mm: add swap cache interface for swap reference
[linux-2.6.git] / mm / vmscan.c
1 /*
2  *  linux/mm/vmscan.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  *
6  *  Swap reorganised 29.12.95, Stephen Tweedie.
7  *  kswapd added: 7.1.96  sct
8  *  Removed kswapd_ctl limits, and swap out as many pages as needed
9  *  to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10  *  Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11  *  Multiqueue VM started 5.8.00, Rik van Riel.
12  */
13
14 #include <linux/mm.h>
15 #include <linux/module.h>
16 #include <linux/slab.h>
17 #include <linux/kernel_stat.h>
18 #include <linux/swap.h>
19 #include <linux/pagemap.h>
20 #include <linux/init.h>
21 #include <linux/highmem.h>
22 #include <linux/vmstat.h>
23 #include <linux/file.h>
24 #include <linux/writeback.h>
25 #include <linux/blkdev.h>
26 #include <linux/buffer_head.h>  /* for try_to_release_page(),
27                                         buffer_heads_over_limit */
28 #include <linux/mm_inline.h>
29 #include <linux/pagevec.h>
30 #include <linux/backing-dev.h>
31 #include <linux/rmap.h>
32 #include <linux/topology.h>
33 #include <linux/cpu.h>
34 #include <linux/cpuset.h>
35 #include <linux/notifier.h>
36 #include <linux/rwsem.h>
37 #include <linux/delay.h>
38 #include <linux/kthread.h>
39 #include <linux/freezer.h>
40 #include <linux/memcontrol.h>
41 #include <linux/delayacct.h>
42 #include <linux/sysctl.h>
43
44 #include <asm/tlbflush.h>
45 #include <asm/div64.h>
46
47 #include <linux/swapops.h>
48
49 #include "internal.h"
50
51 struct scan_control {
52         /* Incremented by the number of inactive pages that were scanned */
53         unsigned long nr_scanned;
54
55         /* Number of pages freed so far during a call to shrink_zones() */
56         unsigned long nr_reclaimed;
57
58         /* This context's GFP mask */
59         gfp_t gfp_mask;
60
61         int may_writepage;
62
63         /* Can mapped pages be reclaimed? */
64         int may_unmap;
65
66         /* Can pages be swapped as part of reclaim? */
67         int may_swap;
68
69         /* This context's SWAP_CLUSTER_MAX. If freeing memory for
70          * suspend, we effectively ignore SWAP_CLUSTER_MAX.
71          * In this context, it doesn't matter that we scan the
72          * whole list at once. */
73         int swap_cluster_max;
74
75         int swappiness;
76
77         int all_unreclaimable;
78
79         int order;
80
81         /* Which cgroup do we reclaim from */
82         struct mem_cgroup *mem_cgroup;
83
84         /*
85          * Nodemask of nodes allowed by the caller. If NULL, all nodes
86          * are scanned.
87          */
88         nodemask_t      *nodemask;
89
90         /* Pluggable isolate pages callback */
91         unsigned long (*isolate_pages)(unsigned long nr, struct list_head *dst,
92                         unsigned long *scanned, int order, int mode,
93                         struct zone *z, struct mem_cgroup *mem_cont,
94                         int active, int file);
95 };
96
97 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
98
99 #ifdef ARCH_HAS_PREFETCH
100 #define prefetch_prev_lru_page(_page, _base, _field)                    \
101         do {                                                            \
102                 if ((_page)->lru.prev != _base) {                       \
103                         struct page *prev;                              \
104                                                                         \
105                         prev = lru_to_page(&(_page->lru));              \
106                         prefetch(&prev->_field);                        \
107                 }                                                       \
108         } while (0)
109 #else
110 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
111 #endif
112
113 #ifdef ARCH_HAS_PREFETCHW
114 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
115         do {                                                            \
116                 if ((_page)->lru.prev != _base) {                       \
117                         struct page *prev;                              \
118                                                                         \
119                         prev = lru_to_page(&(_page->lru));              \
120                         prefetchw(&prev->_field);                       \
121                 }                                                       \
122         } while (0)
123 #else
124 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
125 #endif
126
127 /*
128  * From 0 .. 100.  Higher means more swappy.
129  */
130 int vm_swappiness = 60;
131 long vm_total_pages;    /* The total number of pages which the VM controls */
132
133 static LIST_HEAD(shrinker_list);
134 static DECLARE_RWSEM(shrinker_rwsem);
135
136 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
137 #define scanning_global_lru(sc) (!(sc)->mem_cgroup)
138 #else
139 #define scanning_global_lru(sc) (1)
140 #endif
141
142 static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
143                                                   struct scan_control *sc)
144 {
145         if (!scanning_global_lru(sc))
146                 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
147
148         return &zone->reclaim_stat;
149 }
150
151 static unsigned long zone_nr_pages(struct zone *zone, struct scan_control *sc,
152                                    enum lru_list lru)
153 {
154         if (!scanning_global_lru(sc))
155                 return mem_cgroup_zone_nr_pages(sc->mem_cgroup, zone, lru);
156
157         return zone_page_state(zone, NR_LRU_BASE + lru);
158 }
159
160
161 /*
162  * Add a shrinker callback to be called from the vm
163  */
164 void register_shrinker(struct shrinker *shrinker)
165 {
166         shrinker->nr = 0;
167         down_write(&shrinker_rwsem);
168         list_add_tail(&shrinker->list, &shrinker_list);
169         up_write(&shrinker_rwsem);
170 }
171 EXPORT_SYMBOL(register_shrinker);
172
173 /*
174  * Remove one
175  */
176 void unregister_shrinker(struct shrinker *shrinker)
177 {
178         down_write(&shrinker_rwsem);
179         list_del(&shrinker->list);
180         up_write(&shrinker_rwsem);
181 }
182 EXPORT_SYMBOL(unregister_shrinker);
183
184 #define SHRINK_BATCH 128
185 /*
186  * Call the shrink functions to age shrinkable caches
187  *
188  * Here we assume it costs one seek to replace a lru page and that it also
189  * takes a seek to recreate a cache object.  With this in mind we age equal
190  * percentages of the lru and ageable caches.  This should balance the seeks
191  * generated by these structures.
192  *
193  * If the vm encountered mapped pages on the LRU it increase the pressure on
194  * slab to avoid swapping.
195  *
196  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
197  *
198  * `lru_pages' represents the number of on-LRU pages in all the zones which
199  * are eligible for the caller's allocation attempt.  It is used for balancing
200  * slab reclaim versus page reclaim.
201  *
202  * Returns the number of slab objects which we shrunk.
203  */
204 unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
205                         unsigned long lru_pages)
206 {
207         struct shrinker *shrinker;
208         unsigned long ret = 0;
209
210         if (scanned == 0)
211                 scanned = SWAP_CLUSTER_MAX;
212
213         if (!down_read_trylock(&shrinker_rwsem))
214                 return 1;       /* Assume we'll be able to shrink next time */
215
216         list_for_each_entry(shrinker, &shrinker_list, list) {
217                 unsigned long long delta;
218                 unsigned long total_scan;
219                 unsigned long max_pass = (*shrinker->shrink)(0, gfp_mask);
220
221                 delta = (4 * scanned) / shrinker->seeks;
222                 delta *= max_pass;
223                 do_div(delta, lru_pages + 1);
224                 shrinker->nr += delta;
225                 if (shrinker->nr < 0) {
226                         printk(KERN_ERR "shrink_slab: %pF negative objects to "
227                                "delete nr=%ld\n",
228                                shrinker->shrink, shrinker->nr);
229                         shrinker->nr = max_pass;
230                 }
231
232                 /*
233                  * Avoid risking looping forever due to too large nr value:
234                  * never try to free more than twice the estimate number of
235                  * freeable entries.
236                  */
237                 if (shrinker->nr > max_pass * 2)
238                         shrinker->nr = max_pass * 2;
239
240                 total_scan = shrinker->nr;
241                 shrinker->nr = 0;
242
243                 while (total_scan >= SHRINK_BATCH) {
244                         long this_scan = SHRINK_BATCH;
245                         int shrink_ret;
246                         int nr_before;
247
248                         nr_before = (*shrinker->shrink)(0, gfp_mask);
249                         shrink_ret = (*shrinker->shrink)(this_scan, gfp_mask);
250                         if (shrink_ret == -1)
251                                 break;
252                         if (shrink_ret < nr_before)
253                                 ret += nr_before - shrink_ret;
254                         count_vm_events(SLABS_SCANNED, this_scan);
255                         total_scan -= this_scan;
256
257                         cond_resched();
258                 }
259
260                 shrinker->nr += total_scan;
261         }
262         up_read(&shrinker_rwsem);
263         return ret;
264 }
265
266 /* Called without lock on whether page is mapped, so answer is unstable */
267 static inline int page_mapping_inuse(struct page *page)
268 {
269         struct address_space *mapping;
270
271         /* Page is in somebody's page tables. */
272         if (page_mapped(page))
273                 return 1;
274
275         /* Be more reluctant to reclaim swapcache than pagecache */
276         if (PageSwapCache(page))
277                 return 1;
278
279         mapping = page_mapping(page);
280         if (!mapping)
281                 return 0;
282
283         /* File is mmap'd by somebody? */
284         return mapping_mapped(mapping);
285 }
286
287 static inline int is_page_cache_freeable(struct page *page)
288 {
289         return page_count(page) - !!page_has_private(page) == 2;
290 }
291
292 static int may_write_to_queue(struct backing_dev_info *bdi)
293 {
294         if (current->flags & PF_SWAPWRITE)
295                 return 1;
296         if (!bdi_write_congested(bdi))
297                 return 1;
298         if (bdi == current->backing_dev_info)
299                 return 1;
300         return 0;
301 }
302
303 /*
304  * We detected a synchronous write error writing a page out.  Probably
305  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
306  * fsync(), msync() or close().
307  *
308  * The tricky part is that after writepage we cannot touch the mapping: nothing
309  * prevents it from being freed up.  But we have a ref on the page and once
310  * that page is locked, the mapping is pinned.
311  *
312  * We're allowed to run sleeping lock_page() here because we know the caller has
313  * __GFP_FS.
314  */
315 static void handle_write_error(struct address_space *mapping,
316                                 struct page *page, int error)
317 {
318         lock_page(page);
319         if (page_mapping(page) == mapping)
320                 mapping_set_error(mapping, error);
321         unlock_page(page);
322 }
323
324 /* Request for sync pageout. */
325 enum pageout_io {
326         PAGEOUT_IO_ASYNC,
327         PAGEOUT_IO_SYNC,
328 };
329
330 /* possible outcome of pageout() */
331 typedef enum {
332         /* failed to write page out, page is locked */
333         PAGE_KEEP,
334         /* move page to the active list, page is locked */
335         PAGE_ACTIVATE,
336         /* page has been sent to the disk successfully, page is unlocked */
337         PAGE_SUCCESS,
338         /* page is clean and locked */
339         PAGE_CLEAN,
340 } pageout_t;
341
342 /*
343  * pageout is called by shrink_page_list() for each dirty page.
344  * Calls ->writepage().
345  */
346 static pageout_t pageout(struct page *page, struct address_space *mapping,
347                                                 enum pageout_io sync_writeback)
348 {
349         /*
350          * If the page is dirty, only perform writeback if that write
351          * will be non-blocking.  To prevent this allocation from being
352          * stalled by pagecache activity.  But note that there may be
353          * stalls if we need to run get_block().  We could test
354          * PagePrivate for that.
355          *
356          * If this process is currently in generic_file_write() against
357          * this page's queue, we can perform writeback even if that
358          * will block.
359          *
360          * If the page is swapcache, write it back even if that would
361          * block, for some throttling. This happens by accident, because
362          * swap_backing_dev_info is bust: it doesn't reflect the
363          * congestion state of the swapdevs.  Easy to fix, if needed.
364          * See swapfile.c:page_queue_congested().
365          */
366         if (!is_page_cache_freeable(page))
367                 return PAGE_KEEP;
368         if (!mapping) {
369                 /*
370                  * Some data journaling orphaned pages can have
371                  * page->mapping == NULL while being dirty with clean buffers.
372                  */
373                 if (page_has_private(page)) {
374                         if (try_to_free_buffers(page)) {
375                                 ClearPageDirty(page);
376                                 printk("%s: orphaned page\n", __func__);
377                                 return PAGE_CLEAN;
378                         }
379                 }
380                 return PAGE_KEEP;
381         }
382         if (mapping->a_ops->writepage == NULL)
383                 return PAGE_ACTIVATE;
384         if (!may_write_to_queue(mapping->backing_dev_info))
385                 return PAGE_KEEP;
386
387         if (clear_page_dirty_for_io(page)) {
388                 int res;
389                 struct writeback_control wbc = {
390                         .sync_mode = WB_SYNC_NONE,
391                         .nr_to_write = SWAP_CLUSTER_MAX,
392                         .range_start = 0,
393                         .range_end = LLONG_MAX,
394                         .nonblocking = 1,
395                         .for_reclaim = 1,
396                 };
397
398                 SetPageReclaim(page);
399                 res = mapping->a_ops->writepage(page, &wbc);
400                 if (res < 0)
401                         handle_write_error(mapping, page, res);
402                 if (res == AOP_WRITEPAGE_ACTIVATE) {
403                         ClearPageReclaim(page);
404                         return PAGE_ACTIVATE;
405                 }
406
407                 /*
408                  * Wait on writeback if requested to. This happens when
409                  * direct reclaiming a large contiguous area and the
410                  * first attempt to free a range of pages fails.
411                  */
412                 if (PageWriteback(page) && sync_writeback == PAGEOUT_IO_SYNC)
413                         wait_on_page_writeback(page);
414
415                 if (!PageWriteback(page)) {
416                         /* synchronous write or broken a_ops? */
417                         ClearPageReclaim(page);
418                 }
419                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
420                 return PAGE_SUCCESS;
421         }
422
423         return PAGE_CLEAN;
424 }
425
426 /*
427  * Same as remove_mapping, but if the page is removed from the mapping, it
428  * gets returned with a refcount of 0.
429  */
430 static int __remove_mapping(struct address_space *mapping, struct page *page)
431 {
432         BUG_ON(!PageLocked(page));
433         BUG_ON(mapping != page_mapping(page));
434
435         spin_lock_irq(&mapping->tree_lock);
436         /*
437          * The non racy check for a busy page.
438          *
439          * Must be careful with the order of the tests. When someone has
440          * a ref to the page, it may be possible that they dirty it then
441          * drop the reference. So if PageDirty is tested before page_count
442          * here, then the following race may occur:
443          *
444          * get_user_pages(&page);
445          * [user mapping goes away]
446          * write_to(page);
447          *                              !PageDirty(page)    [good]
448          * SetPageDirty(page);
449          * put_page(page);
450          *                              !page_count(page)   [good, discard it]
451          *
452          * [oops, our write_to data is lost]
453          *
454          * Reversing the order of the tests ensures such a situation cannot
455          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
456          * load is not satisfied before that of page->_count.
457          *
458          * Note that if SetPageDirty is always performed via set_page_dirty,
459          * and thus under tree_lock, then this ordering is not required.
460          */
461         if (!page_freeze_refs(page, 2))
462                 goto cannot_free;
463         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
464         if (unlikely(PageDirty(page))) {
465                 page_unfreeze_refs(page, 2);
466                 goto cannot_free;
467         }
468
469         if (PageSwapCache(page)) {
470                 swp_entry_t swap = { .val = page_private(page) };
471                 __delete_from_swap_cache(page);
472                 spin_unlock_irq(&mapping->tree_lock);
473                 swapcache_free(swap, page);
474         } else {
475                 __remove_from_page_cache(page);
476                 spin_unlock_irq(&mapping->tree_lock);
477                 mem_cgroup_uncharge_cache_page(page);
478         }
479
480         return 1;
481
482 cannot_free:
483         spin_unlock_irq(&mapping->tree_lock);
484         return 0;
485 }
486
487 /*
488  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
489  * someone else has a ref on the page, abort and return 0.  If it was
490  * successfully detached, return 1.  Assumes the caller has a single ref on
491  * this page.
492  */
493 int remove_mapping(struct address_space *mapping, struct page *page)
494 {
495         if (__remove_mapping(mapping, page)) {
496                 /*
497                  * Unfreezing the refcount with 1 rather than 2 effectively
498                  * drops the pagecache ref for us without requiring another
499                  * atomic operation.
500                  */
501                 page_unfreeze_refs(page, 1);
502                 return 1;
503         }
504         return 0;
505 }
506
507 /**
508  * putback_lru_page - put previously isolated page onto appropriate LRU list
509  * @page: page to be put back to appropriate lru list
510  *
511  * Add previously isolated @page to appropriate LRU list.
512  * Page may still be unevictable for other reasons.
513  *
514  * lru_lock must not be held, interrupts must be enabled.
515  */
516 void putback_lru_page(struct page *page)
517 {
518         int lru;
519         int active = !!TestClearPageActive(page);
520         int was_unevictable = PageUnevictable(page);
521
522         VM_BUG_ON(PageLRU(page));
523
524 redo:
525         ClearPageUnevictable(page);
526
527         if (page_evictable(page, NULL)) {
528                 /*
529                  * For evictable pages, we can use the cache.
530                  * In event of a race, worst case is we end up with an
531                  * unevictable page on [in]active list.
532                  * We know how to handle that.
533                  */
534                 lru = active + page_is_file_cache(page);
535                 lru_cache_add_lru(page, lru);
536         } else {
537                 /*
538                  * Put unevictable pages directly on zone's unevictable
539                  * list.
540                  */
541                 lru = LRU_UNEVICTABLE;
542                 add_page_to_unevictable_list(page);
543         }
544
545         /*
546          * page's status can change while we move it among lru. If an evictable
547          * page is on unevictable list, it never be freed. To avoid that,
548          * check after we added it to the list, again.
549          */
550         if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
551                 if (!isolate_lru_page(page)) {
552                         put_page(page);
553                         goto redo;
554                 }
555                 /* This means someone else dropped this page from LRU
556                  * So, it will be freed or putback to LRU again. There is
557                  * nothing to do here.
558                  */
559         }
560
561         if (was_unevictable && lru != LRU_UNEVICTABLE)
562                 count_vm_event(UNEVICTABLE_PGRESCUED);
563         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
564                 count_vm_event(UNEVICTABLE_PGCULLED);
565
566         put_page(page);         /* drop ref from isolate */
567 }
568
569 /*
570  * shrink_page_list() returns the number of reclaimed pages
571  */
572 static unsigned long shrink_page_list(struct list_head *page_list,
573                                         struct scan_control *sc,
574                                         enum pageout_io sync_writeback)
575 {
576         LIST_HEAD(ret_pages);
577         struct pagevec freed_pvec;
578         int pgactivate = 0;
579         unsigned long nr_reclaimed = 0;
580
581         cond_resched();
582
583         pagevec_init(&freed_pvec, 1);
584         while (!list_empty(page_list)) {
585                 struct address_space *mapping;
586                 struct page *page;
587                 int may_enter_fs;
588                 int referenced;
589
590                 cond_resched();
591
592                 page = lru_to_page(page_list);
593                 list_del(&page->lru);
594
595                 if (!trylock_page(page))
596                         goto keep;
597
598                 VM_BUG_ON(PageActive(page));
599
600                 sc->nr_scanned++;
601
602                 if (unlikely(!page_evictable(page, NULL)))
603                         goto cull_mlocked;
604
605                 if (!sc->may_unmap && page_mapped(page))
606                         goto keep_locked;
607
608                 /* Double the slab pressure for mapped and swapcache pages */
609                 if (page_mapped(page) || PageSwapCache(page))
610                         sc->nr_scanned++;
611
612                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
613                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
614
615                 if (PageWriteback(page)) {
616                         /*
617                          * Synchronous reclaim is performed in two passes,
618                          * first an asynchronous pass over the list to
619                          * start parallel writeback, and a second synchronous
620                          * pass to wait for the IO to complete.  Wait here
621                          * for any page for which writeback has already
622                          * started.
623                          */
624                         if (sync_writeback == PAGEOUT_IO_SYNC && may_enter_fs)
625                                 wait_on_page_writeback(page);
626                         else
627                                 goto keep_locked;
628                 }
629
630                 referenced = page_referenced(page, 1, sc->mem_cgroup);
631                 /* In active use or really unfreeable?  Activate it. */
632                 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER &&
633                                         referenced && page_mapping_inuse(page))
634                         goto activate_locked;
635
636                 /*
637                  * Anonymous process memory has backing store?
638                  * Try to allocate it some swap space here.
639                  */
640                 if (PageAnon(page) && !PageSwapCache(page)) {
641                         if (!(sc->gfp_mask & __GFP_IO))
642                                 goto keep_locked;
643                         if (!add_to_swap(page))
644                                 goto activate_locked;
645                         may_enter_fs = 1;
646                 }
647
648                 mapping = page_mapping(page);
649
650                 /*
651                  * The page is mapped into the page tables of one or more
652                  * processes. Try to unmap it here.
653                  */
654                 if (page_mapped(page) && mapping) {
655                         switch (try_to_unmap(page, 0)) {
656                         case SWAP_FAIL:
657                                 goto activate_locked;
658                         case SWAP_AGAIN:
659                                 goto keep_locked;
660                         case SWAP_MLOCK:
661                                 goto cull_mlocked;
662                         case SWAP_SUCCESS:
663                                 ; /* try to free the page below */
664                         }
665                 }
666
667                 if (PageDirty(page)) {
668                         if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && referenced)
669                                 goto keep_locked;
670                         if (!may_enter_fs)
671                                 goto keep_locked;
672                         if (!sc->may_writepage)
673                                 goto keep_locked;
674
675                         /* Page is dirty, try to write it out here */
676                         switch (pageout(page, mapping, sync_writeback)) {
677                         case PAGE_KEEP:
678                                 goto keep_locked;
679                         case PAGE_ACTIVATE:
680                                 goto activate_locked;
681                         case PAGE_SUCCESS:
682                                 if (PageWriteback(page) || PageDirty(page))
683                                         goto keep;
684                                 /*
685                                  * A synchronous write - probably a ramdisk.  Go
686                                  * ahead and try to reclaim the page.
687                                  */
688                                 if (!trylock_page(page))
689                                         goto keep;
690                                 if (PageDirty(page) || PageWriteback(page))
691                                         goto keep_locked;
692                                 mapping = page_mapping(page);
693                         case PAGE_CLEAN:
694                                 ; /* try to free the page below */
695                         }
696                 }
697
698                 /*
699                  * If the page has buffers, try to free the buffer mappings
700                  * associated with this page. If we succeed we try to free
701                  * the page as well.
702                  *
703                  * We do this even if the page is PageDirty().
704                  * try_to_release_page() does not perform I/O, but it is
705                  * possible for a page to have PageDirty set, but it is actually
706                  * clean (all its buffers are clean).  This happens if the
707                  * buffers were written out directly, with submit_bh(). ext3
708                  * will do this, as well as the blockdev mapping.
709                  * try_to_release_page() will discover that cleanness and will
710                  * drop the buffers and mark the page clean - it can be freed.
711                  *
712                  * Rarely, pages can have buffers and no ->mapping.  These are
713                  * the pages which were not successfully invalidated in
714                  * truncate_complete_page().  We try to drop those buffers here
715                  * and if that worked, and the page is no longer mapped into
716                  * process address space (page_count == 1) it can be freed.
717                  * Otherwise, leave the page on the LRU so it is swappable.
718                  */
719                 if (page_has_private(page)) {
720                         if (!try_to_release_page(page, sc->gfp_mask))
721                                 goto activate_locked;
722                         if (!mapping && page_count(page) == 1) {
723                                 unlock_page(page);
724                                 if (put_page_testzero(page))
725                                         goto free_it;
726                                 else {
727                                         /*
728                                          * rare race with speculative reference.
729                                          * the speculative reference will free
730                                          * this page shortly, so we may
731                                          * increment nr_reclaimed here (and
732                                          * leave it off the LRU).
733                                          */
734                                         nr_reclaimed++;
735                                         continue;
736                                 }
737                         }
738                 }
739
740                 if (!mapping || !__remove_mapping(mapping, page))
741                         goto keep_locked;
742
743                 /*
744                  * At this point, we have no other references and there is
745                  * no way to pick any more up (removed from LRU, removed
746                  * from pagecache). Can use non-atomic bitops now (and
747                  * we obviously don't have to worry about waking up a process
748                  * waiting on the page lock, because there are no references.
749                  */
750                 __clear_page_locked(page);
751 free_it:
752                 nr_reclaimed++;
753                 if (!pagevec_add(&freed_pvec, page)) {
754                         __pagevec_free(&freed_pvec);
755                         pagevec_reinit(&freed_pvec);
756                 }
757                 continue;
758
759 cull_mlocked:
760                 if (PageSwapCache(page))
761                         try_to_free_swap(page);
762                 unlock_page(page);
763                 putback_lru_page(page);
764                 continue;
765
766 activate_locked:
767                 /* Not a candidate for swapping, so reclaim swap space. */
768                 if (PageSwapCache(page) && vm_swap_full())
769                         try_to_free_swap(page);
770                 VM_BUG_ON(PageActive(page));
771                 SetPageActive(page);
772                 pgactivate++;
773 keep_locked:
774                 unlock_page(page);
775 keep:
776                 list_add(&page->lru, &ret_pages);
777                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
778         }
779         list_splice(&ret_pages, page_list);
780         if (pagevec_count(&freed_pvec))
781                 __pagevec_free(&freed_pvec);
782         count_vm_events(PGACTIVATE, pgactivate);
783         return nr_reclaimed;
784 }
785
786 /* LRU Isolation modes. */
787 #define ISOLATE_INACTIVE 0      /* Isolate inactive pages. */
788 #define ISOLATE_ACTIVE 1        /* Isolate active pages. */
789 #define ISOLATE_BOTH 2          /* Isolate both active and inactive pages. */
790
791 /*
792  * Attempt to remove the specified page from its LRU.  Only take this page
793  * if it is of the appropriate PageActive status.  Pages which are being
794  * freed elsewhere are also ignored.
795  *
796  * page:        page to consider
797  * mode:        one of the LRU isolation modes defined above
798  *
799  * returns 0 on success, -ve errno on failure.
800  */
801 int __isolate_lru_page(struct page *page, int mode, int file)
802 {
803         int ret = -EINVAL;
804
805         /* Only take pages on the LRU. */
806         if (!PageLRU(page))
807                 return ret;
808
809         /*
810          * When checking the active state, we need to be sure we are
811          * dealing with comparible boolean values.  Take the logical not
812          * of each.
813          */
814         if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
815                 return ret;
816
817         if (mode != ISOLATE_BOTH && (!page_is_file_cache(page) != !file))
818                 return ret;
819
820         /*
821          * When this function is being called for lumpy reclaim, we
822          * initially look into all LRU pages, active, inactive and
823          * unevictable; only give shrink_page_list evictable pages.
824          */
825         if (PageUnevictable(page))
826                 return ret;
827
828         ret = -EBUSY;
829
830         if (likely(get_page_unless_zero(page))) {
831                 /*
832                  * Be careful not to clear PageLRU until after we're
833                  * sure the page is not being freed elsewhere -- the
834                  * page release code relies on it.
835                  */
836                 ClearPageLRU(page);
837                 ret = 0;
838                 mem_cgroup_del_lru(page);
839         }
840
841         return ret;
842 }
843
844 /*
845  * zone->lru_lock is heavily contended.  Some of the functions that
846  * shrink the lists perform better by taking out a batch of pages
847  * and working on them outside the LRU lock.
848  *
849  * For pagecache intensive workloads, this function is the hottest
850  * spot in the kernel (apart from copy_*_user functions).
851  *
852  * Appropriate locks must be held before calling this function.
853  *
854  * @nr_to_scan: The number of pages to look through on the list.
855  * @src:        The LRU list to pull pages off.
856  * @dst:        The temp list to put pages on to.
857  * @scanned:    The number of pages that were scanned.
858  * @order:      The caller's attempted allocation order
859  * @mode:       One of the LRU isolation modes
860  * @file:       True [1] if isolating file [!anon] pages
861  *
862  * returns how many pages were moved onto *@dst.
863  */
864 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
865                 struct list_head *src, struct list_head *dst,
866                 unsigned long *scanned, int order, int mode, int file)
867 {
868         unsigned long nr_taken = 0;
869         unsigned long scan;
870
871         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
872                 struct page *page;
873                 unsigned long pfn;
874                 unsigned long end_pfn;
875                 unsigned long page_pfn;
876                 int zone_id;
877
878                 page = lru_to_page(src);
879                 prefetchw_prev_lru_page(page, src, flags);
880
881                 VM_BUG_ON(!PageLRU(page));
882
883                 switch (__isolate_lru_page(page, mode, file)) {
884                 case 0:
885                         list_move(&page->lru, dst);
886                         nr_taken++;
887                         break;
888
889                 case -EBUSY:
890                         /* else it is being freed elsewhere */
891                         list_move(&page->lru, src);
892                         continue;
893
894                 default:
895                         BUG();
896                 }
897
898                 if (!order)
899                         continue;
900
901                 /*
902                  * Attempt to take all pages in the order aligned region
903                  * surrounding the tag page.  Only take those pages of
904                  * the same active state as that tag page.  We may safely
905                  * round the target page pfn down to the requested order
906                  * as the mem_map is guarenteed valid out to MAX_ORDER,
907                  * where that page is in a different zone we will detect
908                  * it from its zone id and abort this block scan.
909                  */
910                 zone_id = page_zone_id(page);
911                 page_pfn = page_to_pfn(page);
912                 pfn = page_pfn & ~((1 << order) - 1);
913                 end_pfn = pfn + (1 << order);
914                 for (; pfn < end_pfn; pfn++) {
915                         struct page *cursor_page;
916
917                         /* The target page is in the block, ignore it. */
918                         if (unlikely(pfn == page_pfn))
919                                 continue;
920
921                         /* Avoid holes within the zone. */
922                         if (unlikely(!pfn_valid_within(pfn)))
923                                 break;
924
925                         cursor_page = pfn_to_page(pfn);
926
927                         /* Check that we have not crossed a zone boundary. */
928                         if (unlikely(page_zone_id(cursor_page) != zone_id))
929                                 continue;
930                         switch (__isolate_lru_page(cursor_page, mode, file)) {
931                         case 0:
932                                 list_move(&cursor_page->lru, dst);
933                                 nr_taken++;
934                                 scan++;
935                                 break;
936
937                         case -EBUSY:
938                                 /* else it is being freed elsewhere */
939                                 list_move(&cursor_page->lru, src);
940                         default:
941                                 break;  /* ! on LRU or wrong list */
942                         }
943                 }
944         }
945
946         *scanned = scan;
947         return nr_taken;
948 }
949
950 static unsigned long isolate_pages_global(unsigned long nr,
951                                         struct list_head *dst,
952                                         unsigned long *scanned, int order,
953                                         int mode, struct zone *z,
954                                         struct mem_cgroup *mem_cont,
955                                         int active, int file)
956 {
957         int lru = LRU_BASE;
958         if (active)
959                 lru += LRU_ACTIVE;
960         if (file)
961                 lru += LRU_FILE;
962         return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
963                                                                 mode, !!file);
964 }
965
966 /*
967  * clear_active_flags() is a helper for shrink_active_list(), clearing
968  * any active bits from the pages in the list.
969  */
970 static unsigned long clear_active_flags(struct list_head *page_list,
971                                         unsigned int *count)
972 {
973         int nr_active = 0;
974         int lru;
975         struct page *page;
976
977         list_for_each_entry(page, page_list, lru) {
978                 lru = page_is_file_cache(page);
979                 if (PageActive(page)) {
980                         lru += LRU_ACTIVE;
981                         ClearPageActive(page);
982                         nr_active++;
983                 }
984                 count[lru]++;
985         }
986
987         return nr_active;
988 }
989
990 /**
991  * isolate_lru_page - tries to isolate a page from its LRU list
992  * @page: page to isolate from its LRU list
993  *
994  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
995  * vmstat statistic corresponding to whatever LRU list the page was on.
996  *
997  * Returns 0 if the page was removed from an LRU list.
998  * Returns -EBUSY if the page was not on an LRU list.
999  *
1000  * The returned page will have PageLRU() cleared.  If it was found on
1001  * the active list, it will have PageActive set.  If it was found on
1002  * the unevictable list, it will have the PageUnevictable bit set. That flag
1003  * may need to be cleared by the caller before letting the page go.
1004  *
1005  * The vmstat statistic corresponding to the list on which the page was
1006  * found will be decremented.
1007  *
1008  * Restrictions:
1009  * (1) Must be called with an elevated refcount on the page. This is a
1010  *     fundamentnal difference from isolate_lru_pages (which is called
1011  *     without a stable reference).
1012  * (2) the lru_lock must not be held.
1013  * (3) interrupts must be enabled.
1014  */
1015 int isolate_lru_page(struct page *page)
1016 {
1017         int ret = -EBUSY;
1018
1019         if (PageLRU(page)) {
1020                 struct zone *zone = page_zone(page);
1021
1022                 spin_lock_irq(&zone->lru_lock);
1023                 if (PageLRU(page) && get_page_unless_zero(page)) {
1024                         int lru = page_lru(page);
1025                         ret = 0;
1026                         ClearPageLRU(page);
1027
1028                         del_page_from_lru_list(zone, page, lru);
1029                 }
1030                 spin_unlock_irq(&zone->lru_lock);
1031         }
1032         return ret;
1033 }
1034
1035 /*
1036  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1037  * of reclaimed pages
1038  */
1039 static unsigned long shrink_inactive_list(unsigned long max_scan,
1040                         struct zone *zone, struct scan_control *sc,
1041                         int priority, int file)
1042 {
1043         LIST_HEAD(page_list);
1044         struct pagevec pvec;
1045         unsigned long nr_scanned = 0;
1046         unsigned long nr_reclaimed = 0;
1047         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1048         int lumpy_reclaim = 0;
1049
1050         /*
1051          * If we need a large contiguous chunk of memory, or have
1052          * trouble getting a small set of contiguous pages, we
1053          * will reclaim both active and inactive pages.
1054          *
1055          * We use the same threshold as pageout congestion_wait below.
1056          */
1057         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1058                 lumpy_reclaim = 1;
1059         else if (sc->order && priority < DEF_PRIORITY - 2)
1060                 lumpy_reclaim = 1;
1061
1062         pagevec_init(&pvec, 1);
1063
1064         lru_add_drain();
1065         spin_lock_irq(&zone->lru_lock);
1066         do {
1067                 struct page *page;
1068                 unsigned long nr_taken;
1069                 unsigned long nr_scan;
1070                 unsigned long nr_freed;
1071                 unsigned long nr_active;
1072                 unsigned int count[NR_LRU_LISTS] = { 0, };
1073                 int mode = lumpy_reclaim ? ISOLATE_BOTH : ISOLATE_INACTIVE;
1074
1075                 nr_taken = sc->isolate_pages(sc->swap_cluster_max,
1076                              &page_list, &nr_scan, sc->order, mode,
1077                                 zone, sc->mem_cgroup, 0, file);
1078                 nr_active = clear_active_flags(&page_list, count);
1079                 __count_vm_events(PGDEACTIVATE, nr_active);
1080
1081                 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1082                                                 -count[LRU_ACTIVE_FILE]);
1083                 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1084                                                 -count[LRU_INACTIVE_FILE]);
1085                 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1086                                                 -count[LRU_ACTIVE_ANON]);
1087                 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1088                                                 -count[LRU_INACTIVE_ANON]);
1089
1090                 if (scanning_global_lru(sc))
1091                         zone->pages_scanned += nr_scan;
1092
1093                 reclaim_stat->recent_scanned[0] += count[LRU_INACTIVE_ANON];
1094                 reclaim_stat->recent_scanned[0] += count[LRU_ACTIVE_ANON];
1095                 reclaim_stat->recent_scanned[1] += count[LRU_INACTIVE_FILE];
1096                 reclaim_stat->recent_scanned[1] += count[LRU_ACTIVE_FILE];
1097
1098                 spin_unlock_irq(&zone->lru_lock);
1099
1100                 nr_scanned += nr_scan;
1101                 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
1102
1103                 /*
1104                  * If we are direct reclaiming for contiguous pages and we do
1105                  * not reclaim everything in the list, try again and wait
1106                  * for IO to complete. This will stall high-order allocations
1107                  * but that should be acceptable to the caller
1108                  */
1109                 if (nr_freed < nr_taken && !current_is_kswapd() &&
1110                     lumpy_reclaim) {
1111                         congestion_wait(WRITE, HZ/10);
1112
1113                         /*
1114                          * The attempt at page out may have made some
1115                          * of the pages active, mark them inactive again.
1116                          */
1117                         nr_active = clear_active_flags(&page_list, count);
1118                         count_vm_events(PGDEACTIVATE, nr_active);
1119
1120                         nr_freed += shrink_page_list(&page_list, sc,
1121                                                         PAGEOUT_IO_SYNC);
1122                 }
1123
1124                 nr_reclaimed += nr_freed;
1125                 local_irq_disable();
1126                 if (current_is_kswapd()) {
1127                         __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scan);
1128                         __count_vm_events(KSWAPD_STEAL, nr_freed);
1129                 } else if (scanning_global_lru(sc))
1130                         __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scan);
1131
1132                 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
1133
1134                 if (nr_taken == 0)
1135                         goto done;
1136
1137                 spin_lock(&zone->lru_lock);
1138                 /*
1139                  * Put back any unfreeable pages.
1140                  */
1141                 while (!list_empty(&page_list)) {
1142                         int lru;
1143                         page = lru_to_page(&page_list);
1144                         VM_BUG_ON(PageLRU(page));
1145                         list_del(&page->lru);
1146                         if (unlikely(!page_evictable(page, NULL))) {
1147                                 spin_unlock_irq(&zone->lru_lock);
1148                                 putback_lru_page(page);
1149                                 spin_lock_irq(&zone->lru_lock);
1150                                 continue;
1151                         }
1152                         SetPageLRU(page);
1153                         lru = page_lru(page);
1154                         add_page_to_lru_list(zone, page, lru);
1155                         if (PageActive(page)) {
1156                                 int file = !!page_is_file_cache(page);
1157                                 reclaim_stat->recent_rotated[file]++;
1158                         }
1159                         if (!pagevec_add(&pvec, page)) {
1160                                 spin_unlock_irq(&zone->lru_lock);
1161                                 __pagevec_release(&pvec);
1162                                 spin_lock_irq(&zone->lru_lock);
1163                         }
1164                 }
1165         } while (nr_scanned < max_scan);
1166         spin_unlock(&zone->lru_lock);
1167 done:
1168         local_irq_enable();
1169         pagevec_release(&pvec);
1170         return nr_reclaimed;
1171 }
1172
1173 /*
1174  * We are about to scan this zone at a certain priority level.  If that priority
1175  * level is smaller (ie: more urgent) than the previous priority, then note
1176  * that priority level within the zone.  This is done so that when the next
1177  * process comes in to scan this zone, it will immediately start out at this
1178  * priority level rather than having to build up its own scanning priority.
1179  * Here, this priority affects only the reclaim-mapped threshold.
1180  */
1181 static inline void note_zone_scanning_priority(struct zone *zone, int priority)
1182 {
1183         if (priority < zone->prev_priority)
1184                 zone->prev_priority = priority;
1185 }
1186
1187 /*
1188  * This moves pages from the active list to the inactive list.
1189  *
1190  * We move them the other way if the page is referenced by one or more
1191  * processes, from rmap.
1192  *
1193  * If the pages are mostly unmapped, the processing is fast and it is
1194  * appropriate to hold zone->lru_lock across the whole operation.  But if
1195  * the pages are mapped, the processing is slow (page_referenced()) so we
1196  * should drop zone->lru_lock around each page.  It's impossible to balance
1197  * this, so instead we remove the pages from the LRU while processing them.
1198  * It is safe to rely on PG_active against the non-LRU pages in here because
1199  * nobody will play with that bit on a non-LRU page.
1200  *
1201  * The downside is that we have to touch page->_count against each page.
1202  * But we had to alter page->flags anyway.
1203  */
1204
1205
1206 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1207                         struct scan_control *sc, int priority, int file)
1208 {
1209         unsigned long pgmoved;
1210         unsigned long pgscanned;
1211         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1212         LIST_HEAD(l_inactive);
1213         struct page *page;
1214         struct pagevec pvec;
1215         enum lru_list lru;
1216         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1217
1218         lru_add_drain();
1219         spin_lock_irq(&zone->lru_lock);
1220         pgmoved = sc->isolate_pages(nr_pages, &l_hold, &pgscanned, sc->order,
1221                                         ISOLATE_ACTIVE, zone,
1222                                         sc->mem_cgroup, 1, file);
1223         /*
1224          * zone->pages_scanned is used for detect zone's oom
1225          * mem_cgroup remembers nr_scan by itself.
1226          */
1227         if (scanning_global_lru(sc)) {
1228                 zone->pages_scanned += pgscanned;
1229         }
1230         reclaim_stat->recent_scanned[!!file] += pgmoved;
1231
1232         if (file)
1233                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -pgmoved);
1234         else
1235                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -pgmoved);
1236         spin_unlock_irq(&zone->lru_lock);
1237
1238         pgmoved = 0;  /* count referenced (mapping) mapped pages */
1239         while (!list_empty(&l_hold)) {
1240                 cond_resched();
1241                 page = lru_to_page(&l_hold);
1242                 list_del(&page->lru);
1243
1244                 if (unlikely(!page_evictable(page, NULL))) {
1245                         putback_lru_page(page);
1246                         continue;
1247                 }
1248
1249                 /* page_referenced clears PageReferenced */
1250                 if (page_mapping_inuse(page) &&
1251                     page_referenced(page, 0, sc->mem_cgroup))
1252                         pgmoved++;
1253
1254                 list_add(&page->lru, &l_inactive);
1255         }
1256
1257         /*
1258          * Move the pages to the [file or anon] inactive list.
1259          */
1260         pagevec_init(&pvec, 1);
1261         lru = LRU_BASE + file * LRU_FILE;
1262
1263         spin_lock_irq(&zone->lru_lock);
1264         /*
1265          * Count referenced pages from currently used mappings as
1266          * rotated, even though they are moved to the inactive list.
1267          * This helps balance scan pressure between file and anonymous
1268          * pages in get_scan_ratio.
1269          */
1270         reclaim_stat->recent_rotated[!!file] += pgmoved;
1271
1272         pgmoved = 0;  /* count pages moved to inactive list */
1273         while (!list_empty(&l_inactive)) {
1274                 page = lru_to_page(&l_inactive);
1275                 prefetchw_prev_lru_page(page, &l_inactive, flags);
1276                 VM_BUG_ON(PageLRU(page));
1277                 SetPageLRU(page);
1278                 VM_BUG_ON(!PageActive(page));
1279                 ClearPageActive(page);
1280
1281                 list_move(&page->lru, &zone->lru[lru].list);
1282                 mem_cgroup_add_lru_list(page, lru);
1283                 pgmoved++;
1284                 if (!pagevec_add(&pvec, page)) {
1285                         spin_unlock_irq(&zone->lru_lock);
1286                         if (buffer_heads_over_limit)
1287                                 pagevec_strip(&pvec);
1288                         __pagevec_release(&pvec);
1289                         spin_lock_irq(&zone->lru_lock);
1290                 }
1291         }
1292         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1293         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1294         __count_vm_events(PGDEACTIVATE, pgmoved);
1295         spin_unlock_irq(&zone->lru_lock);
1296         if (buffer_heads_over_limit)
1297                 pagevec_strip(&pvec);
1298         pagevec_release(&pvec);
1299 }
1300
1301 static int inactive_anon_is_low_global(struct zone *zone)
1302 {
1303         unsigned long active, inactive;
1304
1305         active = zone_page_state(zone, NR_ACTIVE_ANON);
1306         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1307
1308         if (inactive * zone->inactive_ratio < active)
1309                 return 1;
1310
1311         return 0;
1312 }
1313
1314 /**
1315  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1316  * @zone: zone to check
1317  * @sc:   scan control of this context
1318  *
1319  * Returns true if the zone does not have enough inactive anon pages,
1320  * meaning some active anon pages need to be deactivated.
1321  */
1322 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1323 {
1324         int low;
1325
1326         if (scanning_global_lru(sc))
1327                 low = inactive_anon_is_low_global(zone);
1328         else
1329                 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1330         return low;
1331 }
1332
1333 static int inactive_file_is_low_global(struct zone *zone)
1334 {
1335         unsigned long active, inactive;
1336
1337         active = zone_page_state(zone, NR_ACTIVE_FILE);
1338         inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1339
1340         return (active > inactive);
1341 }
1342
1343 /**
1344  * inactive_file_is_low - check if file pages need to be deactivated
1345  * @zone: zone to check
1346  * @sc:   scan control of this context
1347  *
1348  * When the system is doing streaming IO, memory pressure here
1349  * ensures that active file pages get deactivated, until more
1350  * than half of the file pages are on the inactive list.
1351  *
1352  * Once we get to that situation, protect the system's working
1353  * set from being evicted by disabling active file page aging.
1354  *
1355  * This uses a different ratio than the anonymous pages, because
1356  * the page cache uses a use-once replacement algorithm.
1357  */
1358 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1359 {
1360         int low;
1361
1362         if (scanning_global_lru(sc))
1363                 low = inactive_file_is_low_global(zone);
1364         else
1365                 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1366         return low;
1367 }
1368
1369 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1370         struct zone *zone, struct scan_control *sc, int priority)
1371 {
1372         int file = is_file_lru(lru);
1373
1374         if (lru == LRU_ACTIVE_FILE && inactive_file_is_low(zone, sc)) {
1375                 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1376                 return 0;
1377         }
1378
1379         if (lru == LRU_ACTIVE_ANON && inactive_anon_is_low(zone, sc)) {
1380                 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1381                 return 0;
1382         }
1383         return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1384 }
1385
1386 /*
1387  * Determine how aggressively the anon and file LRU lists should be
1388  * scanned.  The relative value of each set of LRU lists is determined
1389  * by looking at the fraction of the pages scanned we did rotate back
1390  * onto the active list instead of evict.
1391  *
1392  * percent[0] specifies how much pressure to put on ram/swap backed
1393  * memory, while percent[1] determines pressure on the file LRUs.
1394  */
1395 static void get_scan_ratio(struct zone *zone, struct scan_control *sc,
1396                                         unsigned long *percent)
1397 {
1398         unsigned long anon, file, free;
1399         unsigned long anon_prio, file_prio;
1400         unsigned long ap, fp;
1401         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1402
1403         /* If we have no swap space, do not bother scanning anon pages. */
1404         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1405                 percent[0] = 0;
1406                 percent[1] = 100;
1407                 return;
1408         }
1409
1410         anon  = zone_nr_pages(zone, sc, LRU_ACTIVE_ANON) +
1411                 zone_nr_pages(zone, sc, LRU_INACTIVE_ANON);
1412         file  = zone_nr_pages(zone, sc, LRU_ACTIVE_FILE) +
1413                 zone_nr_pages(zone, sc, LRU_INACTIVE_FILE);
1414
1415         if (scanning_global_lru(sc)) {
1416                 free  = zone_page_state(zone, NR_FREE_PAGES);
1417                 /* If we have very few page cache pages,
1418                    force-scan anon pages. */
1419                 if (unlikely(file + free <= high_wmark_pages(zone))) {
1420                         percent[0] = 100;
1421                         percent[1] = 0;
1422                         return;
1423                 }
1424         }
1425
1426         /*
1427          * OK, so we have swap space and a fair amount of page cache
1428          * pages.  We use the recently rotated / recently scanned
1429          * ratios to determine how valuable each cache is.
1430          *
1431          * Because workloads change over time (and to avoid overflow)
1432          * we keep these statistics as a floating average, which ends
1433          * up weighing recent references more than old ones.
1434          *
1435          * anon in [0], file in [1]
1436          */
1437         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1438                 spin_lock_irq(&zone->lru_lock);
1439                 reclaim_stat->recent_scanned[0] /= 2;
1440                 reclaim_stat->recent_rotated[0] /= 2;
1441                 spin_unlock_irq(&zone->lru_lock);
1442         }
1443
1444         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1445                 spin_lock_irq(&zone->lru_lock);
1446                 reclaim_stat->recent_scanned[1] /= 2;
1447                 reclaim_stat->recent_rotated[1] /= 2;
1448                 spin_unlock_irq(&zone->lru_lock);
1449         }
1450
1451         /*
1452          * With swappiness at 100, anonymous and file have the same priority.
1453          * This scanning priority is essentially the inverse of IO cost.
1454          */
1455         anon_prio = sc->swappiness;
1456         file_prio = 200 - sc->swappiness;
1457
1458         /*
1459          * The amount of pressure on anon vs file pages is inversely
1460          * proportional to the fraction of recently scanned pages on
1461          * each list that were recently referenced and in active use.
1462          */
1463         ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1464         ap /= reclaim_stat->recent_rotated[0] + 1;
1465
1466         fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1467         fp /= reclaim_stat->recent_rotated[1] + 1;
1468
1469         /* Normalize to percentages */
1470         percent[0] = 100 * ap / (ap + fp + 1);
1471         percent[1] = 100 - percent[0];
1472 }
1473
1474 /*
1475  * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1476  * until we collected @swap_cluster_max pages to scan.
1477  */
1478 static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
1479                                        unsigned long *nr_saved_scan,
1480                                        unsigned long swap_cluster_max)
1481 {
1482         unsigned long nr;
1483
1484         *nr_saved_scan += nr_to_scan;
1485         nr = *nr_saved_scan;
1486
1487         if (nr >= swap_cluster_max)
1488                 *nr_saved_scan = 0;
1489         else
1490                 nr = 0;
1491
1492         return nr;
1493 }
1494
1495 /*
1496  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1497  */
1498 static void shrink_zone(int priority, struct zone *zone,
1499                                 struct scan_control *sc)
1500 {
1501         unsigned long nr[NR_LRU_LISTS];
1502         unsigned long nr_to_scan;
1503         unsigned long percent[2];       /* anon @ 0; file @ 1 */
1504         enum lru_list l;
1505         unsigned long nr_reclaimed = sc->nr_reclaimed;
1506         unsigned long swap_cluster_max = sc->swap_cluster_max;
1507
1508         get_scan_ratio(zone, sc, percent);
1509
1510         for_each_evictable_lru(l) {
1511                 int file = is_file_lru(l);
1512                 unsigned long scan;
1513
1514                 scan = zone_nr_pages(zone, sc, l);
1515                 if (priority) {
1516                         scan >>= priority;
1517                         scan = (scan * percent[file]) / 100;
1518                 }
1519                 if (scanning_global_lru(sc))
1520                         nr[l] = nr_scan_try_batch(scan,
1521                                                   &zone->lru[l].nr_saved_scan,
1522                                                   swap_cluster_max);
1523                 else
1524                         nr[l] = scan;
1525         }
1526
1527         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1528                                         nr[LRU_INACTIVE_FILE]) {
1529                 for_each_evictable_lru(l) {
1530                         if (nr[l]) {
1531                                 nr_to_scan = min(nr[l], swap_cluster_max);
1532                                 nr[l] -= nr_to_scan;
1533
1534                                 nr_reclaimed += shrink_list(l, nr_to_scan,
1535                                                             zone, sc, priority);
1536                         }
1537                 }
1538                 /*
1539                  * On large memory systems, scan >> priority can become
1540                  * really large. This is fine for the starting priority;
1541                  * we want to put equal scanning pressure on each zone.
1542                  * However, if the VM has a harder time of freeing pages,
1543                  * with multiple processes reclaiming pages, the total
1544                  * freeing target can get unreasonably large.
1545                  */
1546                 if (nr_reclaimed > swap_cluster_max &&
1547                         priority < DEF_PRIORITY && !current_is_kswapd())
1548                         break;
1549         }
1550
1551         sc->nr_reclaimed = nr_reclaimed;
1552
1553         /*
1554          * Even if we did not try to evict anon pages at all, we want to
1555          * rebalance the anon lru active/inactive ratio.
1556          */
1557         if (inactive_anon_is_low(zone, sc) && nr_swap_pages > 0)
1558                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1559
1560         throttle_vm_writeout(sc->gfp_mask);
1561 }
1562
1563 /*
1564  * This is the direct reclaim path, for page-allocating processes.  We only
1565  * try to reclaim pages from zones which will satisfy the caller's allocation
1566  * request.
1567  *
1568  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1569  * Because:
1570  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1571  *    allocation or
1572  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1573  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1574  *    zone defense algorithm.
1575  *
1576  * If a zone is deemed to be full of pinned pages then just give it a light
1577  * scan then give up on it.
1578  */
1579 static void shrink_zones(int priority, struct zonelist *zonelist,
1580                                         struct scan_control *sc)
1581 {
1582         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1583         struct zoneref *z;
1584         struct zone *zone;
1585
1586         sc->all_unreclaimable = 1;
1587         for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1588                                         sc->nodemask) {
1589                 if (!populated_zone(zone))
1590                         continue;
1591                 /*
1592                  * Take care memory controller reclaiming has small influence
1593                  * to global LRU.
1594                  */
1595                 if (scanning_global_lru(sc)) {
1596                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1597                                 continue;
1598                         note_zone_scanning_priority(zone, priority);
1599
1600                         if (zone_is_all_unreclaimable(zone) &&
1601                                                 priority != DEF_PRIORITY)
1602                                 continue;       /* Let kswapd poll it */
1603                         sc->all_unreclaimable = 0;
1604                 } else {
1605                         /*
1606                          * Ignore cpuset limitation here. We just want to reduce
1607                          * # of used pages by us regardless of memory shortage.
1608                          */
1609                         sc->all_unreclaimable = 0;
1610                         mem_cgroup_note_reclaim_priority(sc->mem_cgroup,
1611                                                         priority);
1612                 }
1613
1614                 shrink_zone(priority, zone, sc);
1615         }
1616 }
1617
1618 /*
1619  * This is the main entry point to direct page reclaim.
1620  *
1621  * If a full scan of the inactive list fails to free enough memory then we
1622  * are "out of memory" and something needs to be killed.
1623  *
1624  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1625  * high - the zone may be full of dirty or under-writeback pages, which this
1626  * caller can't do much about.  We kick pdflush and take explicit naps in the
1627  * hope that some of these pages can be written.  But if the allocating task
1628  * holds filesystem locks which prevent writeout this might not work, and the
1629  * allocation attempt will fail.
1630  *
1631  * returns:     0, if no pages reclaimed
1632  *              else, the number of pages reclaimed
1633  */
1634 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
1635                                         struct scan_control *sc)
1636 {
1637         int priority;
1638         unsigned long ret = 0;
1639         unsigned long total_scanned = 0;
1640         struct reclaim_state *reclaim_state = current->reclaim_state;
1641         unsigned long lru_pages = 0;
1642         struct zoneref *z;
1643         struct zone *zone;
1644         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1645
1646         delayacct_freepages_start();
1647
1648         if (scanning_global_lru(sc))
1649                 count_vm_event(ALLOCSTALL);
1650         /*
1651          * mem_cgroup will not do shrink_slab.
1652          */
1653         if (scanning_global_lru(sc)) {
1654                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1655
1656                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1657                                 continue;
1658
1659                         lru_pages += zone_lru_pages(zone);
1660                 }
1661         }
1662
1663         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1664                 sc->nr_scanned = 0;
1665                 if (!priority)
1666                         disable_swap_token();
1667                 shrink_zones(priority, zonelist, sc);
1668                 /*
1669                  * Don't shrink slabs when reclaiming memory from
1670                  * over limit cgroups
1671                  */
1672                 if (scanning_global_lru(sc)) {
1673                         shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
1674                         if (reclaim_state) {
1675                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
1676                                 reclaim_state->reclaimed_slab = 0;
1677                         }
1678                 }
1679                 total_scanned += sc->nr_scanned;
1680                 if (sc->nr_reclaimed >= sc->swap_cluster_max) {
1681                         ret = sc->nr_reclaimed;
1682                         goto out;
1683                 }
1684
1685                 /*
1686                  * Try to write back as many pages as we just scanned.  This
1687                  * tends to cause slow streaming writers to write data to the
1688                  * disk smoothly, at the dirtying rate, which is nice.   But
1689                  * that's undesirable in laptop mode, where we *want* lumpy
1690                  * writeout.  So in laptop mode, write out the whole world.
1691                  */
1692                 if (total_scanned > sc->swap_cluster_max +
1693                                         sc->swap_cluster_max / 2) {
1694                         wakeup_pdflush(laptop_mode ? 0 : total_scanned);
1695                         sc->may_writepage = 1;
1696                 }
1697
1698                 /* Take a nap, wait for some writeback to complete */
1699                 if (sc->nr_scanned && priority < DEF_PRIORITY - 2)
1700                         congestion_wait(WRITE, HZ/10);
1701         }
1702         /* top priority shrink_zones still had more to do? don't OOM, then */
1703         if (!sc->all_unreclaimable && scanning_global_lru(sc))
1704                 ret = sc->nr_reclaimed;
1705 out:
1706         /*
1707          * Now that we've scanned all the zones at this priority level, note
1708          * that level within the zone so that the next thread which performs
1709          * scanning of this zone will immediately start out at this priority
1710          * level.  This affects only the decision whether or not to bring
1711          * mapped pages onto the inactive list.
1712          */
1713         if (priority < 0)
1714                 priority = 0;
1715
1716         if (scanning_global_lru(sc)) {
1717                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1718
1719                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1720                                 continue;
1721
1722                         zone->prev_priority = priority;
1723                 }
1724         } else
1725                 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority);
1726
1727         delayacct_freepages_end();
1728
1729         return ret;
1730 }
1731
1732 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
1733                                 gfp_t gfp_mask, nodemask_t *nodemask)
1734 {
1735         struct scan_control sc = {
1736                 .gfp_mask = gfp_mask,
1737                 .may_writepage = !laptop_mode,
1738                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1739                 .may_unmap = 1,
1740                 .may_swap = 1,
1741                 .swappiness = vm_swappiness,
1742                 .order = order,
1743                 .mem_cgroup = NULL,
1744                 .isolate_pages = isolate_pages_global,
1745                 .nodemask = nodemask,
1746         };
1747
1748         return do_try_to_free_pages(zonelist, &sc);
1749 }
1750
1751 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
1752
1753 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
1754                                            gfp_t gfp_mask,
1755                                            bool noswap,
1756                                            unsigned int swappiness)
1757 {
1758         struct scan_control sc = {
1759                 .may_writepage = !laptop_mode,
1760                 .may_unmap = 1,
1761                 .may_swap = !noswap,
1762                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1763                 .swappiness = swappiness,
1764                 .order = 0,
1765                 .mem_cgroup = mem_cont,
1766                 .isolate_pages = mem_cgroup_isolate_pages,
1767                 .nodemask = NULL, /* we don't care the placement */
1768         };
1769         struct zonelist *zonelist;
1770
1771         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1772                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1773         zonelist = NODE_DATA(numa_node_id())->node_zonelists;
1774         return do_try_to_free_pages(zonelist, &sc);
1775 }
1776 #endif
1777
1778 /*
1779  * For kswapd, balance_pgdat() will work across all this node's zones until
1780  * they are all at high_wmark_pages(zone).
1781  *
1782  * Returns the number of pages which were actually freed.
1783  *
1784  * There is special handling here for zones which are full of pinned pages.
1785  * This can happen if the pages are all mlocked, or if they are all used by
1786  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1787  * What we do is to detect the case where all pages in the zone have been
1788  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1789  * dead and from now on, only perform a short scan.  Basically we're polling
1790  * the zone for when the problem goes away.
1791  *
1792  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
1793  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
1794  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
1795  * lower zones regardless of the number of free pages in the lower zones. This
1796  * interoperates with the page allocator fallback scheme to ensure that aging
1797  * of pages is balanced across the zones.
1798  */
1799 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
1800 {
1801         int all_zones_ok;
1802         int priority;
1803         int i;
1804         unsigned long total_scanned;
1805         struct reclaim_state *reclaim_state = current->reclaim_state;
1806         struct scan_control sc = {
1807                 .gfp_mask = GFP_KERNEL,
1808                 .may_unmap = 1,
1809                 .may_swap = 1,
1810                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1811                 .swappiness = vm_swappiness,
1812                 .order = order,
1813                 .mem_cgroup = NULL,
1814                 .isolate_pages = isolate_pages_global,
1815         };
1816         /*
1817          * temp_priority is used to remember the scanning priority at which
1818          * this zone was successfully refilled to
1819          * free_pages == high_wmark_pages(zone).
1820          */
1821         int temp_priority[MAX_NR_ZONES];
1822
1823 loop_again:
1824         total_scanned = 0;
1825         sc.nr_reclaimed = 0;
1826         sc.may_writepage = !laptop_mode;
1827         count_vm_event(PAGEOUTRUN);
1828
1829         for (i = 0; i < pgdat->nr_zones; i++)
1830                 temp_priority[i] = DEF_PRIORITY;
1831
1832         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1833                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1834                 unsigned long lru_pages = 0;
1835
1836                 /* The swap token gets in the way of swapout... */
1837                 if (!priority)
1838                         disable_swap_token();
1839
1840                 all_zones_ok = 1;
1841
1842                 /*
1843                  * Scan in the highmem->dma direction for the highest
1844                  * zone which needs scanning
1845                  */
1846                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1847                         struct zone *zone = pgdat->node_zones + i;
1848
1849                         if (!populated_zone(zone))
1850                                 continue;
1851
1852                         if (zone_is_all_unreclaimable(zone) &&
1853                             priority != DEF_PRIORITY)
1854                                 continue;
1855
1856                         /*
1857                          * Do some background aging of the anon list, to give
1858                          * pages a chance to be referenced before reclaiming.
1859                          */
1860                         if (inactive_anon_is_low(zone, &sc))
1861                                 shrink_active_list(SWAP_CLUSTER_MAX, zone,
1862                                                         &sc, priority, 0);
1863
1864                         if (!zone_watermark_ok(zone, order,
1865                                         high_wmark_pages(zone), 0, 0)) {
1866                                 end_zone = i;
1867                                 break;
1868                         }
1869                 }
1870                 if (i < 0)
1871                         goto out;
1872
1873                 for (i = 0; i <= end_zone; i++) {
1874                         struct zone *zone = pgdat->node_zones + i;
1875
1876                         lru_pages += zone_lru_pages(zone);
1877                 }
1878
1879                 /*
1880                  * Now scan the zone in the dma->highmem direction, stopping
1881                  * at the last zone which needs scanning.
1882                  *
1883                  * We do this because the page allocator works in the opposite
1884                  * direction.  This prevents the page allocator from allocating
1885                  * pages behind kswapd's direction of progress, which would
1886                  * cause too much scanning of the lower zones.
1887                  */
1888                 for (i = 0; i <= end_zone; i++) {
1889                         struct zone *zone = pgdat->node_zones + i;
1890                         int nr_slab;
1891
1892                         if (!populated_zone(zone))
1893                                 continue;
1894
1895                         if (zone_is_all_unreclaimable(zone) &&
1896                                         priority != DEF_PRIORITY)
1897                                 continue;
1898
1899                         if (!zone_watermark_ok(zone, order,
1900                                         high_wmark_pages(zone), end_zone, 0))
1901                                 all_zones_ok = 0;
1902                         temp_priority[i] = priority;
1903                         sc.nr_scanned = 0;
1904                         note_zone_scanning_priority(zone, priority);
1905                         /*
1906                          * We put equal pressure on every zone, unless one
1907                          * zone has way too many pages free already.
1908                          */
1909                         if (!zone_watermark_ok(zone, order,
1910                                         8*high_wmark_pages(zone), end_zone, 0))
1911                                 shrink_zone(priority, zone, &sc);
1912                         reclaim_state->reclaimed_slab = 0;
1913                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1914                                                 lru_pages);
1915                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1916                         total_scanned += sc.nr_scanned;
1917                         if (zone_is_all_unreclaimable(zone))
1918                                 continue;
1919                         if (nr_slab == 0 && zone->pages_scanned >=
1920                                                 (zone_lru_pages(zone) * 6))
1921                                         zone_set_flag(zone,
1922                                                       ZONE_ALL_UNRECLAIMABLE);
1923                         /*
1924                          * If we've done a decent amount of scanning and
1925                          * the reclaim ratio is low, start doing writepage
1926                          * even in laptop mode
1927                          */
1928                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1929                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
1930                                 sc.may_writepage = 1;
1931                 }
1932                 if (all_zones_ok)
1933                         break;          /* kswapd: all done */
1934                 /*
1935                  * OK, kswapd is getting into trouble.  Take a nap, then take
1936                  * another pass across the zones.
1937                  */
1938                 if (total_scanned && priority < DEF_PRIORITY - 2)
1939                         congestion_wait(WRITE, HZ/10);
1940
1941                 /*
1942                  * We do this so kswapd doesn't build up large priorities for
1943                  * example when it is freeing in parallel with allocators. It
1944                  * matches the direct reclaim path behaviour in terms of impact
1945                  * on zone->*_priority.
1946                  */
1947                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
1948                         break;
1949         }
1950 out:
1951         /*
1952          * Note within each zone the priority level at which this zone was
1953          * brought into a happy state.  So that the next thread which scans this
1954          * zone will start out at that priority level.
1955          */
1956         for (i = 0; i < pgdat->nr_zones; i++) {
1957                 struct zone *zone = pgdat->node_zones + i;
1958
1959                 zone->prev_priority = temp_priority[i];
1960         }
1961         if (!all_zones_ok) {
1962                 cond_resched();
1963
1964                 try_to_freeze();
1965
1966                 /*
1967                  * Fragmentation may mean that the system cannot be
1968                  * rebalanced for high-order allocations in all zones.
1969                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
1970                  * it means the zones have been fully scanned and are still
1971                  * not balanced. For high-order allocations, there is
1972                  * little point trying all over again as kswapd may
1973                  * infinite loop.
1974                  *
1975                  * Instead, recheck all watermarks at order-0 as they
1976                  * are the most important. If watermarks are ok, kswapd will go
1977                  * back to sleep. High-order users can still perform direct
1978                  * reclaim if they wish.
1979                  */
1980                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
1981                         order = sc.order = 0;
1982
1983                 goto loop_again;
1984         }
1985
1986         return sc.nr_reclaimed;
1987 }
1988
1989 /*
1990  * The background pageout daemon, started as a kernel thread
1991  * from the init process.
1992  *
1993  * This basically trickles out pages so that we have _some_
1994  * free memory available even if there is no other activity
1995  * that frees anything up. This is needed for things like routing
1996  * etc, where we otherwise might have all activity going on in
1997  * asynchronous contexts that cannot page things out.
1998  *
1999  * If there are applications that are active memory-allocators
2000  * (most normal use), this basically shouldn't matter.
2001  */
2002 static int kswapd(void *p)
2003 {
2004         unsigned long order;
2005         pg_data_t *pgdat = (pg_data_t*)p;
2006         struct task_struct *tsk = current;
2007         DEFINE_WAIT(wait);
2008         struct reclaim_state reclaim_state = {
2009                 .reclaimed_slab = 0,
2010         };
2011         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2012
2013         lockdep_set_current_reclaim_state(GFP_KERNEL);
2014
2015         if (!cpumask_empty(cpumask))
2016                 set_cpus_allowed_ptr(tsk, cpumask);
2017         current->reclaim_state = &reclaim_state;
2018
2019         /*
2020          * Tell the memory management that we're a "memory allocator",
2021          * and that if we need more memory we should get access to it
2022          * regardless (see "__alloc_pages()"). "kswapd" should
2023          * never get caught in the normal page freeing logic.
2024          *
2025          * (Kswapd normally doesn't need memory anyway, but sometimes
2026          * you need a small amount of memory in order to be able to
2027          * page out something else, and this flag essentially protects
2028          * us from recursively trying to free more memory as we're
2029          * trying to free the first piece of memory in the first place).
2030          */
2031         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2032         set_freezable();
2033
2034         order = 0;
2035         for ( ; ; ) {
2036                 unsigned long new_order;
2037
2038                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2039                 new_order = pgdat->kswapd_max_order;
2040                 pgdat->kswapd_max_order = 0;
2041                 if (order < new_order) {
2042                         /*
2043                          * Don't sleep if someone wants a larger 'order'
2044                          * allocation
2045                          */
2046                         order = new_order;
2047                 } else {
2048                         if (!freezing(current))
2049                                 schedule();
2050
2051                         order = pgdat->kswapd_max_order;
2052                 }
2053                 finish_wait(&pgdat->kswapd_wait, &wait);
2054
2055                 if (!try_to_freeze()) {
2056                         /* We can speed up thawing tasks if we don't call
2057                          * balance_pgdat after returning from the refrigerator
2058                          */
2059                         balance_pgdat(pgdat, order);
2060                 }
2061         }
2062         return 0;
2063 }
2064
2065 /*
2066  * A zone is low on free memory, so wake its kswapd task to service it.
2067  */
2068 void wakeup_kswapd(struct zone *zone, int order)
2069 {
2070         pg_data_t *pgdat;
2071
2072         if (!populated_zone(zone))
2073                 return;
2074
2075         pgdat = zone->zone_pgdat;
2076         if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
2077                 return;
2078         if (pgdat->kswapd_max_order < order)
2079                 pgdat->kswapd_max_order = order;
2080         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2081                 return;
2082         if (!waitqueue_active(&pgdat->kswapd_wait))
2083                 return;
2084         wake_up_interruptible(&pgdat->kswapd_wait);
2085 }
2086
2087 unsigned long global_lru_pages(void)
2088 {
2089         return global_page_state(NR_ACTIVE_ANON)
2090                 + global_page_state(NR_ACTIVE_FILE)
2091                 + global_page_state(NR_INACTIVE_ANON)
2092                 + global_page_state(NR_INACTIVE_FILE);
2093 }
2094
2095 #ifdef CONFIG_HIBERNATION
2096 /*
2097  * Helper function for shrink_all_memory().  Tries to reclaim 'nr_pages' pages
2098  * from LRU lists system-wide, for given pass and priority.
2099  *
2100  * For pass > 3 we also try to shrink the LRU lists that contain a few pages
2101  */
2102 static void shrink_all_zones(unsigned long nr_pages, int prio,
2103                                       int pass, struct scan_control *sc)
2104 {
2105         struct zone *zone;
2106         unsigned long nr_reclaimed = 0;
2107
2108         for_each_populated_zone(zone) {
2109                 enum lru_list l;
2110
2111                 if (zone_is_all_unreclaimable(zone) && prio != DEF_PRIORITY)
2112                         continue;
2113
2114                 for_each_evictable_lru(l) {
2115                         enum zone_stat_item ls = NR_LRU_BASE + l;
2116                         unsigned long lru_pages = zone_page_state(zone, ls);
2117
2118                         /* For pass = 0, we don't shrink the active list */
2119                         if (pass == 0 && (l == LRU_ACTIVE_ANON ||
2120                                                 l == LRU_ACTIVE_FILE))
2121                                 continue;
2122
2123                         zone->lru[l].nr_saved_scan += (lru_pages >> prio) + 1;
2124                         if (zone->lru[l].nr_saved_scan >= nr_pages || pass > 3) {
2125                                 unsigned long nr_to_scan;
2126
2127                                 zone->lru[l].nr_saved_scan = 0;
2128                                 nr_to_scan = min(nr_pages, lru_pages);
2129                                 nr_reclaimed += shrink_list(l, nr_to_scan, zone,
2130                                                                 sc, prio);
2131                                 if (nr_reclaimed >= nr_pages) {
2132                                         sc->nr_reclaimed += nr_reclaimed;
2133                                         return;
2134                                 }
2135                         }
2136                 }
2137         }
2138         sc->nr_reclaimed += nr_reclaimed;
2139 }
2140
2141 /*
2142  * Try to free `nr_pages' of memory, system-wide, and return the number of
2143  * freed pages.
2144  *
2145  * Rather than trying to age LRUs the aim is to preserve the overall
2146  * LRU order by reclaiming preferentially
2147  * inactive > active > active referenced > active mapped
2148  */
2149 unsigned long shrink_all_memory(unsigned long nr_pages)
2150 {
2151         unsigned long lru_pages, nr_slab;
2152         int pass;
2153         struct reclaim_state reclaim_state;
2154         struct scan_control sc = {
2155                 .gfp_mask = GFP_KERNEL,
2156                 .may_unmap = 0,
2157                 .may_writepage = 1,
2158                 .isolate_pages = isolate_pages_global,
2159                 .nr_reclaimed = 0,
2160         };
2161
2162         current->reclaim_state = &reclaim_state;
2163
2164         lru_pages = global_lru_pages();
2165         nr_slab = global_page_state(NR_SLAB_RECLAIMABLE);
2166         /* If slab caches are huge, it's better to hit them first */
2167         while (nr_slab >= lru_pages) {
2168                 reclaim_state.reclaimed_slab = 0;
2169                 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
2170                 if (!reclaim_state.reclaimed_slab)
2171                         break;
2172
2173                 sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2174                 if (sc.nr_reclaimed >= nr_pages)
2175                         goto out;
2176
2177                 nr_slab -= reclaim_state.reclaimed_slab;
2178         }
2179
2180         /*
2181          * We try to shrink LRUs in 5 passes:
2182          * 0 = Reclaim from inactive_list only
2183          * 1 = Reclaim from active list but don't reclaim mapped
2184          * 2 = 2nd pass of type 1
2185          * 3 = Reclaim mapped (normal reclaim)
2186          * 4 = 2nd pass of type 3
2187          */
2188         for (pass = 0; pass < 5; pass++) {
2189                 int prio;
2190
2191                 /* Force reclaiming mapped pages in the passes #3 and #4 */
2192                 if (pass > 2)
2193                         sc.may_unmap = 1;
2194
2195                 for (prio = DEF_PRIORITY; prio >= 0; prio--) {
2196                         unsigned long nr_to_scan = nr_pages - sc.nr_reclaimed;
2197
2198                         sc.nr_scanned = 0;
2199                         sc.swap_cluster_max = nr_to_scan;
2200                         shrink_all_zones(nr_to_scan, prio, pass, &sc);
2201                         if (sc.nr_reclaimed >= nr_pages)
2202                                 goto out;
2203
2204                         reclaim_state.reclaimed_slab = 0;
2205                         shrink_slab(sc.nr_scanned, sc.gfp_mask,
2206                                         global_lru_pages());
2207                         sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2208                         if (sc.nr_reclaimed >= nr_pages)
2209                                 goto out;
2210
2211                         if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
2212                                 congestion_wait(WRITE, HZ / 10);
2213                 }
2214         }
2215
2216         /*
2217          * If sc.nr_reclaimed = 0, we could not shrink LRUs, but there may be
2218          * something in slab caches
2219          */
2220         if (!sc.nr_reclaimed) {
2221                 do {
2222                         reclaim_state.reclaimed_slab = 0;
2223                         shrink_slab(nr_pages, sc.gfp_mask, global_lru_pages());
2224                         sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2225                 } while (sc.nr_reclaimed < nr_pages &&
2226                                 reclaim_state.reclaimed_slab > 0);
2227         }
2228
2229
2230 out:
2231         current->reclaim_state = NULL;
2232
2233         return sc.nr_reclaimed;
2234 }
2235 #endif /* CONFIG_HIBERNATION */
2236
2237 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2238    not required for correctness.  So if the last cpu in a node goes
2239    away, we get changed to run anywhere: as the first one comes back,
2240    restore their cpu bindings. */
2241 static int __devinit cpu_callback(struct notifier_block *nfb,
2242                                   unsigned long action, void *hcpu)
2243 {
2244         int nid;
2245
2246         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2247                 for_each_node_state(nid, N_HIGH_MEMORY) {
2248                         pg_data_t *pgdat = NODE_DATA(nid);
2249                         const struct cpumask *mask;
2250
2251                         mask = cpumask_of_node(pgdat->node_id);
2252
2253                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2254                                 /* One of our CPUs online: restore mask */
2255                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2256                 }
2257         }
2258         return NOTIFY_OK;
2259 }
2260
2261 /*
2262  * This kswapd start function will be called by init and node-hot-add.
2263  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2264  */
2265 int kswapd_run(int nid)
2266 {
2267         pg_data_t *pgdat = NODE_DATA(nid);
2268         int ret = 0;
2269
2270         if (pgdat->kswapd)
2271                 return 0;
2272
2273         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2274         if (IS_ERR(pgdat->kswapd)) {
2275                 /* failure at boot is fatal */
2276                 BUG_ON(system_state == SYSTEM_BOOTING);
2277                 printk("Failed to start kswapd on node %d\n",nid);
2278                 ret = -1;
2279         }
2280         return ret;
2281 }
2282
2283 static int __init kswapd_init(void)
2284 {
2285         int nid;
2286
2287         swap_setup();
2288         for_each_node_state(nid, N_HIGH_MEMORY)
2289                 kswapd_run(nid);
2290         hotcpu_notifier(cpu_callback, 0);
2291         return 0;
2292 }
2293
2294 module_init(kswapd_init)
2295
2296 #ifdef CONFIG_NUMA
2297 /*
2298  * Zone reclaim mode
2299  *
2300  * If non-zero call zone_reclaim when the number of free pages falls below
2301  * the watermarks.
2302  */
2303 int zone_reclaim_mode __read_mostly;
2304
2305 #define RECLAIM_OFF 0
2306 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
2307 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
2308 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
2309
2310 /*
2311  * Priority for ZONE_RECLAIM. This determines the fraction of pages
2312  * of a node considered for each zone_reclaim. 4 scans 1/16th of
2313  * a zone.
2314  */
2315 #define ZONE_RECLAIM_PRIORITY 4
2316
2317 /*
2318  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2319  * occur.
2320  */
2321 int sysctl_min_unmapped_ratio = 1;
2322
2323 /*
2324  * If the number of slab pages in a zone grows beyond this percentage then
2325  * slab reclaim needs to occur.
2326  */
2327 int sysctl_min_slab_ratio = 5;
2328
2329 /*
2330  * Try to free up some pages from this zone through reclaim.
2331  */
2332 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2333 {
2334         /* Minimum pages needed in order to stay on node */
2335         const unsigned long nr_pages = 1 << order;
2336         struct task_struct *p = current;
2337         struct reclaim_state reclaim_state;
2338         int priority;
2339         struct scan_control sc = {
2340                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2341                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2342                 .may_swap = 1,
2343                 .swap_cluster_max = max_t(unsigned long, nr_pages,
2344                                         SWAP_CLUSTER_MAX),
2345                 .gfp_mask = gfp_mask,
2346                 .swappiness = vm_swappiness,
2347                 .order = order,
2348                 .isolate_pages = isolate_pages_global,
2349         };
2350         unsigned long slab_reclaimable;
2351
2352         disable_swap_token();
2353         cond_resched();
2354         /*
2355          * We need to be able to allocate from the reserves for RECLAIM_SWAP
2356          * and we also need to be able to write out pages for RECLAIM_WRITE
2357          * and RECLAIM_SWAP.
2358          */
2359         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2360         reclaim_state.reclaimed_slab = 0;
2361         p->reclaim_state = &reclaim_state;
2362
2363         if (zone_page_state(zone, NR_FILE_PAGES) -
2364                 zone_page_state(zone, NR_FILE_MAPPED) >
2365                 zone->min_unmapped_pages) {
2366                 /*
2367                  * Free memory by calling shrink zone with increasing
2368                  * priorities until we have enough memory freed.
2369                  */
2370                 priority = ZONE_RECLAIM_PRIORITY;
2371                 do {
2372                         note_zone_scanning_priority(zone, priority);
2373                         shrink_zone(priority, zone, &sc);
2374                         priority--;
2375                 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
2376         }
2377
2378         slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2379         if (slab_reclaimable > zone->min_slab_pages) {
2380                 /*
2381                  * shrink_slab() does not currently allow us to determine how
2382                  * many pages were freed in this zone. So we take the current
2383                  * number of slab pages and shake the slab until it is reduced
2384                  * by the same nr_pages that we used for reclaiming unmapped
2385                  * pages.
2386                  *
2387                  * Note that shrink_slab will free memory on all zones and may
2388                  * take a long time.
2389                  */
2390                 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
2391                         zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
2392                                 slab_reclaimable - nr_pages)
2393                         ;
2394
2395                 /*
2396                  * Update nr_reclaimed by the number of slab pages we
2397                  * reclaimed from this zone.
2398                  */
2399                 sc.nr_reclaimed += slab_reclaimable -
2400                         zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2401         }
2402
2403         p->reclaim_state = NULL;
2404         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
2405         return sc.nr_reclaimed >= nr_pages;
2406 }
2407
2408 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2409 {
2410         int node_id;
2411         int ret;
2412
2413         /*
2414          * Zone reclaim reclaims unmapped file backed pages and
2415          * slab pages if we are over the defined limits.
2416          *
2417          * A small portion of unmapped file backed pages is needed for
2418          * file I/O otherwise pages read by file I/O will be immediately
2419          * thrown out if the zone is overallocated. So we do not reclaim
2420          * if less than a specified percentage of the zone is used by
2421          * unmapped file backed pages.
2422          */
2423         if (zone_page_state(zone, NR_FILE_PAGES) -
2424             zone_page_state(zone, NR_FILE_MAPPED) <= zone->min_unmapped_pages
2425             && zone_page_state(zone, NR_SLAB_RECLAIMABLE)
2426                         <= zone->min_slab_pages)
2427                 return 0;
2428
2429         if (zone_is_all_unreclaimable(zone))
2430                 return 0;
2431
2432         /*
2433          * Do not scan if the allocation should not be delayed.
2434          */
2435         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
2436                         return 0;
2437
2438         /*
2439          * Only run zone reclaim on the local zone or on zones that do not
2440          * have associated processors. This will favor the local processor
2441          * over remote processors and spread off node memory allocations
2442          * as wide as possible.
2443          */
2444         node_id = zone_to_nid(zone);
2445         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
2446                 return 0;
2447
2448         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2449                 return 0;
2450         ret = __zone_reclaim(zone, gfp_mask, order);
2451         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2452
2453         return ret;
2454 }
2455 #endif
2456
2457 /*
2458  * page_evictable - test whether a page is evictable
2459  * @page: the page to test
2460  * @vma: the VMA in which the page is or will be mapped, may be NULL
2461  *
2462  * Test whether page is evictable--i.e., should be placed on active/inactive
2463  * lists vs unevictable list.  The vma argument is !NULL when called from the
2464  * fault path to determine how to instantate a new page.
2465  *
2466  * Reasons page might not be evictable:
2467  * (1) page's mapping marked unevictable
2468  * (2) page is part of an mlocked VMA
2469  *
2470  */
2471 int page_evictable(struct page *page, struct vm_area_struct *vma)
2472 {
2473
2474         if (mapping_unevictable(page_mapping(page)))
2475                 return 0;
2476
2477         if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2478                 return 0;
2479
2480         return 1;
2481 }
2482
2483 /**
2484  * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2485  * @page: page to check evictability and move to appropriate lru list
2486  * @zone: zone page is in
2487  *
2488  * Checks a page for evictability and moves the page to the appropriate
2489  * zone lru list.
2490  *
2491  * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2492  * have PageUnevictable set.
2493  */
2494 static void check_move_unevictable_page(struct page *page, struct zone *zone)
2495 {
2496         VM_BUG_ON(PageActive(page));
2497
2498 retry:
2499         ClearPageUnevictable(page);
2500         if (page_evictable(page, NULL)) {
2501                 enum lru_list l = LRU_INACTIVE_ANON + page_is_file_cache(page);
2502
2503                 __dec_zone_state(zone, NR_UNEVICTABLE);
2504                 list_move(&page->lru, &zone->lru[l].list);
2505                 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
2506                 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
2507                 __count_vm_event(UNEVICTABLE_PGRESCUED);
2508         } else {
2509                 /*
2510                  * rotate unevictable list
2511                  */
2512                 SetPageUnevictable(page);
2513                 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
2514                 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
2515                 if (page_evictable(page, NULL))
2516                         goto retry;
2517         }
2518 }
2519
2520 /**
2521  * scan_mapping_unevictable_pages - scan an address space for evictable pages
2522  * @mapping: struct address_space to scan for evictable pages
2523  *
2524  * Scan all pages in mapping.  Check unevictable pages for
2525  * evictability and move them to the appropriate zone lru list.
2526  */
2527 void scan_mapping_unevictable_pages(struct address_space *mapping)
2528 {
2529         pgoff_t next = 0;
2530         pgoff_t end   = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
2531                          PAGE_CACHE_SHIFT;
2532         struct zone *zone;
2533         struct pagevec pvec;
2534
2535         if (mapping->nrpages == 0)
2536                 return;
2537
2538         pagevec_init(&pvec, 0);
2539         while (next < end &&
2540                 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
2541                 int i;
2542                 int pg_scanned = 0;
2543
2544                 zone = NULL;
2545
2546                 for (i = 0; i < pagevec_count(&pvec); i++) {
2547                         struct page *page = pvec.pages[i];
2548                         pgoff_t page_index = page->index;
2549                         struct zone *pagezone = page_zone(page);
2550
2551                         pg_scanned++;
2552                         if (page_index > next)
2553                                 next = page_index;
2554                         next++;
2555
2556                         if (pagezone != zone) {
2557                                 if (zone)
2558                                         spin_unlock_irq(&zone->lru_lock);
2559                                 zone = pagezone;
2560                                 spin_lock_irq(&zone->lru_lock);
2561                         }
2562
2563                         if (PageLRU(page) && PageUnevictable(page))
2564                                 check_move_unevictable_page(page, zone);
2565                 }
2566                 if (zone)
2567                         spin_unlock_irq(&zone->lru_lock);
2568                 pagevec_release(&pvec);
2569
2570                 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
2571         }
2572
2573 }
2574
2575 /**
2576  * scan_zone_unevictable_pages - check unevictable list for evictable pages
2577  * @zone - zone of which to scan the unevictable list
2578  *
2579  * Scan @zone's unevictable LRU lists to check for pages that have become
2580  * evictable.  Move those that have to @zone's inactive list where they
2581  * become candidates for reclaim, unless shrink_inactive_zone() decides
2582  * to reactivate them.  Pages that are still unevictable are rotated
2583  * back onto @zone's unevictable list.
2584  */
2585 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
2586 static void scan_zone_unevictable_pages(struct zone *zone)
2587 {
2588         struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
2589         unsigned long scan;
2590         unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
2591
2592         while (nr_to_scan > 0) {
2593                 unsigned long batch_size = min(nr_to_scan,
2594                                                 SCAN_UNEVICTABLE_BATCH_SIZE);
2595
2596                 spin_lock_irq(&zone->lru_lock);
2597                 for (scan = 0;  scan < batch_size; scan++) {
2598                         struct page *page = lru_to_page(l_unevictable);
2599
2600                         if (!trylock_page(page))
2601                                 continue;
2602
2603                         prefetchw_prev_lru_page(page, l_unevictable, flags);
2604
2605                         if (likely(PageLRU(page) && PageUnevictable(page)))
2606                                 check_move_unevictable_page(page, zone);
2607
2608                         unlock_page(page);
2609                 }
2610                 spin_unlock_irq(&zone->lru_lock);
2611
2612                 nr_to_scan -= batch_size;
2613         }
2614 }
2615
2616
2617 /**
2618  * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
2619  *
2620  * A really big hammer:  scan all zones' unevictable LRU lists to check for
2621  * pages that have become evictable.  Move those back to the zones'
2622  * inactive list where they become candidates for reclaim.
2623  * This occurs when, e.g., we have unswappable pages on the unevictable lists,
2624  * and we add swap to the system.  As such, it runs in the context of a task
2625  * that has possibly/probably made some previously unevictable pages
2626  * evictable.
2627  */
2628 static void scan_all_zones_unevictable_pages(void)
2629 {
2630         struct zone *zone;
2631
2632         for_each_zone(zone) {
2633                 scan_zone_unevictable_pages(zone);
2634         }
2635 }
2636
2637 /*
2638  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
2639  * all nodes' unevictable lists for evictable pages
2640  */
2641 unsigned long scan_unevictable_pages;
2642
2643 int scan_unevictable_handler(struct ctl_table *table, int write,
2644                            struct file *file, void __user *buffer,
2645                            size_t *length, loff_t *ppos)
2646 {
2647         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
2648
2649         if (write && *(unsigned long *)table->data)
2650                 scan_all_zones_unevictable_pages();
2651
2652         scan_unevictable_pages = 0;
2653         return 0;
2654 }
2655
2656 /*
2657  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
2658  * a specified node's per zone unevictable lists for evictable pages.
2659  */
2660
2661 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
2662                                           struct sysdev_attribute *attr,
2663                                           char *buf)
2664 {
2665         return sprintf(buf, "0\n");     /* always zero; should fit... */
2666 }
2667
2668 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
2669                                            struct sysdev_attribute *attr,
2670                                         const char *buf, size_t count)
2671 {
2672         struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
2673         struct zone *zone;
2674         unsigned long res;
2675         unsigned long req = strict_strtoul(buf, 10, &res);
2676
2677         if (!req)
2678                 return 1;       /* zero is no-op */
2679
2680         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2681                 if (!populated_zone(zone))
2682                         continue;
2683                 scan_zone_unevictable_pages(zone);
2684         }
2685         return 1;
2686 }
2687
2688
2689 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
2690                         read_scan_unevictable_node,
2691                         write_scan_unevictable_node);
2692
2693 int scan_unevictable_register_node(struct node *node)
2694 {
2695         return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
2696 }
2697
2698 void scan_unevictable_unregister_node(struct node *node)
2699 {
2700         sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
2701 }
2702