PM/Hibernate: Fix memory shrinking
[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/slab.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/notifier.h>
36 #include <linux/rwsem.h>
37 #include <linux/delay.h>
38 #include <linux/kthread.h>
39 #include <linux/freezer.h>
40 #include <linux/memcontrol.h>
41 #include <linux/delayacct.h>
42 #include <linux/sysctl.h>
43
44 #include <asm/tlbflush.h>
45 #include <asm/div64.h>
46
47 #include <linux/swapops.h>
48
49 #include "internal.h"
50
51 struct scan_control {
52         /* Incremented by the number of inactive pages that were scanned */
53         unsigned long nr_scanned;
54
55         /* Number of pages freed so far during a call to shrink_zones() */
56         unsigned long nr_reclaimed;
57
58         /* This context's GFP mask */
59         gfp_t gfp_mask;
60
61         int may_writepage;
62
63         /* Can mapped pages be reclaimed? */
64         int may_unmap;
65
66         /* This context's SWAP_CLUSTER_MAX. If freeing memory for
67          * suspend, we effectively ignore SWAP_CLUSTER_MAX.
68          * In this context, it doesn't matter that we scan the
69          * whole list at once. */
70         int swap_cluster_max;
71
72         int swappiness;
73
74         int all_unreclaimable;
75
76         int order;
77
78         /* Which cgroup do we reclaim from */
79         struct mem_cgroup *mem_cgroup;
80
81         /*
82          * Nodemask of nodes allowed by the caller. If NULL, all nodes
83          * are scanned.
84          */
85         nodemask_t      *nodemask;
86
87         /* Pluggable isolate pages callback */
88         unsigned long (*isolate_pages)(unsigned long nr, struct list_head *dst,
89                         unsigned long *scanned, int order, int mode,
90                         struct zone *z, struct mem_cgroup *mem_cont,
91                         int active, int file);
92 };
93
94 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
95
96 #ifdef ARCH_HAS_PREFETCH
97 #define prefetch_prev_lru_page(_page, _base, _field)                    \
98         do {                                                            \
99                 if ((_page)->lru.prev != _base) {                       \
100                         struct page *prev;                              \
101                                                                         \
102                         prev = lru_to_page(&(_page->lru));              \
103                         prefetch(&prev->_field);                        \
104                 }                                                       \
105         } while (0)
106 #else
107 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
108 #endif
109
110 #ifdef ARCH_HAS_PREFETCHW
111 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
112         do {                                                            \
113                 if ((_page)->lru.prev != _base) {                       \
114                         struct page *prev;                              \
115                                                                         \
116                         prev = lru_to_page(&(_page->lru));              \
117                         prefetchw(&prev->_field);                       \
118                 }                                                       \
119         } while (0)
120 #else
121 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
122 #endif
123
124 /*
125  * From 0 .. 100.  Higher means more swappy.
126  */
127 int vm_swappiness = 60;
128 long vm_total_pages;    /* The total number of pages which the VM controls */
129
130 static LIST_HEAD(shrinker_list);
131 static DECLARE_RWSEM(shrinker_rwsem);
132
133 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
134 #define scanning_global_lru(sc) (!(sc)->mem_cgroup)
135 #else
136 #define scanning_global_lru(sc) (1)
137 #endif
138
139 static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
140                                                   struct scan_control *sc)
141 {
142         if (!scanning_global_lru(sc))
143                 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
144
145         return &zone->reclaim_stat;
146 }
147
148 static unsigned long zone_nr_pages(struct zone *zone, struct scan_control *sc,
149                                    enum lru_list lru)
150 {
151         if (!scanning_global_lru(sc))
152                 return mem_cgroup_zone_nr_pages(sc->mem_cgroup, zone, lru);
153
154         return zone_page_state(zone, NR_LRU_BASE + lru);
155 }
156
157
158 /*
159  * Add a shrinker callback to be called from the vm
160  */
161 void register_shrinker(struct shrinker *shrinker)
162 {
163         shrinker->nr = 0;
164         down_write(&shrinker_rwsem);
165         list_add_tail(&shrinker->list, &shrinker_list);
166         up_write(&shrinker_rwsem);
167 }
168 EXPORT_SYMBOL(register_shrinker);
169
170 /*
171  * Remove one
172  */
173 void unregister_shrinker(struct shrinker *shrinker)
174 {
175         down_write(&shrinker_rwsem);
176         list_del(&shrinker->list);
177         up_write(&shrinker_rwsem);
178 }
179 EXPORT_SYMBOL(unregister_shrinker);
180
181 #define SHRINK_BATCH 128
182 /*
183  * Call the shrink functions to age shrinkable caches
184  *
185  * Here we assume it costs one seek to replace a lru page and that it also
186  * takes a seek to recreate a cache object.  With this in mind we age equal
187  * percentages of the lru and ageable caches.  This should balance the seeks
188  * generated by these structures.
189  *
190  * If the vm encountered mapped pages on the LRU it increase the pressure on
191  * slab to avoid swapping.
192  *
193  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
194  *
195  * `lru_pages' represents the number of on-LRU pages in all the zones which
196  * are eligible for the caller's allocation attempt.  It is used for balancing
197  * slab reclaim versus page reclaim.
198  *
199  * Returns the number of slab objects which we shrunk.
200  */
201 unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
202                         unsigned long lru_pages)
203 {
204         struct shrinker *shrinker;
205         unsigned long ret = 0;
206
207         if (scanned == 0)
208                 scanned = SWAP_CLUSTER_MAX;
209
210         if (!down_read_trylock(&shrinker_rwsem))
211                 return 1;       /* Assume we'll be able to shrink next time */
212
213         list_for_each_entry(shrinker, &shrinker_list, list) {
214                 unsigned long long delta;
215                 unsigned long total_scan;
216                 unsigned long max_pass = (*shrinker->shrink)(0, gfp_mask);
217
218                 delta = (4 * scanned) / shrinker->seeks;
219                 delta *= max_pass;
220                 do_div(delta, lru_pages + 1);
221                 shrinker->nr += delta;
222                 if (shrinker->nr < 0) {
223                         printk(KERN_ERR "shrink_slab: %pF negative objects to "
224                                "delete nr=%ld\n",
225                                shrinker->shrink, shrinker->nr);
226                         shrinker->nr = max_pass;
227                 }
228
229                 /*
230                  * Avoid risking looping forever due to too large nr value:
231                  * never try to free more than twice the estimate number of
232                  * freeable entries.
233                  */
234                 if (shrinker->nr > max_pass * 2)
235                         shrinker->nr = max_pass * 2;
236
237                 total_scan = shrinker->nr;
238                 shrinker->nr = 0;
239
240                 while (total_scan >= SHRINK_BATCH) {
241                         long this_scan = SHRINK_BATCH;
242                         int shrink_ret;
243                         int nr_before;
244
245                         nr_before = (*shrinker->shrink)(0, gfp_mask);
246                         shrink_ret = (*shrinker->shrink)(this_scan, gfp_mask);
247                         if (shrink_ret == -1)
248                                 break;
249                         if (shrink_ret < nr_before)
250                                 ret += nr_before - shrink_ret;
251                         count_vm_events(SLABS_SCANNED, this_scan);
252                         total_scan -= this_scan;
253
254                         cond_resched();
255                 }
256
257                 shrinker->nr += total_scan;
258         }
259         up_read(&shrinker_rwsem);
260         return ret;
261 }
262
263 /* Called without lock on whether page is mapped, so answer is unstable */
264 static inline int page_mapping_inuse(struct page *page)
265 {
266         struct address_space *mapping;
267
268         /* Page is in somebody's page tables. */
269         if (page_mapped(page))
270                 return 1;
271
272         /* Be more reluctant to reclaim swapcache than pagecache */
273         if (PageSwapCache(page))
274                 return 1;
275
276         mapping = page_mapping(page);
277         if (!mapping)
278                 return 0;
279
280         /* File is mmap'd by somebody? */
281         return mapping_mapped(mapping);
282 }
283
284 static inline int is_page_cache_freeable(struct page *page)
285 {
286         return page_count(page) - !!page_has_private(page) == 2;
287 }
288
289 static int may_write_to_queue(struct backing_dev_info *bdi)
290 {
291         if (current->flags & PF_SWAPWRITE)
292                 return 1;
293         if (!bdi_write_congested(bdi))
294                 return 1;
295         if (bdi == current->backing_dev_info)
296                 return 1;
297         return 0;
298 }
299
300 /*
301  * We detected a synchronous write error writing a page out.  Probably
302  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
303  * fsync(), msync() or close().
304  *
305  * The tricky part is that after writepage we cannot touch the mapping: nothing
306  * prevents it from being freed up.  But we have a ref on the page and once
307  * that page is locked, the mapping is pinned.
308  *
309  * We're allowed to run sleeping lock_page() here because we know the caller has
310  * __GFP_FS.
311  */
312 static void handle_write_error(struct address_space *mapping,
313                                 struct page *page, int error)
314 {
315         lock_page(page);
316         if (page_mapping(page) == mapping)
317                 mapping_set_error(mapping, error);
318         unlock_page(page);
319 }
320
321 /* Request for sync pageout. */
322 enum pageout_io {
323         PAGEOUT_IO_ASYNC,
324         PAGEOUT_IO_SYNC,
325 };
326
327 /* possible outcome of pageout() */
328 typedef enum {
329         /* failed to write page out, page is locked */
330         PAGE_KEEP,
331         /* move page to the active list, page is locked */
332         PAGE_ACTIVATE,
333         /* page has been sent to the disk successfully, page is unlocked */
334         PAGE_SUCCESS,
335         /* page is clean and locked */
336         PAGE_CLEAN,
337 } pageout_t;
338
339 /*
340  * pageout is called by shrink_page_list() for each dirty page.
341  * Calls ->writepage().
342  */
343 static pageout_t pageout(struct page *page, struct address_space *mapping,
344                                                 enum pageout_io sync_writeback)
345 {
346         /*
347          * If the page is dirty, only perform writeback if that write
348          * will be non-blocking.  To prevent this allocation from being
349          * stalled by pagecache activity.  But note that there may be
350          * stalls if we need to run get_block().  We could test
351          * PagePrivate for that.
352          *
353          * If this process is currently in generic_file_write() against
354          * this page's queue, we can perform writeback even if that
355          * will block.
356          *
357          * If the page is swapcache, write it back even if that would
358          * block, for some throttling. This happens by accident, because
359          * swap_backing_dev_info is bust: it doesn't reflect the
360          * congestion state of the swapdevs.  Easy to fix, if needed.
361          * See swapfile.c:page_queue_congested().
362          */
363         if (!is_page_cache_freeable(page))
364                 return PAGE_KEEP;
365         if (!mapping) {
366                 /*
367                  * Some data journaling orphaned pages can have
368                  * page->mapping == NULL while being dirty with clean buffers.
369                  */
370                 if (page_has_private(page)) {
371                         if (try_to_free_buffers(page)) {
372                                 ClearPageDirty(page);
373                                 printk("%s: orphaned page\n", __func__);
374                                 return PAGE_CLEAN;
375                         }
376                 }
377                 return PAGE_KEEP;
378         }
379         if (mapping->a_ops->writepage == NULL)
380                 return PAGE_ACTIVATE;
381         if (!may_write_to_queue(mapping->backing_dev_info))
382                 return PAGE_KEEP;
383
384         if (clear_page_dirty_for_io(page)) {
385                 int res;
386                 struct writeback_control wbc = {
387                         .sync_mode = WB_SYNC_NONE,
388                         .nr_to_write = SWAP_CLUSTER_MAX,
389                         .range_start = 0,
390                         .range_end = LLONG_MAX,
391                         .nonblocking = 1,
392                         .for_reclaim = 1,
393                 };
394
395                 SetPageReclaim(page);
396                 res = mapping->a_ops->writepage(page, &wbc);
397                 if (res < 0)
398                         handle_write_error(mapping, page, res);
399                 if (res == AOP_WRITEPAGE_ACTIVATE) {
400                         ClearPageReclaim(page);
401                         return PAGE_ACTIVATE;
402                 }
403
404                 /*
405                  * Wait on writeback if requested to. This happens when
406                  * direct reclaiming a large contiguous area and the
407                  * first attempt to free a range of pages fails.
408                  */
409                 if (PageWriteback(page) && sync_writeback == PAGEOUT_IO_SYNC)
410                         wait_on_page_writeback(page);
411
412                 if (!PageWriteback(page)) {
413                         /* synchronous write or broken a_ops? */
414                         ClearPageReclaim(page);
415                 }
416                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
417                 return PAGE_SUCCESS;
418         }
419
420         return PAGE_CLEAN;
421 }
422
423 /*
424  * Same as remove_mapping, but if the page is removed from the mapping, it
425  * gets returned with a refcount of 0.
426  */
427 static int __remove_mapping(struct address_space *mapping, struct page *page)
428 {
429         BUG_ON(!PageLocked(page));
430         BUG_ON(mapping != page_mapping(page));
431
432         spin_lock_irq(&mapping->tree_lock);
433         /*
434          * The non racy check for a busy page.
435          *
436          * Must be careful with the order of the tests. When someone has
437          * a ref to the page, it may be possible that they dirty it then
438          * drop the reference. So if PageDirty is tested before page_count
439          * here, then the following race may occur:
440          *
441          * get_user_pages(&page);
442          * [user mapping goes away]
443          * write_to(page);
444          *                              !PageDirty(page)    [good]
445          * SetPageDirty(page);
446          * put_page(page);
447          *                              !page_count(page)   [good, discard it]
448          *
449          * [oops, our write_to data is lost]
450          *
451          * Reversing the order of the tests ensures such a situation cannot
452          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
453          * load is not satisfied before that of page->_count.
454          *
455          * Note that if SetPageDirty is always performed via set_page_dirty,
456          * and thus under tree_lock, then this ordering is not required.
457          */
458         if (!page_freeze_refs(page, 2))
459                 goto cannot_free;
460         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
461         if (unlikely(PageDirty(page))) {
462                 page_unfreeze_refs(page, 2);
463                 goto cannot_free;
464         }
465
466         if (PageSwapCache(page)) {
467                 swp_entry_t swap = { .val = page_private(page) };
468                 __delete_from_swap_cache(page);
469                 spin_unlock_irq(&mapping->tree_lock);
470                 swap_free(swap);
471         } else {
472                 __remove_from_page_cache(page);
473                 spin_unlock_irq(&mapping->tree_lock);
474         }
475
476         return 1;
477
478 cannot_free:
479         spin_unlock_irq(&mapping->tree_lock);
480         return 0;
481 }
482
483 /*
484  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
485  * someone else has a ref on the page, abort and return 0.  If it was
486  * successfully detached, return 1.  Assumes the caller has a single ref on
487  * this page.
488  */
489 int remove_mapping(struct address_space *mapping, struct page *page)
490 {
491         if (__remove_mapping(mapping, page)) {
492                 /*
493                  * Unfreezing the refcount with 1 rather than 2 effectively
494                  * drops the pagecache ref for us without requiring another
495                  * atomic operation.
496                  */
497                 page_unfreeze_refs(page, 1);
498                 return 1;
499         }
500         return 0;
501 }
502
503 /**
504  * putback_lru_page - put previously isolated page onto appropriate LRU list
505  * @page: page to be put back to appropriate lru list
506  *
507  * Add previously isolated @page to appropriate LRU list.
508  * Page may still be unevictable for other reasons.
509  *
510  * lru_lock must not be held, interrupts must be enabled.
511  */
512 #ifdef CONFIG_UNEVICTABLE_LRU
513 void putback_lru_page(struct page *page)
514 {
515         int lru;
516         int active = !!TestClearPageActive(page);
517         int was_unevictable = PageUnevictable(page);
518
519         VM_BUG_ON(PageLRU(page));
520
521 redo:
522         ClearPageUnevictable(page);
523
524         if (page_evictable(page, NULL)) {
525                 /*
526                  * For evictable pages, we can use the cache.
527                  * In event of a race, worst case is we end up with an
528                  * unevictable page on [in]active list.
529                  * We know how to handle that.
530                  */
531                 lru = active + page_is_file_cache(page);
532                 lru_cache_add_lru(page, lru);
533         } else {
534                 /*
535                  * Put unevictable pages directly on zone's unevictable
536                  * list.
537                  */
538                 lru = LRU_UNEVICTABLE;
539                 add_page_to_unevictable_list(page);
540         }
541
542         /*
543          * page's status can change while we move it among lru. If an evictable
544          * page is on unevictable list, it never be freed. To avoid that,
545          * check after we added it to the list, again.
546          */
547         if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
548                 if (!isolate_lru_page(page)) {
549                         put_page(page);
550                         goto redo;
551                 }
552                 /* This means someone else dropped this page from LRU
553                  * So, it will be freed or putback to LRU again. There is
554                  * nothing to do here.
555                  */
556         }
557
558         if (was_unevictable && lru != LRU_UNEVICTABLE)
559                 count_vm_event(UNEVICTABLE_PGRESCUED);
560         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
561                 count_vm_event(UNEVICTABLE_PGCULLED);
562
563         put_page(page);         /* drop ref from isolate */
564 }
565
566 #else /* CONFIG_UNEVICTABLE_LRU */
567
568 void putback_lru_page(struct page *page)
569 {
570         int lru;
571         VM_BUG_ON(PageLRU(page));
572
573         lru = !!TestClearPageActive(page) + page_is_file_cache(page);
574         lru_cache_add_lru(page, lru);
575         put_page(page);
576 }
577 #endif /* CONFIG_UNEVICTABLE_LRU */
578
579
580 /*
581  * shrink_page_list() returns the number of reclaimed pages
582  */
583 static unsigned long shrink_page_list(struct list_head *page_list,
584                                         struct scan_control *sc,
585                                         enum pageout_io sync_writeback)
586 {
587         LIST_HEAD(ret_pages);
588         struct pagevec freed_pvec;
589         int pgactivate = 0;
590         unsigned long nr_reclaimed = 0;
591
592         cond_resched();
593
594         pagevec_init(&freed_pvec, 1);
595         while (!list_empty(page_list)) {
596                 struct address_space *mapping;
597                 struct page *page;
598                 int may_enter_fs;
599                 int referenced;
600
601                 cond_resched();
602
603                 page = lru_to_page(page_list);
604                 list_del(&page->lru);
605
606                 if (!trylock_page(page))
607                         goto keep;
608
609                 VM_BUG_ON(PageActive(page));
610
611                 sc->nr_scanned++;
612
613                 if (unlikely(!page_evictable(page, NULL)))
614                         goto cull_mlocked;
615
616                 if (!sc->may_unmap && page_mapped(page))
617                         goto keep_locked;
618
619                 /* Double the slab pressure for mapped and swapcache pages */
620                 if (page_mapped(page) || PageSwapCache(page))
621                         sc->nr_scanned++;
622
623                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
624                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
625
626                 if (PageWriteback(page)) {
627                         /*
628                          * Synchronous reclaim is performed in two passes,
629                          * first an asynchronous pass over the list to
630                          * start parallel writeback, and a second synchronous
631                          * pass to wait for the IO to complete.  Wait here
632                          * for any page for which writeback has already
633                          * started.
634                          */
635                         if (sync_writeback == PAGEOUT_IO_SYNC && may_enter_fs)
636                                 wait_on_page_writeback(page);
637                         else
638                                 goto keep_locked;
639                 }
640
641                 referenced = page_referenced(page, 1, sc->mem_cgroup);
642                 /* In active use or really unfreeable?  Activate it. */
643                 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER &&
644                                         referenced && page_mapping_inuse(page))
645                         goto activate_locked;
646
647                 /*
648                  * Anonymous process memory has backing store?
649                  * Try to allocate it some swap space here.
650                  */
651                 if (PageAnon(page) && !PageSwapCache(page)) {
652                         if (!(sc->gfp_mask & __GFP_IO))
653                                 goto keep_locked;
654                         if (!add_to_swap(page))
655                                 goto activate_locked;
656                         may_enter_fs = 1;
657                 }
658
659                 mapping = page_mapping(page);
660
661                 /*
662                  * The page is mapped into the page tables of one or more
663                  * processes. Try to unmap it here.
664                  */
665                 if (page_mapped(page) && mapping) {
666                         switch (try_to_unmap(page, 0)) {
667                         case SWAP_FAIL:
668                                 goto activate_locked;
669                         case SWAP_AGAIN:
670                                 goto keep_locked;
671                         case SWAP_MLOCK:
672                                 goto cull_mlocked;
673                         case SWAP_SUCCESS:
674                                 ; /* try to free the page below */
675                         }
676                 }
677
678                 if (PageDirty(page)) {
679                         if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && referenced)
680                                 goto keep_locked;
681                         if (!may_enter_fs)
682                                 goto keep_locked;
683                         if (!sc->may_writepage)
684                                 goto keep_locked;
685
686                         /* Page is dirty, try to write it out here */
687                         switch (pageout(page, mapping, sync_writeback)) {
688                         case PAGE_KEEP:
689                                 goto keep_locked;
690                         case PAGE_ACTIVATE:
691                                 goto activate_locked;
692                         case PAGE_SUCCESS:
693                                 if (PageWriteback(page) || PageDirty(page))
694                                         goto keep;
695                                 /*
696                                  * A synchronous write - probably a ramdisk.  Go
697                                  * ahead and try to reclaim the page.
698                                  */
699                                 if (!trylock_page(page))
700                                         goto keep;
701                                 if (PageDirty(page) || PageWriteback(page))
702                                         goto keep_locked;
703                                 mapping = page_mapping(page);
704                         case PAGE_CLEAN:
705                                 ; /* try to free the page below */
706                         }
707                 }
708
709                 /*
710                  * If the page has buffers, try to free the buffer mappings
711                  * associated with this page. If we succeed we try to free
712                  * the page as well.
713                  *
714                  * We do this even if the page is PageDirty().
715                  * try_to_release_page() does not perform I/O, but it is
716                  * possible for a page to have PageDirty set, but it is actually
717                  * clean (all its buffers are clean).  This happens if the
718                  * buffers were written out directly, with submit_bh(). ext3
719                  * will do this, as well as the blockdev mapping.
720                  * try_to_release_page() will discover that cleanness and will
721                  * drop the buffers and mark the page clean - it can be freed.
722                  *
723                  * Rarely, pages can have buffers and no ->mapping.  These are
724                  * the pages which were not successfully invalidated in
725                  * truncate_complete_page().  We try to drop those buffers here
726                  * and if that worked, and the page is no longer mapped into
727                  * process address space (page_count == 1) it can be freed.
728                  * Otherwise, leave the page on the LRU so it is swappable.
729                  */
730                 if (page_has_private(page)) {
731                         if (!try_to_release_page(page, sc->gfp_mask))
732                                 goto activate_locked;
733                         if (!mapping && page_count(page) == 1) {
734                                 unlock_page(page);
735                                 if (put_page_testzero(page))
736                                         goto free_it;
737                                 else {
738                                         /*
739                                          * rare race with speculative reference.
740                                          * the speculative reference will free
741                                          * this page shortly, so we may
742                                          * increment nr_reclaimed here (and
743                                          * leave it off the LRU).
744                                          */
745                                         nr_reclaimed++;
746                                         continue;
747                                 }
748                         }
749                 }
750
751                 if (!mapping || !__remove_mapping(mapping, page))
752                         goto keep_locked;
753
754                 /*
755                  * At this point, we have no other references and there is
756                  * no way to pick any more up (removed from LRU, removed
757                  * from pagecache). Can use non-atomic bitops now (and
758                  * we obviously don't have to worry about waking up a process
759                  * waiting on the page lock, because there are no references.
760                  */
761                 __clear_page_locked(page);
762 free_it:
763                 nr_reclaimed++;
764                 if (!pagevec_add(&freed_pvec, page)) {
765                         __pagevec_free(&freed_pvec);
766                         pagevec_reinit(&freed_pvec);
767                 }
768                 continue;
769
770 cull_mlocked:
771                 if (PageSwapCache(page))
772                         try_to_free_swap(page);
773                 unlock_page(page);
774                 putback_lru_page(page);
775                 continue;
776
777 activate_locked:
778                 /* Not a candidate for swapping, so reclaim swap space. */
779                 if (PageSwapCache(page) && vm_swap_full())
780                         try_to_free_swap(page);
781                 VM_BUG_ON(PageActive(page));
782                 SetPageActive(page);
783                 pgactivate++;
784 keep_locked:
785                 unlock_page(page);
786 keep:
787                 list_add(&page->lru, &ret_pages);
788                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
789         }
790         list_splice(&ret_pages, page_list);
791         if (pagevec_count(&freed_pvec))
792                 __pagevec_free(&freed_pvec);
793         count_vm_events(PGACTIVATE, pgactivate);
794         return nr_reclaimed;
795 }
796
797 /* LRU Isolation modes. */
798 #define ISOLATE_INACTIVE 0      /* Isolate inactive pages. */
799 #define ISOLATE_ACTIVE 1        /* Isolate active pages. */
800 #define ISOLATE_BOTH 2          /* Isolate both active and inactive pages. */
801
802 /*
803  * Attempt to remove the specified page from its LRU.  Only take this page
804  * if it is of the appropriate PageActive status.  Pages which are being
805  * freed elsewhere are also ignored.
806  *
807  * page:        page to consider
808  * mode:        one of the LRU isolation modes defined above
809  *
810  * returns 0 on success, -ve errno on failure.
811  */
812 int __isolate_lru_page(struct page *page, int mode, int file)
813 {
814         int ret = -EINVAL;
815
816         /* Only take pages on the LRU. */
817         if (!PageLRU(page))
818                 return ret;
819
820         /*
821          * When checking the active state, we need to be sure we are
822          * dealing with comparible boolean values.  Take the logical not
823          * of each.
824          */
825         if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
826                 return ret;
827
828         if (mode != ISOLATE_BOTH && (!page_is_file_cache(page) != !file))
829                 return ret;
830
831         /*
832          * When this function is being called for lumpy reclaim, we
833          * initially look into all LRU pages, active, inactive and
834          * unevictable; only give shrink_page_list evictable pages.
835          */
836         if (PageUnevictable(page))
837                 return ret;
838
839         ret = -EBUSY;
840
841         if (likely(get_page_unless_zero(page))) {
842                 /*
843                  * Be careful not to clear PageLRU until after we're
844                  * sure the page is not being freed elsewhere -- the
845                  * page release code relies on it.
846                  */
847                 ClearPageLRU(page);
848                 ret = 0;
849                 mem_cgroup_del_lru(page);
850         }
851
852         return ret;
853 }
854
855 /*
856  * zone->lru_lock is heavily contended.  Some of the functions that
857  * shrink the lists perform better by taking out a batch of pages
858  * and working on them outside the LRU lock.
859  *
860  * For pagecache intensive workloads, this function is the hottest
861  * spot in the kernel (apart from copy_*_user functions).
862  *
863  * Appropriate locks must be held before calling this function.
864  *
865  * @nr_to_scan: The number of pages to look through on the list.
866  * @src:        The LRU list to pull pages off.
867  * @dst:        The temp list to put pages on to.
868  * @scanned:    The number of pages that were scanned.
869  * @order:      The caller's attempted allocation order
870  * @mode:       One of the LRU isolation modes
871  * @file:       True [1] if isolating file [!anon] pages
872  *
873  * returns how many pages were moved onto *@dst.
874  */
875 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
876                 struct list_head *src, struct list_head *dst,
877                 unsigned long *scanned, int order, int mode, int file)
878 {
879         unsigned long nr_taken = 0;
880         unsigned long scan;
881
882         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
883                 struct page *page;
884                 unsigned long pfn;
885                 unsigned long end_pfn;
886                 unsigned long page_pfn;
887                 int zone_id;
888
889                 page = lru_to_page(src);
890                 prefetchw_prev_lru_page(page, src, flags);
891
892                 VM_BUG_ON(!PageLRU(page));
893
894                 switch (__isolate_lru_page(page, mode, file)) {
895                 case 0:
896                         list_move(&page->lru, dst);
897                         nr_taken++;
898                         break;
899
900                 case -EBUSY:
901                         /* else it is being freed elsewhere */
902                         list_move(&page->lru, src);
903                         continue;
904
905                 default:
906                         BUG();
907                 }
908
909                 if (!order)
910                         continue;
911
912                 /*
913                  * Attempt to take all pages in the order aligned region
914                  * surrounding the tag page.  Only take those pages of
915                  * the same active state as that tag page.  We may safely
916                  * round the target page pfn down to the requested order
917                  * as the mem_map is guarenteed valid out to MAX_ORDER,
918                  * where that page is in a different zone we will detect
919                  * it from its zone id and abort this block scan.
920                  */
921                 zone_id = page_zone_id(page);
922                 page_pfn = page_to_pfn(page);
923                 pfn = page_pfn & ~((1 << order) - 1);
924                 end_pfn = pfn + (1 << order);
925                 for (; pfn < end_pfn; pfn++) {
926                         struct page *cursor_page;
927
928                         /* The target page is in the block, ignore it. */
929                         if (unlikely(pfn == page_pfn))
930                                 continue;
931
932                         /* Avoid holes within the zone. */
933                         if (unlikely(!pfn_valid_within(pfn)))
934                                 break;
935
936                         cursor_page = pfn_to_page(pfn);
937
938                         /* Check that we have not crossed a zone boundary. */
939                         if (unlikely(page_zone_id(cursor_page) != zone_id))
940                                 continue;
941                         switch (__isolate_lru_page(cursor_page, mode, file)) {
942                         case 0:
943                                 list_move(&cursor_page->lru, dst);
944                                 nr_taken++;
945                                 scan++;
946                                 break;
947
948                         case -EBUSY:
949                                 /* else it is being freed elsewhere */
950                                 list_move(&cursor_page->lru, src);
951                         default:
952                                 break;  /* ! on LRU or wrong list */
953                         }
954                 }
955         }
956
957         *scanned = scan;
958         return nr_taken;
959 }
960
961 static unsigned long isolate_pages_global(unsigned long nr,
962                                         struct list_head *dst,
963                                         unsigned long *scanned, int order,
964                                         int mode, struct zone *z,
965                                         struct mem_cgroup *mem_cont,
966                                         int active, int file)
967 {
968         int lru = LRU_BASE;
969         if (active)
970                 lru += LRU_ACTIVE;
971         if (file)
972                 lru += LRU_FILE;
973         return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
974                                                                 mode, !!file);
975 }
976
977 /*
978  * clear_active_flags() is a helper for shrink_active_list(), clearing
979  * any active bits from the pages in the list.
980  */
981 static unsigned long clear_active_flags(struct list_head *page_list,
982                                         unsigned int *count)
983 {
984         int nr_active = 0;
985         int lru;
986         struct page *page;
987
988         list_for_each_entry(page, page_list, lru) {
989                 lru = page_is_file_cache(page);
990                 if (PageActive(page)) {
991                         lru += LRU_ACTIVE;
992                         ClearPageActive(page);
993                         nr_active++;
994                 }
995                 count[lru]++;
996         }
997
998         return nr_active;
999 }
1000
1001 /**
1002  * isolate_lru_page - tries to isolate a page from its LRU list
1003  * @page: page to isolate from its LRU list
1004  *
1005  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1006  * vmstat statistic corresponding to whatever LRU list the page was on.
1007  *
1008  * Returns 0 if the page was removed from an LRU list.
1009  * Returns -EBUSY if the page was not on an LRU list.
1010  *
1011  * The returned page will have PageLRU() cleared.  If it was found on
1012  * the active list, it will have PageActive set.  If it was found on
1013  * the unevictable list, it will have the PageUnevictable bit set. That flag
1014  * may need to be cleared by the caller before letting the page go.
1015  *
1016  * The vmstat statistic corresponding to the list on which the page was
1017  * found will be decremented.
1018  *
1019  * Restrictions:
1020  * (1) Must be called with an elevated refcount on the page. This is a
1021  *     fundamentnal difference from isolate_lru_pages (which is called
1022  *     without a stable reference).
1023  * (2) the lru_lock must not be held.
1024  * (3) interrupts must be enabled.
1025  */
1026 int isolate_lru_page(struct page *page)
1027 {
1028         int ret = -EBUSY;
1029
1030         if (PageLRU(page)) {
1031                 struct zone *zone = page_zone(page);
1032
1033                 spin_lock_irq(&zone->lru_lock);
1034                 if (PageLRU(page) && get_page_unless_zero(page)) {
1035                         int lru = page_lru(page);
1036                         ret = 0;
1037                         ClearPageLRU(page);
1038
1039                         del_page_from_lru_list(zone, page, lru);
1040                 }
1041                 spin_unlock_irq(&zone->lru_lock);
1042         }
1043         return ret;
1044 }
1045
1046 /*
1047  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1048  * of reclaimed pages
1049  */
1050 static unsigned long shrink_inactive_list(unsigned long max_scan,
1051                         struct zone *zone, struct scan_control *sc,
1052                         int priority, int file)
1053 {
1054         LIST_HEAD(page_list);
1055         struct pagevec pvec;
1056         unsigned long nr_scanned = 0;
1057         unsigned long nr_reclaimed = 0;
1058         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1059
1060         pagevec_init(&pvec, 1);
1061
1062         lru_add_drain();
1063         spin_lock_irq(&zone->lru_lock);
1064         do {
1065                 struct page *page;
1066                 unsigned long nr_taken;
1067                 unsigned long nr_scan;
1068                 unsigned long nr_freed;
1069                 unsigned long nr_active;
1070                 unsigned int count[NR_LRU_LISTS] = { 0, };
1071                 int mode = ISOLATE_INACTIVE;
1072
1073                 /*
1074                  * If we need a large contiguous chunk of memory, or have
1075                  * trouble getting a small set of contiguous pages, we
1076                  * will reclaim both active and inactive pages.
1077                  *
1078                  * We use the same threshold as pageout congestion_wait below.
1079                  */
1080                 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1081                         mode = ISOLATE_BOTH;
1082                 else if (sc->order && priority < DEF_PRIORITY - 2)
1083                         mode = ISOLATE_BOTH;
1084
1085                 nr_taken = sc->isolate_pages(sc->swap_cluster_max,
1086                              &page_list, &nr_scan, sc->order, mode,
1087                                 zone, sc->mem_cgroup, 0, file);
1088                 nr_active = clear_active_flags(&page_list, count);
1089                 __count_vm_events(PGDEACTIVATE, nr_active);
1090
1091                 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1092                                                 -count[LRU_ACTIVE_FILE]);
1093                 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1094                                                 -count[LRU_INACTIVE_FILE]);
1095                 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1096                                                 -count[LRU_ACTIVE_ANON]);
1097                 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1098                                                 -count[LRU_INACTIVE_ANON]);
1099
1100                 if (scanning_global_lru(sc))
1101                         zone->pages_scanned += nr_scan;
1102
1103                 reclaim_stat->recent_scanned[0] += count[LRU_INACTIVE_ANON];
1104                 reclaim_stat->recent_scanned[0] += count[LRU_ACTIVE_ANON];
1105                 reclaim_stat->recent_scanned[1] += count[LRU_INACTIVE_FILE];
1106                 reclaim_stat->recent_scanned[1] += count[LRU_ACTIVE_FILE];
1107
1108                 spin_unlock_irq(&zone->lru_lock);
1109
1110                 nr_scanned += nr_scan;
1111                 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
1112
1113                 /*
1114                  * If we are direct reclaiming for contiguous pages and we do
1115                  * not reclaim everything in the list, try again and wait
1116                  * for IO to complete. This will stall high-order allocations
1117                  * but that should be acceptable to the caller
1118                  */
1119                 if (nr_freed < nr_taken && !current_is_kswapd() &&
1120                                         sc->order > PAGE_ALLOC_COSTLY_ORDER) {
1121                         congestion_wait(WRITE, HZ/10);
1122
1123                         /*
1124                          * The attempt at page out may have made some
1125                          * of the pages active, mark them inactive again.
1126                          */
1127                         nr_active = clear_active_flags(&page_list, count);
1128                         count_vm_events(PGDEACTIVATE, nr_active);
1129
1130                         nr_freed += shrink_page_list(&page_list, sc,
1131                                                         PAGEOUT_IO_SYNC);
1132                 }
1133
1134                 nr_reclaimed += nr_freed;
1135                 local_irq_disable();
1136                 if (current_is_kswapd()) {
1137                         __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scan);
1138                         __count_vm_events(KSWAPD_STEAL, nr_freed);
1139                 } else if (scanning_global_lru(sc))
1140                         __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scan);
1141
1142                 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
1143
1144                 if (nr_taken == 0)
1145                         goto done;
1146
1147                 spin_lock(&zone->lru_lock);
1148                 /*
1149                  * Put back any unfreeable pages.
1150                  */
1151                 while (!list_empty(&page_list)) {
1152                         int lru;
1153                         page = lru_to_page(&page_list);
1154                         VM_BUG_ON(PageLRU(page));
1155                         list_del(&page->lru);
1156                         if (unlikely(!page_evictable(page, NULL))) {
1157                                 spin_unlock_irq(&zone->lru_lock);
1158                                 putback_lru_page(page);
1159                                 spin_lock_irq(&zone->lru_lock);
1160                                 continue;
1161                         }
1162                         SetPageLRU(page);
1163                         lru = page_lru(page);
1164                         add_page_to_lru_list(zone, page, lru);
1165                         if (PageActive(page)) {
1166                                 int file = !!page_is_file_cache(page);
1167                                 reclaim_stat->recent_rotated[file]++;
1168                         }
1169                         if (!pagevec_add(&pvec, page)) {
1170                                 spin_unlock_irq(&zone->lru_lock);
1171                                 __pagevec_release(&pvec);
1172                                 spin_lock_irq(&zone->lru_lock);
1173                         }
1174                 }
1175         } while (nr_scanned < max_scan);
1176         spin_unlock(&zone->lru_lock);
1177 done:
1178         local_irq_enable();
1179         pagevec_release(&pvec);
1180         return nr_reclaimed;
1181 }
1182
1183 /*
1184  * We are about to scan this zone at a certain priority level.  If that priority
1185  * level is smaller (ie: more urgent) than the previous priority, then note
1186  * that priority level within the zone.  This is done so that when the next
1187  * process comes in to scan this zone, it will immediately start out at this
1188  * priority level rather than having to build up its own scanning priority.
1189  * Here, this priority affects only the reclaim-mapped threshold.
1190  */
1191 static inline void note_zone_scanning_priority(struct zone *zone, int priority)
1192 {
1193         if (priority < zone->prev_priority)
1194                 zone->prev_priority = priority;
1195 }
1196
1197 /*
1198  * This moves pages from the active list to the inactive list.
1199  *
1200  * We move them the other way if the page is referenced by one or more
1201  * processes, from rmap.
1202  *
1203  * If the pages are mostly unmapped, the processing is fast and it is
1204  * appropriate to hold zone->lru_lock across the whole operation.  But if
1205  * the pages are mapped, the processing is slow (page_referenced()) so we
1206  * should drop zone->lru_lock around each page.  It's impossible to balance
1207  * this, so instead we remove the pages from the LRU while processing them.
1208  * It is safe to rely on PG_active against the non-LRU pages in here because
1209  * nobody will play with that bit on a non-LRU page.
1210  *
1211  * The downside is that we have to touch page->_count against each page.
1212  * But we had to alter page->flags anyway.
1213  */
1214
1215
1216 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1217                         struct scan_control *sc, int priority, int file)
1218 {
1219         unsigned long pgmoved;
1220         int pgdeactivate = 0;
1221         unsigned long pgscanned;
1222         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1223         LIST_HEAD(l_inactive);
1224         struct page *page;
1225         struct pagevec pvec;
1226         enum lru_list lru;
1227         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1228
1229         lru_add_drain();
1230         spin_lock_irq(&zone->lru_lock);
1231         pgmoved = sc->isolate_pages(nr_pages, &l_hold, &pgscanned, sc->order,
1232                                         ISOLATE_ACTIVE, zone,
1233                                         sc->mem_cgroup, 1, file);
1234         /*
1235          * zone->pages_scanned is used for detect zone's oom
1236          * mem_cgroup remembers nr_scan by itself.
1237          */
1238         if (scanning_global_lru(sc)) {
1239                 zone->pages_scanned += pgscanned;
1240         }
1241         reclaim_stat->recent_scanned[!!file] += pgmoved;
1242
1243         if (file)
1244                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -pgmoved);
1245         else
1246                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -pgmoved);
1247         spin_unlock_irq(&zone->lru_lock);
1248
1249         pgmoved = 0;
1250         while (!list_empty(&l_hold)) {
1251                 cond_resched();
1252                 page = lru_to_page(&l_hold);
1253                 list_del(&page->lru);
1254
1255                 if (unlikely(!page_evictable(page, NULL))) {
1256                         putback_lru_page(page);
1257                         continue;
1258                 }
1259
1260                 /* page_referenced clears PageReferenced */
1261                 if (page_mapping_inuse(page) &&
1262                     page_referenced(page, 0, sc->mem_cgroup))
1263                         pgmoved++;
1264
1265                 list_add(&page->lru, &l_inactive);
1266         }
1267
1268         /*
1269          * Move the pages to the [file or anon] inactive list.
1270          */
1271         pagevec_init(&pvec, 1);
1272         lru = LRU_BASE + file * LRU_FILE;
1273
1274         spin_lock_irq(&zone->lru_lock);
1275         /*
1276          * Count referenced pages from currently used mappings as
1277          * rotated, even though they are moved to the inactive list.
1278          * This helps balance scan pressure between file and anonymous
1279          * pages in get_scan_ratio.
1280          */
1281         reclaim_stat->recent_rotated[!!file] += pgmoved;
1282
1283         pgmoved = 0;
1284         while (!list_empty(&l_inactive)) {
1285                 page = lru_to_page(&l_inactive);
1286                 prefetchw_prev_lru_page(page, &l_inactive, flags);
1287                 VM_BUG_ON(PageLRU(page));
1288                 SetPageLRU(page);
1289                 VM_BUG_ON(!PageActive(page));
1290                 ClearPageActive(page);
1291
1292                 list_move(&page->lru, &zone->lru[lru].list);
1293                 mem_cgroup_add_lru_list(page, lru);
1294                 pgmoved++;
1295                 if (!pagevec_add(&pvec, page)) {
1296                         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1297                         spin_unlock_irq(&zone->lru_lock);
1298                         pgdeactivate += pgmoved;
1299                         pgmoved = 0;
1300                         if (buffer_heads_over_limit)
1301                                 pagevec_strip(&pvec);
1302                         __pagevec_release(&pvec);
1303                         spin_lock_irq(&zone->lru_lock);
1304                 }
1305         }
1306         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1307         pgdeactivate += pgmoved;
1308         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1309         __count_vm_events(PGDEACTIVATE, pgdeactivate);
1310         spin_unlock_irq(&zone->lru_lock);
1311         if (buffer_heads_over_limit)
1312                 pagevec_strip(&pvec);
1313         pagevec_release(&pvec);
1314 }
1315
1316 static int inactive_anon_is_low_global(struct zone *zone)
1317 {
1318         unsigned long active, inactive;
1319
1320         active = zone_page_state(zone, NR_ACTIVE_ANON);
1321         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1322
1323         if (inactive * zone->inactive_ratio < active)
1324                 return 1;
1325
1326         return 0;
1327 }
1328
1329 /**
1330  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1331  * @zone: zone to check
1332  * @sc:   scan control of this context
1333  *
1334  * Returns true if the zone does not have enough inactive anon pages,
1335  * meaning some active anon pages need to be deactivated.
1336  */
1337 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1338 {
1339         int low;
1340
1341         if (scanning_global_lru(sc))
1342                 low = inactive_anon_is_low_global(zone);
1343         else
1344                 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1345         return low;
1346 }
1347
1348 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1349         struct zone *zone, struct scan_control *sc, int priority)
1350 {
1351         int file = is_file_lru(lru);
1352
1353         if (lru == LRU_ACTIVE_FILE) {
1354                 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1355                 return 0;
1356         }
1357
1358         if (lru == LRU_ACTIVE_ANON && inactive_anon_is_low(zone, sc)) {
1359                 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1360                 return 0;
1361         }
1362         return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1363 }
1364
1365 /*
1366  * Determine how aggressively the anon and file LRU lists should be
1367  * scanned.  The relative value of each set of LRU lists is determined
1368  * by looking at the fraction of the pages scanned we did rotate back
1369  * onto the active list instead of evict.
1370  *
1371  * percent[0] specifies how much pressure to put on ram/swap backed
1372  * memory, while percent[1] determines pressure on the file LRUs.
1373  */
1374 static void get_scan_ratio(struct zone *zone, struct scan_control *sc,
1375                                         unsigned long *percent)
1376 {
1377         unsigned long anon, file, free;
1378         unsigned long anon_prio, file_prio;
1379         unsigned long ap, fp;
1380         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1381
1382         /* If we have no swap space, do not bother scanning anon pages. */
1383         if (nr_swap_pages <= 0) {
1384                 percent[0] = 0;
1385                 percent[1] = 100;
1386                 return;
1387         }
1388
1389         anon  = zone_nr_pages(zone, sc, LRU_ACTIVE_ANON) +
1390                 zone_nr_pages(zone, sc, LRU_INACTIVE_ANON);
1391         file  = zone_nr_pages(zone, sc, LRU_ACTIVE_FILE) +
1392                 zone_nr_pages(zone, sc, LRU_INACTIVE_FILE);
1393
1394         if (scanning_global_lru(sc)) {
1395                 free  = zone_page_state(zone, NR_FREE_PAGES);
1396                 /* If we have very few page cache pages,
1397                    force-scan anon pages. */
1398                 if (unlikely(file + free <= zone->pages_high)) {
1399                         percent[0] = 100;
1400                         percent[1] = 0;
1401                         return;
1402                 }
1403         }
1404
1405         /*
1406          * OK, so we have swap space and a fair amount of page cache
1407          * pages.  We use the recently rotated / recently scanned
1408          * ratios to determine how valuable each cache is.
1409          *
1410          * Because workloads change over time (and to avoid overflow)
1411          * we keep these statistics as a floating average, which ends
1412          * up weighing recent references more than old ones.
1413          *
1414          * anon in [0], file in [1]
1415          */
1416         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1417                 spin_lock_irq(&zone->lru_lock);
1418                 reclaim_stat->recent_scanned[0] /= 2;
1419                 reclaim_stat->recent_rotated[0] /= 2;
1420                 spin_unlock_irq(&zone->lru_lock);
1421         }
1422
1423         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1424                 spin_lock_irq(&zone->lru_lock);
1425                 reclaim_stat->recent_scanned[1] /= 2;
1426                 reclaim_stat->recent_rotated[1] /= 2;
1427                 spin_unlock_irq(&zone->lru_lock);
1428         }
1429
1430         /*
1431          * With swappiness at 100, anonymous and file have the same priority.
1432          * This scanning priority is essentially the inverse of IO cost.
1433          */
1434         anon_prio = sc->swappiness;
1435         file_prio = 200 - sc->swappiness;
1436
1437         /*
1438          * The amount of pressure on anon vs file pages is inversely
1439          * proportional to the fraction of recently scanned pages on
1440          * each list that were recently referenced and in active use.
1441          */
1442         ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1443         ap /= reclaim_stat->recent_rotated[0] + 1;
1444
1445         fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1446         fp /= reclaim_stat->recent_rotated[1] + 1;
1447
1448         /* Normalize to percentages */
1449         percent[0] = 100 * ap / (ap + fp + 1);
1450         percent[1] = 100 - percent[0];
1451 }
1452
1453
1454 /*
1455  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1456  */
1457 static void shrink_zone(int priority, struct zone *zone,
1458                                 struct scan_control *sc)
1459 {
1460         unsigned long nr[NR_LRU_LISTS];
1461         unsigned long nr_to_scan;
1462         unsigned long percent[2];       /* anon @ 0; file @ 1 */
1463         enum lru_list l;
1464         unsigned long nr_reclaimed = sc->nr_reclaimed;
1465         unsigned long swap_cluster_max = sc->swap_cluster_max;
1466
1467         get_scan_ratio(zone, sc, percent);
1468
1469         for_each_evictable_lru(l) {
1470                 int file = is_file_lru(l);
1471                 int scan;
1472
1473                 scan = zone_nr_pages(zone, sc, l);
1474                 if (priority) {
1475                         scan >>= priority;
1476                         scan = (scan * percent[file]) / 100;
1477                 }
1478                 if (scanning_global_lru(sc)) {
1479                         zone->lru[l].nr_scan += scan;
1480                         nr[l] = zone->lru[l].nr_scan;
1481                         if (nr[l] >= swap_cluster_max)
1482                                 zone->lru[l].nr_scan = 0;
1483                         else
1484                                 nr[l] = 0;
1485                 } else
1486                         nr[l] = scan;
1487         }
1488
1489         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1490                                         nr[LRU_INACTIVE_FILE]) {
1491                 for_each_evictable_lru(l) {
1492                         if (nr[l]) {
1493                                 nr_to_scan = min(nr[l], swap_cluster_max);
1494                                 nr[l] -= nr_to_scan;
1495
1496                                 nr_reclaimed += shrink_list(l, nr_to_scan,
1497                                                             zone, sc, priority);
1498                         }
1499                 }
1500                 /*
1501                  * On large memory systems, scan >> priority can become
1502                  * really large. This is fine for the starting priority;
1503                  * we want to put equal scanning pressure on each zone.
1504                  * However, if the VM has a harder time of freeing pages,
1505                  * with multiple processes reclaiming pages, the total
1506                  * freeing target can get unreasonably large.
1507                  */
1508                 if (nr_reclaimed > swap_cluster_max &&
1509                         priority < DEF_PRIORITY && !current_is_kswapd())
1510                         break;
1511         }
1512
1513         sc->nr_reclaimed = nr_reclaimed;
1514
1515         /*
1516          * Even if we did not try to evict anon pages at all, we want to
1517          * rebalance the anon lru active/inactive ratio.
1518          */
1519         if (inactive_anon_is_low(zone, sc))
1520                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1521
1522         throttle_vm_writeout(sc->gfp_mask);
1523 }
1524
1525 /*
1526  * This is the direct reclaim path, for page-allocating processes.  We only
1527  * try to reclaim pages from zones which will satisfy the caller's allocation
1528  * request.
1529  *
1530  * We reclaim from a zone even if that zone is over pages_high.  Because:
1531  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1532  *    allocation or
1533  * b) The zones may be over pages_high but they must go *over* pages_high to
1534  *    satisfy the `incremental min' zone defense algorithm.
1535  *
1536  * If a zone is deemed to be full of pinned pages then just give it a light
1537  * scan then give up on it.
1538  */
1539 static void shrink_zones(int priority, struct zonelist *zonelist,
1540                                         struct scan_control *sc)
1541 {
1542         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1543         struct zoneref *z;
1544         struct zone *zone;
1545
1546         sc->all_unreclaimable = 1;
1547         for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1548                                         sc->nodemask) {
1549                 if (!populated_zone(zone))
1550                         continue;
1551                 /*
1552                  * Take care memory controller reclaiming has small influence
1553                  * to global LRU.
1554                  */
1555                 if (scanning_global_lru(sc)) {
1556                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1557                                 continue;
1558                         note_zone_scanning_priority(zone, priority);
1559
1560                         if (zone_is_all_unreclaimable(zone) &&
1561                                                 priority != DEF_PRIORITY)
1562                                 continue;       /* Let kswapd poll it */
1563                         sc->all_unreclaimable = 0;
1564                 } else {
1565                         /*
1566                          * Ignore cpuset limitation here. We just want to reduce
1567                          * # of used pages by us regardless of memory shortage.
1568                          */
1569                         sc->all_unreclaimable = 0;
1570                         mem_cgroup_note_reclaim_priority(sc->mem_cgroup,
1571                                                         priority);
1572                 }
1573
1574                 shrink_zone(priority, zone, sc);
1575         }
1576 }
1577
1578 /*
1579  * This is the main entry point to direct page reclaim.
1580  *
1581  * If a full scan of the inactive list fails to free enough memory then we
1582  * are "out of memory" and something needs to be killed.
1583  *
1584  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1585  * high - the zone may be full of dirty or under-writeback pages, which this
1586  * caller can't do much about.  We kick pdflush and take explicit naps in the
1587  * hope that some of these pages can be written.  But if the allocating task
1588  * holds filesystem locks which prevent writeout this might not work, and the
1589  * allocation attempt will fail.
1590  *
1591  * returns:     0, if no pages reclaimed
1592  *              else, the number of pages reclaimed
1593  */
1594 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
1595                                         struct scan_control *sc)
1596 {
1597         int priority;
1598         unsigned long ret = 0;
1599         unsigned long total_scanned = 0;
1600         struct reclaim_state *reclaim_state = current->reclaim_state;
1601         unsigned long lru_pages = 0;
1602         struct zoneref *z;
1603         struct zone *zone;
1604         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1605
1606         delayacct_freepages_start();
1607
1608         if (scanning_global_lru(sc))
1609                 count_vm_event(ALLOCSTALL);
1610         /*
1611          * mem_cgroup will not do shrink_slab.
1612          */
1613         if (scanning_global_lru(sc)) {
1614                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1615
1616                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1617                                 continue;
1618
1619                         lru_pages += zone_lru_pages(zone);
1620                 }
1621         }
1622
1623         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1624                 sc->nr_scanned = 0;
1625                 if (!priority)
1626                         disable_swap_token();
1627                 shrink_zones(priority, zonelist, sc);
1628                 /*
1629                  * Don't shrink slabs when reclaiming memory from
1630                  * over limit cgroups
1631                  */
1632                 if (scanning_global_lru(sc)) {
1633                         shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
1634                         if (reclaim_state) {
1635                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
1636                                 reclaim_state->reclaimed_slab = 0;
1637                         }
1638                 }
1639                 total_scanned += sc->nr_scanned;
1640                 if (sc->nr_reclaimed >= sc->swap_cluster_max) {
1641                         ret = sc->nr_reclaimed;
1642                         goto out;
1643                 }
1644
1645                 /*
1646                  * Try to write back as many pages as we just scanned.  This
1647                  * tends to cause slow streaming writers to write data to the
1648                  * disk smoothly, at the dirtying rate, which is nice.   But
1649                  * that's undesirable in laptop mode, where we *want* lumpy
1650                  * writeout.  So in laptop mode, write out the whole world.
1651                  */
1652                 if (total_scanned > sc->swap_cluster_max +
1653                                         sc->swap_cluster_max / 2) {
1654                         wakeup_pdflush(laptop_mode ? 0 : total_scanned);
1655                         sc->may_writepage = 1;
1656                 }
1657
1658                 /* Take a nap, wait for some writeback to complete */
1659                 if (sc->nr_scanned && priority < DEF_PRIORITY - 2)
1660                         congestion_wait(WRITE, HZ/10);
1661         }
1662         /* top priority shrink_zones still had more to do? don't OOM, then */
1663         if (!sc->all_unreclaimable && scanning_global_lru(sc))
1664                 ret = sc->nr_reclaimed;
1665 out:
1666         /*
1667          * Now that we've scanned all the zones at this priority level, note
1668          * that level within the zone so that the next thread which performs
1669          * scanning of this zone will immediately start out at this priority
1670          * level.  This affects only the decision whether or not to bring
1671          * mapped pages onto the inactive list.
1672          */
1673         if (priority < 0)
1674                 priority = 0;
1675
1676         if (scanning_global_lru(sc)) {
1677                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1678
1679                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1680                                 continue;
1681
1682                         zone->prev_priority = priority;
1683                 }
1684         } else
1685                 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority);
1686
1687         delayacct_freepages_end();
1688
1689         return ret;
1690 }
1691
1692 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
1693                                 gfp_t gfp_mask, nodemask_t *nodemask)
1694 {
1695         struct scan_control sc = {
1696                 .gfp_mask = gfp_mask,
1697                 .may_writepage = !laptop_mode,
1698                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1699                 .may_unmap = 1,
1700                 .swappiness = vm_swappiness,
1701                 .order = order,
1702                 .mem_cgroup = NULL,
1703                 .isolate_pages = isolate_pages_global,
1704                 .nodemask = nodemask,
1705         };
1706
1707         return do_try_to_free_pages(zonelist, &sc);
1708 }
1709
1710 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
1711
1712 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
1713                                            gfp_t gfp_mask,
1714                                            bool noswap,
1715                                            unsigned int swappiness)
1716 {
1717         struct scan_control sc = {
1718                 .may_writepage = !laptop_mode,
1719                 .may_unmap = 1,
1720                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1721                 .swappiness = swappiness,
1722                 .order = 0,
1723                 .mem_cgroup = mem_cont,
1724                 .isolate_pages = mem_cgroup_isolate_pages,
1725                 .nodemask = NULL, /* we don't care the placement */
1726         };
1727         struct zonelist *zonelist;
1728
1729         if (noswap)
1730                 sc.may_unmap = 0;
1731
1732         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1733                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1734         zonelist = NODE_DATA(numa_node_id())->node_zonelists;
1735         return do_try_to_free_pages(zonelist, &sc);
1736 }
1737 #endif
1738
1739 /*
1740  * For kswapd, balance_pgdat() will work across all this node's zones until
1741  * they are all at pages_high.
1742  *
1743  * Returns the number of pages which were actually freed.
1744  *
1745  * There is special handling here for zones which are full of pinned pages.
1746  * This can happen if the pages are all mlocked, or if they are all used by
1747  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1748  * What we do is to detect the case where all pages in the zone have been
1749  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1750  * dead and from now on, only perform a short scan.  Basically we're polling
1751  * the zone for when the problem goes away.
1752  *
1753  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
1754  * zones which have free_pages > pages_high, but once a zone is found to have
1755  * free_pages <= pages_high, we scan that zone and the lower zones regardless
1756  * of the number of free pages in the lower zones.  This interoperates with
1757  * the page allocator fallback scheme to ensure that aging of pages is balanced
1758  * across the zones.
1759  */
1760 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
1761 {
1762         int all_zones_ok;
1763         int priority;
1764         int i;
1765         unsigned long total_scanned;
1766         struct reclaim_state *reclaim_state = current->reclaim_state;
1767         struct scan_control sc = {
1768                 .gfp_mask = GFP_KERNEL,
1769                 .may_unmap = 1,
1770                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1771                 .swappiness = vm_swappiness,
1772                 .order = order,
1773                 .mem_cgroup = NULL,
1774                 .isolate_pages = isolate_pages_global,
1775         };
1776         /*
1777          * temp_priority is used to remember the scanning priority at which
1778          * this zone was successfully refilled to free_pages == pages_high.
1779          */
1780         int temp_priority[MAX_NR_ZONES];
1781
1782 loop_again:
1783         total_scanned = 0;
1784         sc.nr_reclaimed = 0;
1785         sc.may_writepage = !laptop_mode;
1786         count_vm_event(PAGEOUTRUN);
1787
1788         for (i = 0; i < pgdat->nr_zones; i++)
1789                 temp_priority[i] = DEF_PRIORITY;
1790
1791         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1792                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1793                 unsigned long lru_pages = 0;
1794
1795                 /* The swap token gets in the way of swapout... */
1796                 if (!priority)
1797                         disable_swap_token();
1798
1799                 all_zones_ok = 1;
1800
1801                 /*
1802                  * Scan in the highmem->dma direction for the highest
1803                  * zone which needs scanning
1804                  */
1805                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1806                         struct zone *zone = pgdat->node_zones + i;
1807
1808                         if (!populated_zone(zone))
1809                                 continue;
1810
1811                         if (zone_is_all_unreclaimable(zone) &&
1812                             priority != DEF_PRIORITY)
1813                                 continue;
1814
1815                         /*
1816                          * Do some background aging of the anon list, to give
1817                          * pages a chance to be referenced before reclaiming.
1818                          */
1819                         if (inactive_anon_is_low(zone, &sc))
1820                                 shrink_active_list(SWAP_CLUSTER_MAX, zone,
1821                                                         &sc, priority, 0);
1822
1823                         if (!zone_watermark_ok(zone, order, zone->pages_high,
1824                                                0, 0)) {
1825                                 end_zone = i;
1826                                 break;
1827                         }
1828                 }
1829                 if (i < 0)
1830                         goto out;
1831
1832                 for (i = 0; i <= end_zone; i++) {
1833                         struct zone *zone = pgdat->node_zones + i;
1834
1835                         lru_pages += zone_lru_pages(zone);
1836                 }
1837
1838                 /*
1839                  * Now scan the zone in the dma->highmem direction, stopping
1840                  * at the last zone which needs scanning.
1841                  *
1842                  * We do this because the page allocator works in the opposite
1843                  * direction.  This prevents the page allocator from allocating
1844                  * pages behind kswapd's direction of progress, which would
1845                  * cause too much scanning of the lower zones.
1846                  */
1847                 for (i = 0; i <= end_zone; i++) {
1848                         struct zone *zone = pgdat->node_zones + i;
1849                         int nr_slab;
1850
1851                         if (!populated_zone(zone))
1852                                 continue;
1853
1854                         if (zone_is_all_unreclaimable(zone) &&
1855                                         priority != DEF_PRIORITY)
1856                                 continue;
1857
1858                         if (!zone_watermark_ok(zone, order, zone->pages_high,
1859                                                end_zone, 0))
1860                                 all_zones_ok = 0;
1861                         temp_priority[i] = priority;
1862                         sc.nr_scanned = 0;
1863                         note_zone_scanning_priority(zone, priority);
1864                         /*
1865                          * We put equal pressure on every zone, unless one
1866                          * zone has way too many pages free already.
1867                          */
1868                         if (!zone_watermark_ok(zone, order, 8*zone->pages_high,
1869                                                 end_zone, 0))
1870                                 shrink_zone(priority, zone, &sc);
1871                         reclaim_state->reclaimed_slab = 0;
1872                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1873                                                 lru_pages);
1874                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1875                         total_scanned += sc.nr_scanned;
1876                         if (zone_is_all_unreclaimable(zone))
1877                                 continue;
1878                         if (nr_slab == 0 && zone->pages_scanned >=
1879                                                 (zone_lru_pages(zone) * 6))
1880                                         zone_set_flag(zone,
1881                                                       ZONE_ALL_UNRECLAIMABLE);
1882                         /*
1883                          * If we've done a decent amount of scanning and
1884                          * the reclaim ratio is low, start doing writepage
1885                          * even in laptop mode
1886                          */
1887                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1888                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
1889                                 sc.may_writepage = 1;
1890                 }
1891                 if (all_zones_ok)
1892                         break;          /* kswapd: all done */
1893                 /*
1894                  * OK, kswapd is getting into trouble.  Take a nap, then take
1895                  * another pass across the zones.
1896                  */
1897                 if (total_scanned && priority < DEF_PRIORITY - 2)
1898                         congestion_wait(WRITE, HZ/10);
1899
1900                 /*
1901                  * We do this so kswapd doesn't build up large priorities for
1902                  * example when it is freeing in parallel with allocators. It
1903                  * matches the direct reclaim path behaviour in terms of impact
1904                  * on zone->*_priority.
1905                  */
1906                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
1907                         break;
1908         }
1909 out:
1910         /*
1911          * Note within each zone the priority level at which this zone was
1912          * brought into a happy state.  So that the next thread which scans this
1913          * zone will start out at that priority level.
1914          */
1915         for (i = 0; i < pgdat->nr_zones; i++) {
1916                 struct zone *zone = pgdat->node_zones + i;
1917
1918                 zone->prev_priority = temp_priority[i];
1919         }
1920         if (!all_zones_ok) {
1921                 cond_resched();
1922
1923                 try_to_freeze();
1924
1925                 /*
1926                  * Fragmentation may mean that the system cannot be
1927                  * rebalanced for high-order allocations in all zones.
1928                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
1929                  * it means the zones have been fully scanned and are still
1930                  * not balanced. For high-order allocations, there is
1931                  * little point trying all over again as kswapd may
1932                  * infinite loop.
1933                  *
1934                  * Instead, recheck all watermarks at order-0 as they
1935                  * are the most important. If watermarks are ok, kswapd will go
1936                  * back to sleep. High-order users can still perform direct
1937                  * reclaim if they wish.
1938                  */
1939                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
1940                         order = sc.order = 0;
1941
1942                 goto loop_again;
1943         }
1944
1945         return sc.nr_reclaimed;
1946 }
1947
1948 /*
1949  * The background pageout daemon, started as a kernel thread
1950  * from the init process.
1951  *
1952  * This basically trickles out pages so that we have _some_
1953  * free memory available even if there is no other activity
1954  * that frees anything up. This is needed for things like routing
1955  * etc, where we otherwise might have all activity going on in
1956  * asynchronous contexts that cannot page things out.
1957  *
1958  * If there are applications that are active memory-allocators
1959  * (most normal use), this basically shouldn't matter.
1960  */
1961 static int kswapd(void *p)
1962 {
1963         unsigned long order;
1964         pg_data_t *pgdat = (pg_data_t*)p;
1965         struct task_struct *tsk = current;
1966         DEFINE_WAIT(wait);
1967         struct reclaim_state reclaim_state = {
1968                 .reclaimed_slab = 0,
1969         };
1970         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1971
1972         lockdep_set_current_reclaim_state(GFP_KERNEL);
1973
1974         if (!cpumask_empty(cpumask))
1975                 set_cpus_allowed_ptr(tsk, cpumask);
1976         current->reclaim_state = &reclaim_state;
1977
1978         /*
1979          * Tell the memory management that we're a "memory allocator",
1980          * and that if we need more memory we should get access to it
1981          * regardless (see "__alloc_pages()"). "kswapd" should
1982          * never get caught in the normal page freeing logic.
1983          *
1984          * (Kswapd normally doesn't need memory anyway, but sometimes
1985          * you need a small amount of memory in order to be able to
1986          * page out something else, and this flag essentially protects
1987          * us from recursively trying to free more memory as we're
1988          * trying to free the first piece of memory in the first place).
1989          */
1990         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
1991         set_freezable();
1992
1993         order = 0;
1994         for ( ; ; ) {
1995                 unsigned long new_order;
1996
1997                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1998                 new_order = pgdat->kswapd_max_order;
1999                 pgdat->kswapd_max_order = 0;
2000                 if (order < new_order) {
2001                         /*
2002                          * Don't sleep if someone wants a larger 'order'
2003                          * allocation
2004                          */
2005                         order = new_order;
2006                 } else {
2007                         if (!freezing(current))
2008                                 schedule();
2009
2010                         order = pgdat->kswapd_max_order;
2011                 }
2012                 finish_wait(&pgdat->kswapd_wait, &wait);
2013
2014                 if (!try_to_freeze()) {
2015                         /* We can speed up thawing tasks if we don't call
2016                          * balance_pgdat after returning from the refrigerator
2017                          */
2018                         balance_pgdat(pgdat, order);
2019                 }
2020         }
2021         return 0;
2022 }
2023
2024 /*
2025  * A zone is low on free memory, so wake its kswapd task to service it.
2026  */
2027 void wakeup_kswapd(struct zone *zone, int order)
2028 {
2029         pg_data_t *pgdat;
2030
2031         if (!populated_zone(zone))
2032                 return;
2033
2034         pgdat = zone->zone_pgdat;
2035         if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0))
2036                 return;
2037         if (pgdat->kswapd_max_order < order)
2038                 pgdat->kswapd_max_order = order;
2039         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2040                 return;
2041         if (!waitqueue_active(&pgdat->kswapd_wait))
2042                 return;
2043         wake_up_interruptible(&pgdat->kswapd_wait);
2044 }
2045
2046 unsigned long global_lru_pages(void)
2047 {
2048         return global_page_state(NR_ACTIVE_ANON)
2049                 + global_page_state(NR_ACTIVE_FILE)
2050                 + global_page_state(NR_INACTIVE_ANON)
2051                 + global_page_state(NR_INACTIVE_FILE);
2052 }
2053
2054 #ifdef CONFIG_PM
2055 /*
2056  * Helper function for shrink_all_memory().  Tries to reclaim 'nr_pages' pages
2057  * from LRU lists system-wide, for given pass and priority.
2058  *
2059  * For pass > 3 we also try to shrink the LRU lists that contain a few pages
2060  */
2061 static void shrink_all_zones(unsigned long nr_pages, int prio,
2062                                       int pass, struct scan_control *sc)
2063 {
2064         struct zone *zone;
2065         unsigned long nr_reclaimed = 0;
2066
2067         for_each_populated_zone(zone) {
2068                 enum lru_list l;
2069
2070                 if (zone_is_all_unreclaimable(zone) && prio != DEF_PRIORITY)
2071                         continue;
2072
2073                 for_each_evictable_lru(l) {
2074                         enum zone_stat_item ls = NR_LRU_BASE + l;
2075                         unsigned long lru_pages = zone_page_state(zone, ls);
2076
2077                         /* For pass = 0, we don't shrink the active list */
2078                         if (pass == 0 && (l == LRU_ACTIVE_ANON ||
2079                                                 l == LRU_ACTIVE_FILE))
2080                                 continue;
2081
2082                         zone->lru[l].nr_scan += (lru_pages >> prio) + 1;
2083                         if (zone->lru[l].nr_scan >= nr_pages || pass > 3) {
2084                                 unsigned long nr_to_scan;
2085
2086                                 zone->lru[l].nr_scan = 0;
2087                                 nr_to_scan = min(nr_pages, lru_pages);
2088                                 nr_reclaimed += shrink_list(l, nr_to_scan, zone,
2089                                                                 sc, prio);
2090                                 if (nr_reclaimed >= nr_pages) {
2091                                         sc->nr_reclaimed += nr_reclaimed;
2092                                         return;
2093                                 }
2094                         }
2095                 }
2096         }
2097         sc->nr_reclaimed += nr_reclaimed;
2098 }
2099
2100 /*
2101  * Try to free `nr_pages' of memory, system-wide, and return the number of
2102  * freed pages.
2103  *
2104  * Rather than trying to age LRUs the aim is to preserve the overall
2105  * LRU order by reclaiming preferentially
2106  * inactive > active > active referenced > active mapped
2107  */
2108 unsigned long shrink_all_memory(unsigned long nr_pages)
2109 {
2110         unsigned long lru_pages, nr_slab;
2111         int pass;
2112         struct reclaim_state reclaim_state;
2113         struct scan_control sc = {
2114                 .gfp_mask = GFP_KERNEL,
2115                 .may_unmap = 0,
2116                 .may_writepage = 1,
2117                 .isolate_pages = isolate_pages_global,
2118                 .nr_reclaimed = 0,
2119         };
2120
2121         current->reclaim_state = &reclaim_state;
2122
2123         lru_pages = global_lru_pages();
2124         nr_slab = global_page_state(NR_SLAB_RECLAIMABLE);
2125         /* If slab caches are huge, it's better to hit them first */
2126         while (nr_slab >= lru_pages) {
2127                 reclaim_state.reclaimed_slab = 0;
2128                 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
2129                 if (!reclaim_state.reclaimed_slab)
2130                         break;
2131
2132                 sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2133                 if (sc.nr_reclaimed >= nr_pages)
2134                         goto out;
2135
2136                 nr_slab -= reclaim_state.reclaimed_slab;
2137         }
2138
2139         /*
2140          * We try to shrink LRUs in 5 passes:
2141          * 0 = Reclaim from inactive_list only
2142          * 1 = Reclaim from active list but don't reclaim mapped
2143          * 2 = 2nd pass of type 1
2144          * 3 = Reclaim mapped (normal reclaim)
2145          * 4 = 2nd pass of type 3
2146          */
2147         for (pass = 0; pass < 5; pass++) {
2148                 int prio;
2149
2150                 /* Force reclaiming mapped pages in the passes #3 and #4 */
2151                 if (pass > 2)
2152                         sc.may_unmap = 1;
2153
2154                 for (prio = DEF_PRIORITY; prio >= 0; prio--) {
2155                         unsigned long nr_to_scan = nr_pages - sc.nr_reclaimed;
2156
2157                         sc.nr_scanned = 0;
2158                         sc.swap_cluster_max = nr_to_scan;
2159                         shrink_all_zones(nr_to_scan, prio, pass, &sc);
2160                         if (sc.nr_reclaimed >= nr_pages)
2161                                 goto out;
2162
2163                         reclaim_state.reclaimed_slab = 0;
2164                         shrink_slab(sc.nr_scanned, sc.gfp_mask,
2165                                         global_lru_pages());
2166                         sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2167                         if (sc.nr_reclaimed >= nr_pages)
2168                                 goto out;
2169
2170                         if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
2171                                 congestion_wait(WRITE, HZ / 10);
2172                 }
2173         }
2174
2175         /*
2176          * If sc.nr_reclaimed = 0, we could not shrink LRUs, but there may be
2177          * something in slab caches
2178          */
2179         if (!sc.nr_reclaimed) {
2180                 do {
2181                         reclaim_state.reclaimed_slab = 0;
2182                         shrink_slab(nr_pages, sc.gfp_mask, global_lru_pages());
2183                         sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2184                 } while (sc.nr_reclaimed < nr_pages &&
2185                                 reclaim_state.reclaimed_slab > 0);
2186         }
2187
2188
2189 out:
2190         current->reclaim_state = NULL;
2191
2192         return sc.nr_reclaimed;
2193 }
2194 #endif
2195
2196 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2197    not required for correctness.  So if the last cpu in a node goes
2198    away, we get changed to run anywhere: as the first one comes back,
2199    restore their cpu bindings. */
2200 static int __devinit cpu_callback(struct notifier_block *nfb,
2201                                   unsigned long action, void *hcpu)
2202 {
2203         int nid;
2204
2205         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2206                 for_each_node_state(nid, N_HIGH_MEMORY) {
2207                         pg_data_t *pgdat = NODE_DATA(nid);
2208                         const struct cpumask *mask;
2209
2210                         mask = cpumask_of_node(pgdat->node_id);
2211
2212                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2213                                 /* One of our CPUs online: restore mask */
2214                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2215                 }
2216         }
2217         return NOTIFY_OK;
2218 }
2219
2220 /*
2221  * This kswapd start function will be called by init and node-hot-add.
2222  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2223  */
2224 int kswapd_run(int nid)
2225 {
2226         pg_data_t *pgdat = NODE_DATA(nid);
2227         int ret = 0;
2228
2229         if (pgdat->kswapd)
2230                 return 0;
2231
2232         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2233         if (IS_ERR(pgdat->kswapd)) {
2234                 /* failure at boot is fatal */
2235                 BUG_ON(system_state == SYSTEM_BOOTING);
2236                 printk("Failed to start kswapd on node %d\n",nid);
2237                 ret = -1;
2238         }
2239         return ret;
2240 }
2241
2242 static int __init kswapd_init(void)
2243 {
2244         int nid;
2245
2246         swap_setup();
2247         for_each_node_state(nid, N_HIGH_MEMORY)
2248                 kswapd_run(nid);
2249         hotcpu_notifier(cpu_callback, 0);
2250         return 0;
2251 }
2252
2253 module_init(kswapd_init)
2254
2255 #ifdef CONFIG_NUMA
2256 /*
2257  * Zone reclaim mode
2258  *
2259  * If non-zero call zone_reclaim when the number of free pages falls below
2260  * the watermarks.
2261  */
2262 int zone_reclaim_mode __read_mostly;
2263
2264 #define RECLAIM_OFF 0
2265 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
2266 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
2267 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
2268
2269 /*
2270  * Priority for ZONE_RECLAIM. This determines the fraction of pages
2271  * of a node considered for each zone_reclaim. 4 scans 1/16th of
2272  * a zone.
2273  */
2274 #define ZONE_RECLAIM_PRIORITY 4
2275
2276 /*
2277  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2278  * occur.
2279  */
2280 int sysctl_min_unmapped_ratio = 1;
2281
2282 /*
2283  * If the number of slab pages in a zone grows beyond this percentage then
2284  * slab reclaim needs to occur.
2285  */
2286 int sysctl_min_slab_ratio = 5;
2287
2288 /*
2289  * Try to free up some pages from this zone through reclaim.
2290  */
2291 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2292 {
2293         /* Minimum pages needed in order to stay on node */
2294         const unsigned long nr_pages = 1 << order;
2295         struct task_struct *p = current;
2296         struct reclaim_state reclaim_state;
2297         int priority;
2298         struct scan_control sc = {
2299                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2300                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2301                 .swap_cluster_max = max_t(unsigned long, nr_pages,
2302                                         SWAP_CLUSTER_MAX),
2303                 .gfp_mask = gfp_mask,
2304                 .swappiness = vm_swappiness,
2305                 .order = order,
2306                 .isolate_pages = isolate_pages_global,
2307         };
2308         unsigned long slab_reclaimable;
2309
2310         disable_swap_token();
2311         cond_resched();
2312         /*
2313          * We need to be able to allocate from the reserves for RECLAIM_SWAP
2314          * and we also need to be able to write out pages for RECLAIM_WRITE
2315          * and RECLAIM_SWAP.
2316          */
2317         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2318         reclaim_state.reclaimed_slab = 0;
2319         p->reclaim_state = &reclaim_state;
2320
2321         if (zone_page_state(zone, NR_FILE_PAGES) -
2322                 zone_page_state(zone, NR_FILE_MAPPED) >
2323                 zone->min_unmapped_pages) {
2324                 /*
2325                  * Free memory by calling shrink zone with increasing
2326                  * priorities until we have enough memory freed.
2327                  */
2328                 priority = ZONE_RECLAIM_PRIORITY;
2329                 do {
2330                         note_zone_scanning_priority(zone, priority);
2331                         shrink_zone(priority, zone, &sc);
2332                         priority--;
2333                 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
2334         }
2335
2336         slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2337         if (slab_reclaimable > zone->min_slab_pages) {
2338                 /*
2339                  * shrink_slab() does not currently allow us to determine how
2340                  * many pages were freed in this zone. So we take the current
2341                  * number of slab pages and shake the slab until it is reduced
2342                  * by the same nr_pages that we used for reclaiming unmapped
2343                  * pages.
2344                  *
2345                  * Note that shrink_slab will free memory on all zones and may
2346                  * take a long time.
2347                  */
2348                 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
2349                         zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
2350                                 slab_reclaimable - nr_pages)
2351                         ;
2352
2353                 /*
2354                  * Update nr_reclaimed by the number of slab pages we
2355                  * reclaimed from this zone.
2356                  */
2357                 sc.nr_reclaimed += slab_reclaimable -
2358                         zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2359         }
2360
2361         p->reclaim_state = NULL;
2362         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
2363         return sc.nr_reclaimed >= nr_pages;
2364 }
2365
2366 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2367 {
2368         int node_id;
2369         int ret;
2370
2371         /*
2372          * Zone reclaim reclaims unmapped file backed pages and
2373          * slab pages if we are over the defined limits.
2374          *
2375          * A small portion of unmapped file backed pages is needed for
2376          * file I/O otherwise pages read by file I/O will be immediately
2377          * thrown out if the zone is overallocated. So we do not reclaim
2378          * if less than a specified percentage of the zone is used by
2379          * unmapped file backed pages.
2380          */
2381         if (zone_page_state(zone, NR_FILE_PAGES) -
2382             zone_page_state(zone, NR_FILE_MAPPED) <= zone->min_unmapped_pages
2383             && zone_page_state(zone, NR_SLAB_RECLAIMABLE)
2384                         <= zone->min_slab_pages)
2385                 return 0;
2386
2387         if (zone_is_all_unreclaimable(zone))
2388                 return 0;
2389
2390         /*
2391          * Do not scan if the allocation should not be delayed.
2392          */
2393         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
2394                         return 0;
2395
2396         /*
2397          * Only run zone reclaim on the local zone or on zones that do not
2398          * have associated processors. This will favor the local processor
2399          * over remote processors and spread off node memory allocations
2400          * as wide as possible.
2401          */
2402         node_id = zone_to_nid(zone);
2403         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
2404                 return 0;
2405
2406         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2407                 return 0;
2408         ret = __zone_reclaim(zone, gfp_mask, order);
2409         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2410
2411         return ret;
2412 }
2413 #endif
2414
2415 #ifdef CONFIG_UNEVICTABLE_LRU
2416 /*
2417  * page_evictable - test whether a page is evictable
2418  * @page: the page to test
2419  * @vma: the VMA in which the page is or will be mapped, may be NULL
2420  *
2421  * Test whether page is evictable--i.e., should be placed on active/inactive
2422  * lists vs unevictable list.  The vma argument is !NULL when called from the
2423  * fault path to determine how to instantate a new page.
2424  *
2425  * Reasons page might not be evictable:
2426  * (1) page's mapping marked unevictable
2427  * (2) page is part of an mlocked VMA
2428  *
2429  */
2430 int page_evictable(struct page *page, struct vm_area_struct *vma)
2431 {
2432
2433         if (mapping_unevictable(page_mapping(page)))
2434                 return 0;
2435
2436         if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2437                 return 0;
2438
2439         return 1;
2440 }
2441
2442 /**
2443  * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2444  * @page: page to check evictability and move to appropriate lru list
2445  * @zone: zone page is in
2446  *
2447  * Checks a page for evictability and moves the page to the appropriate
2448  * zone lru list.
2449  *
2450  * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2451  * have PageUnevictable set.
2452  */
2453 static void check_move_unevictable_page(struct page *page, struct zone *zone)
2454 {
2455         VM_BUG_ON(PageActive(page));
2456
2457 retry:
2458         ClearPageUnevictable(page);
2459         if (page_evictable(page, NULL)) {
2460                 enum lru_list l = LRU_INACTIVE_ANON + page_is_file_cache(page);
2461
2462                 __dec_zone_state(zone, NR_UNEVICTABLE);
2463                 list_move(&page->lru, &zone->lru[l].list);
2464                 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
2465                 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
2466                 __count_vm_event(UNEVICTABLE_PGRESCUED);
2467         } else {
2468                 /*
2469                  * rotate unevictable list
2470                  */
2471                 SetPageUnevictable(page);
2472                 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
2473                 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
2474                 if (page_evictable(page, NULL))
2475                         goto retry;
2476         }
2477 }
2478
2479 /**
2480  * scan_mapping_unevictable_pages - scan an address space for evictable pages
2481  * @mapping: struct address_space to scan for evictable pages
2482  *
2483  * Scan all pages in mapping.  Check unevictable pages for
2484  * evictability and move them to the appropriate zone lru list.
2485  */
2486 void scan_mapping_unevictable_pages(struct address_space *mapping)
2487 {
2488         pgoff_t next = 0;
2489         pgoff_t end   = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
2490                          PAGE_CACHE_SHIFT;
2491         struct zone *zone;
2492         struct pagevec pvec;
2493
2494         if (mapping->nrpages == 0)
2495                 return;
2496
2497         pagevec_init(&pvec, 0);
2498         while (next < end &&
2499                 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
2500                 int i;
2501                 int pg_scanned = 0;
2502
2503                 zone = NULL;
2504
2505                 for (i = 0; i < pagevec_count(&pvec); i++) {
2506                         struct page *page = pvec.pages[i];
2507                         pgoff_t page_index = page->index;
2508                         struct zone *pagezone = page_zone(page);
2509
2510                         pg_scanned++;
2511                         if (page_index > next)
2512                                 next = page_index;
2513                         next++;
2514
2515                         if (pagezone != zone) {
2516                                 if (zone)
2517                                         spin_unlock_irq(&zone->lru_lock);
2518                                 zone = pagezone;
2519                                 spin_lock_irq(&zone->lru_lock);
2520                         }
2521
2522                         if (PageLRU(page) && PageUnevictable(page))
2523                                 check_move_unevictable_page(page, zone);
2524                 }
2525                 if (zone)
2526                         spin_unlock_irq(&zone->lru_lock);
2527                 pagevec_release(&pvec);
2528
2529                 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
2530         }
2531
2532 }
2533
2534 /**
2535  * scan_zone_unevictable_pages - check unevictable list for evictable pages
2536  * @zone - zone of which to scan the unevictable list
2537  *
2538  * Scan @zone's unevictable LRU lists to check for pages that have become
2539  * evictable.  Move those that have to @zone's inactive list where they
2540  * become candidates for reclaim, unless shrink_inactive_zone() decides
2541  * to reactivate them.  Pages that are still unevictable are rotated
2542  * back onto @zone's unevictable list.
2543  */
2544 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
2545 static void scan_zone_unevictable_pages(struct zone *zone)
2546 {
2547         struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
2548         unsigned long scan;
2549         unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
2550
2551         while (nr_to_scan > 0) {
2552                 unsigned long batch_size = min(nr_to_scan,
2553                                                 SCAN_UNEVICTABLE_BATCH_SIZE);
2554
2555                 spin_lock_irq(&zone->lru_lock);
2556                 for (scan = 0;  scan < batch_size; scan++) {
2557                         struct page *page = lru_to_page(l_unevictable);
2558
2559                         if (!trylock_page(page))
2560                                 continue;
2561
2562                         prefetchw_prev_lru_page(page, l_unevictable, flags);
2563
2564                         if (likely(PageLRU(page) && PageUnevictable(page)))
2565                                 check_move_unevictable_page(page, zone);
2566
2567                         unlock_page(page);
2568                 }
2569                 spin_unlock_irq(&zone->lru_lock);
2570
2571                 nr_to_scan -= batch_size;
2572         }
2573 }
2574
2575
2576 /**
2577  * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
2578  *
2579  * A really big hammer:  scan all zones' unevictable LRU lists to check for
2580  * pages that have become evictable.  Move those back to the zones'
2581  * inactive list where they become candidates for reclaim.
2582  * This occurs when, e.g., we have unswappable pages on the unevictable lists,
2583  * and we add swap to the system.  As such, it runs in the context of a task
2584  * that has possibly/probably made some previously unevictable pages
2585  * evictable.
2586  */
2587 static void scan_all_zones_unevictable_pages(void)
2588 {
2589         struct zone *zone;
2590
2591         for_each_zone(zone) {
2592                 scan_zone_unevictable_pages(zone);
2593         }
2594 }
2595
2596 /*
2597  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
2598  * all nodes' unevictable lists for evictable pages
2599  */
2600 unsigned long scan_unevictable_pages;
2601
2602 int scan_unevictable_handler(struct ctl_table *table, int write,
2603                            struct file *file, void __user *buffer,
2604                            size_t *length, loff_t *ppos)
2605 {
2606         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
2607
2608         if (write && *(unsigned long *)table->data)
2609                 scan_all_zones_unevictable_pages();
2610
2611         scan_unevictable_pages = 0;
2612         return 0;
2613 }
2614
2615 /*
2616  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
2617  * a specified node's per zone unevictable lists for evictable pages.
2618  */
2619
2620 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
2621                                           struct sysdev_attribute *attr,
2622                                           char *buf)
2623 {
2624         return sprintf(buf, "0\n");     /* always zero; should fit... */
2625 }
2626
2627 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
2628                                            struct sysdev_attribute *attr,
2629                                         const char *buf, size_t count)
2630 {
2631         struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
2632         struct zone *zone;
2633         unsigned long res;
2634         unsigned long req = strict_strtoul(buf, 10, &res);
2635
2636         if (!req)
2637                 return 1;       /* zero is no-op */
2638
2639         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2640                 if (!populated_zone(zone))
2641                         continue;
2642                 scan_zone_unevictable_pages(zone);
2643         }
2644         return 1;
2645 }
2646
2647
2648 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
2649                         read_scan_unevictable_node,
2650                         write_scan_unevictable_node);
2651
2652 int scan_unevictable_register_node(struct node *node)
2653 {
2654         return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
2655 }
2656
2657 void scan_unevictable_unregister_node(struct node *node)
2658 {
2659         sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
2660 }
2661
2662 #endif