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