2656                 /*
2657                  * vmstat counters are not perfectly accurate and the estimated
2658                  * value for counters such as NR_FREE_PAGES can deviate from the
2659                  * true value by nr_online_cpus * threshold. To avoid the zone
2660                  * watermarks being breached while under pressure, we reduce the
2661                  * per-cpu vmstat threshold while kswapd is awake and restore
2662                  * them before going back to sleep.
2663                  */
2664                 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2665                 schedule();
2666                 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2667         } else {
2668                 if (remaining)
2669                         count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2670                 else
2671                         count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2672         }
2673         finish_wait(&pgdat->kswapd_wait, &wait);
2674 }
2675
2676 /*
2677  * The background pageout daemon, started as a kernel thread
2678  * from the init process.
2679  *
2680  * This basically trickles out pages so that we have _some_
2681  * free memory available even if there is no other activity
2682  * that frees anything up. This is needed for things like routing
2683  * etc, where we otherwise might have all activity going on in
2684  * asynchronous contexts that cannot page things out.
2685  *
2686  * If there are applications that are active memory-allocators
2687  * (most normal use), this basically shouldn't matter.
2688  */
2689 static int kswapd(void *p)
2690 {
2691         unsigned long order;
2692         int classzone_idx;
2693         pg_data_t *pgdat = (pg_data_t*)p;
2694         struct task_struct *tsk = current;
2695
2696         struct reclaim_state reclaim_state = {
2697                 .reclaimed_slab = 0,
2698         };
2699         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2700
2701         lockdep_set_current_reclaim_state(GFP_KERNEL);
2702
2703         if (!cpumask_empty(cpumask))
2704                 set_cpus_allowed_ptr(tsk, cpumask);
2705         current->reclaim_state = &reclaim_state;
2706
2707         /*
2708          * Tell the memory management that we're a "memory allocator",
2709          * and that if we need more memory we should get access to it
2710          * regardless (see "__alloc_pages()"). "kswapd" should
2711          * never get caught in the normal page freeing logic.
2712          *
2713          * (Kswapd normally doesn't need memory anyway, but sometimes
2714          * you need a small amount of memory in order to be able to
2715          * page out something else, and this flag essentially protects
2716          * us from recursively trying to free more memory as we're
2717          * trying to free the first piece of memory in the first place).
2718          */
2719         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2720         set_freezable();
2721
2722         order = 0;
2723         classzone_idx = MAX_NR_ZONES - 1;
2724         for ( ; ; ) {
2725                 unsigned long new_order;
2726                 int new_classzone_idx;
2727                 int ret;
2728
2729                 new_order = pgdat->kswapd_max_order;
2730                 new_classzone_idx = pgdat->classzone_idx;
2731                 pgdat->kswapd_max_order = 0;
2732                 pgdat->classzone_idx = MAX_NR_ZONES - 1;
2733                 if (order < new_order || classzone_idx > new_classzone_idx) {
2734                         /*
2735                          * Don't sleep if someone wants a larger 'order'
2736                          * allocation or has tigher zone constraints
2737                          */
2738                         order = new_order;
2739                         classzone_idx = new_classzone_idx;
2740                 } else {
2741                         kswapd_try_to_sleep(pgdat, order, classzone_idx);
2742                         order = pgdat->kswapd_max_order;
2743                         classzone_idx = pgdat->classzone_idx;
2744                         pgdat->kswapd_max_order = 0;
2745                         pgdat->classzone_idx = MAX_NR_ZONES - 1;
2746                 }
2747
2748                 ret = try_to_freeze();
2749                 if (kthread_should_stop())
2750                         break;
2751
2752                 /*
2753                  * We can speed up thawing tasks if we don't call balance_pgdat
2754                  * after returning from the refrigerator
2755                  */
2756                 if (!ret) {
2757                         trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2758                         order = balance_pgdat(pgdat, order, &classzone_idx);
2759                 }
2760         }
2761         return 0;
2762 }
2763
2764 /*
2765  * A zone is low on free memory, so wake its kswapd task to service it.
2766  */
2767 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
2768 {
2769         pg_data_t *pgdat;
2770
2771         if (!populated_zone(zone))
2772                 return;
2773
2774         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2775                 return;
2776         pgdat = zone->zone_pgdat;
2777         if (pgdat->kswapd_max_order < order) {
2778                 pgdat->kswapd_max_order = order;
2779                 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
2780         }
2781         if (!waitqueue_active(&pgdat->kswapd_wait))
2782                 return;
2783         if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
2784                 return;
2785
2786         trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
2787         wake_up_interruptible(&pgdat->kswapd_wait);
2788 }
2789
2790 /*
2791  * The reclaimable count would be mostly accurate.
2792  * The less reclaimable pages may be
2793  * - mlocked pages, which will be moved to unevictable list when encountered
2794  * - mapped pages, which may require several travels to be reclaimed
2795  * - dirty pages, which is not "instantly" reclaimable
2796  */
2797 unsigned long global_reclaimable_pages(void)
2798 {
2799         int nr;
2800
2801         nr = global_page_state(NR_ACTIVE_FILE) +
2802              global_page_state(NR_INACTIVE_FILE);
2803
2804         if (nr_swap_pages > 0)
2805                 nr += global_page_state(NR_ACTIVE_ANON) +
2806                       global_page_state(NR_INACTIVE_ANON);
2807
2808         return nr;
2809 }
2810
2811 unsigned long zone_reclaimable_pages(struct zone *zone)
2812 {
2813         int nr;
2814
2815         nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2816              zone_page_state(zone, NR_INACTIVE_FILE);
2817
2818         if (nr_swap_pages > 0)
2819                 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2820                       zone_page_state(zone, NR_INACTIVE_ANON);
2821
2822         return nr;
2823 }
2824
2825 #ifdef CONFIG_HIBERNATION
2826 /*
2827  * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2828  * freed pages.
2829  *
2830  * Rather than trying to age LRUs the aim is to preserve the overall
2831  * LRU order by reclaiming preferentially
2832  * inactive > active > active referenced > active mapped
2833  */
2834 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
2835 {
2836         struct reclaim_state reclaim_state;
2837         struct scan_control sc = {
2838                 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2839                 .may_swap = 1,
2840                 .may_unmap = 1,
2841                 .may_writepage = 1,
2842                 .nr_to_reclaim = nr_to_reclaim,
2843                 .hibernation_mode = 1,
2844                 .swappiness = vm_swappiness,
2845                 .order = 0,
2846         };
2847         struct shrink_control shrink = {
2848                 .gfp_mask = sc.gfp_mask,
2849         };
2850         struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2851         struct task_struct *p = current;
2852         unsigned long nr_reclaimed;
2853
2854         p->flags |= PF_MEMALLOC;
2855         lockdep_set_current_reclaim_state(sc.gfp_mask);
2856         reclaim_state.reclaimed_slab = 0;
2857         p->reclaim_state = &reclaim_state;
2858
2859         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2860
2861         p->reclaim_state = NULL;
2862         lockdep_clear_current_reclaim_state();
2863         p->flags &= ~PF_MEMALLOC;
2864
2865         return nr_reclaimed;
2866 }
2867 #endif /* CONFIG_HIBERNATION */
2868
2869 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2870    not required for correctness.  So if the last cpu in a node goes
2871    away, we get changed to run anywhere: as the first one comes back,
2872    restore their cpu bindings. */
2873 static int __devinit cpu_callback(struct notifier_block *nfb,
2874                                   unsigned long action, void *hcpu)
2875 {
2876         int nid;
2877
2878         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2879                 for_each_node_state(nid, N_HIGH_MEMORY) {
2880                         pg_data_t *pgdat = NODE_DATA(nid);
2881                         const struct cpumask *mask;
2882
2883                         mask = cpumask_of_node(pgdat->node_id);
2884
2885                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2886                                 /* One of our CPUs online: restore mask */
2887                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2888                 }
2889         }
2890         return NOTIFY_OK;
2891 }
2892
2893 /*
2894  * This kswapd start function will be called by init and node-hot-add.
2895  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2896  */
2897 int kswapd_run(int nid)
2898 {
2899         pg_data_t *pgdat = NODE_DATA(nid);
2900         int ret = 0;
2901
2902         if (pgdat->kswapd)
2903                 return 0;
2904
2905         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2906         if (IS_ERR(pgdat->kswapd)) {
2907                 /* failure at boot is fatal */
2908                 BUG_ON(system_state == SYSTEM_BOOTING);
2909                 printk("Failed to start kswapd on node %d\n",nid);
2910                 ret = -1;
2911         }
2912         return ret;
2913 }
2914
2915 /*
2916  * Called by memory hotplug when all memory in a node is offlined.
2917  */
2918 void kswapd_stop(int nid)
2919 {
2920         struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
2921
2922         if (kswapd)
2923                 kthread_stop(kswapd);
2924 }
2925
2926 static int __init kswapd_init(void)
2927 {
2928         int nid;
2929
2930         swap_setup();
2931         for_each_node_state(nid, N_HIGH_MEMORY)
2932                 kswapd_run(nid);
2933         hotcpu_notifier(cpu_callback, 0);
2934         return 0;
2935 }
2936
2937 module_init(kswapd_init)
2938
2939 #ifdef CONFIG_NUMA
2940 /*
2941  * Zone reclaim mode
2942  *
2943  * If non-zero call zone_reclaim when the number of free pages falls below
2944  * the watermarks.
2945  */
2946 int zone_reclaim_mode __read_mostly;
2947
2948 #define RECLAIM_OFF 0
2949 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
2950 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
2951 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
2952
2953 /*
2954  * Priority for ZONE_RECLAIM. This determines the fraction of pages
2955  * of a node considered for each zone_reclaim. 4 scans 1/16th of
2956  * a zone.
2957  */
2958 #define ZONE_RECLAIM_PRIORITY 4
2959
2960 /*
2961  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2962  * occur.
2963  */
2964 int sysctl_min_unmapped_ratio = 1;
2965
2966 /*
2967  * If the number of slab pages in a zone grows beyond this percentage then
2968  * slab reclaim needs to occur.
2969  */
2970 int sysctl_min_slab_ratio = 5;
2971
2972 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
2973 {
2974         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
2975         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
2976                 zone_page_state(zone, NR_ACTIVE_FILE);
2977
2978         /*
2979          * It's possible for there to be more file mapped pages than
2980          * accounted for by the pages on the file LRU lists because
2981          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2982          */
2983         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
2984 }
2985
2986 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
2987 static long zone_pagecache_reclaimable(struct zone *zone)
2988 {
2989         long nr_pagecache_reclaimable;
2990         long delta = 0;
2991
2992         /*
2993          * If RECLAIM_SWAP is set, then all file pages are considered
2994          * potentially reclaimable. Otherwise, we have to worry about
2995          * pages like swapcache and zone_unmapped_file_pages() provides
2996          * a better estimate
2997          */
2998         if (zone_reclaim_mode & RECLAIM_SWAP)
2999                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3000         else
3001                 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3002
3003         /* If we can't clean pages, remove dirty pages from consideration */
3004         if (!(zone_reclaim_mode & RECLAIM_WRITE))
3005                 delta += zone_page_state(zone, NR_FILE_DIRTY);
3006
3007         /* Watch for any possible underflows due to delta */
3008         if (unlikely(delta > nr_pagecache_reclaimable))
3009                 delta = nr_pagecache_reclaimable;
3010
3011         return nr_pagecache_reclaimable - delta;
3012 }
3013
3014 /*
3015  * Try to free up some pages from this zone through reclaim.
3016  */
3017 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3018 {
3019         /* Minimum pages needed in order to stay on node */
3020         const unsigned long nr_pages = 1 << order;
3021         struct task_struct *p = current;
3022         struct reclaim_state reclaim_state;
3023         int priority;
3024         struct scan_control sc = {
3025                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3026                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3027                 .may_swap = 1,
3028                 .nr_to_reclaim = max_t(unsigned long, nr_pages,
3029                                        SWAP_CLUSTER_MAX),
3030                 .gfp_mask = gfp_mask,
3031                 .swappiness = vm_swappiness,
3032                 .order = order,
3033         };
3034         struct shrink_control shrink = {
3035                 .gfp_mask = sc.gfp_mask,
3036         };
3037         unsigned long nr_slab_pages0, nr_slab_pages1;
3038
3039         cond_resched();
3040         /*
3041          * We need to be able to allocate from the reserves for RECLAIM_SWAP
3042          * and we also need to be able to write out pages for RECLAIM_WRITE
3043          * and RECLAIM_SWAP.
3044          */
3045         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3046         lockdep_set_current_reclaim_state(gfp_mask);
3047         reclaim_state.reclaimed_slab = 0;
3048         p->reclaim_state = &reclaim_state;
3049
3050         if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3051                 /*
3052                  * Free memory by calling shrink zone with increasing
3053                  * priorities until we have enough memory freed.
3054                  */
3055                 priority = ZONE_RECLAIM_PRIORITY;
3056                 do {
3057                         shrink_zone(priority, zone, &sc);
3058                         priority--;
3059                 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
3060         }
3061
3062         nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3063         if (nr_slab_pages0 > zone->min_slab_pages) {
3064                 /*
3065                  * shrink_slab() does not currently allow us to determine how
3066                  * many pages were freed in this zone. So we take the current
3067                  * number of slab pages and shake the slab until it is reduced
3068                  * by the same nr_pages that we used for reclaiming unmapped
3069                  * pages.
3070                  *
3071                  * Note that shrink_slab will free memory on all zones and may
3072                  * take a long time.
3073                  */
3074                 for (;;) {
3075                         unsigned long lru_pages = zone_reclaimable_pages(zone);
3076
3077                         /* No reclaimable slab or very low memory pressure */
3078                         if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3079                                 break;
3080
3081                         /* Freed enough memory */
3082                         nr_slab_pages1 = zone_page_state(zone,
3083                                                         NR_SLAB_RECLAIMABLE);
3084                         if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3085                                 break;
3086                 }
3087
3088                 /*
3089                  * Update nr_reclaimed by the number of slab pages we
3090                  * reclaimed from this zone.
3091                  */
3092                 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3093                 if (nr_slab_pages1 < nr_slab_pages0)
3094                         sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3095         }
3096
3097         p->reclaim_state = NULL;
3098         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3099         lockdep_clear_current_reclaim_state();
3100         return sc.nr_reclaimed >= nr_pages;
3101 }
3102
3103 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3104 {
3105         int node_id;
3106         int ret;
3107
3108         /*
3109          * Zone reclaim reclaims unmapped file backed pages and
3110          * slab pages if we are over the defined limits.
3111          *
3112          * A small portion of unmapped file backed pages is needed for
3113          * file I/O otherwise pages read by file I/O will be immediately
3114          * thrown out if the zone is overallocated. So we do not reclaim
3115          * if less than a specified percentage of the zone is used by
3116          * unmapped file backed pages.
3117          */
3118         if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3119             zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3120                 return ZONE_RECLAIM_FULL;
3121
3122         if (zone->all_unreclaimable)
3123                 return ZONE_RECLAIM_FULL;
3124
3125         /*
3126          * Do not scan if the allocation should not be delayed.
3127          */
3128         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3129                 return ZONE_RECLAIM_NOSCAN;
3130
3131         /*
3132          * Only run zone reclaim on the local zone or on zones that do not
3133          * have associated processors. This will favor the local processor
3134          * over remote processors and spread off node memory allocations
3135          * as wide as possible.
3136          */
3137         node_id = zone_to_nid(zone);
3138         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3139                 return ZONE_RECLAIM_NOSCAN;
3140
3141         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3142                 return ZONE_RECLAIM_NOSCAN;
3143
3144         ret = __zone_reclaim(zone, gfp_mask, order);
3145         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3146
3147         if (!ret)
3148                 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3149
3150         return ret;
3151 }
3152 #endif
3153
3154 /*
3155  * page_evictable - test whether a page is evictable
3156  * @page: the page to test
3157  * @vma: the VMA in which the page is or will be mapped, may be NULL
3158  *
3159  * Test whether page is evictable--i.e., should be placed on active/inactive
3160  * lists vs unevictable list.  The vma argument is !NULL when called from the
3161  * fault path to determine how to instantate a new page.
3162  *
3163  * Reasons page might not be evictable:
3164  * (1) page's mapping marked unevictable
3165  * (2) page is part of an mlocked VMA
3166  *
3167  */
3168 int page_evictable(struct page *page, struct vm_area_struct *vma)
3169 {
3170
3171         if (mapping_unevictable(page_mapping(page)))
3172                 return 0;
3173
3174         if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
3175                 return 0;
3176
3177         return 1;
3178 }
3179
3180 /**
3181  * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
3182  * @page: page to check evictability and move to appropriate lru list
3183  * @zone: zone page is in
3184  *
3185  * Checks a page for evictability and moves the page to the appropriate
3186  * zone lru list.
3187  *
3188  * Restrictions: zone->lru_lock must be held, page must be on LRU and must
3189  * have PageUnevictable set.
3190  */
3191 static void check_move_unevictable_page(struct page *page, struct zone *zone)
3192 {
3193         VM_BUG_ON(PageActive(page));
3194
3195 retry:
3196         ClearPageUnevictable(page);
3197         if (page_evictable(page, NULL)) {
3198                 enum lru_list l = page_lru_base_type(page);
3199
3200                 __dec_zone_state(zone, NR_UNEVICTABLE);
3201                 list_move(&page->lru, &zone->lru[l].list);
3202                 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
3203                 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
3204                 __count_vm_event(UNEVICTABLE_PGRESCUED);
3205         } else {
3206                 /*
3207                  * rotate unevictable list
3208                  */
3209                 SetPageUnevictable(page);
3210                 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
3211                 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
3212                 if (page_evictable(page, NULL))
3213                         goto retry;
3214         }
3215 }
3216
3217 /**
3218  * scan_mapping_unevictable_pages - scan an address space for evictable pages
3219  * @mapping: struct address_space to scan for evictable pages
3220  *
3221  * Scan all pages in mapping.  Check unevictable pages for
3222  * evictability and move them to the appropriate zone lru list.
3223  */
3224 void scan_mapping_unevictable_pages(struct address_space *mapping)
3225 {
3226         pgoff_t next = 0;
3227         pgoff_t end   = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
3228                          PAGE_CACHE_SHIFT;
3229         struct zone *zone;
3230         struct pagevec pvec;
3231
3232         if (mapping->nrpages == 0)
3233                 return;
3234
3235         pagevec_init(&pvec, 0);
3236         while (next < end &&
3237                 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
3238                 int i;
3239                 int pg_scanned = 0;
3240
3241                 zone = NULL;
3242
3243                 for (i = 0; i < pagevec_count(&pvec); i++) {
3244                         struct page *page = pvec.pages[i];
3245                         pgoff_t page_index = page->index;
3246                         struct zone *pagezone = page_zone(page);
3247
3248                         pg_scanned++;
3249                         if (page_index > next)
3250                                 next = page_index;
3251                         next++;
3252
3253                         if (pagezone != zone) {
3254                                 if (zone)
3255                                         spin_unlock_irq(&zone->lru_lock);
3256                                 zone = pagezone;
3257                                 spin_lock_irq(&zone->lru_lock);
3258                         }
3259
3260                         if (PageLRU(page) && PageUnevictable(page))
3261                                 check_move_unevictable_page(page, zone);
3262                 }
3263                 if (zone)
3264                         spin_unlock_irq(&zone->lru_lock);
3265                 pagevec_release(&pvec);
3266
3267                 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
3268         }
3269
3270 }
3271
3272 /**
3273  * scan_zone_unevictable_pages - check unevictable list for evictable pages
3274  * @zone - zone of which to scan the unevictable list
3275  *
3276  * Scan @zone's unevictable LRU lists to check for pages that have become
3277  * evictable.  Move those that have to @zone's inactive list where they
3278  * become candidates for reclaim, unless shrink_inactive_zone() decides
3279  * to reactivate them.  Pages that are still unevictable are rotated
3280  * back onto @zone's unevictable list.
3281  */
3282 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
3283 static void scan_zone_unevictable_pages(struct zone *zone)
3284 {
3285         struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
3286         unsigned long scan;
3287         unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
3288
3289         while (nr_to_scan > 0) {
3290                 unsigned long batch_size = min(nr_to_scan,
3291                                                 SCAN_UNEVICTABLE_BATCH_SIZE);
3292
3293                 spin_lock_irq(&zone->lru_lock);
3294                 for (scan = 0;  scan < batch_size; scan++) {
3295                         struct page *page = lru_to_page(l_unevictable);
3296
3297                         if (!trylock_page(page))
3298                                 continue;
3299
3300                         prefetchw_prev_lru_page(page, l_unevictable, flags);
3301
3302                         if (likely(PageLRU(page) && PageUnevictable(page)))
3303                                 check_move_unevictable_page(page, zone);
3304
3305                         unlock_page(page);
3306                 }
3307                 spin_unlock_irq(&zone->lru_lock);
3308
3309                 nr_to_scan -= batch_size;
3310         }
3311 }
3312
3313
3314 /**
3315  * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
3316  *
3317  * A really big hammer:  scan all zones' unevictable LRU lists to check for
3318  * pages that have become evictable.  Move those back to the zones'
3319  * inactive list where they become candidates for reclaim.
3320  * This occurs when, e.g., we have unswappable pages on the unevictable lists,
3321  * and we add swap to the system.  As such, it runs in the context of a task
3322  * that has possibly/probably made some previously unevictable pages
3323  * evictable.
3324  */
3325 static void scan_all_zones_unevictable_pages(void)
3326 {
3327         struct zone *zone;
3328
3329         for_each_zone(zone) {
3330                 scan_zone_unevictable_pages(zone);
3331         }
3332 }
3333
3334 /*
3335  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
3336  * all nodes' unevictable lists for evictable pages
3337  */
3338 unsigned long scan_unevictable_pages;
3339
3340 int scan_unevictable_handler(struct ctl_table *table, int write,
3341                            void __user *buffer,
3342                            size_t *length, loff_t *ppos)
3343 {
3344         proc_doulongvec_minmax(table, write, buffer, length, ppos);
3345
3346         if (write && *(unsigned long *)table->data)
3347                 scan_all_zones_unevictable_pages();
3348
3349         scan_unevictable_pages = 0;
3350         return 0;
3351 }
3352
3353 #ifdef CONFIG_NUMA
3354 /*
3355  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
3356  * a specified node's per zone unevictable lists for evictable pages.
3357  */
3358
3359 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
3360                                           struct sysdev_attribute *attr,
3361                                           char *buf)
3362 {
3363         return sprintf(buf, "0\n");     /* always zero; should fit... */
3364 }
3365
3366 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
3367                                            struct sysdev_attribute *attr,
3368                                         const char *buf, size_t count)
3369 {
3370         struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
3371         struct zone *zone;
3372         unsigned long res;
3373         unsigned long req = strict_strtoul(buf, 10, &res);
3374
3375         if (!req)
3376                 return 1;       /* zero is no-op */
3377
3378         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
3379                 if (!populated_zone(zone))
3380                         continue;
3381                 scan_zone_unevictable_pages(zone);
3382         }
3383         return 1;
3384 }
3385
3386
3387 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3388                         read_scan_unevictable_node,
3389                         write_scan_unevictable_node);
3390
3391 int scan_unevictable_register_node(struct node *node)
3392 {
3393         return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
3394 }
3395
3396 void scan_unevictable_unregister_node(struct node *node)
3397 {
3398         sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
3399 }
3400 #endif