vmscan: low order lumpy reclaim also should use PAGEOUT_IO_SYNC
[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         /* Can pages be swapped as part of reclaim? */
67         int may_swap;
68
69         /* This context's SWAP_CLUSTER_MAX. If freeing memory for
70          * suspend, we effectively ignore SWAP_CLUSTER_MAX.
71          * In this context, it doesn't matter that we scan the
72          * whole list at once. */
73         int swap_cluster_max;
74
75         int swappiness;
76
77         int all_unreclaimable;
78
79         int order;
80
81         /* Which cgroup do we reclaim from */
82         struct mem_cgroup *mem_cgroup;
83
84         /*
85          * Nodemask of nodes allowed by the caller. If NULL, all nodes
86          * are scanned.
87          */
88         nodemask_t      *nodemask;
89
90         /* Pluggable isolate pages callback */
91         unsigned long (*isolate_pages)(unsigned long nr, struct list_head *dst,
92                         unsigned long *scanned, int order, int mode,
93                         struct zone *z, struct mem_cgroup *mem_cont,
94                         int active, int file);
95 };
96
97 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
98
99 #ifdef ARCH_HAS_PREFETCH
100 #define prefetch_prev_lru_page(_page, _base, _field)                    \
101         do {                                                            \
102                 if ((_page)->lru.prev != _base) {                       \
103                         struct page *prev;                              \
104                                                                         \
105                         prev = lru_to_page(&(_page->lru));              \
106                         prefetch(&prev->_field);                        \
107                 }                                                       \
108         } while (0)
109 #else
110 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
111 #endif
112
113 #ifdef ARCH_HAS_PREFETCHW
114 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
115         do {                                                            \
116                 if ((_page)->lru.prev != _base) {                       \
117                         struct page *prev;                              \
118                                                                         \
119                         prev = lru_to_page(&(_page->lru));              \
120                         prefetchw(&prev->_field);                       \
121                 }                                                       \
122         } while (0)
123 #else
124 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
125 #endif
126
127 /*
128  * From 0 .. 100.  Higher means more swappy.
129  */
130 int vm_swappiness = 60;
131 long vm_total_pages;    /* The total number of pages which the VM controls */
132
133 static LIST_HEAD(shrinker_list);
134 static DECLARE_RWSEM(shrinker_rwsem);
135
136 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
137 #define scanning_global_lru(sc) (!(sc)->mem_cgroup)
138 #else
139 #define scanning_global_lru(sc) (1)
140 #endif
141
142 static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
143                                                   struct scan_control *sc)
144 {
145         if (!scanning_global_lru(sc))
146                 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
147
148         return &zone->reclaim_stat;
149 }
150
151 static unsigned long zone_nr_pages(struct zone *zone, struct scan_control *sc,
152                                    enum lru_list lru)
153 {
154         if (!scanning_global_lru(sc))
155                 return mem_cgroup_zone_nr_pages(sc->mem_cgroup, zone, lru);
156
157         return zone_page_state(zone, NR_LRU_BASE + lru);
158 }
159
160
161 /*
162  * Add a shrinker callback to be called from the vm
163  */
164 void register_shrinker(struct shrinker *shrinker)
165 {
166         shrinker->nr = 0;
167         down_write(&shrinker_rwsem);
168         list_add_tail(&shrinker->list, &shrinker_list);
169         up_write(&shrinker_rwsem);
170 }
171 EXPORT_SYMBOL(register_shrinker);
172
173 /*
174  * Remove one
175  */
176 void unregister_shrinker(struct shrinker *shrinker)
177 {
178         down_write(&shrinker_rwsem);
179         list_del(&shrinker->list);
180         up_write(&shrinker_rwsem);
181 }
182 EXPORT_SYMBOL(unregister_shrinker);
183
184 #define SHRINK_BATCH 128
185 /*
186  * Call the shrink functions to age shrinkable caches
187  *
188  * Here we assume it costs one seek to replace a lru page and that it also
189  * takes a seek to recreate a cache object.  With this in mind we age equal
190  * percentages of the lru and ageable caches.  This should balance the seeks
191  * generated by these structures.
192  *
193  * If the vm encountered mapped pages on the LRU it increase the pressure on
194  * slab to avoid swapping.
195  *
196  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
197  *
198  * `lru_pages' represents the number of on-LRU pages in all the zones which
199  * are eligible for the caller's allocation attempt.  It is used for balancing
200  * slab reclaim versus page reclaim.
201  *
202  * Returns the number of slab objects which we shrunk.
203  */
204 unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
205                         unsigned long lru_pages)
206 {
207         struct shrinker *shrinker;
208         unsigned long ret = 0;
209
210         if (scanned == 0)
211                 scanned = SWAP_CLUSTER_MAX;
212
213         if (!down_read_trylock(&shrinker_rwsem))
214                 return 1;       /* Assume we'll be able to shrink next time */
215
216         list_for_each_entry(shrinker, &shrinker_list, list) {
217                 unsigned long long delta;
218                 unsigned long total_scan;
219                 unsigned long max_pass = (*shrinker->shrink)(0, gfp_mask);
220
221                 delta = (4 * scanned) / shrinker->seeks;
222                 delta *= max_pass;
223                 do_div(delta, lru_pages + 1);
224                 shrinker->nr += delta;
225                 if (shrinker->nr < 0) {
226                         printk(KERN_ERR "shrink_slab: %pF negative objects to "
227                                "delete nr=%ld\n",
228                                shrinker->shrink, shrinker->nr);
229                         shrinker->nr = max_pass;
230                 }
231
232                 /*
233                  * Avoid risking looping forever due to too large nr value:
234                  * never try to free more than twice the estimate number of
235                  * freeable entries.
236                  */
237                 if (shrinker->nr > max_pass * 2)
238                         shrinker->nr = max_pass * 2;
239
240                 total_scan = shrinker->nr;
241                 shrinker->nr = 0;
242
243                 while (total_scan >= SHRINK_BATCH) {
244                         long this_scan = SHRINK_BATCH;
245                         int shrink_ret;
246                         int nr_before;
247
248                         nr_before = (*shrinker->shrink)(0, gfp_mask);
249                         shrink_ret = (*shrinker->shrink)(this_scan, gfp_mask);
250                         if (shrink_ret == -1)
251                                 break;
252                         if (shrink_ret < nr_before)
253                                 ret += nr_before - shrink_ret;
254                         count_vm_events(SLABS_SCANNED, this_scan);
255                         total_scan -= this_scan;
256
257                         cond_resched();
258                 }
259
260                 shrinker->nr += total_scan;
261         }
262         up_read(&shrinker_rwsem);
263         return ret;
264 }
265
266 /* Called without lock on whether page is mapped, so answer is unstable */
267 static inline int page_mapping_inuse(struct page *page)
268 {
269         struct address_space *mapping;
270
271         /* Page is in somebody's page tables. */
272         if (page_mapped(page))
273                 return 1;
274
275         /* Be more reluctant to reclaim swapcache than pagecache */
276         if (PageSwapCache(page))
277                 return 1;
278
279         mapping = page_mapping(page);
280         if (!mapping)
281                 return 0;
282
283         /* File is mmap'd by somebody? */
284         return mapping_mapped(mapping);
285 }
286
287 static inline int is_page_cache_freeable(struct page *page)
288 {
289         return page_count(page) - !!page_has_private(page) == 2;
290 }
291
292 static int may_write_to_queue(struct backing_dev_info *bdi)
293 {
294         if (current->flags & PF_SWAPWRITE)
295                 return 1;
296         if (!bdi_write_congested(bdi))
297                 return 1;
298         if (bdi == current->backing_dev_info)
299                 return 1;
300         return 0;
301 }
302
303 /*
304  * We detected a synchronous write error writing a page out.  Probably
305  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
306  * fsync(), msync() or close().
307  *
308  * The tricky part is that after writepage we cannot touch the mapping: nothing
309  * prevents it from being freed up.  But we have a ref on the page and once
310  * that page is locked, the mapping is pinned.
311  *
312  * We're allowed to run sleeping lock_page() here because we know the caller has
313  * __GFP_FS.
314  */
315 static void handle_write_error(struct address_space *mapping,
316                                 struct page *page, int error)
317 {
318         lock_page(page);
319         if (page_mapping(page) == mapping)
320                 mapping_set_error(mapping, error);
321         unlock_page(page);
322 }
323
324 /* Request for sync pageout. */
325 enum pageout_io {
326         PAGEOUT_IO_ASYNC,
327         PAGEOUT_IO_SYNC,
328 };
329
330 /* possible outcome of pageout() */
331 typedef enum {
332         /* failed to write page out, page is locked */
333         PAGE_KEEP,
334         /* move page to the active list, page is locked */
335         PAGE_ACTIVATE,
336         /* page has been sent to the disk successfully, page is unlocked */
337         PAGE_SUCCESS,
338         /* page is clean and locked */
339         PAGE_CLEAN,
340 } pageout_t;
341
342 /*
343  * pageout is called by shrink_page_list() for each dirty page.
344  * Calls ->writepage().
345  */
346 static pageout_t pageout(struct page *page, struct address_space *mapping,
347                                                 enum pageout_io sync_writeback)
348 {
349         /*
350          * If the page is dirty, only perform writeback if that write
351          * will be non-blocking.  To prevent this allocation from being
352          * stalled by pagecache activity.  But note that there may be
353          * stalls if we need to run get_block().  We could test
354          * PagePrivate for that.
355          *
356          * If this process is currently in generic_file_write() against
357          * this page's queue, we can perform writeback even if that
358          * will block.
359          *
360          * If the page is swapcache, write it back even if that would
361          * block, for some throttling. This happens by accident, because
362          * swap_backing_dev_info is bust: it doesn't reflect the
363          * congestion state of the swapdevs.  Easy to fix, if needed.
364          * See swapfile.c:page_queue_congested().
365          */
366         if (!is_page_cache_freeable(page))
367                 return PAGE_KEEP;
368         if (!mapping) {
369                 /*
370                  * Some data journaling orphaned pages can have
371                  * page->mapping == NULL while being dirty with clean buffers.
372                  */
373                 if (page_has_private(page)) {
374                         if (try_to_free_buffers(page)) {
375                                 ClearPageDirty(page);
376                                 printk("%s: orphaned page\n", __func__);
377                                 return PAGE_CLEAN;
378                         }
379                 }
380                 return PAGE_KEEP;
381         }
382         if (mapping->a_ops->writepage == NULL)
383                 return PAGE_ACTIVATE;
384         if (!may_write_to_queue(mapping->backing_dev_info))
385                 return PAGE_KEEP;
386
387         if (clear_page_dirty_for_io(page)) {
388                 int res;
389                 struct writeback_control wbc = {
390                         .sync_mode = WB_SYNC_NONE,
391                         .nr_to_write = SWAP_CLUSTER_MAX,
392                         .range_start = 0,
393                         .range_end = LLONG_MAX,
394                         .nonblocking = 1,
395                         .for_reclaim = 1,
396                 };
397
398                 SetPageReclaim(page);
399                 res = mapping->a_ops->writepage(page, &wbc);
400                 if (res < 0)
401                         handle_write_error(mapping, page, res);
402                 if (res == AOP_WRITEPAGE_ACTIVATE) {
403                         ClearPageReclaim(page);
404                         return PAGE_ACTIVATE;
405                 }
406
407                 /*
408                  * Wait on writeback if requested to. This happens when
409                  * direct reclaiming a large contiguous area and the
410                  * first attempt to free a range of pages fails.
411                  */
412                 if (PageWriteback(page) && sync_writeback == PAGEOUT_IO_SYNC)
413                         wait_on_page_writeback(page);
414
415                 if (!PageWriteback(page)) {
416                         /* synchronous write or broken a_ops? */
417                         ClearPageReclaim(page);
418                 }
419                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
420                 return PAGE_SUCCESS;
421         }
422
423         return PAGE_CLEAN;
424 }
425
426 /*
427  * Same as remove_mapping, but if the page is removed from the mapping, it
428  * gets returned with a refcount of 0.
429  */
430 static int __remove_mapping(struct address_space *mapping, struct page *page)
431 {
432         BUG_ON(!PageLocked(page));
433         BUG_ON(mapping != page_mapping(page));
434
435         spin_lock_irq(&mapping->tree_lock);
436         /*
437          * The non racy check for a busy page.
438          *
439          * Must be careful with the order of the tests. When someone has
440          * a ref to the page, it may be possible that they dirty it then
441          * drop the reference. So if PageDirty is tested before page_count
442          * here, then the following race may occur:
443          *
444          * get_user_pages(&page);
445          * [user mapping goes away]
446          * write_to(page);
447          *                              !PageDirty(page)    [good]
448          * SetPageDirty(page);
449          * put_page(page);
450          *                              !page_count(page)   [good, discard it]
451          *
452          * [oops, our write_to data is lost]
453          *
454          * Reversing the order of the tests ensures such a situation cannot
455          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
456          * load is not satisfied before that of page->_count.
457          *
458          * Note that if SetPageDirty is always performed via set_page_dirty,
459          * and thus under tree_lock, then this ordering is not required.
460          */
461         if (!page_freeze_refs(page, 2))
462                 goto cannot_free;
463         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
464         if (unlikely(PageDirty(page))) {
465                 page_unfreeze_refs(page, 2);
466                 goto cannot_free;
467         }
468
469         if (PageSwapCache(page)) {
470                 swp_entry_t swap = { .val = page_private(page) };
471                 __delete_from_swap_cache(page);
472                 spin_unlock_irq(&mapping->tree_lock);
473                 mem_cgroup_uncharge_swapcache(page, swap);
474                 swap_free(swap);
475         } else {
476                 __remove_from_page_cache(page);
477                 spin_unlock_irq(&mapping->tree_lock);
478                 mem_cgroup_uncharge_cache_page(page);
479         }
480
481         return 1;
482
483 cannot_free:
484         spin_unlock_irq(&mapping->tree_lock);
485         return 0;
486 }
487
488 /*
489  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
490  * someone else has a ref on the page, abort and return 0.  If it was
491  * successfully detached, return 1.  Assumes the caller has a single ref on
492  * this page.
493  */
494 int remove_mapping(struct address_space *mapping, struct page *page)
495 {
496         if (__remove_mapping(mapping, page)) {
497                 /*
498                  * Unfreezing the refcount with 1 rather than 2 effectively
499                  * drops the pagecache ref for us without requiring another
500                  * atomic operation.
501                  */
502                 page_unfreeze_refs(page, 1);
503                 return 1;
504         }
505         return 0;
506 }
507
508 /**
509  * putback_lru_page - put previously isolated page onto appropriate LRU list
510  * @page: page to be put back to appropriate lru list
511  *
512  * Add previously isolated @page to appropriate LRU list.
513  * Page may still be unevictable for other reasons.
514  *
515  * lru_lock must not be held, interrupts must be enabled.
516  */
517 #ifdef CONFIG_UNEVICTABLE_LRU
518 void putback_lru_page(struct page *page)
519 {
520         int lru;
521         int active = !!TestClearPageActive(page);
522         int was_unevictable = PageUnevictable(page);
523
524         VM_BUG_ON(PageLRU(page));
525
526 redo:
527         ClearPageUnevictable(page);
528
529         if (page_evictable(page, NULL)) {
530                 /*
531                  * For evictable pages, we can use the cache.
532                  * In event of a race, worst case is we end up with an
533                  * unevictable page on [in]active list.
534                  * We know how to handle that.
535                  */
536                 lru = active + page_is_file_cache(page);
537                 lru_cache_add_lru(page, lru);
538         } else {
539                 /*
540                  * Put unevictable pages directly on zone's unevictable
541                  * list.
542                  */
543                 lru = LRU_UNEVICTABLE;
544                 add_page_to_unevictable_list(page);
545         }
546
547         /*
548          * page's status can change while we move it among lru. If an evictable
549          * page is on unevictable list, it never be freed. To avoid that,
550          * check after we added it to the list, again.
551          */
552         if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
553                 if (!isolate_lru_page(page)) {
554                         put_page(page);
555                         goto redo;
556                 }
557                 /* This means someone else dropped this page from LRU
558                  * So, it will be freed or putback to LRU again. There is
559                  * nothing to do here.
560                  */
561         }
562
563         if (was_unevictable && lru != LRU_UNEVICTABLE)
564                 count_vm_event(UNEVICTABLE_PGRESCUED);
565         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
566                 count_vm_event(UNEVICTABLE_PGCULLED);
567
568         put_page(page);         /* drop ref from isolate */
569 }
570
571 #else /* CONFIG_UNEVICTABLE_LRU */
572
573 void putback_lru_page(struct page *page)
574 {
575         int lru;
576         VM_BUG_ON(PageLRU(page));
577
578         lru = !!TestClearPageActive(page) + page_is_file_cache(page);
579         lru_cache_add_lru(page, lru);
580         put_page(page);
581 }
582 #endif /* CONFIG_UNEVICTABLE_LRU */
583
584
585 /*
586  * shrink_page_list() returns the number of reclaimed pages
587  */
588 static unsigned long shrink_page_list(struct list_head *page_list,
589                                         struct scan_control *sc,
590                                         enum pageout_io sync_writeback)
591 {
592         LIST_HEAD(ret_pages);
593         struct pagevec freed_pvec;
594         int pgactivate = 0;
595         unsigned long nr_reclaimed = 0;
596
597         cond_resched();
598
599         pagevec_init(&freed_pvec, 1);
600         while (!list_empty(page_list)) {
601                 struct address_space *mapping;
602                 struct page *page;
603                 int may_enter_fs;
604                 int referenced;
605
606                 cond_resched();
607
608                 page = lru_to_page(page_list);
609                 list_del(&page->lru);
610
611                 if (!trylock_page(page))
612                         goto keep;
613
614                 VM_BUG_ON(PageActive(page));
615
616                 sc->nr_scanned++;
617
618                 if (unlikely(!page_evictable(page, NULL)))
619                         goto cull_mlocked;
620
621                 if (!sc->may_unmap && page_mapped(page))
622                         goto keep_locked;
623
624                 /* Double the slab pressure for mapped and swapcache pages */
625                 if (page_mapped(page) || PageSwapCache(page))
626                         sc->nr_scanned++;
627
628                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
629                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
630
631                 if (PageWriteback(page)) {
632                         /*
633                          * Synchronous reclaim is performed in two passes,
634                          * first an asynchronous pass over the list to
635                          * start parallel writeback, and a second synchronous
636                          * pass to wait for the IO to complete.  Wait here
637                          * for any page for which writeback has already
638                          * started.
639                          */
640                         if (sync_writeback == PAGEOUT_IO_SYNC && may_enter_fs)
641                                 wait_on_page_writeback(page);
642                         else
643                                 goto keep_locked;
644                 }
645
646                 referenced = page_referenced(page, 1, sc->mem_cgroup);
647                 /* In active use or really unfreeable?  Activate it. */
648                 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER &&
649                                         referenced && page_mapping_inuse(page))
650                         goto activate_locked;
651
652                 /*
653                  * Anonymous process memory has backing store?
654                  * Try to allocate it some swap space here.
655                  */
656                 if (PageAnon(page) && !PageSwapCache(page)) {
657                         if (!(sc->gfp_mask & __GFP_IO))
658                                 goto keep_locked;
659                         if (!add_to_swap(page))
660                                 goto activate_locked;
661                         may_enter_fs = 1;
662                 }
663
664                 mapping = page_mapping(page);
665
666                 /*
667                  * The page is mapped into the page tables of one or more
668                  * processes. Try to unmap it here.
669                  */
670                 if (page_mapped(page) && mapping) {
671                         switch (try_to_unmap(page, 0)) {
672                         case SWAP_FAIL:
673                                 goto activate_locked;
674                         case SWAP_AGAIN:
675                                 goto keep_locked;
676                         case SWAP_MLOCK:
677                                 goto cull_mlocked;
678                         case SWAP_SUCCESS:
679                                 ; /* try to free the page below */
680                         }
681                 }
682
683                 if (PageDirty(page)) {
684                         if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && referenced)
685                                 goto keep_locked;
686                         if (!may_enter_fs)
687                                 goto keep_locked;
688                         if (!sc->may_writepage)
689                                 goto keep_locked;
690
691                         /* Page is dirty, try to write it out here */
692                         switch (pageout(page, mapping, sync_writeback)) {
693                         case PAGE_KEEP:
694                                 goto keep_locked;
695                         case PAGE_ACTIVATE:
696                                 goto activate_locked;
697                         case PAGE_SUCCESS:
698                                 if (PageWriteback(page) || PageDirty(page))
699                                         goto keep;
700                                 /*
701                                  * A synchronous write - probably a ramdisk.  Go
702                                  * ahead and try to reclaim the page.
703                                  */
704                                 if (!trylock_page(page))
705                                         goto keep;
706                                 if (PageDirty(page) || PageWriteback(page))
707                                         goto keep_locked;
708                                 mapping = page_mapping(page);
709                         case PAGE_CLEAN:
710                                 ; /* try to free the page below */
711                         }
712                 }
713
714                 /*
715                  * If the page has buffers, try to free the buffer mappings
716                  * associated with this page. If we succeed we try to free
717                  * the page as well.
718                  *
719                  * We do this even if the page is PageDirty().
720                  * try_to_release_page() does not perform I/O, but it is
721                  * possible for a page to have PageDirty set, but it is actually
722                  * clean (all its buffers are clean).  This happens if the
723                  * buffers were written out directly, with submit_bh(). ext3
724                  * will do this, as well as the blockdev mapping.
725                  * try_to_release_page() will discover that cleanness and will
726                  * drop the buffers and mark the page clean - it can be freed.
727                  *
728                  * Rarely, pages can have buffers and no ->mapping.  These are
729                  * the pages which were not successfully invalidated in
730                  * truncate_complete_page().  We try to drop those buffers here
731                  * and if that worked, and the page is no longer mapped into
732                  * process address space (page_count == 1) it can be freed.
733                  * Otherwise, leave the page on the LRU so it is swappable.
734                  */
735                 if (page_has_private(page)) {
736                         if (!try_to_release_page(page, sc->gfp_mask))
737                                 goto activate_locked;
738                         if (!mapping && page_count(page) == 1) {
739                                 unlock_page(page);
740                                 if (put_page_testzero(page))
741                                         goto free_it;
742                                 else {
743                                         /*
744                                          * rare race with speculative reference.
745                                          * the speculative reference will free
746                                          * this page shortly, so we may
747                                          * increment nr_reclaimed here (and
748                                          * leave it off the LRU).
749                                          */
750                                         nr_reclaimed++;
751                                         continue;
752                                 }
753                         }
754                 }
755
756                 if (!mapping || !__remove_mapping(mapping, page))
757                         goto keep_locked;
758
759                 /*
760                  * At this point, we have no other references and there is
761                  * no way to pick any more up (removed from LRU, removed
762                  * from pagecache). Can use non-atomic bitops now (and
763                  * we obviously don't have to worry about waking up a process
764                  * waiting on the page lock, because there are no references.
765                  */
766                 __clear_page_locked(page);
767 free_it:
768                 nr_reclaimed++;
769                 if (!pagevec_add(&freed_pvec, page)) {
770                         __pagevec_free(&freed_pvec);
771                         pagevec_reinit(&freed_pvec);
772                 }
773                 continue;
774
775 cull_mlocked:
776                 if (PageSwapCache(page))
777                         try_to_free_swap(page);
778                 unlock_page(page);
779                 putback_lru_page(page);
780                 continue;
781
782 activate_locked:
783                 /* Not a candidate for swapping, so reclaim swap space. */
784                 if (PageSwapCache(page) && vm_swap_full())
785                         try_to_free_swap(page);
786                 VM_BUG_ON(PageActive(page));
787                 SetPageActive(page);
788                 pgactivate++;
789 keep_locked:
790                 unlock_page(page);
791 keep:
792                 list_add(&page->lru, &ret_pages);
793                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
794         }
795         list_splice(&ret_pages, page_list);
796         if (pagevec_count(&freed_pvec))
797                 __pagevec_free(&freed_pvec);
798         count_vm_events(PGACTIVATE, pgactivate);
799         return nr_reclaimed;
800 }
801
802 /* LRU Isolation modes. */
803 #define ISOLATE_INACTIVE 0      /* Isolate inactive pages. */
804 #define ISOLATE_ACTIVE 1        /* Isolate active pages. */
805 #define ISOLATE_BOTH 2          /* Isolate both active and inactive pages. */
806
807 /*
808  * Attempt to remove the specified page from its LRU.  Only take this page
809  * if it is of the appropriate PageActive status.  Pages which are being
810  * freed elsewhere are also ignored.
811  *
812  * page:        page to consider
813  * mode:        one of the LRU isolation modes defined above
814  *
815  * returns 0 on success, -ve errno on failure.
816  */
817 int __isolate_lru_page(struct page *page, int mode, int file)
818 {
819         int ret = -EINVAL;
820
821         /* Only take pages on the LRU. */
822         if (!PageLRU(page))
823                 return ret;
824
825         /*
826          * When checking the active state, we need to be sure we are
827          * dealing with comparible boolean values.  Take the logical not
828          * of each.
829          */
830         if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
831                 return ret;
832
833         if (mode != ISOLATE_BOTH && (!page_is_file_cache(page) != !file))
834                 return ret;
835
836         /*
837          * When this function is being called for lumpy reclaim, we
838          * initially look into all LRU pages, active, inactive and
839          * unevictable; only give shrink_page_list evictable pages.
840          */
841         if (PageUnevictable(page))
842                 return ret;
843
844         ret = -EBUSY;
845
846         if (likely(get_page_unless_zero(page))) {
847                 /*
848                  * Be careful not to clear PageLRU until after we're
849                  * sure the page is not being freed elsewhere -- the
850                  * page release code relies on it.
851                  */
852                 ClearPageLRU(page);
853                 ret = 0;
854                 mem_cgroup_del_lru(page);
855         }
856
857         return ret;
858 }
859
860 /*
861  * zone->lru_lock is heavily contended.  Some of the functions that
862  * shrink the lists perform better by taking out a batch of pages
863  * and working on them outside the LRU lock.
864  *
865  * For pagecache intensive workloads, this function is the hottest
866  * spot in the kernel (apart from copy_*_user functions).
867  *
868  * Appropriate locks must be held before calling this function.
869  *
870  * @nr_to_scan: The number of pages to look through on the list.
871  * @src:        The LRU list to pull pages off.
872  * @dst:        The temp list to put pages on to.
873  * @scanned:    The number of pages that were scanned.
874  * @order:      The caller's attempted allocation order
875  * @mode:       One of the LRU isolation modes
876  * @file:       True [1] if isolating file [!anon] pages
877  *
878  * returns how many pages were moved onto *@dst.
879  */
880 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
881                 struct list_head *src, struct list_head *dst,
882                 unsigned long *scanned, int order, int mode, int file)
883 {
884         unsigned long nr_taken = 0;
885         unsigned long scan;
886
887         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
888                 struct page *page;
889                 unsigned long pfn;
890                 unsigned long end_pfn;
891                 unsigned long page_pfn;
892                 int zone_id;
893
894                 page = lru_to_page(src);
895                 prefetchw_prev_lru_page(page, src, flags);
896
897                 VM_BUG_ON(!PageLRU(page));
898
899                 switch (__isolate_lru_page(page, mode, file)) {
900                 case 0:
901                         list_move(&page->lru, dst);
902                         nr_taken++;
903                         break;
904
905                 case -EBUSY:
906                         /* else it is being freed elsewhere */
907                         list_move(&page->lru, src);
908                         continue;
909
910                 default:
911                         BUG();
912                 }
913
914                 if (!order)
915                         continue;
916
917                 /*
918                  * Attempt to take all pages in the order aligned region
919                  * surrounding the tag page.  Only take those pages of
920                  * the same active state as that tag page.  We may safely
921                  * round the target page pfn down to the requested order
922                  * as the mem_map is guarenteed valid out to MAX_ORDER,
923                  * where that page is in a different zone we will detect
924                  * it from its zone id and abort this block scan.
925                  */
926                 zone_id = page_zone_id(page);
927                 page_pfn = page_to_pfn(page);
928                 pfn = page_pfn & ~((1 << order) - 1);
929                 end_pfn = pfn + (1 << order);
930                 for (; pfn < end_pfn; pfn++) {
931                         struct page *cursor_page;
932
933                         /* The target page is in the block, ignore it. */
934                         if (unlikely(pfn == page_pfn))
935                                 continue;
936
937                         /* Avoid holes within the zone. */
938                         if (unlikely(!pfn_valid_within(pfn)))
939                                 break;
940
941                         cursor_page = pfn_to_page(pfn);
942
943                         /* Check that we have not crossed a zone boundary. */
944                         if (unlikely(page_zone_id(cursor_page) != zone_id))
945                                 continue;
946                         switch (__isolate_lru_page(cursor_page, mode, file)) {
947                         case 0:
948                                 list_move(&cursor_page->lru, dst);
949                                 nr_taken++;
950                                 scan++;
951                                 break;
952
953                         case -EBUSY:
954                                 /* else it is being freed elsewhere */
955                                 list_move(&cursor_page->lru, src);
956                         default:
957                                 break;  /* ! on LRU or wrong list */
958                         }
959                 }
960         }
961
962         *scanned = scan;
963         return nr_taken;
964 }
965
966 static unsigned long isolate_pages_global(unsigned long nr,
967                                         struct list_head *dst,
968                                         unsigned long *scanned, int order,
969                                         int mode, struct zone *z,
970                                         struct mem_cgroup *mem_cont,
971                                         int active, int file)
972 {
973         int lru = LRU_BASE;
974         if (active)
975                 lru += LRU_ACTIVE;
976         if (file)
977                 lru += LRU_FILE;
978         return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
979                                                                 mode, !!file);
980 }
981
982 /*
983  * clear_active_flags() is a helper for shrink_active_list(), clearing
984  * any active bits from the pages in the list.
985  */
986 static unsigned long clear_active_flags(struct list_head *page_list,
987                                         unsigned int *count)
988 {
989         int nr_active = 0;
990         int lru;
991         struct page *page;
992
993         list_for_each_entry(page, page_list, lru) {
994                 lru = page_is_file_cache(page);
995                 if (PageActive(page)) {
996                         lru += LRU_ACTIVE;
997                         ClearPageActive(page);
998                         nr_active++;
999                 }
1000                 count[lru]++;
1001         }
1002
1003         return nr_active;
1004 }
1005
1006 /**
1007  * isolate_lru_page - tries to isolate a page from its LRU list
1008  * @page: page to isolate from its LRU list
1009  *
1010  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1011  * vmstat statistic corresponding to whatever LRU list the page was on.
1012  *
1013  * Returns 0 if the page was removed from an LRU list.
1014  * Returns -EBUSY if the page was not on an LRU list.
1015  *
1016  * The returned page will have PageLRU() cleared.  If it was found on
1017  * the active list, it will have PageActive set.  If it was found on
1018  * the unevictable list, it will have the PageUnevictable bit set. That flag
1019  * may need to be cleared by the caller before letting the page go.
1020  *
1021  * The vmstat statistic corresponding to the list on which the page was
1022  * found will be decremented.
1023  *
1024  * Restrictions:
1025  * (1) Must be called with an elevated refcount on the page. This is a
1026  *     fundamentnal difference from isolate_lru_pages (which is called
1027  *     without a stable reference).
1028  * (2) the lru_lock must not be held.
1029  * (3) interrupts must be enabled.
1030  */
1031 int isolate_lru_page(struct page *page)
1032 {
1033         int ret = -EBUSY;
1034
1035         if (PageLRU(page)) {
1036                 struct zone *zone = page_zone(page);
1037
1038                 spin_lock_irq(&zone->lru_lock);
1039                 if (PageLRU(page) && get_page_unless_zero(page)) {
1040                         int lru = page_lru(page);
1041                         ret = 0;
1042                         ClearPageLRU(page);
1043
1044                         del_page_from_lru_list(zone, page, lru);
1045                 }
1046                 spin_unlock_irq(&zone->lru_lock);
1047         }
1048         return ret;
1049 }
1050
1051 /*
1052  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1053  * of reclaimed pages
1054  */
1055 static unsigned long shrink_inactive_list(unsigned long max_scan,
1056                         struct zone *zone, struct scan_control *sc,
1057                         int priority, int file)
1058 {
1059         LIST_HEAD(page_list);
1060         struct pagevec pvec;
1061         unsigned long nr_scanned = 0;
1062         unsigned long nr_reclaimed = 0;
1063         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1064         int lumpy_reclaim = 0;
1065
1066         /*
1067          * If we need a large contiguous chunk of memory, or have
1068          * trouble getting a small set of contiguous pages, we
1069          * will reclaim both active and inactive pages.
1070          *
1071          * We use the same threshold as pageout congestion_wait below.
1072          */
1073         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1074                 lumpy_reclaim = 1;
1075         else if (sc->order && priority < DEF_PRIORITY - 2)
1076                 lumpy_reclaim = 1;
1077
1078         pagevec_init(&pvec, 1);
1079
1080         lru_add_drain();
1081         spin_lock_irq(&zone->lru_lock);
1082         do {
1083                 struct page *page;
1084                 unsigned long nr_taken;
1085                 unsigned long nr_scan;
1086                 unsigned long nr_freed;
1087                 unsigned long nr_active;
1088                 unsigned int count[NR_LRU_LISTS] = { 0, };
1089                 int mode = lumpy_reclaim ? ISOLATE_BOTH : ISOLATE_INACTIVE;
1090
1091                 nr_taken = sc->isolate_pages(sc->swap_cluster_max,
1092                              &page_list, &nr_scan, sc->order, mode,
1093                                 zone, sc->mem_cgroup, 0, file);
1094                 nr_active = clear_active_flags(&page_list, count);
1095                 __count_vm_events(PGDEACTIVATE, nr_active);
1096
1097                 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1098                                                 -count[LRU_ACTIVE_FILE]);
1099                 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1100                                                 -count[LRU_INACTIVE_FILE]);
1101                 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1102                                                 -count[LRU_ACTIVE_ANON]);
1103                 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1104                                                 -count[LRU_INACTIVE_ANON]);
1105
1106                 if (scanning_global_lru(sc))
1107                         zone->pages_scanned += nr_scan;
1108
1109                 reclaim_stat->recent_scanned[0] += count[LRU_INACTIVE_ANON];
1110                 reclaim_stat->recent_scanned[0] += count[LRU_ACTIVE_ANON];
1111                 reclaim_stat->recent_scanned[1] += count[LRU_INACTIVE_FILE];
1112                 reclaim_stat->recent_scanned[1] += count[LRU_ACTIVE_FILE];
1113
1114                 spin_unlock_irq(&zone->lru_lock);
1115
1116                 nr_scanned += nr_scan;
1117                 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
1118
1119                 /*
1120                  * If we are direct reclaiming for contiguous pages and we do
1121                  * not reclaim everything in the list, try again and wait
1122                  * for IO to complete. This will stall high-order allocations
1123                  * but that should be acceptable to the caller
1124                  */
1125                 if (nr_freed < nr_taken && !current_is_kswapd() &&
1126                     lumpy_reclaim) {
1127                         congestion_wait(WRITE, HZ/10);
1128
1129                         /*
1130                          * The attempt at page out may have made some
1131                          * of the pages active, mark them inactive again.
1132                          */
1133                         nr_active = clear_active_flags(&page_list, count);
1134                         count_vm_events(PGDEACTIVATE, nr_active);
1135
1136                         nr_freed += shrink_page_list(&page_list, sc,
1137                                                         PAGEOUT_IO_SYNC);
1138                 }
1139
1140                 nr_reclaimed += nr_freed;
1141                 local_irq_disable();
1142                 if (current_is_kswapd()) {
1143                         __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scan);
1144                         __count_vm_events(KSWAPD_STEAL, nr_freed);
1145                 } else if (scanning_global_lru(sc))
1146                         __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scan);
1147
1148                 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
1149
1150                 if (nr_taken == 0)
1151                         goto done;
1152
1153                 spin_lock(&zone->lru_lock);
1154                 /*
1155                  * Put back any unfreeable pages.
1156                  */
1157                 while (!list_empty(&page_list)) {
1158                         int lru;
1159                         page = lru_to_page(&page_list);
1160                         VM_BUG_ON(PageLRU(page));
1161                         list_del(&page->lru);
1162                         if (unlikely(!page_evictable(page, NULL))) {
1163                                 spin_unlock_irq(&zone->lru_lock);
1164                                 putback_lru_page(page);
1165                                 spin_lock_irq(&zone->lru_lock);
1166                                 continue;
1167                         }
1168                         SetPageLRU(page);
1169                         lru = page_lru(page);
1170                         add_page_to_lru_list(zone, page, lru);
1171                         if (PageActive(page)) {
1172                                 int file = !!page_is_file_cache(page);
1173                                 reclaim_stat->recent_rotated[file]++;
1174                         }
1175                         if (!pagevec_add(&pvec, page)) {
1176                                 spin_unlock_irq(&zone->lru_lock);
1177                                 __pagevec_release(&pvec);
1178                                 spin_lock_irq(&zone->lru_lock);
1179                         }
1180                 }
1181         } while (nr_scanned < max_scan);
1182         spin_unlock(&zone->lru_lock);
1183 done:
1184         local_irq_enable();
1185         pagevec_release(&pvec);
1186         return nr_reclaimed;
1187 }
1188
1189 /*
1190  * We are about to scan this zone at a certain priority level.  If that priority
1191  * level is smaller (ie: more urgent) than the previous priority, then note
1192  * that priority level within the zone.  This is done so that when the next
1193  * process comes in to scan this zone, it will immediately start out at this
1194  * priority level rather than having to build up its own scanning priority.
1195  * Here, this priority affects only the reclaim-mapped threshold.
1196  */
1197 static inline void note_zone_scanning_priority(struct zone *zone, int priority)
1198 {
1199         if (priority < zone->prev_priority)
1200                 zone->prev_priority = priority;
1201 }
1202
1203 /*
1204  * This moves pages from the active list to the inactive list.
1205  *
1206  * We move them the other way if the page is referenced by one or more
1207  * processes, from rmap.
1208  *
1209  * If the pages are mostly unmapped, the processing is fast and it is
1210  * appropriate to hold zone->lru_lock across the whole operation.  But if
1211  * the pages are mapped, the processing is slow (page_referenced()) so we
1212  * should drop zone->lru_lock around each page.  It's impossible to balance
1213  * this, so instead we remove the pages from the LRU while processing them.
1214  * It is safe to rely on PG_active against the non-LRU pages in here because
1215  * nobody will play with that bit on a non-LRU page.
1216  *
1217  * The downside is that we have to touch page->_count against each page.
1218  * But we had to alter page->flags anyway.
1219  */
1220
1221
1222 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1223                         struct scan_control *sc, int priority, int file)
1224 {
1225         unsigned long pgmoved;
1226         int pgdeactivate = 0;
1227         unsigned long pgscanned;
1228         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1229         LIST_HEAD(l_inactive);
1230         struct page *page;
1231         struct pagevec pvec;
1232         enum lru_list lru;
1233         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1234
1235         lru_add_drain();
1236         spin_lock_irq(&zone->lru_lock);
1237         pgmoved = sc->isolate_pages(nr_pages, &l_hold, &pgscanned, sc->order,
1238                                         ISOLATE_ACTIVE, zone,
1239                                         sc->mem_cgroup, 1, file);
1240         /*
1241          * zone->pages_scanned is used for detect zone's oom
1242          * mem_cgroup remembers nr_scan by itself.
1243          */
1244         if (scanning_global_lru(sc)) {
1245                 zone->pages_scanned += pgscanned;
1246         }
1247         reclaim_stat->recent_scanned[!!file] += pgmoved;
1248
1249         if (file)
1250                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -pgmoved);
1251         else
1252                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -pgmoved);
1253         spin_unlock_irq(&zone->lru_lock);
1254
1255         pgmoved = 0;
1256         while (!list_empty(&l_hold)) {
1257                 cond_resched();
1258                 page = lru_to_page(&l_hold);
1259                 list_del(&page->lru);
1260
1261                 if (unlikely(!page_evictable(page, NULL))) {
1262                         putback_lru_page(page);
1263                         continue;
1264                 }
1265
1266                 /* page_referenced clears PageReferenced */
1267                 if (page_mapping_inuse(page) &&
1268                     page_referenced(page, 0, sc->mem_cgroup))
1269                         pgmoved++;
1270
1271                 list_add(&page->lru, &l_inactive);
1272         }
1273
1274         /*
1275          * Move the pages to the [file or anon] inactive list.
1276          */
1277         pagevec_init(&pvec, 1);
1278         lru = LRU_BASE + file * LRU_FILE;
1279
1280         spin_lock_irq(&zone->lru_lock);
1281         /*
1282          * Count referenced pages from currently used mappings as
1283          * rotated, even though they are moved to the inactive list.
1284          * This helps balance scan pressure between file and anonymous
1285          * pages in get_scan_ratio.
1286          */
1287         reclaim_stat->recent_rotated[!!file] += pgmoved;
1288
1289         pgmoved = 0;
1290         while (!list_empty(&l_inactive)) {
1291                 page = lru_to_page(&l_inactive);
1292                 prefetchw_prev_lru_page(page, &l_inactive, flags);
1293                 VM_BUG_ON(PageLRU(page));
1294                 SetPageLRU(page);
1295                 VM_BUG_ON(!PageActive(page));
1296                 ClearPageActive(page);
1297
1298                 list_move(&page->lru, &zone->lru[lru].list);
1299                 mem_cgroup_add_lru_list(page, lru);
1300                 pgmoved++;
1301                 if (!pagevec_add(&pvec, page)) {
1302                         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1303                         spin_unlock_irq(&zone->lru_lock);
1304                         pgdeactivate += pgmoved;
1305                         pgmoved = 0;
1306                         if (buffer_heads_over_limit)
1307                                 pagevec_strip(&pvec);
1308                         __pagevec_release(&pvec);
1309                         spin_lock_irq(&zone->lru_lock);
1310                 }
1311         }
1312         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1313         pgdeactivate += pgmoved;
1314         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1315         __count_vm_events(PGDEACTIVATE, pgdeactivate);
1316         spin_unlock_irq(&zone->lru_lock);
1317         if (buffer_heads_over_limit)
1318                 pagevec_strip(&pvec);
1319         pagevec_release(&pvec);
1320 }
1321
1322 static int inactive_anon_is_low_global(struct zone *zone)
1323 {
1324         unsigned long active, inactive;
1325
1326         active = zone_page_state(zone, NR_ACTIVE_ANON);
1327         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1328
1329         if (inactive * zone->inactive_ratio < active)
1330                 return 1;
1331
1332         return 0;
1333 }
1334
1335 /**
1336  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1337  * @zone: zone to check
1338  * @sc:   scan control of this context
1339  *
1340  * Returns true if the zone does not have enough inactive anon pages,
1341  * meaning some active anon pages need to be deactivated.
1342  */
1343 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1344 {
1345         int low;
1346
1347         if (scanning_global_lru(sc))
1348                 low = inactive_anon_is_low_global(zone);
1349         else
1350                 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1351         return low;
1352 }
1353
1354 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1355         struct zone *zone, struct scan_control *sc, int priority)
1356 {
1357         int file = is_file_lru(lru);
1358
1359         if (lru == LRU_ACTIVE_FILE) {
1360                 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1361                 return 0;
1362         }
1363
1364         if (lru == LRU_ACTIVE_ANON && inactive_anon_is_low(zone, sc)) {
1365                 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1366                 return 0;
1367         }
1368         return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1369 }
1370
1371 /*
1372  * Determine how aggressively the anon and file LRU lists should be
1373  * scanned.  The relative value of each set of LRU lists is determined
1374  * by looking at the fraction of the pages scanned we did rotate back
1375  * onto the active list instead of evict.
1376  *
1377  * percent[0] specifies how much pressure to put on ram/swap backed
1378  * memory, while percent[1] determines pressure on the file LRUs.
1379  */
1380 static void get_scan_ratio(struct zone *zone, struct scan_control *sc,
1381                                         unsigned long *percent)
1382 {
1383         unsigned long anon, file, free;
1384         unsigned long anon_prio, file_prio;
1385         unsigned long ap, fp;
1386         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1387
1388         /* If we have no swap space, do not bother scanning anon pages. */
1389         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1390                 percent[0] = 0;
1391                 percent[1] = 100;
1392                 return;
1393         }
1394
1395         anon  = zone_nr_pages(zone, sc, LRU_ACTIVE_ANON) +
1396                 zone_nr_pages(zone, sc, LRU_INACTIVE_ANON);
1397         file  = zone_nr_pages(zone, sc, LRU_ACTIVE_FILE) +
1398                 zone_nr_pages(zone, sc, LRU_INACTIVE_FILE);
1399
1400         if (scanning_global_lru(sc)) {
1401                 free  = zone_page_state(zone, NR_FREE_PAGES);
1402                 /* If we have very few page cache pages,
1403                    force-scan anon pages. */
1404                 if (unlikely(file + free <= zone->pages_high)) {
1405                         percent[0] = 100;
1406                         percent[1] = 0;
1407                         return;
1408                 }
1409         }
1410
1411         /*
1412          * OK, so we have swap space and a fair amount of page cache
1413          * pages.  We use the recently rotated / recently scanned
1414          * ratios to determine how valuable each cache is.
1415          *
1416          * Because workloads change over time (and to avoid overflow)
1417          * we keep these statistics as a floating average, which ends
1418          * up weighing recent references more than old ones.
1419          *
1420          * anon in [0], file in [1]
1421          */
1422         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1423                 spin_lock_irq(&zone->lru_lock);
1424                 reclaim_stat->recent_scanned[0] /= 2;
1425                 reclaim_stat->recent_rotated[0] /= 2;
1426                 spin_unlock_irq(&zone->lru_lock);
1427         }
1428
1429         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1430                 spin_lock_irq(&zone->lru_lock);
1431                 reclaim_stat->recent_scanned[1] /= 2;
1432                 reclaim_stat->recent_rotated[1] /= 2;
1433                 spin_unlock_irq(&zone->lru_lock);
1434         }
1435
1436         /*
1437          * With swappiness at 100, anonymous and file have the same priority.
1438          * This scanning priority is essentially the inverse of IO cost.
1439          */
1440         anon_prio = sc->swappiness;
1441         file_prio = 200 - sc->swappiness;
1442
1443         /*
1444          * The amount of pressure on anon vs file pages is inversely
1445          * proportional to the fraction of recently scanned pages on
1446          * each list that were recently referenced and in active use.
1447          */
1448         ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1449         ap /= reclaim_stat->recent_rotated[0] + 1;
1450
1451         fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1452         fp /= reclaim_stat->recent_rotated[1] + 1;
1453
1454         /* Normalize to percentages */
1455         percent[0] = 100 * ap / (ap + fp + 1);
1456         percent[1] = 100 - percent[0];
1457 }
1458
1459
1460 /*
1461  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1462  */
1463 static void shrink_zone(int priority, struct zone *zone,
1464                                 struct scan_control *sc)
1465 {
1466         unsigned long nr[NR_LRU_LISTS];
1467         unsigned long nr_to_scan;
1468         unsigned long percent[2];       /* anon @ 0; file @ 1 */
1469         enum lru_list l;
1470         unsigned long nr_reclaimed = sc->nr_reclaimed;
1471         unsigned long swap_cluster_max = sc->swap_cluster_max;
1472
1473         get_scan_ratio(zone, sc, percent);
1474
1475         for_each_evictable_lru(l) {
1476                 int file = is_file_lru(l);
1477                 unsigned long scan;
1478
1479                 scan = zone_nr_pages(zone, sc, l);
1480                 if (priority) {
1481                         scan >>= priority;
1482                         scan = (scan * percent[file]) / 100;
1483                 }
1484                 if (scanning_global_lru(sc)) {
1485                         zone->lru[l].nr_scan += scan;
1486                         nr[l] = zone->lru[l].nr_scan;
1487                         if (nr[l] >= swap_cluster_max)
1488                                 zone->lru[l].nr_scan = 0;
1489                         else
1490                                 nr[l] = 0;
1491                 } else
1492                         nr[l] = scan;
1493         }
1494
1495         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1496                                         nr[LRU_INACTIVE_FILE]) {
1497                 for_each_evictable_lru(l) {
1498                         if (nr[l]) {
1499                                 nr_to_scan = min(nr[l], swap_cluster_max);
1500                                 nr[l] -= nr_to_scan;
1501
1502                                 nr_reclaimed += shrink_list(l, nr_to_scan,
1503                                                             zone, sc, priority);
1504                         }
1505                 }
1506                 /*
1507                  * On large memory systems, scan >> priority can become
1508                  * really large. This is fine for the starting priority;
1509                  * we want to put equal scanning pressure on each zone.
1510                  * However, if the VM has a harder time of freeing pages,
1511                  * with multiple processes reclaiming pages, the total
1512                  * freeing target can get unreasonably large.
1513                  */
1514                 if (nr_reclaimed > swap_cluster_max &&
1515                         priority < DEF_PRIORITY && !current_is_kswapd())
1516                         break;
1517         }
1518
1519         sc->nr_reclaimed = nr_reclaimed;
1520
1521         /*
1522          * Even if we did not try to evict anon pages at all, we want to
1523          * rebalance the anon lru active/inactive ratio.
1524          */
1525         if (inactive_anon_is_low(zone, sc))
1526                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1527
1528         throttle_vm_writeout(sc->gfp_mask);
1529 }
1530
1531 /*
1532  * This is the direct reclaim path, for page-allocating processes.  We only
1533  * try to reclaim pages from zones which will satisfy the caller's allocation
1534  * request.
1535  *
1536  * We reclaim from a zone even if that zone is over pages_high.  Because:
1537  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1538  *    allocation or
1539  * b) The zones may be over pages_high but they must go *over* pages_high to
1540  *    satisfy the `incremental min' zone defense algorithm.
1541  *
1542  * If a zone is deemed to be full of pinned pages then just give it a light
1543  * scan then give up on it.
1544  */
1545 static void shrink_zones(int priority, struct zonelist *zonelist,
1546                                         struct scan_control *sc)
1547 {
1548         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1549         struct zoneref *z;
1550         struct zone *zone;
1551
1552         sc->all_unreclaimable = 1;
1553         for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1554                                         sc->nodemask) {
1555                 if (!populated_zone(zone))
1556                         continue;
1557                 /*
1558                  * Take care memory controller reclaiming has small influence
1559                  * to global LRU.
1560                  */
1561                 if (scanning_global_lru(sc)) {
1562                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1563                                 continue;
1564                         note_zone_scanning_priority(zone, priority);
1565
1566                         if (zone_is_all_unreclaimable(zone) &&
1567                                                 priority != DEF_PRIORITY)
1568                                 continue;       /* Let kswapd poll it */
1569                         sc->all_unreclaimable = 0;
1570                 } else {
1571                         /*
1572                          * Ignore cpuset limitation here. We just want to reduce
1573                          * # of used pages by us regardless of memory shortage.
1574                          */
1575                         sc->all_unreclaimable = 0;
1576                         mem_cgroup_note_reclaim_priority(sc->mem_cgroup,
1577                                                         priority);
1578                 }
1579
1580                 shrink_zone(priority, zone, sc);
1581         }
1582 }
1583
1584 /*
1585  * This is the main entry point to direct page reclaim.
1586  *
1587  * If a full scan of the inactive list fails to free enough memory then we
1588  * are "out of memory" and something needs to be killed.
1589  *
1590  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1591  * high - the zone may be full of dirty or under-writeback pages, which this
1592  * caller can't do much about.  We kick pdflush and take explicit naps in the
1593  * hope that some of these pages can be written.  But if the allocating task
1594  * holds filesystem locks which prevent writeout this might not work, and the
1595  * allocation attempt will fail.
1596  *
1597  * returns:     0, if no pages reclaimed
1598  *              else, the number of pages reclaimed
1599  */
1600 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
1601                                         struct scan_control *sc)
1602 {
1603         int priority;
1604         unsigned long ret = 0;
1605         unsigned long total_scanned = 0;
1606         struct reclaim_state *reclaim_state = current->reclaim_state;
1607         unsigned long lru_pages = 0;
1608         struct zoneref *z;
1609         struct zone *zone;
1610         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1611
1612         delayacct_freepages_start();
1613
1614         if (scanning_global_lru(sc))
1615                 count_vm_event(ALLOCSTALL);
1616         /*
1617          * mem_cgroup will not do shrink_slab.
1618          */
1619         if (scanning_global_lru(sc)) {
1620                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1621
1622                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1623                                 continue;
1624
1625                         lru_pages += zone_lru_pages(zone);
1626                 }
1627         }
1628
1629         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1630                 sc->nr_scanned = 0;
1631                 if (!priority)
1632                         disable_swap_token();
1633                 shrink_zones(priority, zonelist, sc);
1634                 /*
1635                  * Don't shrink slabs when reclaiming memory from
1636                  * over limit cgroups
1637                  */
1638                 if (scanning_global_lru(sc)) {
1639                         shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
1640                         if (reclaim_state) {
1641                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
1642                                 reclaim_state->reclaimed_slab = 0;
1643                         }
1644                 }
1645                 total_scanned += sc->nr_scanned;
1646                 if (sc->nr_reclaimed >= sc->swap_cluster_max) {
1647                         ret = sc->nr_reclaimed;
1648                         goto out;
1649                 }
1650
1651                 /*
1652                  * Try to write back as many pages as we just scanned.  This
1653                  * tends to cause slow streaming writers to write data to the
1654                  * disk smoothly, at the dirtying rate, which is nice.   But
1655                  * that's undesirable in laptop mode, where we *want* lumpy
1656                  * writeout.  So in laptop mode, write out the whole world.
1657                  */
1658                 if (total_scanned > sc->swap_cluster_max +
1659                                         sc->swap_cluster_max / 2) {
1660                         wakeup_pdflush(laptop_mode ? 0 : total_scanned);
1661                         sc->may_writepage = 1;
1662                 }
1663
1664                 /* Take a nap, wait for some writeback to complete */
1665                 if (sc->nr_scanned && priority < DEF_PRIORITY - 2)
1666                         congestion_wait(WRITE, HZ/10);
1667         }
1668         /* top priority shrink_zones still had more to do? don't OOM, then */
1669         if (!sc->all_unreclaimable && scanning_global_lru(sc))
1670                 ret = sc->nr_reclaimed;
1671 out:
1672         /*
1673          * Now that we've scanned all the zones at this priority level, note
1674          * that level within the zone so that the next thread which performs
1675          * scanning of this zone will immediately start out at this priority
1676          * level.  This affects only the decision whether or not to bring
1677          * mapped pages onto the inactive list.
1678          */
1679         if (priority < 0)
1680                 priority = 0;
1681
1682         if (scanning_global_lru(sc)) {
1683                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1684
1685                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1686                                 continue;
1687
1688                         zone->prev_priority = priority;
1689                 }
1690         } else
1691                 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority);
1692
1693         delayacct_freepages_end();
1694
1695         return ret;
1696 }
1697
1698 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
1699                                 gfp_t gfp_mask, nodemask_t *nodemask)
1700 {
1701         struct scan_control sc = {
1702                 .gfp_mask = gfp_mask,
1703                 .may_writepage = !laptop_mode,
1704                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1705                 .may_unmap = 1,
1706                 .may_swap = 1,
1707                 .swappiness = vm_swappiness,
1708                 .order = order,
1709                 .mem_cgroup = NULL,
1710                 .isolate_pages = isolate_pages_global,
1711                 .nodemask = nodemask,
1712         };
1713
1714         return do_try_to_free_pages(zonelist, &sc);
1715 }
1716
1717 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
1718
1719 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
1720                                            gfp_t gfp_mask,
1721                                            bool noswap,
1722                                            unsigned int swappiness)
1723 {
1724         struct scan_control sc = {
1725                 .may_writepage = !laptop_mode,
1726                 .may_unmap = 1,
1727                 .may_swap = !noswap,
1728                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1729                 .swappiness = swappiness,
1730                 .order = 0,
1731                 .mem_cgroup = mem_cont,
1732                 .isolate_pages = mem_cgroup_isolate_pages,
1733                 .nodemask = NULL, /* we don't care the placement */
1734         };
1735         struct zonelist *zonelist;
1736
1737         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1738                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1739         zonelist = NODE_DATA(numa_node_id())->node_zonelists;
1740         return do_try_to_free_pages(zonelist, &sc);
1741 }
1742 #endif
1743
1744 /*
1745  * For kswapd, balance_pgdat() will work across all this node's zones until
1746  * they are all at pages_high.
1747  *
1748  * Returns the number of pages which were actually freed.
1749  *
1750  * There is special handling here for zones which are full of pinned pages.
1751  * This can happen if the pages are all mlocked, or if they are all used by
1752  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1753  * What we do is to detect the case where all pages in the zone have been
1754  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1755  * dead and from now on, only perform a short scan.  Basically we're polling
1756  * the zone for when the problem goes away.
1757  *
1758  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
1759  * zones which have free_pages > pages_high, but once a zone is found to have
1760  * free_pages <= pages_high, we scan that zone and the lower zones regardless
1761  * of the number of free pages in the lower zones.  This interoperates with
1762  * the page allocator fallback scheme to ensure that aging of pages is balanced
1763  * across the zones.
1764  */
1765 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
1766 {
1767         int all_zones_ok;
1768         int priority;
1769         int i;
1770         unsigned long total_scanned;
1771         struct reclaim_state *reclaim_state = current->reclaim_state;
1772         struct scan_control sc = {
1773                 .gfp_mask = GFP_KERNEL,
1774                 .may_unmap = 1,
1775                 .may_swap = 1,
1776                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1777                 .swappiness = vm_swappiness,
1778                 .order = order,
1779                 .mem_cgroup = NULL,
1780                 .isolate_pages = isolate_pages_global,
1781         };
1782         /*
1783          * temp_priority is used to remember the scanning priority at which
1784          * this zone was successfully refilled to free_pages == pages_high.
1785          */
1786         int temp_priority[MAX_NR_ZONES];
1787
1788 loop_again:
1789         total_scanned = 0;
1790         sc.nr_reclaimed = 0;
1791         sc.may_writepage = !laptop_mode;
1792         count_vm_event(PAGEOUTRUN);
1793
1794         for (i = 0; i < pgdat->nr_zones; i++)
1795                 temp_priority[i] = DEF_PRIORITY;
1796
1797         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1798                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1799                 unsigned long lru_pages = 0;
1800
1801                 /* The swap token gets in the way of swapout... */
1802                 if (!priority)
1803                         disable_swap_token();
1804
1805                 all_zones_ok = 1;
1806
1807                 /*
1808                  * Scan in the highmem->dma direction for the highest
1809                  * zone which needs scanning
1810                  */
1811                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1812                         struct zone *zone = pgdat->node_zones + i;
1813
1814                         if (!populated_zone(zone))
1815                                 continue;
1816
1817                         if (zone_is_all_unreclaimable(zone) &&
1818                             priority != DEF_PRIORITY)
1819                                 continue;
1820
1821                         /*
1822                          * Do some background aging of the anon list, to give
1823                          * pages a chance to be referenced before reclaiming.
1824                          */
1825                         if (inactive_anon_is_low(zone, &sc))
1826                                 shrink_active_list(SWAP_CLUSTER_MAX, zone,
1827                                                         &sc, priority, 0);
1828
1829                         if (!zone_watermark_ok(zone, order, zone->pages_high,
1830                                                0, 0)) {
1831                                 end_zone = i;
1832                                 break;
1833                         }
1834                 }
1835                 if (i < 0)
1836                         goto out;
1837
1838                 for (i = 0; i <= end_zone; i++) {
1839                         struct zone *zone = pgdat->node_zones + i;
1840
1841                         lru_pages += zone_lru_pages(zone);
1842                 }
1843
1844                 /*
1845                  * Now scan the zone in the dma->highmem direction, stopping
1846                  * at the last zone which needs scanning.
1847                  *
1848                  * We do this because the page allocator works in the opposite
1849                  * direction.  This prevents the page allocator from allocating
1850                  * pages behind kswapd's direction of progress, which would
1851                  * cause too much scanning of the lower zones.
1852                  */
1853                 for (i = 0; i <= end_zone; i++) {
1854                         struct zone *zone = pgdat->node_zones + i;
1855                         int nr_slab;
1856
1857                         if (!populated_zone(zone))
1858                                 continue;
1859
1860                         if (zone_is_all_unreclaimable(zone) &&
1861                                         priority != DEF_PRIORITY)
1862                                 continue;
1863
1864                         if (!zone_watermark_ok(zone, order, zone->pages_high,
1865                                                end_zone, 0))
1866                                 all_zones_ok = 0;
1867                         temp_priority[i] = priority;
1868                         sc.nr_scanned = 0;
1869                         note_zone_scanning_priority(zone, priority);
1870                         /*
1871                          * We put equal pressure on every zone, unless one
1872                          * zone has way too many pages free already.
1873                          */
1874                         if (!zone_watermark_ok(zone, order, 8*zone->pages_high,
1875                                                 end_zone, 0))
1876                                 shrink_zone(priority, zone, &sc);
1877                         reclaim_state->reclaimed_slab = 0;
1878                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1879                                                 lru_pages);
1880                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1881                         total_scanned += sc.nr_scanned;
1882                         if (zone_is_all_unreclaimable(zone))
1883                                 continue;
1884                         if (nr_slab == 0 && zone->pages_scanned >=
1885                                                 (zone_lru_pages(zone) * 6))
1886                                         zone_set_flag(zone,
1887                                                       ZONE_ALL_UNRECLAIMABLE);
1888                         /*
1889                          * If we've done a decent amount of scanning and
1890                          * the reclaim ratio is low, start doing writepage
1891                          * even in laptop mode
1892                          */
1893                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1894                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
1895                                 sc.may_writepage = 1;
1896                 }
1897                 if (all_zones_ok)
1898                         break;          /* kswapd: all done */
1899                 /*
1900                  * OK, kswapd is getting into trouble.  Take a nap, then take
1901                  * another pass across the zones.
1902                  */
1903                 if (total_scanned && priority < DEF_PRIORITY - 2)
1904                         congestion_wait(WRITE, HZ/10);
1905
1906                 /*
1907                  * We do this so kswapd doesn't build up large priorities for
1908                  * example when it is freeing in parallel with allocators. It
1909                  * matches the direct reclaim path behaviour in terms of impact
1910                  * on zone->*_priority.
1911                  */
1912                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
1913                         break;
1914         }
1915 out:
1916         /*
1917          * Note within each zone the priority level at which this zone was
1918          * brought into a happy state.  So that the next thread which scans this
1919          * zone will start out at that priority level.
1920          */
1921         for (i = 0; i < pgdat->nr_zones; i++) {
1922                 struct zone *zone = pgdat->node_zones + i;
1923
1924                 zone->prev_priority = temp_priority[i];
1925         }
1926         if (!all_zones_ok) {
1927                 cond_resched();
1928
1929                 try_to_freeze();
1930
1931                 /*
1932                  * Fragmentation may mean that the system cannot be
1933                  * rebalanced for high-order allocations in all zones.
1934                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
1935                  * it means the zones have been fully scanned and are still
1936                  * not balanced. For high-order allocations, there is
1937                  * little point trying all over again as kswapd may
1938                  * infinite loop.
1939                  *
1940                  * Instead, recheck all watermarks at order-0 as they
1941                  * are the most important. If watermarks are ok, kswapd will go
1942                  * back to sleep. High-order users can still perform direct
1943                  * reclaim if they wish.
1944                  */
1945                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
1946                         order = sc.order = 0;
1947
1948                 goto loop_again;
1949         }
1950
1951         return sc.nr_reclaimed;
1952 }
1953
1954 /*
1955  * The background pageout daemon, started as a kernel thread
1956  * from the init process.
1957  *
1958  * This basically trickles out pages so that we have _some_
1959  * free memory available even if there is no other activity
1960  * that frees anything up. This is needed for things like routing
1961  * etc, where we otherwise might have all activity going on in
1962  * asynchronous contexts that cannot page things out.
1963  *
1964  * If there are applications that are active memory-allocators
1965  * (most normal use), this basically shouldn't matter.
1966  */
1967 static int kswapd(void *p)
1968 {
1969         unsigned long order;
1970         pg_data_t *pgdat = (pg_data_t*)p;
1971         struct task_struct *tsk = current;
1972         DEFINE_WAIT(wait);
1973         struct reclaim_state reclaim_state = {
1974                 .reclaimed_slab = 0,
1975         };
1976         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1977
1978         lockdep_set_current_reclaim_state(GFP_KERNEL);
1979
1980         if (!cpumask_empty(cpumask))
1981                 set_cpus_allowed_ptr(tsk, cpumask);
1982         current->reclaim_state = &reclaim_state;
1983
1984         /*
1985          * Tell the memory management that we're a "memory allocator",
1986          * and that if we need more memory we should get access to it
1987          * regardless (see "__alloc_pages()"). "kswapd" should
1988          * never get caught in the normal page freeing logic.
1989          *
1990          * (Kswapd normally doesn't need memory anyway, but sometimes
1991          * you need a small amount of memory in order to be able to
1992          * page out something else, and this flag essentially protects
1993          * us from recursively trying to free more memory as we're
1994          * trying to free the first piece of memory in the first place).
1995          */
1996         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
1997         set_freezable();
1998
1999         order = 0;
2000         for ( ; ; ) {
2001                 unsigned long new_order;
2002
2003                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2004                 new_order = pgdat->kswapd_max_order;
2005                 pgdat->kswapd_max_order = 0;
2006                 if (order < new_order) {
2007                         /*
2008                          * Don't sleep if someone wants a larger 'order'
2009                          * allocation
2010                          */
2011                         order = new_order;
2012                 } else {
2013                         if (!freezing(current))
2014                                 schedule();
2015
2016                         order = pgdat->kswapd_max_order;
2017                 }
2018                 finish_wait(&pgdat->kswapd_wait, &wait);
2019
2020                 if (!try_to_freeze()) {
2021                         /* We can speed up thawing tasks if we don't call
2022                          * balance_pgdat after returning from the refrigerator
2023                          */
2024                         balance_pgdat(pgdat, order);
2025                 }
2026         }
2027         return 0;
2028 }
2029
2030 /*
2031  * A zone is low on free memory, so wake its kswapd task to service it.
2032  */
2033 void wakeup_kswapd(struct zone *zone, int order)
2034 {
2035         pg_data_t *pgdat;
2036
2037         if (!populated_zone(zone))
2038                 return;
2039
2040         pgdat = zone->zone_pgdat;
2041         if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0))
2042                 return;
2043         if (pgdat->kswapd_max_order < order)
2044                 pgdat->kswapd_max_order = order;
2045         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2046                 return;
2047         if (!waitqueue_active(&pgdat->kswapd_wait))
2048                 return;
2049         wake_up_interruptible(&pgdat->kswapd_wait);
2050 }
2051
2052 unsigned long global_lru_pages(void)
2053 {
2054         return global_page_state(NR_ACTIVE_ANON)
2055                 + global_page_state(NR_ACTIVE_FILE)
2056                 + global_page_state(NR_INACTIVE_ANON)
2057                 + global_page_state(NR_INACTIVE_FILE);
2058 }
2059
2060 #ifdef CONFIG_HIBERNATION
2061 /*
2062  * Helper function for shrink_all_memory().  Tries to reclaim 'nr_pages' pages
2063  * from LRU lists system-wide, for given pass and priority.
2064  *
2065  * For pass > 3 we also try to shrink the LRU lists that contain a few pages
2066  */
2067 static void shrink_all_zones(unsigned long nr_pages, int prio,
2068                                       int pass, struct scan_control *sc)
2069 {
2070         struct zone *zone;
2071         unsigned long nr_reclaimed = 0;
2072
2073         for_each_populated_zone(zone) {
2074                 enum lru_list l;
2075
2076                 if (zone_is_all_unreclaimable(zone) && prio != DEF_PRIORITY)
2077                         continue;
2078
2079                 for_each_evictable_lru(l) {
2080                         enum zone_stat_item ls = NR_LRU_BASE + l;
2081                         unsigned long lru_pages = zone_page_state(zone, ls);
2082
2083                         /* For pass = 0, we don't shrink the active list */
2084                         if (pass == 0 && (l == LRU_ACTIVE_ANON ||
2085                                                 l == LRU_ACTIVE_FILE))
2086                                 continue;
2087
2088                         zone->lru[l].nr_scan += (lru_pages >> prio) + 1;
2089                         if (zone->lru[l].nr_scan >= nr_pages || pass > 3) {
2090                                 unsigned long nr_to_scan;
2091
2092                                 zone->lru[l].nr_scan = 0;
2093                                 nr_to_scan = min(nr_pages, lru_pages);
2094                                 nr_reclaimed += shrink_list(l, nr_to_scan, zone,
2095                                                                 sc, prio);
2096                                 if (nr_reclaimed >= nr_pages) {
2097                                         sc->nr_reclaimed += nr_reclaimed;
2098                                         return;
2099                                 }
2100                         }
2101                 }
2102         }
2103         sc->nr_reclaimed += nr_reclaimed;
2104 }
2105
2106 /*
2107  * Try to free `nr_pages' of memory, system-wide, and return the number of
2108  * freed pages.
2109  *
2110  * Rather than trying to age LRUs the aim is to preserve the overall
2111  * LRU order by reclaiming preferentially
2112  * inactive > active > active referenced > active mapped
2113  */
2114 unsigned long shrink_all_memory(unsigned long nr_pages)
2115 {
2116         unsigned long lru_pages, nr_slab;
2117         int pass;
2118         struct reclaim_state reclaim_state;
2119         struct scan_control sc = {
2120                 .gfp_mask = GFP_KERNEL,
2121                 .may_unmap = 0,
2122                 .may_writepage = 1,
2123                 .isolate_pages = isolate_pages_global,
2124                 .nr_reclaimed = 0,
2125         };
2126
2127         current->reclaim_state = &reclaim_state;
2128
2129         lru_pages = global_lru_pages();
2130         nr_slab = global_page_state(NR_SLAB_RECLAIMABLE);
2131         /* If slab caches are huge, it's better to hit them first */
2132         while (nr_slab >= lru_pages) {
2133                 reclaim_state.reclaimed_slab = 0;
2134                 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
2135                 if (!reclaim_state.reclaimed_slab)
2136                         break;
2137
2138                 sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2139                 if (sc.nr_reclaimed >= nr_pages)
2140                         goto out;
2141
2142                 nr_slab -= reclaim_state.reclaimed_slab;
2143         }
2144
2145         /*
2146          * We try to shrink LRUs in 5 passes:
2147          * 0 = Reclaim from inactive_list only
2148          * 1 = Reclaim from active list but don't reclaim mapped
2149          * 2 = 2nd pass of type 1
2150          * 3 = Reclaim mapped (normal reclaim)
2151          * 4 = 2nd pass of type 3
2152          */
2153         for (pass = 0; pass < 5; pass++) {
2154                 int prio;
2155
2156                 /* Force reclaiming mapped pages in the passes #3 and #4 */
2157                 if (pass > 2)
2158                         sc.may_unmap = 1;
2159
2160                 for (prio = DEF_PRIORITY; prio >= 0; prio--) {
2161                         unsigned long nr_to_scan = nr_pages - sc.nr_reclaimed;
2162
2163                         sc.nr_scanned = 0;
2164                         sc.swap_cluster_max = nr_to_scan;
2165                         shrink_all_zones(nr_to_scan, prio, pass, &sc);
2166                         if (sc.nr_reclaimed >= nr_pages)
2167                                 goto out;
2168
2169                         reclaim_state.reclaimed_slab = 0;
2170                         shrink_slab(sc.nr_scanned, sc.gfp_mask,
2171                                         global_lru_pages());
2172                         sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2173                         if (sc.nr_reclaimed >= nr_pages)
2174                                 goto out;
2175
2176                         if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
2177                                 congestion_wait(WRITE, HZ / 10);
2178                 }
2179         }
2180
2181         /*
2182          * If sc.nr_reclaimed = 0, we could not shrink LRUs, but there may be
2183          * something in slab caches
2184          */
2185         if (!sc.nr_reclaimed) {
2186                 do {
2187                         reclaim_state.reclaimed_slab = 0;
2188                         shrink_slab(nr_pages, sc.gfp_mask, global_lru_pages());
2189                         sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2190                 } while (sc.nr_reclaimed < nr_pages &&
2191                                 reclaim_state.reclaimed_slab > 0);
2192         }
2193
2194
2195 out:
2196         current->reclaim_state = NULL;
2197
2198         return sc.nr_reclaimed;
2199 }
2200 #endif /* CONFIG_HIBERNATION */
2201
2202 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2203    not required for correctness.  So if the last cpu in a node goes
2204    away, we get changed to run anywhere: as the first one comes back,
2205    restore their cpu bindings. */
2206 static int __devinit cpu_callback(struct notifier_block *nfb,
2207                                   unsigned long action, void *hcpu)
2208 {
2209         int nid;
2210
2211         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2212                 for_each_node_state(nid, N_HIGH_MEMORY) {
2213                         pg_data_t *pgdat = NODE_DATA(nid);
2214                         const struct cpumask *mask;
2215
2216                         mask = cpumask_of_node(pgdat->node_id);
2217
2218                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2219                                 /* One of our CPUs online: restore mask */
2220                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2221                 }
2222         }
2223         return NOTIFY_OK;
2224 }
2225
2226 /*
2227  * This kswapd start function will be called by init and node-hot-add.
2228  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2229  */
2230 int kswapd_run(int nid)
2231 {
2232         pg_data_t *pgdat = NODE_DATA(nid);
2233         int ret = 0;
2234
2235         if (pgdat->kswapd)
2236                 return 0;
2237
2238         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2239         if (IS_ERR(pgdat->kswapd)) {
2240                 /* failure at boot is fatal */
2241                 BUG_ON(system_state == SYSTEM_BOOTING);
2242                 printk("Failed to start kswapd on node %d\n",nid);
2243                 ret = -1;
2244         }
2245         return ret;
2246 }
2247
2248 static int __init kswapd_init(void)
2249 {
2250         int nid;
2251
2252         swap_setup();
2253         for_each_node_state(nid, N_HIGH_MEMORY)
2254                 kswapd_run(nid);
2255         hotcpu_notifier(cpu_callback, 0);
2256         return 0;
2257 }
2258
2259 module_init(kswapd_init)
2260
2261 #ifdef CONFIG_NUMA
2262 /*
2263  * Zone reclaim mode
2264  *
2265  * If non-zero call zone_reclaim when the number of free pages falls below
2266  * the watermarks.
2267  */
2268 int zone_reclaim_mode __read_mostly;
2269
2270 #define RECLAIM_OFF 0
2271 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
2272 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
2273 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
2274
2275 /*
2276  * Priority for ZONE_RECLAIM. This determines the fraction of pages
2277  * of a node considered for each zone_reclaim. 4 scans 1/16th of
2278  * a zone.
2279  */
2280 #define ZONE_RECLAIM_PRIORITY 4
2281
2282 /*
2283  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2284  * occur.
2285  */
2286 int sysctl_min_unmapped_ratio = 1;
2287
2288 /*
2289  * If the number of slab pages in a zone grows beyond this percentage then
2290  * slab reclaim needs to occur.
2291  */
2292 int sysctl_min_slab_ratio = 5;
2293
2294 /*
2295  * Try to free up some pages from this zone through reclaim.
2296  */
2297 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2298 {
2299         /* Minimum pages needed in order to stay on node */
2300         const unsigned long nr_pages = 1 << order;
2301         struct task_struct *p = current;
2302         struct reclaim_state reclaim_state;
2303         int priority;
2304         struct scan_control sc = {
2305                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2306                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2307                 .may_swap = 1,
2308                 .swap_cluster_max = max_t(unsigned long, nr_pages,
2309                                         SWAP_CLUSTER_MAX),
2310                 .gfp_mask = gfp_mask,
2311                 .swappiness = vm_swappiness,
2312                 .order = order,
2313                 .isolate_pages = isolate_pages_global,
2314         };
2315         unsigned long slab_reclaimable;
2316
2317         disable_swap_token();
2318         cond_resched();
2319         /*
2320          * We need to be able to allocate from the reserves for RECLAIM_SWAP
2321          * and we also need to be able to write out pages for RECLAIM_WRITE
2322          * and RECLAIM_SWAP.
2323          */
2324         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2325         reclaim_state.reclaimed_slab = 0;
2326         p->reclaim_state = &reclaim_state;
2327
2328         if (zone_page_state(zone, NR_FILE_PAGES) -
2329                 zone_page_state(zone, NR_FILE_MAPPED) >
2330                 zone->min_unmapped_pages) {
2331                 /*
2332                  * Free memory by calling shrink zone with increasing
2333                  * priorities until we have enough memory freed.
2334                  */
2335                 priority = ZONE_RECLAIM_PRIORITY;
2336                 do {
2337                         note_zone_scanning_priority(zone, priority);
2338                         shrink_zone(priority, zone, &sc);
2339                         priority--;
2340                 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
2341         }
2342
2343         slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2344         if (slab_reclaimable > zone->min_slab_pages) {
2345                 /*
2346                  * shrink_slab() does not currently allow us to determine how
2347                  * many pages were freed in this zone. So we take the current
2348                  * number of slab pages and shake the slab until it is reduced
2349                  * by the same nr_pages that we used for reclaiming unmapped
2350                  * pages.
2351                  *
2352                  * Note that shrink_slab will free memory on all zones and may
2353                  * take a long time.
2354                  */
2355                 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
2356                         zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
2357                                 slab_reclaimable - nr_pages)
2358                         ;
2359
2360                 /*
2361                  * Update nr_reclaimed by the number of slab pages we
2362                  * reclaimed from this zone.
2363                  */
2364                 sc.nr_reclaimed += slab_reclaimable -
2365                         zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2366         }
2367
2368         p->reclaim_state = NULL;
2369         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
2370         return sc.nr_reclaimed >= nr_pages;
2371 }
2372
2373 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2374 {
2375         int node_id;
2376         int ret;
2377
2378         /*
2379          * Zone reclaim reclaims unmapped file backed pages and
2380          * slab pages if we are over the defined limits.
2381          *
2382          * A small portion of unmapped file backed pages is needed for
2383          * file I/O otherwise pages read by file I/O will be immediately
2384          * thrown out if the zone is overallocated. So we do not reclaim
2385          * if less than a specified percentage of the zone is used by
2386          * unmapped file backed pages.
2387          */
2388         if (zone_page_state(zone, NR_FILE_PAGES) -
2389             zone_page_state(zone, NR_FILE_MAPPED) <= zone->min_unmapped_pages
2390             && zone_page_state(zone, NR_SLAB_RECLAIMABLE)
2391                         <= zone->min_slab_pages)
2392                 return 0;
2393
2394         if (zone_is_all_unreclaimable(zone))
2395                 return 0;
2396
2397         /*
2398          * Do not scan if the allocation should not be delayed.
2399          */
2400         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
2401                         return 0;
2402
2403         /*
2404          * Only run zone reclaim on the local zone or on zones that do not
2405          * have associated processors. This will favor the local processor
2406          * over remote processors and spread off node memory allocations
2407          * as wide as possible.
2408          */
2409         node_id = zone_to_nid(zone);
2410         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
2411                 return 0;
2412
2413         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2414                 return 0;
2415         ret = __zone_reclaim(zone, gfp_mask, order);
2416         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2417
2418         return ret;
2419 }
2420 #endif
2421
2422 #ifdef CONFIG_UNEVICTABLE_LRU
2423 /*
2424  * page_evictable - test whether a page is evictable
2425  * @page: the page to test
2426  * @vma: the VMA in which the page is or will be mapped, may be NULL
2427  *
2428  * Test whether page is evictable--i.e., should be placed on active/inactive
2429  * lists vs unevictable list.  The vma argument is !NULL when called from the
2430  * fault path to determine how to instantate a new page.
2431  *
2432  * Reasons page might not be evictable:
2433  * (1) page's mapping marked unevictable
2434  * (2) page is part of an mlocked VMA
2435  *
2436  */
2437 int page_evictable(struct page *page, struct vm_area_struct *vma)
2438 {
2439
2440         if (mapping_unevictable(page_mapping(page)))
2441                 return 0;
2442
2443         if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2444                 return 0;
2445
2446         return 1;
2447 }
2448
2449 /**
2450  * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2451  * @page: page to check evictability and move to appropriate lru list
2452  * @zone: zone page is in
2453  *
2454  * Checks a page for evictability and moves the page to the appropriate
2455  * zone lru list.
2456  *
2457  * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2458  * have PageUnevictable set.
2459  */
2460 static void check_move_unevictable_page(struct page *page, struct zone *zone)
2461 {
2462         VM_BUG_ON(PageActive(page));
2463
2464 retry:
2465         ClearPageUnevictable(page);
2466         if (page_evictable(page, NULL)) {
2467                 enum lru_list l = LRU_INACTIVE_ANON + page_is_file_cache(page);
2468
2469                 __dec_zone_state(zone, NR_UNEVICTABLE);
2470                 list_move(&page->lru, &zone->lru[l].list);
2471                 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
2472                 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
2473                 __count_vm_event(UNEVICTABLE_PGRESCUED);
2474         } else {
2475                 /*
2476                  * rotate unevictable list
2477                  */
2478                 SetPageUnevictable(page);
2479                 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
2480                 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
2481                 if (page_evictable(page, NULL))
2482                         goto retry;
2483         }
2484 }
2485
2486 /**
2487  * scan_mapping_unevictable_pages - scan an address space for evictable pages
2488  * @mapping: struct address_space to scan for evictable pages
2489  *
2490  * Scan all pages in mapping.  Check unevictable pages for
2491  * evictability and move them to the appropriate zone lru list.
2492  */
2493 void scan_mapping_unevictable_pages(struct address_space *mapping)
2494 {
2495         pgoff_t next = 0;
2496         pgoff_t end   = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
2497                          PAGE_CACHE_SHIFT;
2498         struct zone *zone;
2499         struct pagevec pvec;
2500
2501         if (mapping->nrpages == 0)
2502                 return;
2503
2504         pagevec_init(&pvec, 0);
2505         while (next < end &&
2506                 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
2507                 int i;
2508                 int pg_scanned = 0;
2509
2510                 zone = NULL;
2511
2512                 for (i = 0; i < pagevec_count(&pvec); i++) {
2513                         struct page *page = pvec.pages[i];
2514                         pgoff_t page_index = page->index;
2515                         struct zone *pagezone = page_zone(page);
2516
2517                         pg_scanned++;
2518                         if (page_index > next)
2519                                 next = page_index;
2520                         next++;
2521
2522                         if (pagezone != zone) {
2523                                 if (zone)
2524                                         spin_unlock_irq(&zone->lru_lock);
2525                                 zone = pagezone;
2526                                 spin_lock_irq(&zone->lru_lock);
2527                         }
2528
2529                         if (PageLRU(page) && PageUnevictable(page))
2530                                 check_move_unevictable_page(page, zone);
2531                 }
2532                 if (zone)
2533                         spin_unlock_irq(&zone->lru_lock);
2534                 pagevec_release(&pvec);
2535
2536                 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
2537         }
2538
2539 }
2540
2541 /**
2542  * scan_zone_unevictable_pages - check unevictable list for evictable pages
2543  * @zone - zone of which to scan the unevictable list
2544  *
2545  * Scan @zone's unevictable LRU lists to check for pages that have become
2546  * evictable.  Move those that have to @zone's inactive list where they
2547  * become candidates for reclaim, unless shrink_inactive_zone() decides
2548  * to reactivate them.  Pages that are still unevictable are rotated
2549  * back onto @zone's unevictable list.
2550  */
2551 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
2552 static void scan_zone_unevictable_pages(struct zone *zone)
2553 {
2554         struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
2555         unsigned long scan;
2556         unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
2557
2558         while (nr_to_scan > 0) {
2559                 unsigned long batch_size = min(nr_to_scan,
2560                                                 SCAN_UNEVICTABLE_BATCH_SIZE);
2561
2562                 spin_lock_irq(&zone->lru_lock);
2563                 for (scan = 0;  scan < batch_size; scan++) {
2564                         struct page *page = lru_to_page(l_unevictable);
2565
2566                         if (!trylock_page(page))
2567                                 continue;
2568
2569                         prefetchw_prev_lru_page(page, l_unevictable, flags);
2570
2571                         if (likely(PageLRU(page) && PageUnevictable(page)))
2572                                 check_move_unevictable_page(page, zone);
2573
2574                         unlock_page(page);
2575                 }
2576                 spin_unlock_irq(&zone->lru_lock);
2577
2578                 nr_to_scan -= batch_size;
2579         }
2580 }
2581
2582
2583 /**
2584  * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
2585  *
2586  * A really big hammer:  scan all zones' unevictable LRU lists to check for
2587  * pages that have become evictable.  Move those back to the zones'
2588  * inactive list where they become candidates for reclaim.
2589  * This occurs when, e.g., we have unswappable pages on the unevictable lists,
2590  * and we add swap to the system.  As such, it runs in the context of a task
2591  * that has possibly/probably made some previously unevictable pages
2592  * evictable.
2593  */
2594 static void scan_all_zones_unevictable_pages(void)
2595 {
2596         struct zone *zone;
2597
2598         for_each_zone(zone) {
2599                 scan_zone_unevictable_pages(zone);
2600         }
2601 }
2602
2603 /*
2604  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
2605  * all nodes' unevictable lists for evictable pages
2606  */
2607 unsigned long scan_unevictable_pages;
2608
2609 int scan_unevictable_handler(struct ctl_table *table, int write,
2610                            struct file *file, void __user *buffer,
2611                            size_t *length, loff_t *ppos)
2612 {
2613         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
2614
2615         if (write && *(unsigned long *)table->data)
2616                 scan_all_zones_unevictable_pages();
2617
2618         scan_unevictable_pages = 0;
2619         return 0;
2620 }
2621
2622 /*
2623  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
2624  * a specified node's per zone unevictable lists for evictable pages.
2625  */
2626
2627 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
2628                                           struct sysdev_attribute *attr,
2629                                           char *buf)
2630 {
2631         return sprintf(buf, "0\n");     /* always zero; should fit... */
2632 }
2633
2634 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
2635                                            struct sysdev_attribute *attr,
2636                                         const char *buf, size_t count)
2637 {
2638         struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
2639         struct zone *zone;
2640         unsigned long res;
2641         unsigned long req = strict_strtoul(buf, 10, &res);
2642
2643         if (!req)
2644                 return 1;       /* zero is no-op */
2645
2646         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2647                 if (!populated_zone(zone))
2648                         continue;
2649                 scan_zone_unevictable_pages(zone);
2650         }
2651         return 1;
2652 }
2653
2654
2655 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
2656                         read_scan_unevictable_node,
2657                         write_scan_unevictable_node);
2658
2659 int scan_unevictable_register_node(struct node *node)
2660 {
2661         return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
2662 }
2663
2664 void scan_unevictable_unregister_node(struct node *node)
2665 {
2666         sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
2667 }
2668
2669 #endif