mm: vmstat: add isolate pages
[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                 swapcache_free(swap, page);
474         } else {
475                 __remove_from_page_cache(page);
476                 spin_unlock_irq(&mapping->tree_lock);
477                 mem_cgroup_uncharge_cache_page(page);
478         }
479
480         return 1;
481
482 cannot_free:
483         spin_unlock_irq(&mapping->tree_lock);
484         return 0;
485 }
486
487 /*
488  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
489  * someone else has a ref on the page, abort and return 0.  If it was
490  * successfully detached, return 1.  Assumes the caller has a single ref on
491  * this page.
492  */
493 int remove_mapping(struct address_space *mapping, struct page *page)
494 {
495         if (__remove_mapping(mapping, page)) {
496                 /*
497                  * Unfreezing the refcount with 1 rather than 2 effectively
498                  * drops the pagecache ref for us without requiring another
499                  * atomic operation.
500                  */
501                 page_unfreeze_refs(page, 1);
502                 return 1;
503         }
504         return 0;
505 }
506
507 /**
508  * putback_lru_page - put previously isolated page onto appropriate LRU list
509  * @page: page to be put back to appropriate lru list
510  *
511  * Add previously isolated @page to appropriate LRU list.
512  * Page may still be unevictable for other reasons.
513  *
514  * lru_lock must not be held, interrupts must be enabled.
515  */
516 void putback_lru_page(struct page *page)
517 {
518         int lru;
519         int active = !!TestClearPageActive(page);
520         int was_unevictable = PageUnevictable(page);
521
522         VM_BUG_ON(PageLRU(page));
523
524 redo:
525         ClearPageUnevictable(page);
526
527         if (page_evictable(page, NULL)) {
528                 /*
529                  * For evictable pages, we can use the cache.
530                  * In event of a race, worst case is we end up with an
531                  * unevictable page on [in]active list.
532                  * We know how to handle that.
533                  */
534                 lru = active + page_is_file_cache(page);
535                 lru_cache_add_lru(page, lru);
536         } else {
537                 /*
538                  * Put unevictable pages directly on zone's unevictable
539                  * list.
540                  */
541                 lru = LRU_UNEVICTABLE;
542                 add_page_to_unevictable_list(page);
543         }
544
545         /*
546          * page's status can change while we move it among lru. If an evictable
547          * page is on unevictable list, it never be freed. To avoid that,
548          * check after we added it to the list, again.
549          */
550         if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
551                 if (!isolate_lru_page(page)) {
552                         put_page(page);
553                         goto redo;
554                 }
555                 /* This means someone else dropped this page from LRU
556                  * So, it will be freed or putback to LRU again. There is
557                  * nothing to do here.
558                  */
559         }
560
561         if (was_unevictable && lru != LRU_UNEVICTABLE)
562                 count_vm_event(UNEVICTABLE_PGRESCUED);
563         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
564                 count_vm_event(UNEVICTABLE_PGCULLED);
565
566         put_page(page);         /* drop ref from isolate */
567 }
568
569 /*
570  * shrink_page_list() returns the number of reclaimed pages
571  */
572 static unsigned long shrink_page_list(struct list_head *page_list,
573                                         struct scan_control *sc,
574                                         enum pageout_io sync_writeback)
575 {
576         LIST_HEAD(ret_pages);
577         struct pagevec freed_pvec;
578         int pgactivate = 0;
579         unsigned long nr_reclaimed = 0;
580         unsigned long vm_flags;
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,
632                                                 sc->mem_cgroup, &vm_flags);
633                 /*
634                  * In active use or really unfreeable?  Activate it.
635                  * If page which have PG_mlocked lost isoltation race,
636                  * try_to_unmap moves it to unevictable list
637                  */
638                 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER &&
639                                         referenced && page_mapping_inuse(page)
640                                         && !(vm_flags & VM_LOCKED))
641                         goto activate_locked;
642
643                 /*
644                  * Anonymous process memory has backing store?
645                  * Try to allocate it some swap space here.
646                  */
647                 if (PageAnon(page) && !PageSwapCache(page)) {
648                         if (!(sc->gfp_mask & __GFP_IO))
649                                 goto keep_locked;
650                         if (!add_to_swap(page))
651                                 goto activate_locked;
652                         may_enter_fs = 1;
653                 }
654
655                 mapping = page_mapping(page);
656
657                 /*
658                  * The page is mapped into the page tables of one or more
659                  * processes. Try to unmap it here.
660                  */
661                 if (page_mapped(page) && mapping) {
662                         switch (try_to_unmap(page, 0)) {
663                         case SWAP_FAIL:
664                                 goto activate_locked;
665                         case SWAP_AGAIN:
666                                 goto keep_locked;
667                         case SWAP_MLOCK:
668                                 goto cull_mlocked;
669                         case SWAP_SUCCESS:
670                                 ; /* try to free the page below */
671                         }
672                 }
673
674                 if (PageDirty(page)) {
675                         if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && referenced)
676                                 goto keep_locked;
677                         if (!may_enter_fs)
678                                 goto keep_locked;
679                         if (!sc->may_writepage)
680                                 goto keep_locked;
681
682                         /* Page is dirty, try to write it out here */
683                         switch (pageout(page, mapping, sync_writeback)) {
684                         case PAGE_KEEP:
685                                 goto keep_locked;
686                         case PAGE_ACTIVATE:
687                                 goto activate_locked;
688                         case PAGE_SUCCESS:
689                                 if (PageWriteback(page) || PageDirty(page))
690                                         goto keep;
691                                 /*
692                                  * A synchronous write - probably a ramdisk.  Go
693                                  * ahead and try to reclaim the page.
694                                  */
695                                 if (!trylock_page(page))
696                                         goto keep;
697                                 if (PageDirty(page) || PageWriteback(page))
698                                         goto keep_locked;
699                                 mapping = page_mapping(page);
700                         case PAGE_CLEAN:
701                                 ; /* try to free the page below */
702                         }
703                 }
704
705                 /*
706                  * If the page has buffers, try to free the buffer mappings
707                  * associated with this page. If we succeed we try to free
708                  * the page as well.
709                  *
710                  * We do this even if the page is PageDirty().
711                  * try_to_release_page() does not perform I/O, but it is
712                  * possible for a page to have PageDirty set, but it is actually
713                  * clean (all its buffers are clean).  This happens if the
714                  * buffers were written out directly, with submit_bh(). ext3
715                  * will do this, as well as the blockdev mapping.
716                  * try_to_release_page() will discover that cleanness and will
717                  * drop the buffers and mark the page clean - it can be freed.
718                  *
719                  * Rarely, pages can have buffers and no ->mapping.  These are
720                  * the pages which were not successfully invalidated in
721                  * truncate_complete_page().  We try to drop those buffers here
722                  * and if that worked, and the page is no longer mapped into
723                  * process address space (page_count == 1) it can be freed.
724                  * Otherwise, leave the page on the LRU so it is swappable.
725                  */
726                 if (page_has_private(page)) {
727                         if (!try_to_release_page(page, sc->gfp_mask))
728                                 goto activate_locked;
729                         if (!mapping && page_count(page) == 1) {
730                                 unlock_page(page);
731                                 if (put_page_testzero(page))
732                                         goto free_it;
733                                 else {
734                                         /*
735                                          * rare race with speculative reference.
736                                          * the speculative reference will free
737                                          * this page shortly, so we may
738                                          * increment nr_reclaimed here (and
739                                          * leave it off the LRU).
740                                          */
741                                         nr_reclaimed++;
742                                         continue;
743                                 }
744                         }
745                 }
746
747                 if (!mapping || !__remove_mapping(mapping, page))
748                         goto keep_locked;
749
750                 /*
751                  * At this point, we have no other references and there is
752                  * no way to pick any more up (removed from LRU, removed
753                  * from pagecache). Can use non-atomic bitops now (and
754                  * we obviously don't have to worry about waking up a process
755                  * waiting on the page lock, because there are no references.
756                  */
757                 __clear_page_locked(page);
758 free_it:
759                 nr_reclaimed++;
760                 if (!pagevec_add(&freed_pvec, page)) {
761                         __pagevec_free(&freed_pvec);
762                         pagevec_reinit(&freed_pvec);
763                 }
764                 continue;
765
766 cull_mlocked:
767                 if (PageSwapCache(page))
768                         try_to_free_swap(page);
769                 unlock_page(page);
770                 putback_lru_page(page);
771                 continue;
772
773 activate_locked:
774                 /* Not a candidate for swapping, so reclaim swap space. */
775                 if (PageSwapCache(page) && vm_swap_full())
776                         try_to_free_swap(page);
777                 VM_BUG_ON(PageActive(page));
778                 SetPageActive(page);
779                 pgactivate++;
780 keep_locked:
781                 unlock_page(page);
782 keep:
783                 list_add(&page->lru, &ret_pages);
784                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
785         }
786         list_splice(&ret_pages, page_list);
787         if (pagevec_count(&freed_pvec))
788                 __pagevec_free(&freed_pvec);
789         count_vm_events(PGACTIVATE, pgactivate);
790         return nr_reclaimed;
791 }
792
793 /* LRU Isolation modes. */
794 #define ISOLATE_INACTIVE 0      /* Isolate inactive pages. */
795 #define ISOLATE_ACTIVE 1        /* Isolate active pages. */
796 #define ISOLATE_BOTH 2          /* Isolate both active and inactive pages. */
797
798 /*
799  * Attempt to remove the specified page from its LRU.  Only take this page
800  * if it is of the appropriate PageActive status.  Pages which are being
801  * freed elsewhere are also ignored.
802  *
803  * page:        page to consider
804  * mode:        one of the LRU isolation modes defined above
805  *
806  * returns 0 on success, -ve errno on failure.
807  */
808 int __isolate_lru_page(struct page *page, int mode, int file)
809 {
810         int ret = -EINVAL;
811
812         /* Only take pages on the LRU. */
813         if (!PageLRU(page))
814                 return ret;
815
816         /*
817          * When checking the active state, we need to be sure we are
818          * dealing with comparible boolean values.  Take the logical not
819          * of each.
820          */
821         if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
822                 return ret;
823
824         if (mode != ISOLATE_BOTH && (!page_is_file_cache(page) != !file))
825                 return ret;
826
827         /*
828          * When this function is being called for lumpy reclaim, we
829          * initially look into all LRU pages, active, inactive and
830          * unevictable; only give shrink_page_list evictable pages.
831          */
832         if (PageUnevictable(page))
833                 return ret;
834
835         ret = -EBUSY;
836
837         if (likely(get_page_unless_zero(page))) {
838                 /*
839                  * Be careful not to clear PageLRU until after we're
840                  * sure the page is not being freed elsewhere -- the
841                  * page release code relies on it.
842                  */
843                 ClearPageLRU(page);
844                 ret = 0;
845         }
846
847         return ret;
848 }
849
850 /*
851  * zone->lru_lock is heavily contended.  Some of the functions that
852  * shrink the lists perform better by taking out a batch of pages
853  * and working on them outside the LRU lock.
854  *
855  * For pagecache intensive workloads, this function is the hottest
856  * spot in the kernel (apart from copy_*_user functions).
857  *
858  * Appropriate locks must be held before calling this function.
859  *
860  * @nr_to_scan: The number of pages to look through on the list.
861  * @src:        The LRU list to pull pages off.
862  * @dst:        The temp list to put pages on to.
863  * @scanned:    The number of pages that were scanned.
864  * @order:      The caller's attempted allocation order
865  * @mode:       One of the LRU isolation modes
866  * @file:       True [1] if isolating file [!anon] pages
867  *
868  * returns how many pages were moved onto *@dst.
869  */
870 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
871                 struct list_head *src, struct list_head *dst,
872                 unsigned long *scanned, int order, int mode, int file)
873 {
874         unsigned long nr_taken = 0;
875         unsigned long scan;
876
877         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
878                 struct page *page;
879                 unsigned long pfn;
880                 unsigned long end_pfn;
881                 unsigned long page_pfn;
882                 int zone_id;
883
884                 page = lru_to_page(src);
885                 prefetchw_prev_lru_page(page, src, flags);
886
887                 VM_BUG_ON(!PageLRU(page));
888
889                 switch (__isolate_lru_page(page, mode, file)) {
890                 case 0:
891                         list_move(&page->lru, dst);
892                         mem_cgroup_del_lru(page);
893                         nr_taken++;
894                         break;
895
896                 case -EBUSY:
897                         /* else it is being freed elsewhere */
898                         list_move(&page->lru, src);
899                         mem_cgroup_rotate_lru_list(page, page_lru(page));
900                         continue;
901
902                 default:
903                         BUG();
904                 }
905
906                 if (!order)
907                         continue;
908
909                 /*
910                  * Attempt to take all pages in the order aligned region
911                  * surrounding the tag page.  Only take those pages of
912                  * the same active state as that tag page.  We may safely
913                  * round the target page pfn down to the requested order
914                  * as the mem_map is guarenteed valid out to MAX_ORDER,
915                  * where that page is in a different zone we will detect
916                  * it from its zone id and abort this block scan.
917                  */
918                 zone_id = page_zone_id(page);
919                 page_pfn = page_to_pfn(page);
920                 pfn = page_pfn & ~((1 << order) - 1);
921                 end_pfn = pfn + (1 << order);
922                 for (; pfn < end_pfn; pfn++) {
923                         struct page *cursor_page;
924
925                         /* The target page is in the block, ignore it. */
926                         if (unlikely(pfn == page_pfn))
927                                 continue;
928
929                         /* Avoid holes within the zone. */
930                         if (unlikely(!pfn_valid_within(pfn)))
931                                 break;
932
933                         cursor_page = pfn_to_page(pfn);
934
935                         /* Check that we have not crossed a zone boundary. */
936                         if (unlikely(page_zone_id(cursor_page) != zone_id))
937                                 continue;
938                         if (__isolate_lru_page(cursor_page, mode, file) == 0) {
939                                 list_move(&cursor_page->lru, dst);
940                                 mem_cgroup_del_lru(cursor_page);
941                                 nr_taken++;
942                                 scan++;
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                 unsigned long nr_anon;
1076                 unsigned long nr_file;
1077
1078                 nr_taken = sc->isolate_pages(sc->swap_cluster_max,
1079                              &page_list, &nr_scan, sc->order, mode,
1080                                 zone, sc->mem_cgroup, 0, file);
1081
1082                 if (scanning_global_lru(sc)) {
1083                         zone->pages_scanned += nr_scan;
1084                         if (current_is_kswapd())
1085                                 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1086                                                        nr_scan);
1087                         else
1088                                 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1089                                                        nr_scan);
1090                 }
1091
1092                 if (nr_taken == 0)
1093                         goto done;
1094
1095                 nr_active = clear_active_flags(&page_list, count);
1096                 __count_vm_events(PGDEACTIVATE, nr_active);
1097
1098                 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1099                                                 -count[LRU_ACTIVE_FILE]);
1100                 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1101                                                 -count[LRU_INACTIVE_FILE]);
1102                 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1103                                                 -count[LRU_ACTIVE_ANON]);
1104                 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1105                                                 -count[LRU_INACTIVE_ANON]);
1106
1107                 nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1108                 nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1109                 __mod_zone_page_state(zone, NR_ISOLATED_ANON, nr_anon);
1110                 __mod_zone_page_state(zone, NR_ISOLATED_FILE, nr_file);
1111
1112                 reclaim_stat->recent_scanned[0] += count[LRU_INACTIVE_ANON];
1113                 reclaim_stat->recent_scanned[0] += count[LRU_ACTIVE_ANON];
1114                 reclaim_stat->recent_scanned[1] += count[LRU_INACTIVE_FILE];
1115                 reclaim_stat->recent_scanned[1] += count[LRU_ACTIVE_FILE];
1116
1117                 spin_unlock_irq(&zone->lru_lock);
1118
1119                 nr_scanned += nr_scan;
1120                 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
1121
1122                 /*
1123                  * If we are direct reclaiming for contiguous pages and we do
1124                  * not reclaim everything in the list, try again and wait
1125                  * for IO to complete. This will stall high-order allocations
1126                  * but that should be acceptable to the caller
1127                  */
1128                 if (nr_freed < nr_taken && !current_is_kswapd() &&
1129                     lumpy_reclaim) {
1130                         congestion_wait(BLK_RW_ASYNC, HZ/10);
1131
1132                         /*
1133                          * The attempt at page out may have made some
1134                          * of the pages active, mark them inactive again.
1135                          */
1136                         nr_active = clear_active_flags(&page_list, count);
1137                         count_vm_events(PGDEACTIVATE, nr_active);
1138
1139                         nr_freed += shrink_page_list(&page_list, sc,
1140                                                         PAGEOUT_IO_SYNC);
1141                 }
1142
1143                 nr_reclaimed += nr_freed;
1144
1145                 local_irq_disable();
1146                 if (current_is_kswapd())
1147                         __count_vm_events(KSWAPD_STEAL, nr_freed);
1148                 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
1149
1150                 spin_lock(&zone->lru_lock);
1151                 /*
1152                  * Put back any unfreeable pages.
1153                  */
1154                 while (!list_empty(&page_list)) {
1155                         int lru;
1156                         page = lru_to_page(&page_list);
1157                         VM_BUG_ON(PageLRU(page));
1158                         list_del(&page->lru);
1159                         if (unlikely(!page_evictable(page, NULL))) {
1160                                 spin_unlock_irq(&zone->lru_lock);
1161                                 putback_lru_page(page);
1162                                 spin_lock_irq(&zone->lru_lock);
1163                                 continue;
1164                         }
1165                         SetPageLRU(page);
1166                         lru = page_lru(page);
1167                         add_page_to_lru_list(zone, page, lru);
1168                         if (PageActive(page)) {
1169                                 int file = !!page_is_file_cache(page);
1170                                 reclaim_stat->recent_rotated[file]++;
1171                         }
1172                         if (!pagevec_add(&pvec, page)) {
1173                                 spin_unlock_irq(&zone->lru_lock);
1174                                 __pagevec_release(&pvec);
1175                                 spin_lock_irq(&zone->lru_lock);
1176                         }
1177                 }
1178                 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1179                 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1180
1181         } while (nr_scanned < max_scan);
1182
1183 done:
1184         spin_unlock_irq(&zone->lru_lock);
1185         pagevec_release(&pvec);
1186         return nr_reclaimed;
1187 }
1188
1189 /*
1190  * We are about to scan this zone at a certain priority level.  If that priority
1191  * level is smaller (ie: more urgent) than the previous priority, then note
1192  * that priority level within the zone.  This is done so that when the next
1193  * process comes in to scan this zone, it will immediately start out at this
1194  * priority level rather than having to build up its own scanning priority.
1195  * Here, this priority affects only the reclaim-mapped threshold.
1196  */
1197 static inline void note_zone_scanning_priority(struct zone *zone, int priority)
1198 {
1199         if (priority < zone->prev_priority)
1200                 zone->prev_priority = priority;
1201 }
1202
1203 /*
1204  * This moves pages from the active list to the inactive list.
1205  *
1206  * We move them the other way if the page is referenced by one or more
1207  * processes, from rmap.
1208  *
1209  * If the pages are mostly unmapped, the processing is fast and it is
1210  * appropriate to hold zone->lru_lock across the whole operation.  But if
1211  * the pages are mapped, the processing is slow (page_referenced()) so we
1212  * should drop zone->lru_lock around each page.  It's impossible to balance
1213  * this, so instead we remove the pages from the LRU while processing them.
1214  * It is safe to rely on PG_active against the non-LRU pages in here because
1215  * nobody will play with that bit on a non-LRU page.
1216  *
1217  * The downside is that we have to touch page->_count against each page.
1218  * But we had to alter page->flags anyway.
1219  */
1220
1221 static void move_active_pages_to_lru(struct zone *zone,
1222                                      struct list_head *list,
1223                                      enum lru_list lru)
1224 {
1225         unsigned long pgmoved = 0;
1226         struct pagevec pvec;
1227         struct page *page;
1228
1229         pagevec_init(&pvec, 1);
1230
1231         while (!list_empty(list)) {
1232                 page = lru_to_page(list);
1233                 prefetchw_prev_lru_page(page, list, flags);
1234
1235                 VM_BUG_ON(PageLRU(page));
1236                 SetPageLRU(page);
1237
1238                 VM_BUG_ON(!PageActive(page));
1239                 if (!is_active_lru(lru))
1240                         ClearPageActive(page);  /* we are de-activating */
1241
1242                 list_move(&page->lru, &zone->lru[lru].list);
1243                 mem_cgroup_add_lru_list(page, lru);
1244                 pgmoved++;
1245
1246                 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1247                         spin_unlock_irq(&zone->lru_lock);
1248                         if (buffer_heads_over_limit)
1249                                 pagevec_strip(&pvec);
1250                         __pagevec_release(&pvec);
1251                         spin_lock_irq(&zone->lru_lock);
1252                 }
1253         }
1254         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1255         if (!is_active_lru(lru))
1256                 __count_vm_events(PGDEACTIVATE, pgmoved);
1257 }
1258
1259 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1260                         struct scan_control *sc, int priority, int file)
1261 {
1262         unsigned long nr_taken;
1263         unsigned long pgscanned;
1264         unsigned long vm_flags;
1265         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1266         LIST_HEAD(l_active);
1267         LIST_HEAD(l_inactive);
1268         struct page *page;
1269         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1270         unsigned long nr_rotated = 0;
1271
1272         lru_add_drain();
1273         spin_lock_irq(&zone->lru_lock);
1274         nr_taken = sc->isolate_pages(nr_pages, &l_hold, &pgscanned, sc->order,
1275                                         ISOLATE_ACTIVE, zone,
1276                                         sc->mem_cgroup, 1, file);
1277         /*
1278          * zone->pages_scanned is used for detect zone's oom
1279          * mem_cgroup remembers nr_scan by itself.
1280          */
1281         if (scanning_global_lru(sc)) {
1282                 zone->pages_scanned += pgscanned;
1283         }
1284         reclaim_stat->recent_scanned[!!file] += nr_taken;
1285
1286         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1287         if (file)
1288                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1289         else
1290                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1291         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1292         spin_unlock_irq(&zone->lru_lock);
1293
1294         while (!list_empty(&l_hold)) {
1295                 cond_resched();
1296                 page = lru_to_page(&l_hold);
1297                 list_del(&page->lru);
1298
1299                 if (unlikely(!page_evictable(page, NULL))) {
1300                         putback_lru_page(page);
1301                         continue;
1302                 }
1303
1304                 /* page_referenced clears PageReferenced */
1305                 if (page_mapping_inuse(page) &&
1306                     page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1307                         nr_rotated++;
1308                         /*
1309                          * Identify referenced, file-backed active pages and
1310                          * give them one more trip around the active list. So
1311                          * that executable code get better chances to stay in
1312                          * memory under moderate memory pressure.  Anon pages
1313                          * are not likely to be evicted by use-once streaming
1314                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1315                          * so we ignore them here.
1316                          */
1317                         if ((vm_flags & VM_EXEC) && !PageAnon(page)) {
1318                                 list_add(&page->lru, &l_active);
1319                                 continue;
1320                         }
1321                 }
1322
1323                 list_add(&page->lru, &l_inactive);
1324         }
1325
1326         /*
1327          * Move pages back to the lru list.
1328          */
1329         spin_lock_irq(&zone->lru_lock);
1330         /*
1331          * Count referenced pages from currently used mappings as rotated,
1332          * even though only some of them are actually re-activated.  This
1333          * helps balance scan pressure between file and anonymous pages in
1334          * get_scan_ratio.
1335          */
1336         reclaim_stat->recent_rotated[!!file] += nr_rotated;
1337
1338         move_active_pages_to_lru(zone, &l_active,
1339                                                 LRU_ACTIVE + file * LRU_FILE);
1340         move_active_pages_to_lru(zone, &l_inactive,
1341                                                 LRU_BASE   + file * LRU_FILE);
1342         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1343         spin_unlock_irq(&zone->lru_lock);
1344 }
1345
1346 static int inactive_anon_is_low_global(struct zone *zone)
1347 {
1348         unsigned long active, inactive;
1349
1350         active = zone_page_state(zone, NR_ACTIVE_ANON);
1351         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1352
1353         if (inactive * zone->inactive_ratio < active)
1354                 return 1;
1355
1356         return 0;
1357 }
1358
1359 /**
1360  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1361  * @zone: zone to check
1362  * @sc:   scan control of this context
1363  *
1364  * Returns true if the zone does not have enough inactive anon pages,
1365  * meaning some active anon pages need to be deactivated.
1366  */
1367 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1368 {
1369         int low;
1370
1371         if (scanning_global_lru(sc))
1372                 low = inactive_anon_is_low_global(zone);
1373         else
1374                 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1375         return low;
1376 }
1377
1378 static int inactive_file_is_low_global(struct zone *zone)
1379 {
1380         unsigned long active, inactive;
1381
1382         active = zone_page_state(zone, NR_ACTIVE_FILE);
1383         inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1384
1385         return (active > inactive);
1386 }
1387
1388 /**
1389  * inactive_file_is_low - check if file pages need to be deactivated
1390  * @zone: zone to check
1391  * @sc:   scan control of this context
1392  *
1393  * When the system is doing streaming IO, memory pressure here
1394  * ensures that active file pages get deactivated, until more
1395  * than half of the file pages are on the inactive list.
1396  *
1397  * Once we get to that situation, protect the system's working
1398  * set from being evicted by disabling active file page aging.
1399  *
1400  * This uses a different ratio than the anonymous pages, because
1401  * the page cache uses a use-once replacement algorithm.
1402  */
1403 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1404 {
1405         int low;
1406
1407         if (scanning_global_lru(sc))
1408                 low = inactive_file_is_low_global(zone);
1409         else
1410                 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1411         return low;
1412 }
1413
1414 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1415         struct zone *zone, struct scan_control *sc, int priority)
1416 {
1417         int file = is_file_lru(lru);
1418
1419         if (lru == LRU_ACTIVE_FILE && inactive_file_is_low(zone, sc)) {
1420                 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1421                 return 0;
1422         }
1423
1424         if (lru == LRU_ACTIVE_ANON && inactive_anon_is_low(zone, sc)) {
1425                 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1426                 return 0;
1427         }
1428         return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1429 }
1430
1431 /*
1432  * Determine how aggressively the anon and file LRU lists should be
1433  * scanned.  The relative value of each set of LRU lists is determined
1434  * by looking at the fraction of the pages scanned we did rotate back
1435  * onto the active list instead of evict.
1436  *
1437  * percent[0] specifies how much pressure to put on ram/swap backed
1438  * memory, while percent[1] determines pressure on the file LRUs.
1439  */
1440 static void get_scan_ratio(struct zone *zone, struct scan_control *sc,
1441                                         unsigned long *percent)
1442 {
1443         unsigned long anon, file, free;
1444         unsigned long anon_prio, file_prio;
1445         unsigned long ap, fp;
1446         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1447
1448         anon  = zone_nr_pages(zone, sc, LRU_ACTIVE_ANON) +
1449                 zone_nr_pages(zone, sc, LRU_INACTIVE_ANON);
1450         file  = zone_nr_pages(zone, sc, LRU_ACTIVE_FILE) +
1451                 zone_nr_pages(zone, sc, LRU_INACTIVE_FILE);
1452
1453         if (scanning_global_lru(sc)) {
1454                 free  = zone_page_state(zone, NR_FREE_PAGES);
1455                 /* If we have very few page cache pages,
1456                    force-scan anon pages. */
1457                 if (unlikely(file + free <= high_wmark_pages(zone))) {
1458                         percent[0] = 100;
1459                         percent[1] = 0;
1460                         return;
1461                 }
1462         }
1463
1464         /*
1465          * OK, so we have swap space and a fair amount of page cache
1466          * pages.  We use the recently rotated / recently scanned
1467          * ratios to determine how valuable each cache is.
1468          *
1469          * Because workloads change over time (and to avoid overflow)
1470          * we keep these statistics as a floating average, which ends
1471          * up weighing recent references more than old ones.
1472          *
1473          * anon in [0], file in [1]
1474          */
1475         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1476                 spin_lock_irq(&zone->lru_lock);
1477                 reclaim_stat->recent_scanned[0] /= 2;
1478                 reclaim_stat->recent_rotated[0] /= 2;
1479                 spin_unlock_irq(&zone->lru_lock);
1480         }
1481
1482         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1483                 spin_lock_irq(&zone->lru_lock);
1484                 reclaim_stat->recent_scanned[1] /= 2;
1485                 reclaim_stat->recent_rotated[1] /= 2;
1486                 spin_unlock_irq(&zone->lru_lock);
1487         }
1488
1489         /*
1490          * With swappiness at 100, anonymous and file have the same priority.
1491          * This scanning priority is essentially the inverse of IO cost.
1492          */
1493         anon_prio = sc->swappiness;
1494         file_prio = 200 - sc->swappiness;
1495
1496         /*
1497          * The amount of pressure on anon vs file pages is inversely
1498          * proportional to the fraction of recently scanned pages on
1499          * each list that were recently referenced and in active use.
1500          */
1501         ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1502         ap /= reclaim_stat->recent_rotated[0] + 1;
1503
1504         fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1505         fp /= reclaim_stat->recent_rotated[1] + 1;
1506
1507         /* Normalize to percentages */
1508         percent[0] = 100 * ap / (ap + fp + 1);
1509         percent[1] = 100 - percent[0];
1510 }
1511
1512 /*
1513  * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1514  * until we collected @swap_cluster_max pages to scan.
1515  */
1516 static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
1517                                        unsigned long *nr_saved_scan,
1518                                        unsigned long swap_cluster_max)
1519 {
1520         unsigned long nr;
1521
1522         *nr_saved_scan += nr_to_scan;
1523         nr = *nr_saved_scan;
1524
1525         if (nr >= swap_cluster_max)
1526                 *nr_saved_scan = 0;
1527         else
1528                 nr = 0;
1529
1530         return nr;
1531 }
1532
1533 /*
1534  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1535  */
1536 static void shrink_zone(int priority, struct zone *zone,
1537                                 struct scan_control *sc)
1538 {
1539         unsigned long nr[NR_LRU_LISTS];
1540         unsigned long nr_to_scan;
1541         unsigned long percent[2];       /* anon @ 0; file @ 1 */
1542         enum lru_list l;
1543         unsigned long nr_reclaimed = sc->nr_reclaimed;
1544         unsigned long swap_cluster_max = sc->swap_cluster_max;
1545         int noswap = 0;
1546
1547         /* If we have no swap space, do not bother scanning anon pages. */
1548         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1549                 noswap = 1;
1550                 percent[0] = 0;
1551                 percent[1] = 100;
1552         } else
1553                 get_scan_ratio(zone, sc, percent);
1554
1555         for_each_evictable_lru(l) {
1556                 int file = is_file_lru(l);
1557                 unsigned long scan;
1558
1559                 scan = zone_nr_pages(zone, sc, l);
1560                 if (priority || noswap) {
1561                         scan >>= priority;
1562                         scan = (scan * percent[file]) / 100;
1563                 }
1564                 if (scanning_global_lru(sc))
1565                         nr[l] = nr_scan_try_batch(scan,
1566                                                   &zone->lru[l].nr_saved_scan,
1567                                                   swap_cluster_max);
1568                 else
1569                         nr[l] = scan;
1570         }
1571
1572         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1573                                         nr[LRU_INACTIVE_FILE]) {
1574                 for_each_evictable_lru(l) {
1575                         if (nr[l]) {
1576                                 nr_to_scan = min(nr[l], swap_cluster_max);
1577                                 nr[l] -= nr_to_scan;
1578
1579                                 nr_reclaimed += shrink_list(l, nr_to_scan,
1580                                                             zone, sc, priority);
1581                         }
1582                 }
1583                 /*
1584                  * On large memory systems, scan >> priority can become
1585                  * really large. This is fine for the starting priority;
1586                  * we want to put equal scanning pressure on each zone.
1587                  * However, if the VM has a harder time of freeing pages,
1588                  * with multiple processes reclaiming pages, the total
1589                  * freeing target can get unreasonably large.
1590                  */
1591                 if (nr_reclaimed > swap_cluster_max &&
1592                         priority < DEF_PRIORITY && !current_is_kswapd())
1593                         break;
1594         }
1595
1596         sc->nr_reclaimed = nr_reclaimed;
1597
1598         /*
1599          * Even if we did not try to evict anon pages at all, we want to
1600          * rebalance the anon lru active/inactive ratio.
1601          */
1602         if (inactive_anon_is_low(zone, sc) && nr_swap_pages > 0)
1603                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1604
1605         throttle_vm_writeout(sc->gfp_mask);
1606 }
1607
1608 /*
1609  * This is the direct reclaim path, for page-allocating processes.  We only
1610  * try to reclaim pages from zones which will satisfy the caller's allocation
1611  * request.
1612  *
1613  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1614  * Because:
1615  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1616  *    allocation or
1617  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1618  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1619  *    zone defense algorithm.
1620  *
1621  * If a zone is deemed to be full of pinned pages then just give it a light
1622  * scan then give up on it.
1623  */
1624 static void shrink_zones(int priority, struct zonelist *zonelist,
1625                                         struct scan_control *sc)
1626 {
1627         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1628         struct zoneref *z;
1629         struct zone *zone;
1630
1631         sc->all_unreclaimable = 1;
1632         for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1633                                         sc->nodemask) {
1634                 if (!populated_zone(zone))
1635                         continue;
1636                 /*
1637                  * Take care memory controller reclaiming has small influence
1638                  * to global LRU.
1639                  */
1640                 if (scanning_global_lru(sc)) {
1641                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1642                                 continue;
1643                         note_zone_scanning_priority(zone, priority);
1644
1645                         if (zone_is_all_unreclaimable(zone) &&
1646                                                 priority != DEF_PRIORITY)
1647                                 continue;       /* Let kswapd poll it */
1648                         sc->all_unreclaimable = 0;
1649                 } else {
1650                         /*
1651                          * Ignore cpuset limitation here. We just want to reduce
1652                          * # of used pages by us regardless of memory shortage.
1653                          */
1654                         sc->all_unreclaimable = 0;
1655                         mem_cgroup_note_reclaim_priority(sc->mem_cgroup,
1656                                                         priority);
1657                 }
1658
1659                 shrink_zone(priority, zone, sc);
1660         }
1661 }
1662
1663 /*
1664  * This is the main entry point to direct page reclaim.
1665  *
1666  * If a full scan of the inactive list fails to free enough memory then we
1667  * are "out of memory" and something needs to be killed.
1668  *
1669  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1670  * high - the zone may be full of dirty or under-writeback pages, which this
1671  * caller can't do much about.  We kick pdflush and take explicit naps in the
1672  * hope that some of these pages can be written.  But if the allocating task
1673  * holds filesystem locks which prevent writeout this might not work, and the
1674  * allocation attempt will fail.
1675  *
1676  * returns:     0, if no pages reclaimed
1677  *              else, the number of pages reclaimed
1678  */
1679 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
1680                                         struct scan_control *sc)
1681 {
1682         int priority;
1683         unsigned long ret = 0;
1684         unsigned long total_scanned = 0;
1685         struct reclaim_state *reclaim_state = current->reclaim_state;
1686         unsigned long lru_pages = 0;
1687         struct zoneref *z;
1688         struct zone *zone;
1689         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1690
1691         delayacct_freepages_start();
1692
1693         if (scanning_global_lru(sc))
1694                 count_vm_event(ALLOCSTALL);
1695         /*
1696          * mem_cgroup will not do shrink_slab.
1697          */
1698         if (scanning_global_lru(sc)) {
1699                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1700
1701                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1702                                 continue;
1703
1704                         lru_pages += zone_lru_pages(zone);
1705                 }
1706         }
1707
1708         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1709                 sc->nr_scanned = 0;
1710                 if (!priority)
1711                         disable_swap_token();
1712                 shrink_zones(priority, zonelist, sc);
1713                 /*
1714                  * Don't shrink slabs when reclaiming memory from
1715                  * over limit cgroups
1716                  */
1717                 if (scanning_global_lru(sc)) {
1718                         shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
1719                         if (reclaim_state) {
1720                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
1721                                 reclaim_state->reclaimed_slab = 0;
1722                         }
1723                 }
1724                 total_scanned += sc->nr_scanned;
1725                 if (sc->nr_reclaimed >= sc->swap_cluster_max) {
1726                         ret = sc->nr_reclaimed;
1727                         goto out;
1728                 }
1729
1730                 /*
1731                  * Try to write back as many pages as we just scanned.  This
1732                  * tends to cause slow streaming writers to write data to the
1733                  * disk smoothly, at the dirtying rate, which is nice.   But
1734                  * that's undesirable in laptop mode, where we *want* lumpy
1735                  * writeout.  So in laptop mode, write out the whole world.
1736                  */
1737                 if (total_scanned > sc->swap_cluster_max +
1738                                         sc->swap_cluster_max / 2) {
1739                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
1740                         sc->may_writepage = 1;
1741                 }
1742
1743                 /* Take a nap, wait for some writeback to complete */
1744                 if (sc->nr_scanned && priority < DEF_PRIORITY - 2)
1745                         congestion_wait(BLK_RW_ASYNC, HZ/10);
1746         }
1747         /* top priority shrink_zones still had more to do? don't OOM, then */
1748         if (!sc->all_unreclaimable && scanning_global_lru(sc))
1749                 ret = sc->nr_reclaimed;
1750 out:
1751         /*
1752          * Now that we've scanned all the zones at this priority level, note
1753          * that level within the zone so that the next thread which performs
1754          * scanning of this zone will immediately start out at this priority
1755          * level.  This affects only the decision whether or not to bring
1756          * mapped pages onto the inactive list.
1757          */
1758         if (priority < 0)
1759                 priority = 0;
1760
1761         if (scanning_global_lru(sc)) {
1762                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1763
1764                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1765                                 continue;
1766
1767                         zone->prev_priority = priority;
1768                 }
1769         } else
1770                 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority);
1771
1772         delayacct_freepages_end();
1773
1774         return ret;
1775 }
1776
1777 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
1778                                 gfp_t gfp_mask, nodemask_t *nodemask)
1779 {
1780         struct scan_control sc = {
1781                 .gfp_mask = gfp_mask,
1782                 .may_writepage = !laptop_mode,
1783                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1784                 .may_unmap = 1,
1785                 .may_swap = 1,
1786                 .swappiness = vm_swappiness,
1787                 .order = order,
1788                 .mem_cgroup = NULL,
1789                 .isolate_pages = isolate_pages_global,
1790                 .nodemask = nodemask,
1791         };
1792
1793         return do_try_to_free_pages(zonelist, &sc);
1794 }
1795
1796 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
1797
1798 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
1799                                            gfp_t gfp_mask,
1800                                            bool noswap,
1801                                            unsigned int swappiness)
1802 {
1803         struct scan_control sc = {
1804                 .may_writepage = !laptop_mode,
1805                 .may_unmap = 1,
1806                 .may_swap = !noswap,
1807                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1808                 .swappiness = swappiness,
1809                 .order = 0,
1810                 .mem_cgroup = mem_cont,
1811                 .isolate_pages = mem_cgroup_isolate_pages,
1812                 .nodemask = NULL, /* we don't care the placement */
1813         };
1814         struct zonelist *zonelist;
1815
1816         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1817                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1818         zonelist = NODE_DATA(numa_node_id())->node_zonelists;
1819         return do_try_to_free_pages(zonelist, &sc);
1820 }
1821 #endif
1822
1823 /*
1824  * For kswapd, balance_pgdat() will work across all this node's zones until
1825  * they are all at high_wmark_pages(zone).
1826  *
1827  * Returns the number of pages which were actually freed.
1828  *
1829  * There is special handling here for zones which are full of pinned pages.
1830  * This can happen if the pages are all mlocked, or if they are all used by
1831  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1832  * What we do is to detect the case where all pages in the zone have been
1833  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1834  * dead and from now on, only perform a short scan.  Basically we're polling
1835  * the zone for when the problem goes away.
1836  *
1837  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
1838  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
1839  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
1840  * lower zones regardless of the number of free pages in the lower zones. This
1841  * interoperates with the page allocator fallback scheme to ensure that aging
1842  * of pages is balanced across the zones.
1843  */
1844 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
1845 {
1846         int all_zones_ok;
1847         int priority;
1848         int i;
1849         unsigned long total_scanned;
1850         struct reclaim_state *reclaim_state = current->reclaim_state;
1851         struct scan_control sc = {
1852                 .gfp_mask = GFP_KERNEL,
1853                 .may_unmap = 1,
1854                 .may_swap = 1,
1855                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1856                 .swappiness = vm_swappiness,
1857                 .order = order,
1858                 .mem_cgroup = NULL,
1859                 .isolate_pages = isolate_pages_global,
1860         };
1861         /*
1862          * temp_priority is used to remember the scanning priority at which
1863          * this zone was successfully refilled to
1864          * free_pages == high_wmark_pages(zone).
1865          */
1866         int temp_priority[MAX_NR_ZONES];
1867
1868 loop_again:
1869         total_scanned = 0;
1870         sc.nr_reclaimed = 0;
1871         sc.may_writepage = !laptop_mode;
1872         count_vm_event(PAGEOUTRUN);
1873
1874         for (i = 0; i < pgdat->nr_zones; i++)
1875                 temp_priority[i] = DEF_PRIORITY;
1876
1877         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1878                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1879                 unsigned long lru_pages = 0;
1880
1881                 /* The swap token gets in the way of swapout... */
1882                 if (!priority)
1883                         disable_swap_token();
1884
1885                 all_zones_ok = 1;
1886
1887                 /*
1888                  * Scan in the highmem->dma direction for the highest
1889                  * zone which needs scanning
1890                  */
1891                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1892                         struct zone *zone = pgdat->node_zones + i;
1893
1894                         if (!populated_zone(zone))
1895                                 continue;
1896
1897                         if (zone_is_all_unreclaimable(zone) &&
1898                             priority != DEF_PRIORITY)
1899                                 continue;
1900
1901                         /*
1902                          * Do some background aging of the anon list, to give
1903                          * pages a chance to be referenced before reclaiming.
1904                          */
1905                         if (inactive_anon_is_low(zone, &sc))
1906                                 shrink_active_list(SWAP_CLUSTER_MAX, zone,
1907                                                         &sc, priority, 0);
1908
1909                         if (!zone_watermark_ok(zone, order,
1910                                         high_wmark_pages(zone), 0, 0)) {
1911                                 end_zone = i;
1912                                 break;
1913                         }
1914                 }
1915                 if (i < 0)
1916                         goto out;
1917
1918                 for (i = 0; i <= end_zone; i++) {
1919                         struct zone *zone = pgdat->node_zones + i;
1920
1921                         lru_pages += zone_lru_pages(zone);
1922                 }
1923
1924                 /*
1925                  * Now scan the zone in the dma->highmem direction, stopping
1926                  * at the last zone which needs scanning.
1927                  *
1928                  * We do this because the page allocator works in the opposite
1929                  * direction.  This prevents the page allocator from allocating
1930                  * pages behind kswapd's direction of progress, which would
1931                  * cause too much scanning of the lower zones.
1932                  */
1933                 for (i = 0; i <= end_zone; i++) {
1934                         struct zone *zone = pgdat->node_zones + i;
1935                         int nr_slab;
1936
1937                         if (!populated_zone(zone))
1938                                 continue;
1939
1940                         if (zone_is_all_unreclaimable(zone) &&
1941                                         priority != DEF_PRIORITY)
1942                                 continue;
1943
1944                         if (!zone_watermark_ok(zone, order,
1945                                         high_wmark_pages(zone), end_zone, 0))
1946                                 all_zones_ok = 0;
1947                         temp_priority[i] = priority;
1948                         sc.nr_scanned = 0;
1949                         note_zone_scanning_priority(zone, priority);
1950                         /*
1951                          * We put equal pressure on every zone, unless one
1952                          * zone has way too many pages free already.
1953                          */
1954                         if (!zone_watermark_ok(zone, order,
1955                                         8*high_wmark_pages(zone), end_zone, 0))
1956                                 shrink_zone(priority, zone, &sc);
1957                         reclaim_state->reclaimed_slab = 0;
1958                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1959                                                 lru_pages);
1960                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1961                         total_scanned += sc.nr_scanned;
1962                         if (zone_is_all_unreclaimable(zone))
1963                                 continue;
1964                         if (nr_slab == 0 && zone->pages_scanned >=
1965                                                 (zone_lru_pages(zone) * 6))
1966                                         zone_set_flag(zone,
1967                                                       ZONE_ALL_UNRECLAIMABLE);
1968                         /*
1969                          * If we've done a decent amount of scanning and
1970                          * the reclaim ratio is low, start doing writepage
1971                          * even in laptop mode
1972                          */
1973                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1974                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
1975                                 sc.may_writepage = 1;
1976                 }
1977                 if (all_zones_ok)
1978                         break;          /* kswapd: all done */
1979                 /*
1980                  * OK, kswapd is getting into trouble.  Take a nap, then take
1981                  * another pass across the zones.
1982                  */
1983                 if (total_scanned && priority < DEF_PRIORITY - 2)
1984                         congestion_wait(BLK_RW_ASYNC, HZ/10);
1985
1986                 /*
1987                  * We do this so kswapd doesn't build up large priorities for
1988                  * example when it is freeing in parallel with allocators. It
1989                  * matches the direct reclaim path behaviour in terms of impact
1990                  * on zone->*_priority.
1991                  */
1992                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
1993                         break;
1994         }
1995 out:
1996         /*
1997          * Note within each zone the priority level at which this zone was
1998          * brought into a happy state.  So that the next thread which scans this
1999          * zone will start out at that priority level.
2000          */
2001         for (i = 0; i < pgdat->nr_zones; i++) {
2002                 struct zone *zone = pgdat->node_zones + i;
2003
2004                 zone->prev_priority = temp_priority[i];
2005         }
2006         if (!all_zones_ok) {
2007                 cond_resched();
2008
2009                 try_to_freeze();
2010
2011                 /*
2012                  * Fragmentation may mean that the system cannot be
2013                  * rebalanced for high-order allocations in all zones.
2014                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2015                  * it means the zones have been fully scanned and are still
2016                  * not balanced. For high-order allocations, there is
2017                  * little point trying all over again as kswapd may
2018                  * infinite loop.
2019                  *
2020                  * Instead, recheck all watermarks at order-0 as they
2021                  * are the most important. If watermarks are ok, kswapd will go
2022                  * back to sleep. High-order users can still perform direct
2023                  * reclaim if they wish.
2024                  */
2025                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2026                         order = sc.order = 0;
2027
2028                 goto loop_again;
2029         }
2030
2031         return sc.nr_reclaimed;
2032 }
2033
2034 /*
2035  * The background pageout daemon, started as a kernel thread
2036  * from the init process.
2037  *
2038  * This basically trickles out pages so that we have _some_
2039  * free memory available even if there is no other activity
2040  * that frees anything up. This is needed for things like routing
2041  * etc, where we otherwise might have all activity going on in
2042  * asynchronous contexts that cannot page things out.
2043  *
2044  * If there are applications that are active memory-allocators
2045  * (most normal use), this basically shouldn't matter.
2046  */
2047 static int kswapd(void *p)
2048 {
2049         unsigned long order;
2050         pg_data_t *pgdat = (pg_data_t*)p;
2051         struct task_struct *tsk = current;
2052         DEFINE_WAIT(wait);
2053         struct reclaim_state reclaim_state = {
2054                 .reclaimed_slab = 0,
2055         };
2056         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2057
2058         lockdep_set_current_reclaim_state(GFP_KERNEL);
2059
2060         if (!cpumask_empty(cpumask))
2061                 set_cpus_allowed_ptr(tsk, cpumask);
2062         current->reclaim_state = &reclaim_state;
2063
2064         /*
2065          * Tell the memory management that we're a "memory allocator",
2066          * and that if we need more memory we should get access to it
2067          * regardless (see "__alloc_pages()"). "kswapd" should
2068          * never get caught in the normal page freeing logic.
2069          *
2070          * (Kswapd normally doesn't need memory anyway, but sometimes
2071          * you need a small amount of memory in order to be able to
2072          * page out something else, and this flag essentially protects
2073          * us from recursively trying to free more memory as we're
2074          * trying to free the first piece of memory in the first place).
2075          */
2076         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2077         set_freezable();
2078
2079         order = 0;
2080         for ( ; ; ) {
2081                 unsigned long new_order;
2082
2083                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2084                 new_order = pgdat->kswapd_max_order;
2085                 pgdat->kswapd_max_order = 0;
2086                 if (order < new_order) {
2087                         /*
2088                          * Don't sleep if someone wants a larger 'order'
2089                          * allocation
2090                          */
2091                         order = new_order;
2092                 } else {
2093                         if (!freezing(current))
2094                                 schedule();
2095
2096                         order = pgdat->kswapd_max_order;
2097                 }
2098                 finish_wait(&pgdat->kswapd_wait, &wait);
2099
2100                 if (!try_to_freeze()) {
2101                         /* We can speed up thawing tasks if we don't call
2102                          * balance_pgdat after returning from the refrigerator
2103                          */
2104                         balance_pgdat(pgdat, order);
2105                 }
2106         }
2107         return 0;
2108 }
2109
2110 /*
2111  * A zone is low on free memory, so wake its kswapd task to service it.
2112  */
2113 void wakeup_kswapd(struct zone *zone, int order)
2114 {
2115         pg_data_t *pgdat;
2116
2117         if (!populated_zone(zone))
2118                 return;
2119
2120         pgdat = zone->zone_pgdat;
2121         if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
2122                 return;
2123         if (pgdat->kswapd_max_order < order)
2124                 pgdat->kswapd_max_order = order;
2125         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2126                 return;
2127         if (!waitqueue_active(&pgdat->kswapd_wait))
2128                 return;
2129         wake_up_interruptible(&pgdat->kswapd_wait);
2130 }
2131
2132 unsigned long global_lru_pages(void)
2133 {
2134         return global_page_state(NR_ACTIVE_ANON)
2135                 + global_page_state(NR_ACTIVE_FILE)
2136                 + global_page_state(NR_INACTIVE_ANON)
2137                 + global_page_state(NR_INACTIVE_FILE);
2138 }
2139
2140 #ifdef CONFIG_HIBERNATION
2141 /*
2142  * Helper function for shrink_all_memory().  Tries to reclaim 'nr_pages' pages
2143  * from LRU lists system-wide, for given pass and priority.
2144  *
2145  * For pass > 3 we also try to shrink the LRU lists that contain a few pages
2146  */
2147 static void shrink_all_zones(unsigned long nr_pages, int prio,
2148                                       int pass, struct scan_control *sc)
2149 {
2150         struct zone *zone;
2151         unsigned long nr_reclaimed = 0;
2152
2153         for_each_populated_zone(zone) {
2154                 enum lru_list l;
2155
2156                 if (zone_is_all_unreclaimable(zone) && prio != DEF_PRIORITY)
2157                         continue;
2158
2159                 for_each_evictable_lru(l) {
2160                         enum zone_stat_item ls = NR_LRU_BASE + l;
2161                         unsigned long lru_pages = zone_page_state(zone, ls);
2162
2163                         /* For pass = 0, we don't shrink the active list */
2164                         if (pass == 0 && (l == LRU_ACTIVE_ANON ||
2165                                                 l == LRU_ACTIVE_FILE))
2166                                 continue;
2167
2168                         zone->lru[l].nr_saved_scan += (lru_pages >> prio) + 1;
2169                         if (zone->lru[l].nr_saved_scan >= nr_pages || pass > 3) {
2170                                 unsigned long nr_to_scan;
2171
2172                                 zone->lru[l].nr_saved_scan = 0;
2173                                 nr_to_scan = min(nr_pages, lru_pages);
2174                                 nr_reclaimed += shrink_list(l, nr_to_scan, zone,
2175                                                                 sc, prio);
2176                                 if (nr_reclaimed >= nr_pages) {
2177                                         sc->nr_reclaimed += nr_reclaimed;
2178                                         return;
2179                                 }
2180                         }
2181                 }
2182         }
2183         sc->nr_reclaimed += nr_reclaimed;
2184 }
2185
2186 /*
2187  * Try to free `nr_pages' of memory, system-wide, and return the number of
2188  * freed pages.
2189  *
2190  * Rather than trying to age LRUs the aim is to preserve the overall
2191  * LRU order by reclaiming preferentially
2192  * inactive > active > active referenced > active mapped
2193  */
2194 unsigned long shrink_all_memory(unsigned long nr_pages)
2195 {
2196         unsigned long lru_pages, nr_slab;
2197         int pass;
2198         struct reclaim_state reclaim_state;
2199         struct scan_control sc = {
2200                 .gfp_mask = GFP_KERNEL,
2201                 .may_unmap = 0,
2202                 .may_writepage = 1,
2203                 .isolate_pages = isolate_pages_global,
2204                 .nr_reclaimed = 0,
2205         };
2206
2207         current->reclaim_state = &reclaim_state;
2208
2209         lru_pages = global_lru_pages();
2210         nr_slab = global_page_state(NR_SLAB_RECLAIMABLE);
2211         /* If slab caches are huge, it's better to hit them first */
2212         while (nr_slab >= lru_pages) {
2213                 reclaim_state.reclaimed_slab = 0;
2214                 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
2215                 if (!reclaim_state.reclaimed_slab)
2216                         break;
2217
2218                 sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2219                 if (sc.nr_reclaimed >= nr_pages)
2220                         goto out;
2221
2222                 nr_slab -= reclaim_state.reclaimed_slab;
2223         }
2224
2225         /*
2226          * We try to shrink LRUs in 5 passes:
2227          * 0 = Reclaim from inactive_list only
2228          * 1 = Reclaim from active list but don't reclaim mapped
2229          * 2 = 2nd pass of type 1
2230          * 3 = Reclaim mapped (normal reclaim)
2231          * 4 = 2nd pass of type 3
2232          */
2233         for (pass = 0; pass < 5; pass++) {
2234                 int prio;
2235
2236                 /* Force reclaiming mapped pages in the passes #3 and #4 */
2237                 if (pass > 2)
2238                         sc.may_unmap = 1;
2239
2240                 for (prio = DEF_PRIORITY; prio >= 0; prio--) {
2241                         unsigned long nr_to_scan = nr_pages - sc.nr_reclaimed;
2242
2243                         sc.nr_scanned = 0;
2244                         sc.swap_cluster_max = nr_to_scan;
2245                         shrink_all_zones(nr_to_scan, prio, pass, &sc);
2246                         if (sc.nr_reclaimed >= nr_pages)
2247                                 goto out;
2248
2249                         reclaim_state.reclaimed_slab = 0;
2250                         shrink_slab(sc.nr_scanned, sc.gfp_mask,
2251                                         global_lru_pages());
2252                         sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2253                         if (sc.nr_reclaimed >= nr_pages)
2254                                 goto out;
2255
2256                         if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
2257                                 congestion_wait(BLK_RW_ASYNC, HZ / 10);
2258                 }
2259         }
2260
2261         /*
2262          * If sc.nr_reclaimed = 0, we could not shrink LRUs, but there may be
2263          * something in slab caches
2264          */
2265         if (!sc.nr_reclaimed) {
2266                 do {
2267                         reclaim_state.reclaimed_slab = 0;
2268                         shrink_slab(nr_pages, sc.gfp_mask, global_lru_pages());
2269                         sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2270                 } while (sc.nr_reclaimed < nr_pages &&
2271                                 reclaim_state.reclaimed_slab > 0);
2272         }
2273
2274
2275 out:
2276         current->reclaim_state = NULL;
2277
2278         return sc.nr_reclaimed;
2279 }
2280 #endif /* CONFIG_HIBERNATION */
2281
2282 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2283    not required for correctness.  So if the last cpu in a node goes
2284    away, we get changed to run anywhere: as the first one comes back,
2285    restore their cpu bindings. */
2286 static int __devinit cpu_callback(struct notifier_block *nfb,
2287                                   unsigned long action, void *hcpu)
2288 {
2289         int nid;
2290
2291         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2292                 for_each_node_state(nid, N_HIGH_MEMORY) {
2293                         pg_data_t *pgdat = NODE_DATA(nid);
2294                         const struct cpumask *mask;
2295
2296                         mask = cpumask_of_node(pgdat->node_id);
2297
2298                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2299                                 /* One of our CPUs online: restore mask */
2300                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2301                 }
2302         }
2303         return NOTIFY_OK;
2304 }
2305
2306 /*
2307  * This kswapd start function will be called by init and node-hot-add.
2308  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2309  */
2310 int kswapd_run(int nid)
2311 {
2312         pg_data_t *pgdat = NODE_DATA(nid);
2313         int ret = 0;
2314
2315         if (pgdat->kswapd)
2316                 return 0;
2317
2318         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2319         if (IS_ERR(pgdat->kswapd)) {
2320                 /* failure at boot is fatal */
2321                 BUG_ON(system_state == SYSTEM_BOOTING);
2322                 printk("Failed to start kswapd on node %d\n",nid);
2323                 ret = -1;
2324         }
2325         return ret;
2326 }
2327
2328 static int __init kswapd_init(void)
2329 {
2330         int nid;
2331
2332         swap_setup();
2333         for_each_node_state(nid, N_HIGH_MEMORY)
2334                 kswapd_run(nid);
2335         hotcpu_notifier(cpu_callback, 0);
2336         return 0;
2337 }
2338
2339 module_init(kswapd_init)
2340
2341 #ifdef CONFIG_NUMA
2342 /*
2343  * Zone reclaim mode
2344  *
2345  * If non-zero call zone_reclaim when the number of free pages falls below
2346  * the watermarks.
2347  */
2348 int zone_reclaim_mode __read_mostly;
2349
2350 #define RECLAIM_OFF 0
2351 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
2352 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
2353 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
2354
2355 /*
2356  * Priority for ZONE_RECLAIM. This determines the fraction of pages
2357  * of a node considered for each zone_reclaim. 4 scans 1/16th of
2358  * a zone.
2359  */
2360 #define ZONE_RECLAIM_PRIORITY 4
2361
2362 /*
2363  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2364  * occur.
2365  */
2366 int sysctl_min_unmapped_ratio = 1;
2367
2368 /*
2369  * If the number of slab pages in a zone grows beyond this percentage then
2370  * slab reclaim needs to occur.
2371  */
2372 int sysctl_min_slab_ratio = 5;
2373
2374 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
2375 {
2376         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
2377         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
2378                 zone_page_state(zone, NR_ACTIVE_FILE);
2379
2380         /*
2381          * It's possible for there to be more file mapped pages than
2382          * accounted for by the pages on the file LRU lists because
2383          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2384          */
2385         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
2386 }
2387
2388 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
2389 static long zone_pagecache_reclaimable(struct zone *zone)
2390 {
2391         long nr_pagecache_reclaimable;
2392         long delta = 0;
2393
2394         /*
2395          * If RECLAIM_SWAP is set, then all file pages are considered
2396          * potentially reclaimable. Otherwise, we have to worry about
2397          * pages like swapcache and zone_unmapped_file_pages() provides
2398          * a better estimate
2399          */
2400         if (zone_reclaim_mode & RECLAIM_SWAP)
2401                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
2402         else
2403                 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
2404
2405         /* If we can't clean pages, remove dirty pages from consideration */
2406         if (!(zone_reclaim_mode & RECLAIM_WRITE))
2407                 delta += zone_page_state(zone, NR_FILE_DIRTY);
2408
2409         /* Watch for any possible underflows due to delta */
2410         if (unlikely(delta > nr_pagecache_reclaimable))
2411                 delta = nr_pagecache_reclaimable;
2412
2413         return nr_pagecache_reclaimable - delta;
2414 }
2415
2416 /*
2417  * Try to free up some pages from this zone through reclaim.
2418  */
2419 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2420 {
2421         /* Minimum pages needed in order to stay on node */
2422         const unsigned long nr_pages = 1 << order;
2423         struct task_struct *p = current;
2424         struct reclaim_state reclaim_state;
2425         int priority;
2426         struct scan_control sc = {
2427                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2428                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2429                 .may_swap = 1,
2430                 .swap_cluster_max = max_t(unsigned long, nr_pages,
2431                                         SWAP_CLUSTER_MAX),
2432                 .gfp_mask = gfp_mask,
2433                 .swappiness = vm_swappiness,
2434                 .order = order,
2435                 .isolate_pages = isolate_pages_global,
2436         };
2437         unsigned long slab_reclaimable;
2438
2439         disable_swap_token();
2440         cond_resched();
2441         /*
2442          * We need to be able to allocate from the reserves for RECLAIM_SWAP
2443          * and we also need to be able to write out pages for RECLAIM_WRITE
2444          * and RECLAIM_SWAP.
2445          */
2446         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2447         reclaim_state.reclaimed_slab = 0;
2448         p->reclaim_state = &reclaim_state;
2449
2450         if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
2451                 /*
2452                  * Free memory by calling shrink zone with increasing
2453                  * priorities until we have enough memory freed.
2454                  */
2455                 priority = ZONE_RECLAIM_PRIORITY;
2456                 do {
2457                         note_zone_scanning_priority(zone, priority);
2458                         shrink_zone(priority, zone, &sc);
2459                         priority--;
2460                 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
2461         }
2462
2463         slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2464         if (slab_reclaimable > zone->min_slab_pages) {
2465                 /*
2466                  * shrink_slab() does not currently allow us to determine how
2467                  * many pages were freed in this zone. So we take the current
2468                  * number of slab pages and shake the slab until it is reduced
2469                  * by the same nr_pages that we used for reclaiming unmapped
2470                  * pages.
2471                  *
2472                  * Note that shrink_slab will free memory on all zones and may
2473                  * take a long time.
2474                  */
2475                 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
2476                         zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
2477                                 slab_reclaimable - nr_pages)
2478                         ;
2479
2480                 /*
2481                  * Update nr_reclaimed by the number of slab pages we
2482                  * reclaimed from this zone.
2483                  */
2484                 sc.nr_reclaimed += slab_reclaimable -
2485                         zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2486         }
2487
2488         p->reclaim_state = NULL;
2489         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
2490         return sc.nr_reclaimed >= nr_pages;
2491 }
2492
2493 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2494 {
2495         int node_id;
2496         int ret;
2497
2498         /*
2499          * Zone reclaim reclaims unmapped file backed pages and
2500          * slab pages if we are over the defined limits.
2501          *
2502          * A small portion of unmapped file backed pages is needed for
2503          * file I/O otherwise pages read by file I/O will be immediately
2504          * thrown out if the zone is overallocated. So we do not reclaim
2505          * if less than a specified percentage of the zone is used by
2506          * unmapped file backed pages.
2507          */
2508         if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
2509             zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
2510                 return ZONE_RECLAIM_FULL;
2511
2512         if (zone_is_all_unreclaimable(zone))
2513                 return ZONE_RECLAIM_FULL;
2514
2515         /*
2516          * Do not scan if the allocation should not be delayed.
2517          */
2518         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
2519                 return ZONE_RECLAIM_NOSCAN;
2520
2521         /*
2522          * Only run zone reclaim on the local zone or on zones that do not
2523          * have associated processors. This will favor the local processor
2524          * over remote processors and spread off node memory allocations
2525          * as wide as possible.
2526          */
2527         node_id = zone_to_nid(zone);
2528         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
2529                 return ZONE_RECLAIM_NOSCAN;
2530
2531         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2532                 return ZONE_RECLAIM_NOSCAN;
2533
2534         ret = __zone_reclaim(zone, gfp_mask, order);
2535         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2536
2537         if (!ret)
2538                 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
2539
2540         return ret;
2541 }
2542 #endif
2543
2544 /*
2545  * page_evictable - test whether a page is evictable
2546  * @page: the page to test
2547  * @vma: the VMA in which the page is or will be mapped, may be NULL
2548  *
2549  * Test whether page is evictable--i.e., should be placed on active/inactive
2550  * lists vs unevictable list.  The vma argument is !NULL when called from the
2551  * fault path to determine how to instantate a new page.
2552  *
2553  * Reasons page might not be evictable:
2554  * (1) page's mapping marked unevictable
2555  * (2) page is part of an mlocked VMA
2556  *
2557  */
2558 int page_evictable(struct page *page, struct vm_area_struct *vma)
2559 {
2560
2561         if (mapping_unevictable(page_mapping(page)))
2562                 return 0;
2563
2564         if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2565                 return 0;
2566
2567         return 1;
2568 }
2569
2570 /**
2571  * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2572  * @page: page to check evictability and move to appropriate lru list
2573  * @zone: zone page is in
2574  *
2575  * Checks a page for evictability and moves the page to the appropriate
2576  * zone lru list.
2577  *
2578  * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2579  * have PageUnevictable set.
2580  */
2581 static void check_move_unevictable_page(struct page *page, struct zone *zone)
2582 {
2583         VM_BUG_ON(PageActive(page));
2584
2585 retry:
2586         ClearPageUnevictable(page);
2587         if (page_evictable(page, NULL)) {
2588                 enum lru_list l = LRU_INACTIVE_ANON + page_is_file_cache(page);
2589
2590                 __dec_zone_state(zone, NR_UNEVICTABLE);
2591                 list_move(&page->lru, &zone->lru[l].list);
2592                 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
2593                 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
2594                 __count_vm_event(UNEVICTABLE_PGRESCUED);
2595         } else {
2596                 /*
2597                  * rotate unevictable list
2598                  */
2599                 SetPageUnevictable(page);
2600                 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
2601                 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
2602                 if (page_evictable(page, NULL))
2603                         goto retry;
2604         }
2605 }
2606
2607 /**
2608  * scan_mapping_unevictable_pages - scan an address space for evictable pages
2609  * @mapping: struct address_space to scan for evictable pages
2610  *
2611  * Scan all pages in mapping.  Check unevictable pages for
2612  * evictability and move them to the appropriate zone lru list.
2613  */
2614 void scan_mapping_unevictable_pages(struct address_space *mapping)
2615 {
2616         pgoff_t next = 0;
2617         pgoff_t end   = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
2618                          PAGE_CACHE_SHIFT;
2619         struct zone *zone;
2620         struct pagevec pvec;
2621
2622         if (mapping->nrpages == 0)
2623                 return;
2624
2625         pagevec_init(&pvec, 0);
2626         while (next < end &&
2627                 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
2628                 int i;
2629                 int pg_scanned = 0;
2630
2631                 zone = NULL;
2632
2633                 for (i = 0; i < pagevec_count(&pvec); i++) {
2634                         struct page *page = pvec.pages[i];
2635                         pgoff_t page_index = page->index;
2636                         struct zone *pagezone = page_zone(page);
2637
2638                         pg_scanned++;
2639                         if (page_index > next)
2640                                 next = page_index;
2641                         next++;
2642
2643                         if (pagezone != zone) {
2644                                 if (zone)
2645                                         spin_unlock_irq(&zone->lru_lock);
2646                                 zone = pagezone;
2647                                 spin_lock_irq(&zone->lru_lock);
2648                         }
2649
2650                         if (PageLRU(page) && PageUnevictable(page))
2651                                 check_move_unevictable_page(page, zone);
2652                 }
2653                 if (zone)
2654                         spin_unlock_irq(&zone->lru_lock);
2655                 pagevec_release(&pvec);
2656
2657                 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
2658         }
2659
2660 }
2661
2662 /**
2663  * scan_zone_unevictable_pages - check unevictable list for evictable pages
2664  * @zone - zone of which to scan the unevictable list
2665  *
2666  * Scan @zone's unevictable LRU lists to check for pages that have become
2667  * evictable.  Move those that have to @zone's inactive list where they
2668  * become candidates for reclaim, unless shrink_inactive_zone() decides
2669  * to reactivate them.  Pages that are still unevictable are rotated
2670  * back onto @zone's unevictable list.
2671  */
2672 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
2673 static void scan_zone_unevictable_pages(struct zone *zone)
2674 {
2675         struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
2676         unsigned long scan;
2677         unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
2678
2679         while (nr_to_scan > 0) {
2680                 unsigned long batch_size = min(nr_to_scan,
2681                                                 SCAN_UNEVICTABLE_BATCH_SIZE);
2682
2683                 spin_lock_irq(&zone->lru_lock);
2684                 for (scan = 0;  scan < batch_size; scan++) {
2685                         struct page *page = lru_to_page(l_unevictable);
2686
2687                         if (!trylock_page(page))
2688                                 continue;
2689
2690                         prefetchw_prev_lru_page(page, l_unevictable, flags);
2691
2692                         if (likely(PageLRU(page) && PageUnevictable(page)))
2693                                 check_move_unevictable_page(page, zone);
2694
2695                         unlock_page(page);
2696                 }
2697                 spin_unlock_irq(&zone->lru_lock);
2698
2699                 nr_to_scan -= batch_size;
2700         }
2701 }
2702
2703
2704 /**
2705  * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
2706  *
2707  * A really big hammer:  scan all zones' unevictable LRU lists to check for
2708  * pages that have become evictable.  Move those back to the zones'
2709  * inactive list where they become candidates for reclaim.
2710  * This occurs when, e.g., we have unswappable pages on the unevictable lists,
2711  * and we add swap to the system.  As such, it runs in the context of a task
2712  * that has possibly/probably made some previously unevictable pages
2713  * evictable.
2714  */
2715 static void scan_all_zones_unevictable_pages(void)
2716 {
2717         struct zone *zone;
2718
2719         for_each_zone(zone) {
2720                 scan_zone_unevictable_pages(zone);
2721         }
2722 }
2723
2724 /*
2725  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
2726  * all nodes' unevictable lists for evictable pages
2727  */
2728 unsigned long scan_unevictable_pages;
2729
2730 int scan_unevictable_handler(struct ctl_table *table, int write,
2731                            struct file *file, void __user *buffer,
2732                            size_t *length, loff_t *ppos)
2733 {
2734         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
2735
2736         if (write && *(unsigned long *)table->data)
2737                 scan_all_zones_unevictable_pages();
2738
2739         scan_unevictable_pages = 0;
2740         return 0;
2741 }
2742
2743 /*
2744  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
2745  * a specified node's per zone unevictable lists for evictable pages.
2746  */
2747
2748 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
2749                                           struct sysdev_attribute *attr,
2750                                           char *buf)
2751 {
2752         return sprintf(buf, "0\n");     /* always zero; should fit... */
2753 }
2754
2755 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
2756                                            struct sysdev_attribute *attr,
2757                                         const char *buf, size_t count)
2758 {
2759         struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
2760         struct zone *zone;
2761         unsigned long res;
2762         unsigned long req = strict_strtoul(buf, 10, &res);
2763
2764         if (!req)
2765                 return 1;       /* zero is no-op */
2766
2767         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2768                 if (!populated_zone(zone))
2769                         continue;
2770                 scan_zone_unevictable_pages(zone);
2771         }
2772         return 1;
2773 }
2774
2775
2776 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
2777                         read_scan_unevictable_node,
2778                         write_scan_unevictable_node);
2779
2780 int scan_unevictable_register_node(struct node *node)
2781 {
2782         return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
2783 }
2784
2785 void scan_unevictable_unregister_node(struct node *node)
2786 {
2787         sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
2788 }
2789