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