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