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