mm: remove CONFIG_UNEVICTABLE_LRU config option
[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 void putback_lru_page(struct page *page)
518 {
519         int lru;
520         int active = !!TestClearPageActive(page);
521         int was_unevictable = PageUnevictable(page);
522
523         VM_BUG_ON(PageLRU(page));
524
525 redo:
526         ClearPageUnevictable(page);
527
528         if (page_evictable(page, NULL)) {
529                 /*
530                  * For evictable pages, we can use the cache.
531                  * In event of a race, worst case is we end up with an
532                  * unevictable page on [in]active list.
533                  * We know how to handle that.
534                  */
535                 lru = active + page_is_file_cache(page);
536                 lru_cache_add_lru(page, lru);
537         } else {
538                 /*
539                  * Put unevictable pages directly on zone's unevictable
540                  * list.
541                  */
542                 lru = LRU_UNEVICTABLE;
543                 add_page_to_unevictable_list(page);
544         }
545
546         /*
547          * page's status can change while we move it among lru. If an evictable
548          * page is on unevictable list, it never be freed. To avoid that,
549          * check after we added it to the list, again.
550          */
551         if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
552                 if (!isolate_lru_page(page)) {
553                         put_page(page);
554                         goto redo;
555                 }
556                 /* This means someone else dropped this page from LRU
557                  * So, it will be freed or putback to LRU again. There is
558                  * nothing to do here.
559                  */
560         }
561
562         if (was_unevictable && lru != LRU_UNEVICTABLE)
563                 count_vm_event(UNEVICTABLE_PGRESCUED);
564         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
565                 count_vm_event(UNEVICTABLE_PGCULLED);
566
567         put_page(page);         /* drop ref from isolate */
568 }
569
570 /*
571  * shrink_page_list() returns the number of reclaimed pages
572  */
573 static unsigned long shrink_page_list(struct list_head *page_list,
574                                         struct scan_control *sc,
575                                         enum pageout_io sync_writeback)
576 {
577         LIST_HEAD(ret_pages);
578         struct pagevec freed_pvec;
579         int pgactivate = 0;
580         unsigned long nr_reclaimed = 0;
581
582         cond_resched();
583
584         pagevec_init(&freed_pvec, 1);
585         while (!list_empty(page_list)) {
586                 struct address_space *mapping;
587                 struct page *page;
588                 int may_enter_fs;
589                 int referenced;
590
591                 cond_resched();
592
593                 page = lru_to_page(page_list);
594                 list_del(&page->lru);
595
596                 if (!trylock_page(page))
597                         goto keep;
598
599                 VM_BUG_ON(PageActive(page));
600
601                 sc->nr_scanned++;
602
603                 if (unlikely(!page_evictable(page, NULL)))
604                         goto cull_mlocked;
605
606                 if (!sc->may_unmap && page_mapped(page))
607                         goto keep_locked;
608
609                 /* Double the slab pressure for mapped and swapcache pages */
610                 if (page_mapped(page) || PageSwapCache(page))
611                         sc->nr_scanned++;
612
613                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
614                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
615
616                 if (PageWriteback(page)) {
617                         /*
618                          * Synchronous reclaim is performed in two passes,
619                          * first an asynchronous pass over the list to
620                          * start parallel writeback, and a second synchronous
621                          * pass to wait for the IO to complete.  Wait here
622                          * for any page for which writeback has already
623                          * started.
624                          */
625                         if (sync_writeback == PAGEOUT_IO_SYNC && may_enter_fs)
626                                 wait_on_page_writeback(page);
627                         else
628                                 goto keep_locked;
629                 }
630
631                 referenced = page_referenced(page, 1, sc->mem_cgroup);
632                 /* In active use or really unfreeable?  Activate it. */
633                 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER &&
634                                         referenced && page_mapping_inuse(page))
635                         goto activate_locked;
636
637                 /*
638                  * Anonymous process memory has backing store?
639                  * Try to allocate it some swap space here.
640                  */
641                 if (PageAnon(page) && !PageSwapCache(page)) {
642                         if (!(sc->gfp_mask & __GFP_IO))
643                                 goto keep_locked;
644                         if (!add_to_swap(page))
645                                 goto activate_locked;
646                         may_enter_fs = 1;
647                 }
648
649                 mapping = page_mapping(page);
650
651                 /*
652                  * The page is mapped into the page tables of one or more
653                  * processes. Try to unmap it here.
654                  */
655                 if (page_mapped(page) && mapping) {
656                         switch (try_to_unmap(page, 0)) {
657                         case SWAP_FAIL:
658                                 goto activate_locked;
659                         case SWAP_AGAIN:
660                                 goto keep_locked;
661                         case SWAP_MLOCK:
662                                 goto cull_mlocked;
663                         case SWAP_SUCCESS:
664                                 ; /* try to free the page below */
665                         }
666                 }
667
668                 if (PageDirty(page)) {
669                         if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && referenced)
670                                 goto keep_locked;
671                         if (!may_enter_fs)
672                                 goto keep_locked;
673                         if (!sc->may_writepage)
674                                 goto keep_locked;
675
676                         /* Page is dirty, try to write it out here */
677                         switch (pageout(page, mapping, sync_writeback)) {
678                         case PAGE_KEEP:
679                                 goto keep_locked;
680                         case PAGE_ACTIVATE:
681                                 goto activate_locked;
682                         case PAGE_SUCCESS:
683                                 if (PageWriteback(page) || PageDirty(page))
684                                         goto keep;
685                                 /*
686                                  * A synchronous write - probably a ramdisk.  Go
687                                  * ahead and try to reclaim the page.
688                                  */
689                                 if (!trylock_page(page))
690                                         goto keep;
691                                 if (PageDirty(page) || PageWriteback(page))
692                                         goto keep_locked;
693                                 mapping = page_mapping(page);
694                         case PAGE_CLEAN:
695                                 ; /* try to free the page below */
696                         }
697                 }
698
699                 /*
700                  * If the page has buffers, try to free the buffer mappings
701                  * associated with this page. If we succeed we try to free
702                  * the page as well.
703                  *
704                  * We do this even if the page is PageDirty().
705                  * try_to_release_page() does not perform I/O, but it is
706                  * possible for a page to have PageDirty set, but it is actually
707                  * clean (all its buffers are clean).  This happens if the
708                  * buffers were written out directly, with submit_bh(). ext3
709                  * will do this, as well as the blockdev mapping.
710                  * try_to_release_page() will discover that cleanness and will
711                  * drop the buffers and mark the page clean - it can be freed.
712                  *
713                  * Rarely, pages can have buffers and no ->mapping.  These are
714                  * the pages which were not successfully invalidated in
715                  * truncate_complete_page().  We try to drop those buffers here
716                  * and if that worked, and the page is no longer mapped into
717                  * process address space (page_count == 1) it can be freed.
718                  * Otherwise, leave the page on the LRU so it is swappable.
719                  */
720                 if (page_has_private(page)) {
721                         if (!try_to_release_page(page, sc->gfp_mask))
722                                 goto activate_locked;
723                         if (!mapping && page_count(page) == 1) {
724                                 unlock_page(page);
725                                 if (put_page_testzero(page))
726                                         goto free_it;
727                                 else {
728                                         /*
729                                          * rare race with speculative reference.
730                                          * the speculative reference will free
731                                          * this page shortly, so we may
732                                          * increment nr_reclaimed here (and
733                                          * leave it off the LRU).
734                                          */
735                                         nr_reclaimed++;
736                                         continue;
737                                 }
738                         }
739                 }
740
741                 if (!mapping || !__remove_mapping(mapping, page))
742                         goto keep_locked;
743
744                 /*
745                  * At this point, we have no other references and there is
746                  * no way to pick any more up (removed from LRU, removed
747                  * from pagecache). Can use non-atomic bitops now (and
748                  * we obviously don't have to worry about waking up a process
749                  * waiting on the page lock, because there are no references.
750                  */
751                 __clear_page_locked(page);
752 free_it:
753                 nr_reclaimed++;
754                 if (!pagevec_add(&freed_pvec, page)) {
755                         __pagevec_free(&freed_pvec);
756                         pagevec_reinit(&freed_pvec);
757                 }
758                 continue;
759
760 cull_mlocked:
761                 if (PageSwapCache(page))
762                         try_to_free_swap(page);
763                 unlock_page(page);
764                 putback_lru_page(page);
765                 continue;
766
767 activate_locked:
768                 /* Not a candidate for swapping, so reclaim swap space. */
769                 if (PageSwapCache(page) && vm_swap_full())
770                         try_to_free_swap(page);
771                 VM_BUG_ON(PageActive(page));
772                 SetPageActive(page);
773                 pgactivate++;
774 keep_locked:
775                 unlock_page(page);
776 keep:
777                 list_add(&page->lru, &ret_pages);
778                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
779         }
780         list_splice(&ret_pages, page_list);
781         if (pagevec_count(&freed_pvec))
782                 __pagevec_free(&freed_pvec);
783         count_vm_events(PGACTIVATE, pgactivate);
784         return nr_reclaimed;
785 }
786
787 /* LRU Isolation modes. */
788 #define ISOLATE_INACTIVE 0      /* Isolate inactive pages. */
789 #define ISOLATE_ACTIVE 1        /* Isolate active pages. */
790 #define ISOLATE_BOTH 2          /* Isolate both active and inactive pages. */
791
792 /*
793  * Attempt to remove the specified page from its LRU.  Only take this page
794  * if it is of the appropriate PageActive status.  Pages which are being
795  * freed elsewhere are also ignored.
796  *
797  * page:        page to consider
798  * mode:        one of the LRU isolation modes defined above
799  *
800  * returns 0 on success, -ve errno on failure.
801  */
802 int __isolate_lru_page(struct page *page, int mode, int file)
803 {
804         int ret = -EINVAL;
805
806         /* Only take pages on the LRU. */
807         if (!PageLRU(page))
808                 return ret;
809
810         /*
811          * When checking the active state, we need to be sure we are
812          * dealing with comparible boolean values.  Take the logical not
813          * of each.
814          */
815         if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
816                 return ret;
817
818         if (mode != ISOLATE_BOTH && (!page_is_file_cache(page) != !file))
819                 return ret;
820
821         /*
822          * When this function is being called for lumpy reclaim, we
823          * initially look into all LRU pages, active, inactive and
824          * unevictable; only give shrink_page_list evictable pages.
825          */
826         if (PageUnevictable(page))
827                 return ret;
828
829         ret = -EBUSY;
830
831         if (likely(get_page_unless_zero(page))) {
832                 /*
833                  * Be careful not to clear PageLRU until after we're
834                  * sure the page is not being freed elsewhere -- the
835                  * page release code relies on it.
836                  */
837                 ClearPageLRU(page);
838                 ret = 0;
839                 mem_cgroup_del_lru(page);
840         }
841
842         return ret;
843 }
844
845 /*
846  * zone->lru_lock is heavily contended.  Some of the functions that
847  * shrink the lists perform better by taking out a batch of pages
848  * and working on them outside the LRU lock.
849  *
850  * For pagecache intensive workloads, this function is the hottest
851  * spot in the kernel (apart from copy_*_user functions).
852  *
853  * Appropriate locks must be held before calling this function.
854  *
855  * @nr_to_scan: The number of pages to look through on the list.
856  * @src:        The LRU list to pull pages off.
857  * @dst:        The temp list to put pages on to.
858  * @scanned:    The number of pages that were scanned.
859  * @order:      The caller's attempted allocation order
860  * @mode:       One of the LRU isolation modes
861  * @file:       True [1] if isolating file [!anon] pages
862  *
863  * returns how many pages were moved onto *@dst.
864  */
865 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
866                 struct list_head *src, struct list_head *dst,
867                 unsigned long *scanned, int order, int mode, int file)
868 {
869         unsigned long nr_taken = 0;
870         unsigned long scan;
871
872         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
873                 struct page *page;
874                 unsigned long pfn;
875                 unsigned long end_pfn;
876                 unsigned long page_pfn;
877                 int zone_id;
878
879                 page = lru_to_page(src);
880                 prefetchw_prev_lru_page(page, src, flags);
881
882                 VM_BUG_ON(!PageLRU(page));
883
884                 switch (__isolate_lru_page(page, mode, file)) {
885                 case 0:
886                         list_move(&page->lru, dst);
887                         nr_taken++;
888                         break;
889
890                 case -EBUSY:
891                         /* else it is being freed elsewhere */
892                         list_move(&page->lru, src);
893                         continue;
894
895                 default:
896                         BUG();
897                 }
898
899                 if (!order)
900                         continue;
901
902                 /*
903                  * Attempt to take all pages in the order aligned region
904                  * surrounding the tag page.  Only take those pages of
905                  * the same active state as that tag page.  We may safely
906                  * round the target page pfn down to the requested order
907                  * as the mem_map is guarenteed valid out to MAX_ORDER,
908                  * where that page is in a different zone we will detect
909                  * it from its zone id and abort this block scan.
910                  */
911                 zone_id = page_zone_id(page);
912                 page_pfn = page_to_pfn(page);
913                 pfn = page_pfn & ~((1 << order) - 1);
914                 end_pfn = pfn + (1 << order);
915                 for (; pfn < end_pfn; pfn++) {
916                         struct page *cursor_page;
917
918                         /* The target page is in the block, ignore it. */
919                         if (unlikely(pfn == page_pfn))
920                                 continue;
921
922                         /* Avoid holes within the zone. */
923                         if (unlikely(!pfn_valid_within(pfn)))
924                                 break;
925
926                         cursor_page = pfn_to_page(pfn);
927
928                         /* Check that we have not crossed a zone boundary. */
929                         if (unlikely(page_zone_id(cursor_page) != zone_id))
930                                 continue;
931                         switch (__isolate_lru_page(cursor_page, mode, file)) {
932                         case 0:
933                                 list_move(&cursor_page->lru, dst);
934                                 nr_taken++;
935                                 scan++;
936                                 break;
937
938                         case -EBUSY:
939                                 /* else it is being freed elsewhere */
940                                 list_move(&cursor_page->lru, src);
941                         default:
942                                 break;  /* ! on LRU or wrong list */
943                         }
944                 }
945         }
946
947         *scanned = scan;
948         return nr_taken;
949 }
950
951 static unsigned long isolate_pages_global(unsigned long nr,
952                                         struct list_head *dst,
953                                         unsigned long *scanned, int order,
954                                         int mode, struct zone *z,
955                                         struct mem_cgroup *mem_cont,
956                                         int active, int file)
957 {
958         int lru = LRU_BASE;
959         if (active)
960                 lru += LRU_ACTIVE;
961         if (file)
962                 lru += LRU_FILE;
963         return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
964                                                                 mode, !!file);
965 }
966
967 /*
968  * clear_active_flags() is a helper for shrink_active_list(), clearing
969  * any active bits from the pages in the list.
970  */
971 static unsigned long clear_active_flags(struct list_head *page_list,
972                                         unsigned int *count)
973 {
974         int nr_active = 0;
975         int lru;
976         struct page *page;
977
978         list_for_each_entry(page, page_list, lru) {
979                 lru = page_is_file_cache(page);
980                 if (PageActive(page)) {
981                         lru += LRU_ACTIVE;
982                         ClearPageActive(page);
983                         nr_active++;
984                 }
985                 count[lru]++;
986         }
987
988         return nr_active;
989 }
990
991 /**
992  * isolate_lru_page - tries to isolate a page from its LRU list
993  * @page: page to isolate from its LRU list
994  *
995  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
996  * vmstat statistic corresponding to whatever LRU list the page was on.
997  *
998  * Returns 0 if the page was removed from an LRU list.
999  * Returns -EBUSY if the page was not on an LRU list.
1000  *
1001  * The returned page will have PageLRU() cleared.  If it was found on
1002  * the active list, it will have PageActive set.  If it was found on
1003  * the unevictable list, it will have the PageUnevictable bit set. That flag
1004  * may need to be cleared by the caller before letting the page go.
1005  *
1006  * The vmstat statistic corresponding to the list on which the page was
1007  * found will be decremented.
1008  *
1009  * Restrictions:
1010  * (1) Must be called with an elevated refcount on the page. This is a
1011  *     fundamentnal difference from isolate_lru_pages (which is called
1012  *     without a stable reference).
1013  * (2) the lru_lock must not be held.
1014  * (3) interrupts must be enabled.
1015  */
1016 int isolate_lru_page(struct page *page)
1017 {
1018         int ret = -EBUSY;
1019
1020         if (PageLRU(page)) {
1021                 struct zone *zone = page_zone(page);
1022
1023                 spin_lock_irq(&zone->lru_lock);
1024                 if (PageLRU(page) && get_page_unless_zero(page)) {
1025                         int lru = page_lru(page);
1026                         ret = 0;
1027                         ClearPageLRU(page);
1028
1029                         del_page_from_lru_list(zone, page, lru);
1030                 }
1031                 spin_unlock_irq(&zone->lru_lock);
1032         }
1033         return ret;
1034 }
1035
1036 /*
1037  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1038  * of reclaimed pages
1039  */
1040 static unsigned long shrink_inactive_list(unsigned long max_scan,
1041                         struct zone *zone, struct scan_control *sc,
1042                         int priority, int file)
1043 {
1044         LIST_HEAD(page_list);
1045         struct pagevec pvec;
1046         unsigned long nr_scanned = 0;
1047         unsigned long nr_reclaimed = 0;
1048         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1049         int lumpy_reclaim = 0;
1050
1051         /*
1052          * If we need a large contiguous chunk of memory, or have
1053          * trouble getting a small set of contiguous pages, we
1054          * will reclaim both active and inactive pages.
1055          *
1056          * We use the same threshold as pageout congestion_wait below.
1057          */
1058         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1059                 lumpy_reclaim = 1;
1060         else if (sc->order && priority < DEF_PRIORITY - 2)
1061                 lumpy_reclaim = 1;
1062
1063         pagevec_init(&pvec, 1);
1064
1065         lru_add_drain();
1066         spin_lock_irq(&zone->lru_lock);
1067         do {
1068                 struct page *page;
1069                 unsigned long nr_taken;
1070                 unsigned long nr_scan;
1071                 unsigned long nr_freed;
1072                 unsigned long nr_active;
1073                 unsigned int count[NR_LRU_LISTS] = { 0, };
1074                 int mode = lumpy_reclaim ? ISOLATE_BOTH : ISOLATE_INACTIVE;
1075
1076                 nr_taken = sc->isolate_pages(sc->swap_cluster_max,
1077                              &page_list, &nr_scan, sc->order, mode,
1078                                 zone, sc->mem_cgroup, 0, file);
1079                 nr_active = clear_active_flags(&page_list, count);
1080                 __count_vm_events(PGDEACTIVATE, nr_active);
1081
1082                 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1083                                                 -count[LRU_ACTIVE_FILE]);
1084                 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1085                                                 -count[LRU_INACTIVE_FILE]);
1086                 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1087                                                 -count[LRU_ACTIVE_ANON]);
1088                 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1089                                                 -count[LRU_INACTIVE_ANON]);
1090
1091                 if (scanning_global_lru(sc))
1092                         zone->pages_scanned += nr_scan;
1093
1094                 reclaim_stat->recent_scanned[0] += count[LRU_INACTIVE_ANON];
1095                 reclaim_stat->recent_scanned[0] += count[LRU_ACTIVE_ANON];
1096                 reclaim_stat->recent_scanned[1] += count[LRU_INACTIVE_FILE];
1097                 reclaim_stat->recent_scanned[1] += count[LRU_ACTIVE_FILE];
1098
1099                 spin_unlock_irq(&zone->lru_lock);
1100
1101                 nr_scanned += nr_scan;
1102                 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
1103
1104                 /*
1105                  * If we are direct reclaiming for contiguous pages and we do
1106                  * not reclaim everything in the list, try again and wait
1107                  * for IO to complete. This will stall high-order allocations
1108                  * but that should be acceptable to the caller
1109                  */
1110                 if (nr_freed < nr_taken && !current_is_kswapd() &&
1111                     lumpy_reclaim) {
1112                         congestion_wait(WRITE, HZ/10);
1113
1114                         /*
1115                          * The attempt at page out may have made some
1116                          * of the pages active, mark them inactive again.
1117                          */
1118                         nr_active = clear_active_flags(&page_list, count);
1119                         count_vm_events(PGDEACTIVATE, nr_active);
1120
1121                         nr_freed += shrink_page_list(&page_list, sc,
1122                                                         PAGEOUT_IO_SYNC);
1123                 }
1124
1125                 nr_reclaimed += nr_freed;
1126                 local_irq_disable();
1127                 if (current_is_kswapd()) {
1128                         __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scan);
1129                         __count_vm_events(KSWAPD_STEAL, nr_freed);
1130                 } else if (scanning_global_lru(sc))
1131                         __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scan);
1132
1133                 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
1134
1135                 if (nr_taken == 0)
1136                         goto done;
1137
1138                 spin_lock(&zone->lru_lock);
1139                 /*
1140                  * Put back any unfreeable pages.
1141                  */
1142                 while (!list_empty(&page_list)) {
1143                         int lru;
1144                         page = lru_to_page(&page_list);
1145                         VM_BUG_ON(PageLRU(page));
1146                         list_del(&page->lru);
1147                         if (unlikely(!page_evictable(page, NULL))) {
1148                                 spin_unlock_irq(&zone->lru_lock);
1149                                 putback_lru_page(page);
1150                                 spin_lock_irq(&zone->lru_lock);
1151                                 continue;
1152                         }
1153                         SetPageLRU(page);
1154                         lru = page_lru(page);
1155                         add_page_to_lru_list(zone, page, lru);
1156                         if (PageActive(page)) {
1157                                 int file = !!page_is_file_cache(page);
1158                                 reclaim_stat->recent_rotated[file]++;
1159                         }
1160                         if (!pagevec_add(&pvec, page)) {
1161                                 spin_unlock_irq(&zone->lru_lock);
1162                                 __pagevec_release(&pvec);
1163                                 spin_lock_irq(&zone->lru_lock);
1164                         }
1165                 }
1166         } while (nr_scanned < max_scan);
1167         spin_unlock(&zone->lru_lock);
1168 done:
1169         local_irq_enable();
1170         pagevec_release(&pvec);
1171         return nr_reclaimed;
1172 }
1173
1174 /*
1175  * We are about to scan this zone at a certain priority level.  If that priority
1176  * level is smaller (ie: more urgent) than the previous priority, then note
1177  * that priority level within the zone.  This is done so that when the next
1178  * process comes in to scan this zone, it will immediately start out at this
1179  * priority level rather than having to build up its own scanning priority.
1180  * Here, this priority affects only the reclaim-mapped threshold.
1181  */
1182 static inline void note_zone_scanning_priority(struct zone *zone, int priority)
1183 {
1184         if (priority < zone->prev_priority)
1185                 zone->prev_priority = priority;
1186 }
1187
1188 /*
1189  * This moves pages from the active list to the inactive list.
1190  *
1191  * We move them the other way if the page is referenced by one or more
1192  * processes, from rmap.
1193  *
1194  * If the pages are mostly unmapped, the processing is fast and it is
1195  * appropriate to hold zone->lru_lock across the whole operation.  But if
1196  * the pages are mapped, the processing is slow (page_referenced()) so we
1197  * should drop zone->lru_lock around each page.  It's impossible to balance
1198  * this, so instead we remove the pages from the LRU while processing them.
1199  * It is safe to rely on PG_active against the non-LRU pages in here because
1200  * nobody will play with that bit on a non-LRU page.
1201  *
1202  * The downside is that we have to touch page->_count against each page.
1203  * But we had to alter page->flags anyway.
1204  */
1205
1206
1207 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1208                         struct scan_control *sc, int priority, int file)
1209 {
1210         unsigned long pgmoved;
1211         unsigned long pgscanned;
1212         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1213         LIST_HEAD(l_inactive);
1214         struct page *page;
1215         struct pagevec pvec;
1216         enum lru_list lru;
1217         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1218
1219         lru_add_drain();
1220         spin_lock_irq(&zone->lru_lock);
1221         pgmoved = sc->isolate_pages(nr_pages, &l_hold, &pgscanned, sc->order,
1222                                         ISOLATE_ACTIVE, zone,
1223                                         sc->mem_cgroup, 1, file);
1224         /*
1225          * zone->pages_scanned is used for detect zone's oom
1226          * mem_cgroup remembers nr_scan by itself.
1227          */
1228         if (scanning_global_lru(sc)) {
1229                 zone->pages_scanned += pgscanned;
1230         }
1231         reclaim_stat->recent_scanned[!!file] += pgmoved;
1232
1233         if (file)
1234                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -pgmoved);
1235         else
1236                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -pgmoved);
1237         spin_unlock_irq(&zone->lru_lock);
1238
1239         pgmoved = 0;  /* count referenced (mapping) mapped pages */
1240         while (!list_empty(&l_hold)) {
1241                 cond_resched();
1242                 page = lru_to_page(&l_hold);
1243                 list_del(&page->lru);
1244
1245                 if (unlikely(!page_evictable(page, NULL))) {
1246                         putback_lru_page(page);
1247                         continue;
1248                 }
1249
1250                 /* page_referenced clears PageReferenced */
1251                 if (page_mapping_inuse(page) &&
1252                     page_referenced(page, 0, sc->mem_cgroup))
1253                         pgmoved++;
1254
1255                 list_add(&page->lru, &l_inactive);
1256         }
1257
1258         /*
1259          * Move the pages to the [file or anon] inactive list.
1260          */
1261         pagevec_init(&pvec, 1);
1262         lru = LRU_BASE + file * LRU_FILE;
1263
1264         spin_lock_irq(&zone->lru_lock);
1265         /*
1266          * Count referenced pages from currently used mappings as
1267          * rotated, even though they are moved to the inactive list.
1268          * This helps balance scan pressure between file and anonymous
1269          * pages in get_scan_ratio.
1270          */
1271         reclaim_stat->recent_rotated[!!file] += pgmoved;
1272
1273         pgmoved = 0;  /* count pages moved to inactive list */
1274         while (!list_empty(&l_inactive)) {
1275                 page = lru_to_page(&l_inactive);
1276                 prefetchw_prev_lru_page(page, &l_inactive, flags);
1277                 VM_BUG_ON(PageLRU(page));
1278                 SetPageLRU(page);
1279                 VM_BUG_ON(!PageActive(page));
1280                 ClearPageActive(page);
1281
1282                 list_move(&page->lru, &zone->lru[lru].list);
1283                 mem_cgroup_add_lru_list(page, lru);
1284                 pgmoved++;
1285                 if (!pagevec_add(&pvec, page)) {
1286                         spin_unlock_irq(&zone->lru_lock);
1287                         if (buffer_heads_over_limit)
1288                                 pagevec_strip(&pvec);
1289                         __pagevec_release(&pvec);
1290                         spin_lock_irq(&zone->lru_lock);
1291                 }
1292         }
1293         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1294         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1295         __count_vm_events(PGDEACTIVATE, pgmoved);
1296         spin_unlock_irq(&zone->lru_lock);
1297         if (buffer_heads_over_limit)
1298                 pagevec_strip(&pvec);
1299         pagevec_release(&pvec);
1300 }
1301
1302 static int inactive_anon_is_low_global(struct zone *zone)
1303 {
1304         unsigned long active, inactive;
1305
1306         active = zone_page_state(zone, NR_ACTIVE_ANON);
1307         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1308
1309         if (inactive * zone->inactive_ratio < active)
1310                 return 1;
1311
1312         return 0;
1313 }
1314
1315 /**
1316  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1317  * @zone: zone to check
1318  * @sc:   scan control of this context
1319  *
1320  * Returns true if the zone does not have enough inactive anon pages,
1321  * meaning some active anon pages need to be deactivated.
1322  */
1323 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1324 {
1325         int low;
1326
1327         if (scanning_global_lru(sc))
1328                 low = inactive_anon_is_low_global(zone);
1329         else
1330                 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1331         return low;
1332 }
1333
1334 static int inactive_file_is_low_global(struct zone *zone)
1335 {
1336         unsigned long active, inactive;
1337
1338         active = zone_page_state(zone, NR_ACTIVE_FILE);
1339         inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1340
1341         return (active > inactive);
1342 }
1343
1344 /**
1345  * inactive_file_is_low - check if file pages need to be deactivated
1346  * @zone: zone to check
1347  * @sc:   scan control of this context
1348  *
1349  * When the system is doing streaming IO, memory pressure here
1350  * ensures that active file pages get deactivated, until more
1351  * than half of the file pages are on the inactive list.
1352  *
1353  * Once we get to that situation, protect the system's working
1354  * set from being evicted by disabling active file page aging.
1355  *
1356  * This uses a different ratio than the anonymous pages, because
1357  * the page cache uses a use-once replacement algorithm.
1358  */
1359 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1360 {
1361         int low;
1362
1363         if (scanning_global_lru(sc))
1364                 low = inactive_file_is_low_global(zone);
1365         else
1366                 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1367         return low;
1368 }
1369
1370 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1371         struct zone *zone, struct scan_control *sc, int priority)
1372 {
1373         int file = is_file_lru(lru);
1374
1375         if (lru == LRU_ACTIVE_FILE && inactive_file_is_low(zone, sc)) {
1376                 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1377                 return 0;
1378         }
1379
1380         if (lru == LRU_ACTIVE_ANON && inactive_anon_is_low(zone, sc)) {
1381                 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1382                 return 0;
1383         }
1384         return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1385 }
1386
1387 /*
1388  * Determine how aggressively the anon and file LRU lists should be
1389  * scanned.  The relative value of each set of LRU lists is determined
1390  * by looking at the fraction of the pages scanned we did rotate back
1391  * onto the active list instead of evict.
1392  *
1393  * percent[0] specifies how much pressure to put on ram/swap backed
1394  * memory, while percent[1] determines pressure on the file LRUs.
1395  */
1396 static void get_scan_ratio(struct zone *zone, struct scan_control *sc,
1397                                         unsigned long *percent)
1398 {
1399         unsigned long anon, file, free;
1400         unsigned long anon_prio, file_prio;
1401         unsigned long ap, fp;
1402         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1403
1404         /* If we have no swap space, do not bother scanning anon pages. */
1405         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1406                 percent[0] = 0;
1407                 percent[1] = 100;
1408                 return;
1409         }
1410
1411         anon  = zone_nr_pages(zone, sc, LRU_ACTIVE_ANON) +
1412                 zone_nr_pages(zone, sc, LRU_INACTIVE_ANON);
1413         file  = zone_nr_pages(zone, sc, LRU_ACTIVE_FILE) +
1414                 zone_nr_pages(zone, sc, LRU_INACTIVE_FILE);
1415
1416         if (scanning_global_lru(sc)) {
1417                 free  = zone_page_state(zone, NR_FREE_PAGES);
1418                 /* If we have very few page cache pages,
1419                    force-scan anon pages. */
1420                 if (unlikely(file + free <= high_wmark_pages(zone))) {
1421                         percent[0] = 100;
1422                         percent[1] = 0;
1423                         return;
1424                 }
1425         }
1426
1427         /*
1428          * OK, so we have swap space and a fair amount of page cache
1429          * pages.  We use the recently rotated / recently scanned
1430          * ratios to determine how valuable each cache is.
1431          *
1432          * Because workloads change over time (and to avoid overflow)
1433          * we keep these statistics as a floating average, which ends
1434          * up weighing recent references more than old ones.
1435          *
1436          * anon in [0], file in [1]
1437          */
1438         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1439                 spin_lock_irq(&zone->lru_lock);
1440                 reclaim_stat->recent_scanned[0] /= 2;
1441                 reclaim_stat->recent_rotated[0] /= 2;
1442                 spin_unlock_irq(&zone->lru_lock);
1443         }
1444
1445         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1446                 spin_lock_irq(&zone->lru_lock);
1447                 reclaim_stat->recent_scanned[1] /= 2;
1448                 reclaim_stat->recent_rotated[1] /= 2;
1449                 spin_unlock_irq(&zone->lru_lock);
1450         }
1451
1452         /*
1453          * With swappiness at 100, anonymous and file have the same priority.
1454          * This scanning priority is essentially the inverse of IO cost.
1455          */
1456         anon_prio = sc->swappiness;
1457         file_prio = 200 - sc->swappiness;
1458
1459         /*
1460          * The amount of pressure on anon vs file pages is inversely
1461          * proportional to the fraction of recently scanned pages on
1462          * each list that were recently referenced and in active use.
1463          */
1464         ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1465         ap /= reclaim_stat->recent_rotated[0] + 1;
1466
1467         fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1468         fp /= reclaim_stat->recent_rotated[1] + 1;
1469
1470         /* Normalize to percentages */
1471         percent[0] = 100 * ap / (ap + fp + 1);
1472         percent[1] = 100 - percent[0];
1473 }
1474
1475 /*
1476  * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1477  * until we collected @swap_cluster_max pages to scan.
1478  */
1479 static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
1480                                        unsigned long *nr_saved_scan,
1481                                        unsigned long swap_cluster_max)
1482 {
1483         unsigned long nr;
1484
1485         *nr_saved_scan += nr_to_scan;
1486         nr = *nr_saved_scan;
1487
1488         if (nr >= swap_cluster_max)
1489                 *nr_saved_scan = 0;
1490         else
1491                 nr = 0;
1492
1493         return nr;
1494 }
1495
1496 /*
1497  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1498  */
1499 static void shrink_zone(int priority, struct zone *zone,
1500                                 struct scan_control *sc)
1501 {
1502         unsigned long nr[NR_LRU_LISTS];
1503         unsigned long nr_to_scan;
1504         unsigned long percent[2];       /* anon @ 0; file @ 1 */
1505         enum lru_list l;
1506         unsigned long nr_reclaimed = sc->nr_reclaimed;
1507         unsigned long swap_cluster_max = sc->swap_cluster_max;
1508
1509         get_scan_ratio(zone, sc, percent);
1510
1511         for_each_evictable_lru(l) {
1512                 int file = is_file_lru(l);
1513                 unsigned long scan;
1514
1515                 scan = zone_nr_pages(zone, sc, l);
1516                 if (priority) {
1517                         scan >>= priority;
1518                         scan = (scan * percent[file]) / 100;
1519                 }
1520                 if (scanning_global_lru(sc))
1521                         nr[l] = nr_scan_try_batch(scan,
1522                                                   &zone->lru[l].nr_saved_scan,
1523                                                   swap_cluster_max);
1524                 else
1525                         nr[l] = scan;
1526         }
1527
1528         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1529                                         nr[LRU_INACTIVE_FILE]) {
1530                 for_each_evictable_lru(l) {
1531                         if (nr[l]) {
1532                                 nr_to_scan = min(nr[l], swap_cluster_max);
1533                                 nr[l] -= nr_to_scan;
1534
1535                                 nr_reclaimed += shrink_list(l, nr_to_scan,
1536                                                             zone, sc, priority);
1537                         }
1538                 }
1539                 /*
1540                  * On large memory systems, scan >> priority can become
1541                  * really large. This is fine for the starting priority;
1542                  * we want to put equal scanning pressure on each zone.
1543                  * However, if the VM has a harder time of freeing pages,
1544                  * with multiple processes reclaiming pages, the total
1545                  * freeing target can get unreasonably large.
1546                  */
1547                 if (nr_reclaimed > swap_cluster_max &&
1548                         priority < DEF_PRIORITY && !current_is_kswapd())
1549                         break;
1550         }
1551
1552         sc->nr_reclaimed = nr_reclaimed;
1553
1554         /*
1555          * Even if we did not try to evict anon pages at all, we want to
1556          * rebalance the anon lru active/inactive ratio.
1557          */
1558         if (inactive_anon_is_low(zone, sc) && nr_swap_pages > 0)
1559                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1560
1561         throttle_vm_writeout(sc->gfp_mask);
1562 }
1563
1564 /*
1565  * This is the direct reclaim path, for page-allocating processes.  We only
1566  * try to reclaim pages from zones which will satisfy the caller's allocation
1567  * request.
1568  *
1569  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1570  * Because:
1571  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1572  *    allocation or
1573  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1574  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1575  *    zone defense algorithm.
1576  *
1577  * If a zone is deemed to be full of pinned pages then just give it a light
1578  * scan then give up on it.
1579  */
1580 static void shrink_zones(int priority, struct zonelist *zonelist,
1581                                         struct scan_control *sc)
1582 {
1583         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1584         struct zoneref *z;
1585         struct zone *zone;
1586
1587         sc->all_unreclaimable = 1;
1588         for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1589                                         sc->nodemask) {
1590                 if (!populated_zone(zone))
1591                         continue;
1592                 /*
1593                  * Take care memory controller reclaiming has small influence
1594                  * to global LRU.
1595                  */
1596                 if (scanning_global_lru(sc)) {
1597                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1598                                 continue;
1599                         note_zone_scanning_priority(zone, priority);
1600
1601                         if (zone_is_all_unreclaimable(zone) &&
1602                                                 priority != DEF_PRIORITY)
1603                                 continue;       /* Let kswapd poll it */
1604                         sc->all_unreclaimable = 0;
1605                 } else {
1606                         /*
1607                          * Ignore cpuset limitation here. We just want to reduce
1608                          * # of used pages by us regardless of memory shortage.
1609                          */
1610                         sc->all_unreclaimable = 0;
1611                         mem_cgroup_note_reclaim_priority(sc->mem_cgroup,
1612                                                         priority);
1613                 }
1614
1615                 shrink_zone(priority, zone, sc);
1616         }
1617 }
1618
1619 /*
1620  * This is the main entry point to direct page reclaim.
1621  *
1622  * If a full scan of the inactive list fails to free enough memory then we
1623  * are "out of memory" and something needs to be killed.
1624  *
1625  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1626  * high - the zone may be full of dirty or under-writeback pages, which this
1627  * caller can't do much about.  We kick pdflush and take explicit naps in the
1628  * hope that some of these pages can be written.  But if the allocating task
1629  * holds filesystem locks which prevent writeout this might not work, and the
1630  * allocation attempt will fail.
1631  *
1632  * returns:     0, if no pages reclaimed
1633  *              else, the number of pages reclaimed
1634  */
1635 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
1636                                         struct scan_control *sc)
1637 {
1638         int priority;
1639         unsigned long ret = 0;
1640         unsigned long total_scanned = 0;
1641         struct reclaim_state *reclaim_state = current->reclaim_state;
1642         unsigned long lru_pages = 0;
1643         struct zoneref *z;
1644         struct zone *zone;
1645         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1646
1647         delayacct_freepages_start();
1648
1649         if (scanning_global_lru(sc))
1650                 count_vm_event(ALLOCSTALL);
1651         /*
1652          * mem_cgroup will not do shrink_slab.
1653          */
1654         if (scanning_global_lru(sc)) {
1655                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1656
1657                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1658                                 continue;
1659
1660                         lru_pages += zone_lru_pages(zone);
1661                 }
1662         }
1663
1664         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1665                 sc->nr_scanned = 0;
1666                 if (!priority)
1667                         disable_swap_token();
1668                 shrink_zones(priority, zonelist, sc);
1669                 /*
1670                  * Don't shrink slabs when reclaiming memory from
1671                  * over limit cgroups
1672                  */
1673                 if (scanning_global_lru(sc)) {
1674                         shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
1675                         if (reclaim_state) {
1676                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
1677                                 reclaim_state->reclaimed_slab = 0;
1678                         }
1679                 }
1680                 total_scanned += sc->nr_scanned;
1681                 if (sc->nr_reclaimed >= sc->swap_cluster_max) {
1682                         ret = sc->nr_reclaimed;
1683                         goto out;
1684                 }
1685
1686                 /*
1687                  * Try to write back as many pages as we just scanned.  This
1688                  * tends to cause slow streaming writers to write data to the
1689                  * disk smoothly, at the dirtying rate, which is nice.   But
1690                  * that's undesirable in laptop mode, where we *want* lumpy
1691                  * writeout.  So in laptop mode, write out the whole world.
1692                  */
1693                 if (total_scanned > sc->swap_cluster_max +
1694                                         sc->swap_cluster_max / 2) {
1695                         wakeup_pdflush(laptop_mode ? 0 : total_scanned);
1696                         sc->may_writepage = 1;
1697                 }
1698
1699                 /* Take a nap, wait for some writeback to complete */
1700                 if (sc->nr_scanned && priority < DEF_PRIORITY - 2)
1701                         congestion_wait(WRITE, HZ/10);
1702         }
1703         /* top priority shrink_zones still had more to do? don't OOM, then */
1704         if (!sc->all_unreclaimable && scanning_global_lru(sc))
1705                 ret = sc->nr_reclaimed;
1706 out:
1707         /*
1708          * Now that we've scanned all the zones at this priority level, note
1709          * that level within the zone so that the next thread which performs
1710          * scanning of this zone will immediately start out at this priority
1711          * level.  This affects only the decision whether or not to bring
1712          * mapped pages onto the inactive list.
1713          */
1714         if (priority < 0)
1715                 priority = 0;
1716
1717         if (scanning_global_lru(sc)) {
1718                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1719
1720                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1721                                 continue;
1722
1723                         zone->prev_priority = priority;
1724                 }
1725         } else
1726                 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority);
1727
1728         delayacct_freepages_end();
1729
1730         return ret;
1731 }
1732
1733 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
1734                                 gfp_t gfp_mask, nodemask_t *nodemask)
1735 {
1736         struct scan_control sc = {
1737                 .gfp_mask = gfp_mask,
1738                 .may_writepage = !laptop_mode,
1739                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1740                 .may_unmap = 1,
1741                 .may_swap = 1,
1742                 .swappiness = vm_swappiness,
1743                 .order = order,
1744                 .mem_cgroup = NULL,
1745                 .isolate_pages = isolate_pages_global,
1746                 .nodemask = nodemask,
1747         };
1748
1749         return do_try_to_free_pages(zonelist, &sc);
1750 }
1751
1752 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
1753
1754 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
1755                                            gfp_t gfp_mask,
1756                                            bool noswap,
1757                                            unsigned int swappiness)
1758 {
1759         struct scan_control sc = {
1760                 .may_writepage = !laptop_mode,
1761                 .may_unmap = 1,
1762                 .may_swap = !noswap,
1763                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1764                 .swappiness = swappiness,
1765                 .order = 0,
1766                 .mem_cgroup = mem_cont,
1767                 .isolate_pages = mem_cgroup_isolate_pages,
1768                 .nodemask = NULL, /* we don't care the placement */
1769         };
1770         struct zonelist *zonelist;
1771
1772         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1773                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1774         zonelist = NODE_DATA(numa_node_id())->node_zonelists;
1775         return do_try_to_free_pages(zonelist, &sc);
1776 }
1777 #endif
1778
1779 /*
1780  * For kswapd, balance_pgdat() will work across all this node's zones until
1781  * they are all at high_wmark_pages(zone).
1782  *
1783  * Returns the number of pages which were actually freed.
1784  *
1785  * There is special handling here for zones which are full of pinned pages.
1786  * This can happen if the pages are all mlocked, or if they are all used by
1787  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1788  * What we do is to detect the case where all pages in the zone have been
1789  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1790  * dead and from now on, only perform a short scan.  Basically we're polling
1791  * the zone for when the problem goes away.
1792  *
1793  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
1794  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
1795  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
1796  * lower zones regardless of the number of free pages in the lower zones. This
1797  * interoperates with the page allocator fallback scheme to ensure that aging
1798  * of pages is balanced across the zones.
1799  */
1800 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
1801 {
1802         int all_zones_ok;
1803         int priority;
1804         int i;
1805         unsigned long total_scanned;
1806         struct reclaim_state *reclaim_state = current->reclaim_state;
1807         struct scan_control sc = {
1808                 .gfp_mask = GFP_KERNEL,
1809                 .may_unmap = 1,
1810                 .may_swap = 1,
1811                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1812                 .swappiness = vm_swappiness,
1813                 .order = order,
1814                 .mem_cgroup = NULL,
1815                 .isolate_pages = isolate_pages_global,
1816         };
1817         /*
1818          * temp_priority is used to remember the scanning priority at which
1819          * this zone was successfully refilled to
1820          * free_pages == high_wmark_pages(zone).
1821          */
1822         int temp_priority[MAX_NR_ZONES];
1823
1824 loop_again:
1825         total_scanned = 0;
1826         sc.nr_reclaimed = 0;
1827         sc.may_writepage = !laptop_mode;
1828         count_vm_event(PAGEOUTRUN);
1829
1830         for (i = 0; i < pgdat->nr_zones; i++)
1831                 temp_priority[i] = DEF_PRIORITY;
1832
1833         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1834                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1835                 unsigned long lru_pages = 0;
1836
1837                 /* The swap token gets in the way of swapout... */
1838                 if (!priority)
1839                         disable_swap_token();
1840
1841                 all_zones_ok = 1;
1842
1843                 /*
1844                  * Scan in the highmem->dma direction for the highest
1845                  * zone which needs scanning
1846                  */
1847                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1848                         struct zone *zone = pgdat->node_zones + i;
1849
1850                         if (!populated_zone(zone))
1851                                 continue;
1852
1853                         if (zone_is_all_unreclaimable(zone) &&
1854                             priority != DEF_PRIORITY)
1855                                 continue;
1856
1857                         /*
1858                          * Do some background aging of the anon list, to give
1859                          * pages a chance to be referenced before reclaiming.
1860                          */
1861                         if (inactive_anon_is_low(zone, &sc))
1862                                 shrink_active_list(SWAP_CLUSTER_MAX, zone,
1863                                                         &sc, priority, 0);
1864
1865                         if (!zone_watermark_ok(zone, order,
1866                                         high_wmark_pages(zone), 0, 0)) {
1867                                 end_zone = i;
1868                                 break;
1869                         }
1870                 }
1871                 if (i < 0)
1872                         goto out;
1873
1874                 for (i = 0; i <= end_zone; i++) {
1875                         struct zone *zone = pgdat->node_zones + i;
1876
1877                         lru_pages += zone_lru_pages(zone);
1878                 }
1879
1880                 /*
1881                  * Now scan the zone in the dma->highmem direction, stopping
1882                  * at the last zone which needs scanning.
1883                  *
1884                  * We do this because the page allocator works in the opposite
1885                  * direction.  This prevents the page allocator from allocating
1886                  * pages behind kswapd's direction of progress, which would
1887                  * cause too much scanning of the lower zones.
1888                  */
1889                 for (i = 0; i <= end_zone; i++) {
1890                         struct zone *zone = pgdat->node_zones + i;
1891                         int nr_slab;
1892
1893                         if (!populated_zone(zone))
1894                                 continue;
1895
1896                         if (zone_is_all_unreclaimable(zone) &&
1897                                         priority != DEF_PRIORITY)
1898                                 continue;
1899
1900                         if (!zone_watermark_ok(zone, order,
1901                                         high_wmark_pages(zone), end_zone, 0))
1902                                 all_zones_ok = 0;
1903                         temp_priority[i] = priority;
1904                         sc.nr_scanned = 0;
1905                         note_zone_scanning_priority(zone, priority);
1906                         /*
1907                          * We put equal pressure on every zone, unless one
1908                          * zone has way too many pages free already.
1909                          */
1910                         if (!zone_watermark_ok(zone, order,
1911                                         8*high_wmark_pages(zone), end_zone, 0))
1912                                 shrink_zone(priority, zone, &sc);
1913                         reclaim_state->reclaimed_slab = 0;
1914                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1915                                                 lru_pages);
1916                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1917                         total_scanned += sc.nr_scanned;
1918                         if (zone_is_all_unreclaimable(zone))
1919                                 continue;
1920                         if (nr_slab == 0 && zone->pages_scanned >=
1921                                                 (zone_lru_pages(zone) * 6))
1922                                         zone_set_flag(zone,
1923                                                       ZONE_ALL_UNRECLAIMABLE);
1924                         /*
1925                          * If we've done a decent amount of scanning and
1926                          * the reclaim ratio is low, start doing writepage
1927                          * even in laptop mode
1928                          */
1929                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1930                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
1931                                 sc.may_writepage = 1;
1932                 }
1933                 if (all_zones_ok)
1934                         break;          /* kswapd: all done */
1935                 /*
1936                  * OK, kswapd is getting into trouble.  Take a nap, then take
1937                  * another pass across the zones.
1938                  */
1939                 if (total_scanned && priority < DEF_PRIORITY - 2)
1940                         congestion_wait(WRITE, HZ/10);
1941
1942                 /*
1943                  * We do this so kswapd doesn't build up large priorities for
1944                  * example when it is freeing in parallel with allocators. It
1945                  * matches the direct reclaim path behaviour in terms of impact
1946                  * on zone->*_priority.
1947                  */
1948                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
1949                         break;
1950         }
1951 out:
1952         /*
1953          * Note within each zone the priority level at which this zone was
1954          * brought into a happy state.  So that the next thread which scans this
1955          * zone will start out at that priority level.
1956          */
1957         for (i = 0; i < pgdat->nr_zones; i++) {
1958                 struct zone *zone = pgdat->node_zones + i;
1959
1960                 zone->prev_priority = temp_priority[i];
1961         }
1962         if (!all_zones_ok) {
1963                 cond_resched();
1964
1965                 try_to_freeze();
1966
1967                 /*
1968                  * Fragmentation may mean that the system cannot be
1969                  * rebalanced for high-order allocations in all zones.
1970                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
1971                  * it means the zones have been fully scanned and are still
1972                  * not balanced. For high-order allocations, there is
1973                  * little point trying all over again as kswapd may
1974                  * infinite loop.
1975                  *
1976                  * Instead, recheck all watermarks at order-0 as they
1977                  * are the most important. If watermarks are ok, kswapd will go
1978                  * back to sleep. High-order users can still perform direct
1979                  * reclaim if they wish.
1980                  */
1981                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
1982                         order = sc.order = 0;
1983
1984                 goto loop_again;
1985         }
1986
1987         return sc.nr_reclaimed;
1988 }
1989
1990 /*
1991  * The background pageout daemon, started as a kernel thread
1992  * from the init process.
1993  *
1994  * This basically trickles out pages so that we have _some_
1995  * free memory available even if there is no other activity
1996  * that frees anything up. This is needed for things like routing
1997  * etc, where we otherwise might have all activity going on in
1998  * asynchronous contexts that cannot page things out.
1999  *
2000  * If there are applications that are active memory-allocators
2001  * (most normal use), this basically shouldn't matter.
2002  */
2003 static int kswapd(void *p)
2004 {
2005         unsigned long order;
2006         pg_data_t *pgdat = (pg_data_t*)p;
2007         struct task_struct *tsk = current;
2008         DEFINE_WAIT(wait);
2009         struct reclaim_state reclaim_state = {
2010                 .reclaimed_slab = 0,
2011         };
2012         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2013
2014         lockdep_set_current_reclaim_state(GFP_KERNEL);
2015
2016         if (!cpumask_empty(cpumask))
2017                 set_cpus_allowed_ptr(tsk, cpumask);
2018         current->reclaim_state = &reclaim_state;
2019
2020         /*
2021          * Tell the memory management that we're a "memory allocator",
2022          * and that if we need more memory we should get access to it
2023          * regardless (see "__alloc_pages()"). "kswapd" should
2024          * never get caught in the normal page freeing logic.
2025          *
2026          * (Kswapd normally doesn't need memory anyway, but sometimes
2027          * you need a small amount of memory in order to be able to
2028          * page out something else, and this flag essentially protects
2029          * us from recursively trying to free more memory as we're
2030          * trying to free the first piece of memory in the first place).
2031          */
2032         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2033         set_freezable();
2034
2035         order = 0;
2036         for ( ; ; ) {
2037                 unsigned long new_order;
2038
2039                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2040                 new_order = pgdat->kswapd_max_order;
2041                 pgdat->kswapd_max_order = 0;
2042                 if (order < new_order) {
2043                         /*
2044                          * Don't sleep if someone wants a larger 'order'
2045                          * allocation
2046                          */
2047                         order = new_order;
2048                 } else {
2049                         if (!freezing(current))
2050                                 schedule();
2051
2052                         order = pgdat->kswapd_max_order;
2053                 }
2054                 finish_wait(&pgdat->kswapd_wait, &wait);
2055
2056                 if (!try_to_freeze()) {
2057                         /* We can speed up thawing tasks if we don't call
2058                          * balance_pgdat after returning from the refrigerator
2059                          */
2060                         balance_pgdat(pgdat, order);
2061                 }
2062         }
2063         return 0;
2064 }
2065
2066 /*
2067  * A zone is low on free memory, so wake its kswapd task to service it.
2068  */
2069 void wakeup_kswapd(struct zone *zone, int order)
2070 {
2071         pg_data_t *pgdat;
2072
2073         if (!populated_zone(zone))
2074                 return;
2075
2076         pgdat = zone->zone_pgdat;
2077         if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
2078                 return;
2079         if (pgdat->kswapd_max_order < order)
2080                 pgdat->kswapd_max_order = order;
2081         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2082                 return;
2083         if (!waitqueue_active(&pgdat->kswapd_wait))
2084                 return;
2085         wake_up_interruptible(&pgdat->kswapd_wait);
2086 }
2087
2088 unsigned long global_lru_pages(void)
2089 {
2090         return global_page_state(NR_ACTIVE_ANON)
2091                 + global_page_state(NR_ACTIVE_FILE)
2092                 + global_page_state(NR_INACTIVE_ANON)
2093                 + global_page_state(NR_INACTIVE_FILE);
2094 }
2095
2096 #ifdef CONFIG_HIBERNATION
2097 /*
2098  * Helper function for shrink_all_memory().  Tries to reclaim 'nr_pages' pages
2099  * from LRU lists system-wide, for given pass and priority.
2100  *
2101  * For pass > 3 we also try to shrink the LRU lists that contain a few pages
2102  */
2103 static void shrink_all_zones(unsigned long nr_pages, int prio,
2104                                       int pass, struct scan_control *sc)
2105 {
2106         struct zone *zone;
2107         unsigned long nr_reclaimed = 0;
2108
2109         for_each_populated_zone(zone) {
2110                 enum lru_list l;
2111
2112                 if (zone_is_all_unreclaimable(zone) && prio != DEF_PRIORITY)
2113                         continue;
2114
2115                 for_each_evictable_lru(l) {
2116                         enum zone_stat_item ls = NR_LRU_BASE + l;
2117                         unsigned long lru_pages = zone_page_state(zone, ls);
2118
2119                         /* For pass = 0, we don't shrink the active list */
2120                         if (pass == 0 && (l == LRU_ACTIVE_ANON ||
2121                                                 l == LRU_ACTIVE_FILE))
2122                                 continue;
2123
2124                         zone->lru[l].nr_saved_scan += (lru_pages >> prio) + 1;
2125                         if (zone->lru[l].nr_saved_scan >= nr_pages || pass > 3) {
2126                                 unsigned long nr_to_scan;
2127
2128                                 zone->lru[l].nr_saved_scan = 0;
2129                                 nr_to_scan = min(nr_pages, lru_pages);
2130                                 nr_reclaimed += shrink_list(l, nr_to_scan, zone,
2131                                                                 sc, prio);
2132                                 if (nr_reclaimed >= nr_pages) {
2133                                         sc->nr_reclaimed += nr_reclaimed;
2134                                         return;
2135                                 }
2136                         }
2137                 }
2138         }
2139         sc->nr_reclaimed += nr_reclaimed;
2140 }
2141
2142 /*
2143  * Try to free `nr_pages' of memory, system-wide, and return the number of
2144  * freed pages.
2145  *
2146  * Rather than trying to age LRUs the aim is to preserve the overall
2147  * LRU order by reclaiming preferentially
2148  * inactive > active > active referenced > active mapped
2149  */
2150 unsigned long shrink_all_memory(unsigned long nr_pages)
2151 {
2152         unsigned long lru_pages, nr_slab;
2153         int pass;
2154         struct reclaim_state reclaim_state;
2155         struct scan_control sc = {
2156                 .gfp_mask = GFP_KERNEL,
2157                 .may_unmap = 0,
2158                 .may_writepage = 1,
2159                 .isolate_pages = isolate_pages_global,
2160                 .nr_reclaimed = 0,
2161         };
2162
2163         current->reclaim_state = &reclaim_state;
2164
2165         lru_pages = global_lru_pages();
2166         nr_slab = global_page_state(NR_SLAB_RECLAIMABLE);
2167         /* If slab caches are huge, it's better to hit them first */
2168         while (nr_slab >= lru_pages) {
2169                 reclaim_state.reclaimed_slab = 0;
2170                 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
2171                 if (!reclaim_state.reclaimed_slab)
2172                         break;
2173
2174                 sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2175                 if (sc.nr_reclaimed >= nr_pages)
2176                         goto out;
2177
2178                 nr_slab -= reclaim_state.reclaimed_slab;
2179         }
2180
2181         /*
2182          * We try to shrink LRUs in 5 passes:
2183          * 0 = Reclaim from inactive_list only
2184          * 1 = Reclaim from active list but don't reclaim mapped
2185          * 2 = 2nd pass of type 1
2186          * 3 = Reclaim mapped (normal reclaim)
2187          * 4 = 2nd pass of type 3
2188          */
2189         for (pass = 0; pass < 5; pass++) {
2190                 int prio;
2191
2192                 /* Force reclaiming mapped pages in the passes #3 and #4 */
2193                 if (pass > 2)
2194                         sc.may_unmap = 1;
2195
2196                 for (prio = DEF_PRIORITY; prio >= 0; prio--) {
2197                         unsigned long nr_to_scan = nr_pages - sc.nr_reclaimed;
2198
2199                         sc.nr_scanned = 0;
2200                         sc.swap_cluster_max = nr_to_scan;
2201                         shrink_all_zones(nr_to_scan, prio, pass, &sc);
2202                         if (sc.nr_reclaimed >= nr_pages)
2203                                 goto out;
2204
2205                         reclaim_state.reclaimed_slab = 0;
2206                         shrink_slab(sc.nr_scanned, sc.gfp_mask,
2207                                         global_lru_pages());
2208                         sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2209                         if (sc.nr_reclaimed >= nr_pages)
2210                                 goto out;
2211
2212                         if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
2213                                 congestion_wait(WRITE, HZ / 10);
2214                 }
2215         }
2216
2217         /*
2218          * If sc.nr_reclaimed = 0, we could not shrink LRUs, but there may be
2219          * something in slab caches
2220          */
2221         if (!sc.nr_reclaimed) {
2222                 do {
2223                         reclaim_state.reclaimed_slab = 0;
2224                         shrink_slab(nr_pages, sc.gfp_mask, global_lru_pages());
2225                         sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2226                 } while (sc.nr_reclaimed < nr_pages &&
2227                                 reclaim_state.reclaimed_slab > 0);
2228         }
2229
2230
2231 out:
2232         current->reclaim_state = NULL;
2233
2234         return sc.nr_reclaimed;
2235 }
2236 #endif /* CONFIG_HIBERNATION */
2237
2238 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2239    not required for correctness.  So if the last cpu in a node goes
2240    away, we get changed to run anywhere: as the first one comes back,
2241    restore their cpu bindings. */
2242 static int __devinit cpu_callback(struct notifier_block *nfb,
2243                                   unsigned long action, void *hcpu)
2244 {
2245         int nid;
2246
2247         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2248                 for_each_node_state(nid, N_HIGH_MEMORY) {
2249                         pg_data_t *pgdat = NODE_DATA(nid);
2250                         const struct cpumask *mask;
2251
2252                         mask = cpumask_of_node(pgdat->node_id);
2253
2254                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2255                                 /* One of our CPUs online: restore mask */
2256                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2257                 }
2258         }
2259         return NOTIFY_OK;
2260 }
2261
2262 /*
2263  * This kswapd start function will be called by init and node-hot-add.
2264  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2265  */
2266 int kswapd_run(int nid)
2267 {
2268         pg_data_t *pgdat = NODE_DATA(nid);
2269         int ret = 0;
2270
2271         if (pgdat->kswapd)
2272                 return 0;
2273
2274         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2275         if (IS_ERR(pgdat->kswapd)) {
2276                 /* failure at boot is fatal */
2277                 BUG_ON(system_state == SYSTEM_BOOTING);
2278                 printk("Failed to start kswapd on node %d\n",nid);
2279                 ret = -1;
2280         }
2281         return ret;
2282 }
2283
2284 static int __init kswapd_init(void)
2285 {
2286         int nid;
2287
2288         swap_setup();
2289         for_each_node_state(nid, N_HIGH_MEMORY)
2290                 kswapd_run(nid);
2291         hotcpu_notifier(cpu_callback, 0);
2292         return 0;
2293 }
2294
2295 module_init(kswapd_init)
2296
2297 #ifdef CONFIG_NUMA
2298 /*
2299  * Zone reclaim mode
2300  *
2301  * If non-zero call zone_reclaim when the number of free pages falls below
2302  * the watermarks.
2303  */
2304 int zone_reclaim_mode __read_mostly;
2305
2306 #define RECLAIM_OFF 0
2307 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
2308 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
2309 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
2310
2311 /*
2312  * Priority for ZONE_RECLAIM. This determines the fraction of pages
2313  * of a node considered for each zone_reclaim. 4 scans 1/16th of
2314  * a zone.
2315  */
2316 #define ZONE_RECLAIM_PRIORITY 4
2317
2318 /*
2319  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2320  * occur.
2321  */
2322 int sysctl_min_unmapped_ratio = 1;
2323
2324 /*
2325  * If the number of slab pages in a zone grows beyond this percentage then
2326  * slab reclaim needs to occur.
2327  */
2328 int sysctl_min_slab_ratio = 5;
2329
2330 /*
2331  * Try to free up some pages from this zone through reclaim.
2332  */
2333 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2334 {
2335         /* Minimum pages needed in order to stay on node */
2336         const unsigned long nr_pages = 1 << order;
2337         struct task_struct *p = current;
2338         struct reclaim_state reclaim_state;
2339         int priority;
2340         struct scan_control sc = {
2341                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2342                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2343                 .may_swap = 1,
2344                 .swap_cluster_max = max_t(unsigned long, nr_pages,
2345                                         SWAP_CLUSTER_MAX),
2346                 .gfp_mask = gfp_mask,
2347                 .swappiness = vm_swappiness,
2348                 .order = order,
2349                 .isolate_pages = isolate_pages_global,
2350         };
2351         unsigned long slab_reclaimable;
2352
2353         disable_swap_token();
2354         cond_resched();
2355         /*
2356          * We need to be able to allocate from the reserves for RECLAIM_SWAP
2357          * and we also need to be able to write out pages for RECLAIM_WRITE
2358          * and RECLAIM_SWAP.
2359          */
2360         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2361         reclaim_state.reclaimed_slab = 0;
2362         p->reclaim_state = &reclaim_state;
2363
2364         if (zone_page_state(zone, NR_FILE_PAGES) -
2365                 zone_page_state(zone, NR_FILE_MAPPED) >
2366                 zone->min_unmapped_pages) {
2367                 /*
2368                  * Free memory by calling shrink zone with increasing
2369                  * priorities until we have enough memory freed.
2370                  */
2371                 priority = ZONE_RECLAIM_PRIORITY;
2372                 do {
2373                         note_zone_scanning_priority(zone, priority);
2374                         shrink_zone(priority, zone, &sc);
2375                         priority--;
2376                 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
2377         }
2378
2379         slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2380         if (slab_reclaimable > zone->min_slab_pages) {
2381                 /*
2382                  * shrink_slab() does not currently allow us to determine how
2383                  * many pages were freed in this zone. So we take the current
2384                  * number of slab pages and shake the slab until it is reduced
2385                  * by the same nr_pages that we used for reclaiming unmapped
2386                  * pages.
2387                  *
2388                  * Note that shrink_slab will free memory on all zones and may
2389                  * take a long time.
2390                  */
2391                 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
2392                         zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
2393                                 slab_reclaimable - nr_pages)
2394                         ;
2395
2396                 /*
2397                  * Update nr_reclaimed by the number of slab pages we
2398                  * reclaimed from this zone.
2399                  */
2400                 sc.nr_reclaimed += slab_reclaimable -
2401                         zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2402         }
2403
2404         p->reclaim_state = NULL;
2405         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
2406         return sc.nr_reclaimed >= nr_pages;
2407 }
2408
2409 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2410 {
2411         int node_id;
2412         int ret;
2413
2414         /*
2415          * Zone reclaim reclaims unmapped file backed pages and
2416          * slab pages if we are over the defined limits.
2417          *
2418          * A small portion of unmapped file backed pages is needed for
2419          * file I/O otherwise pages read by file I/O will be immediately
2420          * thrown out if the zone is overallocated. So we do not reclaim
2421          * if less than a specified percentage of the zone is used by
2422          * unmapped file backed pages.
2423          */
2424         if (zone_page_state(zone, NR_FILE_PAGES) -
2425             zone_page_state(zone, NR_FILE_MAPPED) <= zone->min_unmapped_pages
2426             && zone_page_state(zone, NR_SLAB_RECLAIMABLE)
2427                         <= zone->min_slab_pages)
2428                 return 0;
2429
2430         if (zone_is_all_unreclaimable(zone))
2431                 return 0;
2432
2433         /*
2434          * Do not scan if the allocation should not be delayed.
2435          */
2436         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
2437                         return 0;
2438
2439         /*
2440          * Only run zone reclaim on the local zone or on zones that do not
2441          * have associated processors. This will favor the local processor
2442          * over remote processors and spread off node memory allocations
2443          * as wide as possible.
2444          */
2445         node_id = zone_to_nid(zone);
2446         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
2447                 return 0;
2448
2449         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2450                 return 0;
2451         ret = __zone_reclaim(zone, gfp_mask, order);
2452         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2453
2454         return ret;
2455 }
2456 #endif
2457
2458 /*
2459  * page_evictable - test whether a page is evictable
2460  * @page: the page to test
2461  * @vma: the VMA in which the page is or will be mapped, may be NULL
2462  *
2463  * Test whether page is evictable--i.e., should be placed on active/inactive
2464  * lists vs unevictable list.  The vma argument is !NULL when called from the
2465  * fault path to determine how to instantate a new page.
2466  *
2467  * Reasons page might not be evictable:
2468  * (1) page's mapping marked unevictable
2469  * (2) page is part of an mlocked VMA
2470  *
2471  */
2472 int page_evictable(struct page *page, struct vm_area_struct *vma)
2473 {
2474
2475         if (mapping_unevictable(page_mapping(page)))
2476                 return 0;
2477
2478         if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2479                 return 0;
2480
2481         return 1;
2482 }
2483
2484 /**
2485  * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2486  * @page: page to check evictability and move to appropriate lru list
2487  * @zone: zone page is in
2488  *
2489  * Checks a page for evictability and moves the page to the appropriate
2490  * zone lru list.
2491  *
2492  * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2493  * have PageUnevictable set.
2494  */
2495 static void check_move_unevictable_page(struct page *page, struct zone *zone)
2496 {
2497         VM_BUG_ON(PageActive(page));
2498
2499 retry:
2500         ClearPageUnevictable(page);
2501         if (page_evictable(page, NULL)) {
2502                 enum lru_list l = LRU_INACTIVE_ANON + page_is_file_cache(page);
2503
2504                 __dec_zone_state(zone, NR_UNEVICTABLE);
2505                 list_move(&page->lru, &zone->lru[l].list);
2506                 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
2507                 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
2508                 __count_vm_event(UNEVICTABLE_PGRESCUED);
2509         } else {
2510                 /*
2511                  * rotate unevictable list
2512                  */
2513                 SetPageUnevictable(page);
2514                 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
2515                 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
2516                 if (page_evictable(page, NULL))
2517                         goto retry;
2518         }
2519 }
2520
2521 /**
2522  * scan_mapping_unevictable_pages - scan an address space for evictable pages
2523  * @mapping: struct address_space to scan for evictable pages
2524  *
2525  * Scan all pages in mapping.  Check unevictable pages for
2526  * evictability and move them to the appropriate zone lru list.
2527  */
2528 void scan_mapping_unevictable_pages(struct address_space *mapping)
2529 {
2530         pgoff_t next = 0;
2531         pgoff_t end   = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
2532                          PAGE_CACHE_SHIFT;
2533         struct zone *zone;
2534         struct pagevec pvec;
2535
2536         if (mapping->nrpages == 0)
2537                 return;
2538
2539         pagevec_init(&pvec, 0);
2540         while (next < end &&
2541                 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
2542                 int i;
2543                 int pg_scanned = 0;
2544
2545                 zone = NULL;
2546
2547                 for (i = 0; i < pagevec_count(&pvec); i++) {
2548                         struct page *page = pvec.pages[i];
2549                         pgoff_t page_index = page->index;
2550                         struct zone *pagezone = page_zone(page);
2551
2552                         pg_scanned++;
2553                         if (page_index > next)
2554                                 next = page_index;
2555                         next++;
2556
2557                         if (pagezone != zone) {
2558                                 if (zone)
2559                                         spin_unlock_irq(&zone->lru_lock);
2560                                 zone = pagezone;
2561                                 spin_lock_irq(&zone->lru_lock);
2562                         }
2563
2564                         if (PageLRU(page) && PageUnevictable(page))
2565                                 check_move_unevictable_page(page, zone);
2566                 }
2567                 if (zone)
2568                         spin_unlock_irq(&zone->lru_lock);
2569                 pagevec_release(&pvec);
2570
2571                 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
2572         }
2573
2574 }
2575
2576 /**
2577  * scan_zone_unevictable_pages - check unevictable list for evictable pages
2578  * @zone - zone of which to scan the unevictable list
2579  *
2580  * Scan @zone's unevictable LRU lists to check for pages that have become
2581  * evictable.  Move those that have to @zone's inactive list where they
2582  * become candidates for reclaim, unless shrink_inactive_zone() decides
2583  * to reactivate them.  Pages that are still unevictable are rotated
2584  * back onto @zone's unevictable list.
2585  */
2586 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
2587 static void scan_zone_unevictable_pages(struct zone *zone)
2588 {
2589         struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
2590         unsigned long scan;
2591         unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
2592
2593         while (nr_to_scan > 0) {
2594                 unsigned long batch_size = min(nr_to_scan,
2595                                                 SCAN_UNEVICTABLE_BATCH_SIZE);
2596
2597                 spin_lock_irq(&zone->lru_lock);
2598                 for (scan = 0;  scan < batch_size; scan++) {
2599                         struct page *page = lru_to_page(l_unevictable);
2600
2601                         if (!trylock_page(page))
2602                                 continue;
2603
2604                         prefetchw_prev_lru_page(page, l_unevictable, flags);
2605
2606                         if (likely(PageLRU(page) && PageUnevictable(page)))
2607                                 check_move_unevictable_page(page, zone);
2608
2609                         unlock_page(page);
2610                 }
2611                 spin_unlock_irq(&zone->lru_lock);
2612
2613                 nr_to_scan -= batch_size;
2614         }
2615 }
2616
2617
2618 /**
2619  * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
2620  *
2621  * A really big hammer:  scan all zones' unevictable LRU lists to check for
2622  * pages that have become evictable.  Move those back to the zones'
2623  * inactive list where they become candidates for reclaim.
2624  * This occurs when, e.g., we have unswappable pages on the unevictable lists,
2625  * and we add swap to the system.  As such, it runs in the context of a task
2626  * that has possibly/probably made some previously unevictable pages
2627  * evictable.
2628  */
2629 static void scan_all_zones_unevictable_pages(void)
2630 {
2631         struct zone *zone;
2632
2633         for_each_zone(zone) {
2634                 scan_zone_unevictable_pages(zone);
2635         }
2636 }
2637
2638 /*
2639  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
2640  * all nodes' unevictable lists for evictable pages
2641  */
2642 unsigned long scan_unevictable_pages;
2643
2644 int scan_unevictable_handler(struct ctl_table *table, int write,
2645                            struct file *file, void __user *buffer,
2646                            size_t *length, loff_t *ppos)
2647 {
2648         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
2649
2650         if (write && *(unsigned long *)table->data)
2651                 scan_all_zones_unevictable_pages();
2652
2653         scan_unevictable_pages = 0;
2654         return 0;
2655 }
2656
2657 /*
2658  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
2659  * a specified node's per zone unevictable lists for evictable pages.
2660  */
2661
2662 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
2663                                           struct sysdev_attribute *attr,
2664                                           char *buf)
2665 {
2666         return sprintf(buf, "0\n");     /* always zero; should fit... */
2667 }
2668
2669 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
2670                                            struct sysdev_attribute *attr,
2671                                         const char *buf, size_t count)
2672 {
2673         struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
2674         struct zone *zone;
2675         unsigned long res;
2676         unsigned long req = strict_strtoul(buf, 10, &res);
2677
2678         if (!req)
2679                 return 1;       /* zero is no-op */
2680
2681         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2682                 if (!populated_zone(zone))
2683                         continue;
2684                 scan_zone_unevictable_pages(zone);
2685         }
2686         return 1;
2687 }
2688
2689
2690 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
2691                         read_scan_unevictable_node,
2692                         write_scan_unevictable_node);
2693
2694 int scan_unevictable_register_node(struct node *node)
2695 {
2696         return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
2697 }
2698
2699 void scan_unevictable_unregister_node(struct node *node)
2700 {
2701         sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
2702 }
2703