writeback: switch to per-bdi threads for flushing data
[linux-3.10.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
1076                 nr_taken = sc->isolate_pages(sc->swap_cluster_max,
1077                              &page_list, &nr_scan, sc->order, mode,
1078                                 zone, sc->mem_cgroup, 0, file);
1079                 nr_active = clear_active_flags(&page_list, count);
1080                 __count_vm_events(PGDEACTIVATE, nr_active);
1081
1082                 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1083                                                 -count[LRU_ACTIVE_FILE]);
1084                 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1085                                                 -count[LRU_INACTIVE_FILE]);
1086                 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1087                                                 -count[LRU_ACTIVE_ANON]);
1088                 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1089                                                 -count[LRU_INACTIVE_ANON]);
1090
1091                 if (scanning_global_lru(sc))
1092                         zone->pages_scanned += nr_scan;
1093
1094                 reclaim_stat->recent_scanned[0] += count[LRU_INACTIVE_ANON];
1095                 reclaim_stat->recent_scanned[0] += count[LRU_ACTIVE_ANON];
1096                 reclaim_stat->recent_scanned[1] += count[LRU_INACTIVE_FILE];
1097                 reclaim_stat->recent_scanned[1] += count[LRU_ACTIVE_FILE];
1098
1099                 spin_unlock_irq(&zone->lru_lock);
1100
1101                 nr_scanned += nr_scan;
1102                 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
1103
1104                 /*
1105                  * If we are direct reclaiming for contiguous pages and we do
1106                  * not reclaim everything in the list, try again and wait
1107                  * for IO to complete. This will stall high-order allocations
1108                  * but that should be acceptable to the caller
1109                  */
1110                 if (nr_freed < nr_taken && !current_is_kswapd() &&
1111                     lumpy_reclaim) {
1112                         congestion_wait(BLK_RW_ASYNC, HZ/10);
1113
1114                         /*
1115                          * The attempt at page out may have made some
1116                          * of the pages active, mark them inactive again.
1117                          */
1118                         nr_active = clear_active_flags(&page_list, count);
1119                         count_vm_events(PGDEACTIVATE, nr_active);
1120
1121                         nr_freed += shrink_page_list(&page_list, sc,
1122                                                         PAGEOUT_IO_SYNC);
1123                 }
1124
1125                 nr_reclaimed += nr_freed;
1126                 local_irq_disable();
1127                 if (current_is_kswapd()) {
1128                         __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scan);
1129                         __count_vm_events(KSWAPD_STEAL, nr_freed);
1130                 } else if (scanning_global_lru(sc))
1131                         __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scan);
1132
1133                 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
1134
1135                 if (nr_taken == 0)
1136                         goto done;
1137
1138                 spin_lock(&zone->lru_lock);
1139                 /*
1140                  * Put back any unfreeable pages.
1141                  */
1142                 while (!list_empty(&page_list)) {
1143                         int lru;
1144                         page = lru_to_page(&page_list);
1145                         VM_BUG_ON(PageLRU(page));
1146                         list_del(&page->lru);
1147                         if (unlikely(!page_evictable(page, NULL))) {
1148                                 spin_unlock_irq(&zone->lru_lock);
1149                                 putback_lru_page(page);
1150                                 spin_lock_irq(&zone->lru_lock);
1151                                 continue;
1152                         }
1153                         SetPageLRU(page);
1154                         lru = page_lru(page);
1155                         add_page_to_lru_list(zone, page, lru);
1156                         if (PageActive(page)) {
1157                                 int file = !!page_is_file_cache(page);
1158                                 reclaim_stat->recent_rotated[file]++;
1159                         }
1160                         if (!pagevec_add(&pvec, page)) {
1161                                 spin_unlock_irq(&zone->lru_lock);
1162                                 __pagevec_release(&pvec);
1163                                 spin_lock_irq(&zone->lru_lock);
1164                         }
1165                 }
1166         } while (nr_scanned < max_scan);
1167         spin_unlock(&zone->lru_lock);
1168 done:
1169         local_irq_enable();
1170         pagevec_release(&pvec);
1171         return nr_reclaimed;
1172 }
1173
1174 /*
1175  * We are about to scan this zone at a certain priority level.  If that priority
1176  * level is smaller (ie: more urgent) than the previous priority, then note
1177  * that priority level within the zone.  This is done so that when the next
1178  * process comes in to scan this zone, it will immediately start out at this
1179  * priority level rather than having to build up its own scanning priority.
1180  * Here, this priority affects only the reclaim-mapped threshold.
1181  */
1182 static inline void note_zone_scanning_priority(struct zone *zone, int priority)
1183 {
1184         if (priority < zone->prev_priority)
1185                 zone->prev_priority = priority;
1186 }
1187
1188 /*
1189  * This moves pages from the active list to the inactive list.
1190  *
1191  * We move them the other way if the page is referenced by one or more
1192  * processes, from rmap.
1193  *
1194  * If the pages are mostly unmapped, the processing is fast and it is
1195  * appropriate to hold zone->lru_lock across the whole operation.  But if
1196  * the pages are mapped, the processing is slow (page_referenced()) so we
1197  * should drop zone->lru_lock around each page.  It's impossible to balance
1198  * this, so instead we remove the pages from the LRU while processing them.
1199  * It is safe to rely on PG_active against the non-LRU pages in here because
1200  * nobody will play with that bit on a non-LRU page.
1201  *
1202  * The downside is that we have to touch page->_count against each page.
1203  * But we had to alter page->flags anyway.
1204  */
1205
1206 static void move_active_pages_to_lru(struct zone *zone,
1207                                      struct list_head *list,
1208                                      enum lru_list lru)
1209 {
1210         unsigned long pgmoved = 0;
1211         struct pagevec pvec;
1212         struct page *page;
1213
1214         pagevec_init(&pvec, 1);
1215
1216         while (!list_empty(list)) {
1217                 page = lru_to_page(list);
1218                 prefetchw_prev_lru_page(page, list, flags);
1219
1220                 VM_BUG_ON(PageLRU(page));
1221                 SetPageLRU(page);
1222
1223                 VM_BUG_ON(!PageActive(page));
1224                 if (!is_active_lru(lru))
1225                         ClearPageActive(page);  /* we are de-activating */
1226
1227                 list_move(&page->lru, &zone->lru[lru].list);
1228                 mem_cgroup_add_lru_list(page, lru);
1229                 pgmoved++;
1230
1231                 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1232                         spin_unlock_irq(&zone->lru_lock);
1233                         if (buffer_heads_over_limit)
1234                                 pagevec_strip(&pvec);
1235                         __pagevec_release(&pvec);
1236                         spin_lock_irq(&zone->lru_lock);
1237                 }
1238         }
1239         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1240         if (!is_active_lru(lru))
1241                 __count_vm_events(PGDEACTIVATE, pgmoved);
1242 }
1243
1244 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1245                         struct scan_control *sc, int priority, int file)
1246 {
1247         unsigned long pgmoved;
1248         unsigned long pgscanned;
1249         unsigned long vm_flags;
1250         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1251         LIST_HEAD(l_active);
1252         LIST_HEAD(l_inactive);
1253         struct page *page;
1254         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1255
1256         lru_add_drain();
1257         spin_lock_irq(&zone->lru_lock);
1258         pgmoved = sc->isolate_pages(nr_pages, &l_hold, &pgscanned, sc->order,
1259                                         ISOLATE_ACTIVE, zone,
1260                                         sc->mem_cgroup, 1, file);
1261         /*
1262          * zone->pages_scanned is used for detect zone's oom
1263          * mem_cgroup remembers nr_scan by itself.
1264          */
1265         if (scanning_global_lru(sc)) {
1266                 zone->pages_scanned += pgscanned;
1267         }
1268         reclaim_stat->recent_scanned[!!file] += pgmoved;
1269
1270         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1271         if (file)
1272                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -pgmoved);
1273         else
1274                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -pgmoved);
1275         spin_unlock_irq(&zone->lru_lock);
1276
1277         pgmoved = 0;  /* count referenced (mapping) mapped pages */
1278         while (!list_empty(&l_hold)) {
1279                 cond_resched();
1280                 page = lru_to_page(&l_hold);
1281                 list_del(&page->lru);
1282
1283                 if (unlikely(!page_evictable(page, NULL))) {
1284                         putback_lru_page(page);
1285                         continue;
1286                 }
1287
1288                 /* page_referenced clears PageReferenced */
1289                 if (page_mapping_inuse(page) &&
1290                     page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1291                         pgmoved++;
1292                         /*
1293                          * Identify referenced, file-backed active pages and
1294                          * give them one more trip around the active list. So
1295                          * that executable code get better chances to stay in
1296                          * memory under moderate memory pressure.  Anon pages
1297                          * are not likely to be evicted by use-once streaming
1298                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1299                          * so we ignore them here.
1300                          */
1301                         if ((vm_flags & VM_EXEC) && !PageAnon(page)) {
1302                                 list_add(&page->lru, &l_active);
1303                                 continue;
1304                         }
1305                 }
1306
1307                 list_add(&page->lru, &l_inactive);
1308         }
1309
1310         /*
1311          * Move pages back to the lru list.
1312          */
1313         spin_lock_irq(&zone->lru_lock);
1314         /*
1315          * Count referenced pages from currently used mappings as rotated,
1316          * even though only some of them are actually re-activated.  This
1317          * helps balance scan pressure between file and anonymous pages in
1318          * get_scan_ratio.
1319          */
1320         reclaim_stat->recent_rotated[!!file] += pgmoved;
1321
1322         move_active_pages_to_lru(zone, &l_active,
1323                                                 LRU_ACTIVE + file * LRU_FILE);
1324         move_active_pages_to_lru(zone, &l_inactive,
1325                                                 LRU_BASE   + file * LRU_FILE);
1326
1327         spin_unlock_irq(&zone->lru_lock);
1328 }
1329
1330 static int inactive_anon_is_low_global(struct zone *zone)
1331 {
1332         unsigned long active, inactive;
1333
1334         active = zone_page_state(zone, NR_ACTIVE_ANON);
1335         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1336
1337         if (inactive * zone->inactive_ratio < active)
1338                 return 1;
1339
1340         return 0;
1341 }
1342
1343 /**
1344  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1345  * @zone: zone to check
1346  * @sc:   scan control of this context
1347  *
1348  * Returns true if the zone does not have enough inactive anon pages,
1349  * meaning some active anon pages need to be deactivated.
1350  */
1351 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1352 {
1353         int low;
1354
1355         if (scanning_global_lru(sc))
1356                 low = inactive_anon_is_low_global(zone);
1357         else
1358                 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1359         return low;
1360 }
1361
1362 static int inactive_file_is_low_global(struct zone *zone)
1363 {
1364         unsigned long active, inactive;
1365
1366         active = zone_page_state(zone, NR_ACTIVE_FILE);
1367         inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1368
1369         return (active > inactive);
1370 }
1371
1372 /**
1373  * inactive_file_is_low - check if file pages need to be deactivated
1374  * @zone: zone to check
1375  * @sc:   scan control of this context
1376  *
1377  * When the system is doing streaming IO, memory pressure here
1378  * ensures that active file pages get deactivated, until more
1379  * than half of the file pages are on the inactive list.
1380  *
1381  * Once we get to that situation, protect the system's working
1382  * set from being evicted by disabling active file page aging.
1383  *
1384  * This uses a different ratio than the anonymous pages, because
1385  * the page cache uses a use-once replacement algorithm.
1386  */
1387 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1388 {
1389         int low;
1390
1391         if (scanning_global_lru(sc))
1392                 low = inactive_file_is_low_global(zone);
1393         else
1394                 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1395         return low;
1396 }
1397
1398 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1399         struct zone *zone, struct scan_control *sc, int priority)
1400 {
1401         int file = is_file_lru(lru);
1402
1403         if (lru == LRU_ACTIVE_FILE && inactive_file_is_low(zone, sc)) {
1404                 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1405                 return 0;
1406         }
1407
1408         if (lru == LRU_ACTIVE_ANON && inactive_anon_is_low(zone, sc)) {
1409                 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1410                 return 0;
1411         }
1412         return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1413 }
1414
1415 /*
1416  * Determine how aggressively the anon and file LRU lists should be
1417  * scanned.  The relative value of each set of LRU lists is determined
1418  * by looking at the fraction of the pages scanned we did rotate back
1419  * onto the active list instead of evict.
1420  *
1421  * percent[0] specifies how much pressure to put on ram/swap backed
1422  * memory, while percent[1] determines pressure on the file LRUs.
1423  */
1424 static void get_scan_ratio(struct zone *zone, struct scan_control *sc,
1425                                         unsigned long *percent)
1426 {
1427         unsigned long anon, file, free;
1428         unsigned long anon_prio, file_prio;
1429         unsigned long ap, fp;
1430         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1431
1432         anon  = zone_nr_pages(zone, sc, LRU_ACTIVE_ANON) +
1433                 zone_nr_pages(zone, sc, LRU_INACTIVE_ANON);
1434         file  = zone_nr_pages(zone, sc, LRU_ACTIVE_FILE) +
1435                 zone_nr_pages(zone, sc, LRU_INACTIVE_FILE);
1436
1437         if (scanning_global_lru(sc)) {
1438                 free  = zone_page_state(zone, NR_FREE_PAGES);
1439                 /* If we have very few page cache pages,
1440                    force-scan anon pages. */
1441                 if (unlikely(file + free <= high_wmark_pages(zone))) {
1442                         percent[0] = 100;
1443                         percent[1] = 0;
1444                         return;
1445                 }
1446         }
1447
1448         /*
1449          * OK, so we have swap space and a fair amount of page cache
1450          * pages.  We use the recently rotated / recently scanned
1451          * ratios to determine how valuable each cache is.
1452          *
1453          * Because workloads change over time (and to avoid overflow)
1454          * we keep these statistics as a floating average, which ends
1455          * up weighing recent references more than old ones.
1456          *
1457          * anon in [0], file in [1]
1458          */
1459         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1460                 spin_lock_irq(&zone->lru_lock);
1461                 reclaim_stat->recent_scanned[0] /= 2;
1462                 reclaim_stat->recent_rotated[0] /= 2;
1463                 spin_unlock_irq(&zone->lru_lock);
1464         }
1465
1466         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1467                 spin_lock_irq(&zone->lru_lock);
1468                 reclaim_stat->recent_scanned[1] /= 2;
1469                 reclaim_stat->recent_rotated[1] /= 2;
1470                 spin_unlock_irq(&zone->lru_lock);
1471         }
1472
1473         /*
1474          * With swappiness at 100, anonymous and file have the same priority.
1475          * This scanning priority is essentially the inverse of IO cost.
1476          */
1477         anon_prio = sc->swappiness;
1478         file_prio = 200 - sc->swappiness;
1479
1480         /*
1481          * The amount of pressure on anon vs file pages is inversely
1482          * proportional to the fraction of recently scanned pages on
1483          * each list that were recently referenced and in active use.
1484          */
1485         ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1486         ap /= reclaim_stat->recent_rotated[0] + 1;
1487
1488         fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1489         fp /= reclaim_stat->recent_rotated[1] + 1;
1490
1491         /* Normalize to percentages */
1492         percent[0] = 100 * ap / (ap + fp + 1);
1493         percent[1] = 100 - percent[0];
1494 }
1495
1496 /*
1497  * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1498  * until we collected @swap_cluster_max pages to scan.
1499  */
1500 static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
1501                                        unsigned long *nr_saved_scan,
1502                                        unsigned long swap_cluster_max)
1503 {
1504         unsigned long nr;
1505
1506         *nr_saved_scan += nr_to_scan;
1507         nr = *nr_saved_scan;
1508
1509         if (nr >= swap_cluster_max)
1510                 *nr_saved_scan = 0;
1511         else
1512                 nr = 0;
1513
1514         return nr;
1515 }
1516
1517 /*
1518  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1519  */
1520 static void shrink_zone(int priority, struct zone *zone,
1521                                 struct scan_control *sc)
1522 {
1523         unsigned long nr[NR_LRU_LISTS];
1524         unsigned long nr_to_scan;
1525         unsigned long percent[2];       /* anon @ 0; file @ 1 */
1526         enum lru_list l;
1527         unsigned long nr_reclaimed = sc->nr_reclaimed;
1528         unsigned long swap_cluster_max = sc->swap_cluster_max;
1529         int noswap = 0;
1530
1531         /* If we have no swap space, do not bother scanning anon pages. */
1532         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1533                 noswap = 1;
1534                 percent[0] = 0;
1535                 percent[1] = 100;
1536         } else
1537                 get_scan_ratio(zone, sc, percent);
1538
1539         for_each_evictable_lru(l) {
1540                 int file = is_file_lru(l);
1541                 unsigned long scan;
1542
1543                 scan = zone_nr_pages(zone, sc, l);
1544                 if (priority || noswap) {
1545                         scan >>= priority;
1546                         scan = (scan * percent[file]) / 100;
1547                 }
1548                 if (scanning_global_lru(sc))
1549                         nr[l] = nr_scan_try_batch(scan,
1550                                                   &zone->lru[l].nr_saved_scan,
1551                                                   swap_cluster_max);
1552                 else
1553                         nr[l] = scan;
1554         }
1555
1556         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1557                                         nr[LRU_INACTIVE_FILE]) {
1558                 for_each_evictable_lru(l) {
1559                         if (nr[l]) {
1560                                 nr_to_scan = min(nr[l], swap_cluster_max);
1561                                 nr[l] -= nr_to_scan;
1562
1563                                 nr_reclaimed += shrink_list(l, nr_to_scan,
1564                                                             zone, sc, priority);
1565                         }
1566                 }
1567                 /*
1568                  * On large memory systems, scan >> priority can become
1569                  * really large. This is fine for the starting priority;
1570                  * we want to put equal scanning pressure on each zone.
1571                  * However, if the VM has a harder time of freeing pages,
1572                  * with multiple processes reclaiming pages, the total
1573                  * freeing target can get unreasonably large.
1574                  */
1575                 if (nr_reclaimed > swap_cluster_max &&
1576                         priority < DEF_PRIORITY && !current_is_kswapd())
1577                         break;
1578         }
1579
1580         sc->nr_reclaimed = nr_reclaimed;
1581
1582         /*
1583          * Even if we did not try to evict anon pages at all, we want to
1584          * rebalance the anon lru active/inactive ratio.
1585          */
1586         if (inactive_anon_is_low(zone, sc) && nr_swap_pages > 0)
1587                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1588
1589         throttle_vm_writeout(sc->gfp_mask);
1590 }
1591
1592 /*
1593  * This is the direct reclaim path, for page-allocating processes.  We only
1594  * try to reclaim pages from zones which will satisfy the caller's allocation
1595  * request.
1596  *
1597  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1598  * Because:
1599  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1600  *    allocation or
1601  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1602  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1603  *    zone defense algorithm.
1604  *
1605  * If a zone is deemed to be full of pinned pages then just give it a light
1606  * scan then give up on it.
1607  */
1608 static void shrink_zones(int priority, struct zonelist *zonelist,
1609                                         struct scan_control *sc)
1610 {
1611         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1612         struct zoneref *z;
1613         struct zone *zone;
1614
1615         sc->all_unreclaimable = 1;
1616         for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1617                                         sc->nodemask) {
1618                 if (!populated_zone(zone))
1619                         continue;
1620                 /*
1621                  * Take care memory controller reclaiming has small influence
1622                  * to global LRU.
1623                  */
1624                 if (scanning_global_lru(sc)) {
1625                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1626                                 continue;
1627                         note_zone_scanning_priority(zone, priority);
1628
1629                         if (zone_is_all_unreclaimable(zone) &&
1630                                                 priority != DEF_PRIORITY)
1631                                 continue;       /* Let kswapd poll it */
1632                         sc->all_unreclaimable = 0;
1633                 } else {
1634                         /*
1635                          * Ignore cpuset limitation here. We just want to reduce
1636                          * # of used pages by us regardless of memory shortage.
1637                          */
1638                         sc->all_unreclaimable = 0;
1639                         mem_cgroup_note_reclaim_priority(sc->mem_cgroup,
1640                                                         priority);
1641                 }
1642
1643                 shrink_zone(priority, zone, sc);
1644         }
1645 }
1646
1647 /*
1648  * This is the main entry point to direct page reclaim.
1649  *
1650  * If a full scan of the inactive list fails to free enough memory then we
1651  * are "out of memory" and something needs to be killed.
1652  *
1653  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1654  * high - the zone may be full of dirty or under-writeback pages, which this
1655  * caller can't do much about.  We kick pdflush and take explicit naps in the
1656  * hope that some of these pages can be written.  But if the allocating task
1657  * holds filesystem locks which prevent writeout this might not work, and the
1658  * allocation attempt will fail.
1659  *
1660  * returns:     0, if no pages reclaimed
1661  *              else, the number of pages reclaimed
1662  */
1663 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
1664                                         struct scan_control *sc)
1665 {
1666         int priority;
1667         unsigned long ret = 0;
1668         unsigned long total_scanned = 0;
1669         struct reclaim_state *reclaim_state = current->reclaim_state;
1670         unsigned long lru_pages = 0;
1671         struct zoneref *z;
1672         struct zone *zone;
1673         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1674
1675         delayacct_freepages_start();
1676
1677         if (scanning_global_lru(sc))
1678                 count_vm_event(ALLOCSTALL);
1679         /*
1680          * mem_cgroup will not do shrink_slab.
1681          */
1682         if (scanning_global_lru(sc)) {
1683                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1684
1685                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1686                                 continue;
1687
1688                         lru_pages += zone_lru_pages(zone);
1689                 }
1690         }
1691
1692         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1693                 sc->nr_scanned = 0;
1694                 if (!priority)
1695                         disable_swap_token();
1696                 shrink_zones(priority, zonelist, sc);
1697                 /*
1698                  * Don't shrink slabs when reclaiming memory from
1699                  * over limit cgroups
1700                  */
1701                 if (scanning_global_lru(sc)) {
1702                         shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
1703                         if (reclaim_state) {
1704                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
1705                                 reclaim_state->reclaimed_slab = 0;
1706                         }
1707                 }
1708                 total_scanned += sc->nr_scanned;
1709                 if (sc->nr_reclaimed >= sc->swap_cluster_max) {
1710                         ret = sc->nr_reclaimed;
1711                         goto out;
1712                 }
1713
1714                 /*
1715                  * Try to write back as many pages as we just scanned.  This
1716                  * tends to cause slow streaming writers to write data to the
1717                  * disk smoothly, at the dirtying rate, which is nice.   But
1718                  * that's undesirable in laptop mode, where we *want* lumpy
1719                  * writeout.  So in laptop mode, write out the whole world.
1720                  */
1721                 if (total_scanned > sc->swap_cluster_max +
1722                                         sc->swap_cluster_max / 2) {
1723                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
1724                         sc->may_writepage = 1;
1725                 }
1726
1727                 /* Take a nap, wait for some writeback to complete */
1728                 if (sc->nr_scanned && priority < DEF_PRIORITY - 2)
1729                         congestion_wait(BLK_RW_ASYNC, HZ/10);
1730         }
1731         /* top priority shrink_zones still had more to do? don't OOM, then */
1732         if (!sc->all_unreclaimable && scanning_global_lru(sc))
1733                 ret = sc->nr_reclaimed;
1734 out:
1735         /*
1736          * Now that we've scanned all the zones at this priority level, note
1737          * that level within the zone so that the next thread which performs
1738          * scanning of this zone will immediately start out at this priority
1739          * level.  This affects only the decision whether or not to bring
1740          * mapped pages onto the inactive list.
1741          */
1742         if (priority < 0)
1743                 priority = 0;
1744
1745         if (scanning_global_lru(sc)) {
1746                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1747
1748                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1749                                 continue;
1750
1751                         zone->prev_priority = priority;
1752                 }
1753         } else
1754                 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority);
1755
1756         delayacct_freepages_end();
1757
1758         return ret;
1759 }
1760
1761 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
1762                                 gfp_t gfp_mask, nodemask_t *nodemask)
1763 {
1764         struct scan_control sc = {
1765                 .gfp_mask = gfp_mask,
1766                 .may_writepage = !laptop_mode,
1767                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1768                 .may_unmap = 1,
1769                 .may_swap = 1,
1770                 .swappiness = vm_swappiness,
1771                 .order = order,
1772                 .mem_cgroup = NULL,
1773                 .isolate_pages = isolate_pages_global,
1774                 .nodemask = nodemask,
1775         };
1776
1777         return do_try_to_free_pages(zonelist, &sc);
1778 }
1779
1780 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
1781
1782 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
1783                                            gfp_t gfp_mask,
1784                                            bool noswap,
1785                                            unsigned int swappiness)
1786 {
1787         struct scan_control sc = {
1788                 .may_writepage = !laptop_mode,
1789                 .may_unmap = 1,
1790                 .may_swap = !noswap,
1791                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1792                 .swappiness = swappiness,
1793                 .order = 0,
1794                 .mem_cgroup = mem_cont,
1795                 .isolate_pages = mem_cgroup_isolate_pages,
1796                 .nodemask = NULL, /* we don't care the placement */
1797         };
1798         struct zonelist *zonelist;
1799
1800         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1801                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1802         zonelist = NODE_DATA(numa_node_id())->node_zonelists;
1803         return do_try_to_free_pages(zonelist, &sc);
1804 }
1805 #endif
1806
1807 /*
1808  * For kswapd, balance_pgdat() will work across all this node's zones until
1809  * they are all at high_wmark_pages(zone).
1810  *
1811  * Returns the number of pages which were actually freed.
1812  *
1813  * There is special handling here for zones which are full of pinned pages.
1814  * This can happen if the pages are all mlocked, or if they are all used by
1815  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1816  * What we do is to detect the case where all pages in the zone have been
1817  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1818  * dead and from now on, only perform a short scan.  Basically we're polling
1819  * the zone for when the problem goes away.
1820  *
1821  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
1822  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
1823  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
1824  * lower zones regardless of the number of free pages in the lower zones. This
1825  * interoperates with the page allocator fallback scheme to ensure that aging
1826  * of pages is balanced across the zones.
1827  */
1828 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
1829 {
1830         int all_zones_ok;
1831         int priority;
1832         int i;
1833         unsigned long total_scanned;
1834         struct reclaim_state *reclaim_state = current->reclaim_state;
1835         struct scan_control sc = {
1836                 .gfp_mask = GFP_KERNEL,
1837                 .may_unmap = 1,
1838                 .may_swap = 1,
1839                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1840                 .swappiness = vm_swappiness,
1841                 .order = order,
1842                 .mem_cgroup = NULL,
1843                 .isolate_pages = isolate_pages_global,
1844         };
1845         /*
1846          * temp_priority is used to remember the scanning priority at which
1847          * this zone was successfully refilled to
1848          * free_pages == high_wmark_pages(zone).
1849          */
1850         int temp_priority[MAX_NR_ZONES];
1851
1852 loop_again:
1853         total_scanned = 0;
1854         sc.nr_reclaimed = 0;
1855         sc.may_writepage = !laptop_mode;
1856         count_vm_event(PAGEOUTRUN);
1857
1858         for (i = 0; i < pgdat->nr_zones; i++)
1859                 temp_priority[i] = DEF_PRIORITY;
1860
1861         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1862                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1863                 unsigned long lru_pages = 0;
1864
1865                 /* The swap token gets in the way of swapout... */
1866                 if (!priority)
1867                         disable_swap_token();
1868
1869                 all_zones_ok = 1;
1870
1871                 /*
1872                  * Scan in the highmem->dma direction for the highest
1873                  * zone which needs scanning
1874                  */
1875                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1876                         struct zone *zone = pgdat->node_zones + i;
1877
1878                         if (!populated_zone(zone))
1879                                 continue;
1880
1881                         if (zone_is_all_unreclaimable(zone) &&
1882                             priority != DEF_PRIORITY)
1883                                 continue;
1884
1885                         /*
1886                          * Do some background aging of the anon list, to give
1887                          * pages a chance to be referenced before reclaiming.
1888                          */
1889                         if (inactive_anon_is_low(zone, &sc))
1890                                 shrink_active_list(SWAP_CLUSTER_MAX, zone,
1891                                                         &sc, priority, 0);
1892
1893                         if (!zone_watermark_ok(zone, order,
1894                                         high_wmark_pages(zone), 0, 0)) {
1895                                 end_zone = i;
1896                                 break;
1897                         }
1898                 }
1899                 if (i < 0)
1900                         goto out;
1901
1902                 for (i = 0; i <= end_zone; i++) {
1903                         struct zone *zone = pgdat->node_zones + i;
1904
1905                         lru_pages += zone_lru_pages(zone);
1906                 }
1907
1908                 /*
1909                  * Now scan the zone in the dma->highmem direction, stopping
1910                  * at the last zone which needs scanning.
1911                  *
1912                  * We do this because the page allocator works in the opposite
1913                  * direction.  This prevents the page allocator from allocating
1914                  * pages behind kswapd's direction of progress, which would
1915                  * cause too much scanning of the lower zones.
1916                  */
1917                 for (i = 0; i <= end_zone; i++) {
1918                         struct zone *zone = pgdat->node_zones + i;
1919                         int nr_slab;
1920
1921                         if (!populated_zone(zone))
1922                                 continue;
1923
1924                         if (zone_is_all_unreclaimable(zone) &&
1925                                         priority != DEF_PRIORITY)
1926                                 continue;
1927
1928                         if (!zone_watermark_ok(zone, order,
1929                                         high_wmark_pages(zone), end_zone, 0))
1930                                 all_zones_ok = 0;
1931                         temp_priority[i] = priority;
1932                         sc.nr_scanned = 0;
1933                         note_zone_scanning_priority(zone, priority);
1934                         /*
1935                          * We put equal pressure on every zone, unless one
1936                          * zone has way too many pages free already.
1937                          */
1938                         if (!zone_watermark_ok(zone, order,
1939                                         8*high_wmark_pages(zone), end_zone, 0))
1940                                 shrink_zone(priority, zone, &sc);
1941                         reclaim_state->reclaimed_slab = 0;
1942                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1943                                                 lru_pages);
1944                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1945                         total_scanned += sc.nr_scanned;
1946                         if (zone_is_all_unreclaimable(zone))
1947                                 continue;
1948                         if (nr_slab == 0 && zone->pages_scanned >=
1949                                                 (zone_lru_pages(zone) * 6))
1950                                         zone_set_flag(zone,
1951                                                       ZONE_ALL_UNRECLAIMABLE);
1952                         /*
1953                          * If we've done a decent amount of scanning and
1954                          * the reclaim ratio is low, start doing writepage
1955                          * even in laptop mode
1956                          */
1957                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1958                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
1959                                 sc.may_writepage = 1;
1960                 }
1961                 if (all_zones_ok)
1962                         break;          /* kswapd: all done */
1963                 /*
1964                  * OK, kswapd is getting into trouble.  Take a nap, then take
1965                  * another pass across the zones.
1966                  */
1967                 if (total_scanned && priority < DEF_PRIORITY - 2)
1968                         congestion_wait(BLK_RW_ASYNC, HZ/10);
1969
1970                 /*
1971                  * We do this so kswapd doesn't build up large priorities for
1972                  * example when it is freeing in parallel with allocators. It
1973                  * matches the direct reclaim path behaviour in terms of impact
1974                  * on zone->*_priority.
1975                  */
1976                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
1977                         break;
1978         }
1979 out:
1980         /*
1981          * Note within each zone the priority level at which this zone was
1982          * brought into a happy state.  So that the next thread which scans this
1983          * zone will start out at that priority level.
1984          */
1985         for (i = 0; i < pgdat->nr_zones; i++) {
1986                 struct zone *zone = pgdat->node_zones + i;
1987
1988                 zone->prev_priority = temp_priority[i];
1989         }
1990         if (!all_zones_ok) {
1991                 cond_resched();
1992
1993                 try_to_freeze();
1994
1995                 /*
1996                  * Fragmentation may mean that the system cannot be
1997                  * rebalanced for high-order allocations in all zones.
1998                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
1999                  * it means the zones have been fully scanned and are still
2000                  * not balanced. For high-order allocations, there is
2001                  * little point trying all over again as kswapd may
2002                  * infinite loop.
2003                  *
2004                  * Instead, recheck all watermarks at order-0 as they
2005                  * are the most important. If watermarks are ok, kswapd will go
2006                  * back to sleep. High-order users can still perform direct
2007                  * reclaim if they wish.
2008                  */
2009                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2010                         order = sc.order = 0;
2011
2012                 goto loop_again;
2013         }
2014
2015         return sc.nr_reclaimed;
2016 }
2017
2018 /*
2019  * The background pageout daemon, started as a kernel thread
2020  * from the init process.
2021  *
2022  * This basically trickles out pages so that we have _some_
2023  * free memory available even if there is no other activity
2024  * that frees anything up. This is needed for things like routing
2025  * etc, where we otherwise might have all activity going on in
2026  * asynchronous contexts that cannot page things out.
2027  *
2028  * If there are applications that are active memory-allocators
2029  * (most normal use), this basically shouldn't matter.
2030  */
2031 static int kswapd(void *p)
2032 {
2033         unsigned long order;
2034         pg_data_t *pgdat = (pg_data_t*)p;
2035         struct task_struct *tsk = current;
2036         DEFINE_WAIT(wait);
2037         struct reclaim_state reclaim_state = {
2038                 .reclaimed_slab = 0,
2039         };
2040         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2041
2042         lockdep_set_current_reclaim_state(GFP_KERNEL);
2043
2044         if (!cpumask_empty(cpumask))
2045                 set_cpus_allowed_ptr(tsk, cpumask);
2046         current->reclaim_state = &reclaim_state;
2047
2048         /*
2049          * Tell the memory management that we're a "memory allocator",
2050          * and that if we need more memory we should get access to it
2051          * regardless (see "__alloc_pages()"). "kswapd" should
2052          * never get caught in the normal page freeing logic.
2053          *
2054          * (Kswapd normally doesn't need memory anyway, but sometimes
2055          * you need a small amount of memory in order to be able to
2056          * page out something else, and this flag essentially protects
2057          * us from recursively trying to free more memory as we're
2058          * trying to free the first piece of memory in the first place).
2059          */
2060         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2061         set_freezable();
2062
2063         order = 0;
2064         for ( ; ; ) {
2065                 unsigned long new_order;
2066
2067                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2068                 new_order = pgdat->kswapd_max_order;
2069                 pgdat->kswapd_max_order = 0;
2070                 if (order < new_order) {
2071                         /*
2072                          * Don't sleep if someone wants a larger 'order'
2073                          * allocation
2074                          */
2075                         order = new_order;
2076                 } else {
2077                         if (!freezing(current))
2078                                 schedule();
2079
2080                         order = pgdat->kswapd_max_order;
2081                 }
2082                 finish_wait(&pgdat->kswapd_wait, &wait);
2083
2084                 if (!try_to_freeze()) {
2085                         /* We can speed up thawing tasks if we don't call
2086                          * balance_pgdat after returning from the refrigerator
2087                          */
2088                         balance_pgdat(pgdat, order);
2089                 }
2090         }
2091         return 0;
2092 }
2093
2094 /*
2095  * A zone is low on free memory, so wake its kswapd task to service it.
2096  */
2097 void wakeup_kswapd(struct zone *zone, int order)
2098 {
2099         pg_data_t *pgdat;
2100
2101         if (!populated_zone(zone))
2102                 return;
2103
2104         pgdat = zone->zone_pgdat;
2105         if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
2106                 return;
2107         if (pgdat->kswapd_max_order < order)
2108                 pgdat->kswapd_max_order = order;
2109         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2110                 return;
2111         if (!waitqueue_active(&pgdat->kswapd_wait))
2112                 return;
2113         wake_up_interruptible(&pgdat->kswapd_wait);
2114 }
2115
2116 unsigned long global_lru_pages(void)
2117 {
2118         return global_page_state(NR_ACTIVE_ANON)
2119                 + global_page_state(NR_ACTIVE_FILE)
2120                 + global_page_state(NR_INACTIVE_ANON)
2121                 + global_page_state(NR_INACTIVE_FILE);
2122 }
2123
2124 #ifdef CONFIG_HIBERNATION
2125 /*
2126  * Helper function for shrink_all_memory().  Tries to reclaim 'nr_pages' pages
2127  * from LRU lists system-wide, for given pass and priority.
2128  *
2129  * For pass > 3 we also try to shrink the LRU lists that contain a few pages
2130  */
2131 static void shrink_all_zones(unsigned long nr_pages, int prio,
2132                                       int pass, struct scan_control *sc)
2133 {
2134         struct zone *zone;
2135         unsigned long nr_reclaimed = 0;
2136
2137         for_each_populated_zone(zone) {
2138                 enum lru_list l;
2139
2140                 if (zone_is_all_unreclaimable(zone) && prio != DEF_PRIORITY)
2141                         continue;
2142
2143                 for_each_evictable_lru(l) {
2144                         enum zone_stat_item ls = NR_LRU_BASE + l;
2145                         unsigned long lru_pages = zone_page_state(zone, ls);
2146
2147                         /* For pass = 0, we don't shrink the active list */
2148                         if (pass == 0 && (l == LRU_ACTIVE_ANON ||
2149                                                 l == LRU_ACTIVE_FILE))
2150                                 continue;
2151
2152                         zone->lru[l].nr_saved_scan += (lru_pages >> prio) + 1;
2153                         if (zone->lru[l].nr_saved_scan >= nr_pages || pass > 3) {
2154                                 unsigned long nr_to_scan;
2155
2156                                 zone->lru[l].nr_saved_scan = 0;
2157                                 nr_to_scan = min(nr_pages, lru_pages);
2158                                 nr_reclaimed += shrink_list(l, nr_to_scan, zone,
2159                                                                 sc, prio);
2160                                 if (nr_reclaimed >= nr_pages) {
2161                                         sc->nr_reclaimed += nr_reclaimed;
2162                                         return;
2163                                 }
2164                         }
2165                 }
2166         }
2167         sc->nr_reclaimed += nr_reclaimed;
2168 }
2169
2170 /*
2171  * Try to free `nr_pages' of memory, system-wide, and return the number of
2172  * freed pages.
2173  *
2174  * Rather than trying to age LRUs the aim is to preserve the overall
2175  * LRU order by reclaiming preferentially
2176  * inactive > active > active referenced > active mapped
2177  */
2178 unsigned long shrink_all_memory(unsigned long nr_pages)
2179 {
2180         unsigned long lru_pages, nr_slab;
2181         int pass;
2182         struct reclaim_state reclaim_state;
2183         struct scan_control sc = {
2184                 .gfp_mask = GFP_KERNEL,
2185                 .may_unmap = 0,
2186                 .may_writepage = 1,
2187                 .isolate_pages = isolate_pages_global,
2188                 .nr_reclaimed = 0,
2189         };
2190
2191         current->reclaim_state = &reclaim_state;
2192
2193         lru_pages = global_lru_pages();
2194         nr_slab = global_page_state(NR_SLAB_RECLAIMABLE);
2195         /* If slab caches are huge, it's better to hit them first */
2196         while (nr_slab >= lru_pages) {
2197                 reclaim_state.reclaimed_slab = 0;
2198                 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
2199                 if (!reclaim_state.reclaimed_slab)
2200                         break;
2201
2202                 sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2203                 if (sc.nr_reclaimed >= nr_pages)
2204                         goto out;
2205
2206                 nr_slab -= reclaim_state.reclaimed_slab;
2207         }
2208
2209         /*
2210          * We try to shrink LRUs in 5 passes:
2211          * 0 = Reclaim from inactive_list only
2212          * 1 = Reclaim from active list but don't reclaim mapped
2213          * 2 = 2nd pass of type 1
2214          * 3 = Reclaim mapped (normal reclaim)
2215          * 4 = 2nd pass of type 3
2216          */
2217         for (pass = 0; pass < 5; pass++) {
2218                 int prio;
2219
2220                 /* Force reclaiming mapped pages in the passes #3 and #4 */
2221                 if (pass > 2)
2222                         sc.may_unmap = 1;
2223
2224                 for (prio = DEF_PRIORITY; prio >= 0; prio--) {
2225                         unsigned long nr_to_scan = nr_pages - sc.nr_reclaimed;
2226
2227                         sc.nr_scanned = 0;
2228                         sc.swap_cluster_max = nr_to_scan;
2229                         shrink_all_zones(nr_to_scan, prio, pass, &sc);
2230                         if (sc.nr_reclaimed >= nr_pages)
2231                                 goto out;
2232
2233                         reclaim_state.reclaimed_slab = 0;
2234                         shrink_slab(sc.nr_scanned, sc.gfp_mask,
2235                                         global_lru_pages());
2236                         sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2237                         if (sc.nr_reclaimed >= nr_pages)
2238                                 goto out;
2239
2240                         if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
2241                                 congestion_wait(BLK_RW_ASYNC, HZ / 10);
2242                 }
2243         }
2244
2245         /*
2246          * If sc.nr_reclaimed = 0, we could not shrink LRUs, but there may be
2247          * something in slab caches
2248          */
2249         if (!sc.nr_reclaimed) {
2250                 do {
2251                         reclaim_state.reclaimed_slab = 0;
2252                         shrink_slab(nr_pages, sc.gfp_mask, global_lru_pages());
2253                         sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2254                 } while (sc.nr_reclaimed < nr_pages &&
2255                                 reclaim_state.reclaimed_slab > 0);
2256         }
2257
2258
2259 out:
2260         current->reclaim_state = NULL;
2261
2262         return sc.nr_reclaimed;
2263 }
2264 #endif /* CONFIG_HIBERNATION */
2265
2266 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2267    not required for correctness.  So if the last cpu in a node goes
2268    away, we get changed to run anywhere: as the first one comes back,
2269    restore their cpu bindings. */
2270 static int __devinit cpu_callback(struct notifier_block *nfb,
2271                                   unsigned long action, void *hcpu)
2272 {
2273         int nid;
2274
2275         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2276                 for_each_node_state(nid, N_HIGH_MEMORY) {
2277                         pg_data_t *pgdat = NODE_DATA(nid);
2278                         const struct cpumask *mask;
2279
2280                         mask = cpumask_of_node(pgdat->node_id);
2281
2282                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2283                                 /* One of our CPUs online: restore mask */
2284                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2285                 }
2286         }
2287         return NOTIFY_OK;
2288 }
2289
2290 /*
2291  * This kswapd start function will be called by init and node-hot-add.
2292  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2293  */
2294 int kswapd_run(int nid)
2295 {
2296         pg_data_t *pgdat = NODE_DATA(nid);
2297         int ret = 0;
2298
2299         if (pgdat->kswapd)
2300                 return 0;
2301
2302         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2303         if (IS_ERR(pgdat->kswapd)) {
2304                 /* failure at boot is fatal */
2305                 BUG_ON(system_state == SYSTEM_BOOTING);
2306                 printk("Failed to start kswapd on node %d\n",nid);
2307                 ret = -1;
2308         }
2309         return ret;
2310 }
2311
2312 static int __init kswapd_init(void)
2313 {
2314         int nid;
2315
2316         swap_setup();
2317         for_each_node_state(nid, N_HIGH_MEMORY)
2318                 kswapd_run(nid);
2319         hotcpu_notifier(cpu_callback, 0);
2320         return 0;
2321 }
2322
2323 module_init(kswapd_init)
2324
2325 #ifdef CONFIG_NUMA
2326 /*
2327  * Zone reclaim mode
2328  *
2329  * If non-zero call zone_reclaim when the number of free pages falls below
2330  * the watermarks.
2331  */
2332 int zone_reclaim_mode __read_mostly;
2333
2334 #define RECLAIM_OFF 0
2335 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
2336 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
2337 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
2338
2339 /*
2340  * Priority for ZONE_RECLAIM. This determines the fraction of pages
2341  * of a node considered for each zone_reclaim. 4 scans 1/16th of
2342  * a zone.
2343  */
2344 #define ZONE_RECLAIM_PRIORITY 4
2345
2346 /*
2347  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2348  * occur.
2349  */
2350 int sysctl_min_unmapped_ratio = 1;
2351
2352 /*
2353  * If the number of slab pages in a zone grows beyond this percentage then
2354  * slab reclaim needs to occur.
2355  */
2356 int sysctl_min_slab_ratio = 5;
2357
2358 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
2359 {
2360         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
2361         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
2362                 zone_page_state(zone, NR_ACTIVE_FILE);
2363
2364         /*
2365          * It's possible for there to be more file mapped pages than
2366          * accounted for by the pages on the file LRU lists because
2367          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2368          */
2369         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
2370 }
2371
2372 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
2373 static long zone_pagecache_reclaimable(struct zone *zone)
2374 {
2375         long nr_pagecache_reclaimable;
2376         long delta = 0;
2377
2378         /*
2379          * If RECLAIM_SWAP is set, then all file pages are considered
2380          * potentially reclaimable. Otherwise, we have to worry about
2381          * pages like swapcache and zone_unmapped_file_pages() provides
2382          * a better estimate
2383          */
2384         if (zone_reclaim_mode & RECLAIM_SWAP)
2385                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
2386         else
2387                 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
2388
2389         /* If we can't clean pages, remove dirty pages from consideration */
2390         if (!(zone_reclaim_mode & RECLAIM_WRITE))
2391                 delta += zone_page_state(zone, NR_FILE_DIRTY);
2392
2393         /* Watch for any possible underflows due to delta */
2394         if (unlikely(delta > nr_pagecache_reclaimable))
2395                 delta = nr_pagecache_reclaimable;
2396
2397         return nr_pagecache_reclaimable - delta;
2398 }
2399
2400 /*
2401  * Try to free up some pages from this zone through reclaim.
2402  */
2403 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2404 {
2405         /* Minimum pages needed in order to stay on node */
2406         const unsigned long nr_pages = 1 << order;
2407         struct task_struct *p = current;
2408         struct reclaim_state reclaim_state;
2409         int priority;
2410         struct scan_control sc = {
2411                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2412                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2413                 .may_swap = 1,
2414                 .swap_cluster_max = max_t(unsigned long, nr_pages,
2415                                         SWAP_CLUSTER_MAX),
2416                 .gfp_mask = gfp_mask,
2417                 .swappiness = vm_swappiness,
2418                 .order = order,
2419                 .isolate_pages = isolate_pages_global,
2420         };
2421         unsigned long slab_reclaimable;
2422
2423         disable_swap_token();
2424         cond_resched();
2425         /*
2426          * We need to be able to allocate from the reserves for RECLAIM_SWAP
2427          * and we also need to be able to write out pages for RECLAIM_WRITE
2428          * and RECLAIM_SWAP.
2429          */
2430         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2431         reclaim_state.reclaimed_slab = 0;
2432         p->reclaim_state = &reclaim_state;
2433
2434         if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
2435                 /*
2436                  * Free memory by calling shrink zone with increasing
2437                  * priorities until we have enough memory freed.
2438                  */
2439                 priority = ZONE_RECLAIM_PRIORITY;
2440                 do {
2441                         note_zone_scanning_priority(zone, priority);
2442                         shrink_zone(priority, zone, &sc);
2443                         priority--;
2444                 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
2445         }
2446
2447         slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2448         if (slab_reclaimable > zone->min_slab_pages) {
2449                 /*
2450                  * shrink_slab() does not currently allow us to determine how
2451                  * many pages were freed in this zone. So we take the current
2452                  * number of slab pages and shake the slab until it is reduced
2453                  * by the same nr_pages that we used for reclaiming unmapped
2454                  * pages.
2455                  *
2456                  * Note that shrink_slab will free memory on all zones and may
2457                  * take a long time.
2458                  */
2459                 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
2460                         zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
2461                                 slab_reclaimable - nr_pages)
2462                         ;
2463
2464                 /*
2465                  * Update nr_reclaimed by the number of slab pages we
2466                  * reclaimed from this zone.
2467                  */
2468                 sc.nr_reclaimed += slab_reclaimable -
2469                         zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2470         }
2471
2472         p->reclaim_state = NULL;
2473         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
2474         return sc.nr_reclaimed >= nr_pages;
2475 }
2476
2477 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2478 {
2479         int node_id;
2480         int ret;
2481
2482         /*
2483          * Zone reclaim reclaims unmapped file backed pages and
2484          * slab pages if we are over the defined limits.
2485          *
2486          * A small portion of unmapped file backed pages is needed for
2487          * file I/O otherwise pages read by file I/O will be immediately
2488          * thrown out if the zone is overallocated. So we do not reclaim
2489          * if less than a specified percentage of the zone is used by
2490          * unmapped file backed pages.
2491          */
2492         if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
2493             zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
2494                 return ZONE_RECLAIM_FULL;
2495
2496         if (zone_is_all_unreclaimable(zone))
2497                 return ZONE_RECLAIM_FULL;
2498
2499         /*
2500          * Do not scan if the allocation should not be delayed.
2501          */
2502         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
2503                 return ZONE_RECLAIM_NOSCAN;
2504
2505         /*
2506          * Only run zone reclaim on the local zone or on zones that do not
2507          * have associated processors. This will favor the local processor
2508          * over remote processors and spread off node memory allocations
2509          * as wide as possible.
2510          */
2511         node_id = zone_to_nid(zone);
2512         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
2513                 return ZONE_RECLAIM_NOSCAN;
2514
2515         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2516                 return ZONE_RECLAIM_NOSCAN;
2517
2518         ret = __zone_reclaim(zone, gfp_mask, order);
2519         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2520
2521         if (!ret)
2522                 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
2523
2524         return ret;
2525 }
2526 #endif
2527
2528 /*
2529  * page_evictable - test whether a page is evictable
2530  * @page: the page to test
2531  * @vma: the VMA in which the page is or will be mapped, may be NULL
2532  *
2533  * Test whether page is evictable--i.e., should be placed on active/inactive
2534  * lists vs unevictable list.  The vma argument is !NULL when called from the
2535  * fault path to determine how to instantate a new page.
2536  *
2537  * Reasons page might not be evictable:
2538  * (1) page's mapping marked unevictable
2539  * (2) page is part of an mlocked VMA
2540  *
2541  */
2542 int page_evictable(struct page *page, struct vm_area_struct *vma)
2543 {
2544
2545         if (mapping_unevictable(page_mapping(page)))
2546                 return 0;
2547
2548         if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2549                 return 0;
2550
2551         return 1;
2552 }
2553
2554 /**
2555  * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2556  * @page: page to check evictability and move to appropriate lru list
2557  * @zone: zone page is in
2558  *
2559  * Checks a page for evictability and moves the page to the appropriate
2560  * zone lru list.
2561  *
2562  * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2563  * have PageUnevictable set.
2564  */
2565 static void check_move_unevictable_page(struct page *page, struct zone *zone)
2566 {
2567         VM_BUG_ON(PageActive(page));
2568
2569 retry:
2570         ClearPageUnevictable(page);
2571         if (page_evictable(page, NULL)) {
2572                 enum lru_list l = LRU_INACTIVE_ANON + page_is_file_cache(page);
2573
2574                 __dec_zone_state(zone, NR_UNEVICTABLE);
2575                 list_move(&page->lru, &zone->lru[l].list);
2576                 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
2577                 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
2578                 __count_vm_event(UNEVICTABLE_PGRESCUED);
2579         } else {
2580                 /*
2581                  * rotate unevictable list
2582                  */
2583                 SetPageUnevictable(page);
2584                 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
2585                 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
2586                 if (page_evictable(page, NULL))
2587                         goto retry;
2588         }
2589 }
2590
2591 /**
2592  * scan_mapping_unevictable_pages - scan an address space for evictable pages
2593  * @mapping: struct address_space to scan for evictable pages
2594  *
2595  * Scan all pages in mapping.  Check unevictable pages for
2596  * evictability and move them to the appropriate zone lru list.
2597  */
2598 void scan_mapping_unevictable_pages(struct address_space *mapping)
2599 {
2600         pgoff_t next = 0;
2601         pgoff_t end   = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
2602                          PAGE_CACHE_SHIFT;
2603         struct zone *zone;
2604         struct pagevec pvec;
2605
2606         if (mapping->nrpages == 0)
2607                 return;
2608
2609         pagevec_init(&pvec, 0);
2610         while (next < end &&
2611                 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
2612                 int i;
2613                 int pg_scanned = 0;
2614
2615                 zone = NULL;
2616
2617                 for (i = 0; i < pagevec_count(&pvec); i++) {
2618                         struct page *page = pvec.pages[i];
2619                         pgoff_t page_index = page->index;
2620                         struct zone *pagezone = page_zone(page);
2621
2622                         pg_scanned++;
2623                         if (page_index > next)
2624                                 next = page_index;
2625                         next++;
2626
2627                         if (pagezone != zone) {
2628                                 if (zone)
2629                                         spin_unlock_irq(&zone->lru_lock);
2630                                 zone = pagezone;
2631                                 spin_lock_irq(&zone->lru_lock);
2632                         }
2633
2634                         if (PageLRU(page) && PageUnevictable(page))
2635                                 check_move_unevictable_page(page, zone);
2636                 }
2637                 if (zone)
2638                         spin_unlock_irq(&zone->lru_lock);
2639                 pagevec_release(&pvec);
2640
2641                 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
2642         }
2643
2644 }
2645
2646 /**
2647  * scan_zone_unevictable_pages - check unevictable list for evictable pages
2648  * @zone - zone of which to scan the unevictable list
2649  *
2650  * Scan @zone's unevictable LRU lists to check for pages that have become
2651  * evictable.  Move those that have to @zone's inactive list where they
2652  * become candidates for reclaim, unless shrink_inactive_zone() decides
2653  * to reactivate them.  Pages that are still unevictable are rotated
2654  * back onto @zone's unevictable list.
2655  */
2656 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
2657 static void scan_zone_unevictable_pages(struct zone *zone)
2658 {
2659         struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
2660         unsigned long scan;
2661         unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
2662
2663         while (nr_to_scan > 0) {
2664                 unsigned long batch_size = min(nr_to_scan,
2665                                                 SCAN_UNEVICTABLE_BATCH_SIZE);
2666
2667                 spin_lock_irq(&zone->lru_lock);
2668                 for (scan = 0;  scan < batch_size; scan++) {
2669                         struct page *page = lru_to_page(l_unevictable);
2670
2671                         if (!trylock_page(page))
2672                                 continue;
2673
2674                         prefetchw_prev_lru_page(page, l_unevictable, flags);
2675
2676                         if (likely(PageLRU(page) && PageUnevictable(page)))
2677                                 check_move_unevictable_page(page, zone);
2678
2679                         unlock_page(page);
2680                 }
2681                 spin_unlock_irq(&zone->lru_lock);
2682
2683                 nr_to_scan -= batch_size;
2684         }
2685 }
2686
2687
2688 /**
2689  * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
2690  *
2691  * A really big hammer:  scan all zones' unevictable LRU lists to check for
2692  * pages that have become evictable.  Move those back to the zones'
2693  * inactive list where they become candidates for reclaim.
2694  * This occurs when, e.g., we have unswappable pages on the unevictable lists,
2695  * and we add swap to the system.  As such, it runs in the context of a task
2696  * that has possibly/probably made some previously unevictable pages
2697  * evictable.
2698  */
2699 static void scan_all_zones_unevictable_pages(void)
2700 {
2701         struct zone *zone;
2702
2703         for_each_zone(zone) {
2704                 scan_zone_unevictable_pages(zone);
2705         }
2706 }
2707
2708 /*
2709  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
2710  * all nodes' unevictable lists for evictable pages
2711  */
2712 unsigned long scan_unevictable_pages;
2713
2714 int scan_unevictable_handler(struct ctl_table *table, int write,
2715                            struct file *file, void __user *buffer,
2716                            size_t *length, loff_t *ppos)
2717 {
2718         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
2719
2720         if (write && *(unsigned long *)table->data)
2721                 scan_all_zones_unevictable_pages();
2722
2723         scan_unevictable_pages = 0;
2724         return 0;
2725 }
2726
2727 /*
2728  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
2729  * a specified node's per zone unevictable lists for evictable pages.
2730  */
2731
2732 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
2733                                           struct sysdev_attribute *attr,
2734                                           char *buf)
2735 {
2736         return sprintf(buf, "0\n");     /* always zero; should fit... */
2737 }
2738
2739 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
2740                                            struct sysdev_attribute *attr,
2741                                         const char *buf, size_t count)
2742 {
2743         struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
2744         struct zone *zone;
2745         unsigned long res;
2746         unsigned long req = strict_strtoul(buf, 10, &res);
2747
2748         if (!req)
2749                 return 1;       /* zero is no-op */
2750
2751         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2752                 if (!populated_zone(zone))
2753                         continue;
2754                 scan_zone_unevictable_pages(zone);
2755         }
2756         return 1;
2757 }
2758
2759
2760 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
2761                         read_scan_unevictable_node,
2762                         write_scan_unevictable_node);
2763
2764 int scan_unevictable_register_node(struct node *node)
2765 {
2766         return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
2767 }
2768
2769 void scan_unevictable_unregister_node(struct node *node)
2770 {
2771         sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
2772 }
2773