mm: fix incorrect page removal from LRU
[linux-2.6.git] / mm / vmscan.c
1 /*
2  *  linux/mm/vmscan.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  *
6  *  Swap reorganised 29.12.95, Stephen Tweedie.
7  *  kswapd added: 7.1.96  sct
8  *  Removed kswapd_ctl limits, and swap out as many pages as needed
9  *  to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10  *  Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11  *  Multiqueue VM started 5.8.00, Rik van Riel.
12  */
13
14 #include <linux/mm.h>
15 #include <linux/module.h>
16 #include <linux/slab.h>
17 #include <linux/kernel_stat.h>
18 #include <linux/swap.h>
19 #include <linux/pagemap.h>
20 #include <linux/init.h>
21 #include <linux/highmem.h>
22 #include <linux/vmstat.h>
23 #include <linux/file.h>
24 #include <linux/writeback.h>
25 #include <linux/blkdev.h>
26 #include <linux/buffer_head.h>  /* for try_to_release_page(),
27                                         buffer_heads_over_limit */
28 #include <linux/mm_inline.h>
29 #include <linux/pagevec.h>
30 #include <linux/backing-dev.h>
31 #include <linux/rmap.h>
32 #include <linux/topology.h>
33 #include <linux/cpu.h>
34 #include <linux/cpuset.h>
35 #include <linux/notifier.h>
36 #include <linux/rwsem.h>
37 #include <linux/delay.h>
38 #include <linux/kthread.h>
39 #include <linux/freezer.h>
40 #include <linux/memcontrol.h>
41 #include <linux/delayacct.h>
42 #include <linux/sysctl.h>
43
44 #include <asm/tlbflush.h>
45 #include <asm/div64.h>
46
47 #include <linux/swapops.h>
48
49 #include "internal.h"
50
51 struct scan_control {
52         /* Incremented by the number of inactive pages that were scanned */
53         unsigned long nr_scanned;
54
55         /* Number of pages freed so far during a call to shrink_zones() */
56         unsigned long nr_reclaimed;
57
58         /* This context's GFP mask */
59         gfp_t gfp_mask;
60
61         int may_writepage;
62
63         /* Can mapped pages be reclaimed? */
64         int may_unmap;
65
66         /* Can pages be swapped as part of reclaim? */
67         int may_swap;
68
69         /* This context's SWAP_CLUSTER_MAX. If freeing memory for
70          * suspend, we effectively ignore SWAP_CLUSTER_MAX.
71          * In this context, it doesn't matter that we scan the
72          * whole list at once. */
73         int swap_cluster_max;
74
75         int swappiness;
76
77         int all_unreclaimable;
78
79         int order;
80
81         /* Which cgroup do we reclaim from */
82         struct mem_cgroup *mem_cgroup;
83
84         /*
85          * Nodemask of nodes allowed by the caller. If NULL, all nodes
86          * are scanned.
87          */
88         nodemask_t      *nodemask;
89
90         /* Pluggable isolate pages callback */
91         unsigned long (*isolate_pages)(unsigned long nr, struct list_head *dst,
92                         unsigned long *scanned, int order, int mode,
93                         struct zone *z, struct mem_cgroup *mem_cont,
94                         int active, int file);
95 };
96
97 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
98
99 #ifdef ARCH_HAS_PREFETCH
100 #define prefetch_prev_lru_page(_page, _base, _field)                    \
101         do {                                                            \
102                 if ((_page)->lru.prev != _base) {                       \
103                         struct page *prev;                              \
104                                                                         \
105                         prev = lru_to_page(&(_page->lru));              \
106                         prefetch(&prev->_field);                        \
107                 }                                                       \
108         } while (0)
109 #else
110 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
111 #endif
112
113 #ifdef ARCH_HAS_PREFETCHW
114 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
115         do {                                                            \
116                 if ((_page)->lru.prev != _base) {                       \
117                         struct page *prev;                              \
118                                                                         \
119                         prev = lru_to_page(&(_page->lru));              \
120                         prefetchw(&prev->_field);                       \
121                 }                                                       \
122         } while (0)
123 #else
124 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
125 #endif
126
127 /*
128  * From 0 .. 100.  Higher means more swappy.
129  */
130 int vm_swappiness = 60;
131 long vm_total_pages;    /* The total number of pages which the VM controls */
132
133 static LIST_HEAD(shrinker_list);
134 static DECLARE_RWSEM(shrinker_rwsem);
135
136 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
137 #define scanning_global_lru(sc) (!(sc)->mem_cgroup)
138 #else
139 #define scanning_global_lru(sc) (1)
140 #endif
141
142 static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
143                                                   struct scan_control *sc)
144 {
145         if (!scanning_global_lru(sc))
146                 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
147
148         return &zone->reclaim_stat;
149 }
150
151 static unsigned long zone_nr_pages(struct zone *zone, struct scan_control *sc,
152                                    enum lru_list lru)
153 {
154         if (!scanning_global_lru(sc))
155                 return mem_cgroup_zone_nr_pages(sc->mem_cgroup, zone, lru);
156
157         return zone_page_state(zone, NR_LRU_BASE + lru);
158 }
159
160
161 /*
162  * Add a shrinker callback to be called from the vm
163  */
164 void register_shrinker(struct shrinker *shrinker)
165 {
166         shrinker->nr = 0;
167         down_write(&shrinker_rwsem);
168         list_add_tail(&shrinker->list, &shrinker_list);
169         up_write(&shrinker_rwsem);
170 }
171 EXPORT_SYMBOL(register_shrinker);
172
173 /*
174  * Remove one
175  */
176 void unregister_shrinker(struct shrinker *shrinker)
177 {
178         down_write(&shrinker_rwsem);
179         list_del(&shrinker->list);
180         up_write(&shrinker_rwsem);
181 }
182 EXPORT_SYMBOL(unregister_shrinker);
183
184 #define SHRINK_BATCH 128
185 /*
186  * Call the shrink functions to age shrinkable caches
187  *
188  * Here we assume it costs one seek to replace a lru page and that it also
189  * takes a seek to recreate a cache object.  With this in mind we age equal
190  * percentages of the lru and ageable caches.  This should balance the seeks
191  * generated by these structures.
192  *
193  * If the vm encountered mapped pages on the LRU it increase the pressure on
194  * slab to avoid swapping.
195  *
196  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
197  *
198  * `lru_pages' represents the number of on-LRU pages in all the zones which
199  * are eligible for the caller's allocation attempt.  It is used for balancing
200  * slab reclaim versus page reclaim.
201  *
202  * Returns the number of slab objects which we shrunk.
203  */
204 unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
205                         unsigned long lru_pages)
206 {
207         struct shrinker *shrinker;
208         unsigned long ret = 0;
209
210         if (scanned == 0)
211                 scanned = SWAP_CLUSTER_MAX;
212
213         if (!down_read_trylock(&shrinker_rwsem))
214                 return 1;       /* Assume we'll be able to shrink next time */
215
216         list_for_each_entry(shrinker, &shrinker_list, list) {
217                 unsigned long long delta;
218                 unsigned long total_scan;
219                 unsigned long max_pass = (*shrinker->shrink)(0, gfp_mask);
220
221                 delta = (4 * scanned) / shrinker->seeks;
222                 delta *= max_pass;
223                 do_div(delta, lru_pages + 1);
224                 shrinker->nr += delta;
225                 if (shrinker->nr < 0) {
226                         printk(KERN_ERR "shrink_slab: %pF negative objects to "
227                                "delete nr=%ld\n",
228                                shrinker->shrink, shrinker->nr);
229                         shrinker->nr = max_pass;
230                 }
231
232                 /*
233                  * Avoid risking looping forever due to too large nr value:
234                  * never try to free more than twice the estimate number of
235                  * freeable entries.
236                  */
237                 if (shrinker->nr > max_pass * 2)
238                         shrinker->nr = max_pass * 2;
239
240                 total_scan = shrinker->nr;
241                 shrinker->nr = 0;
242
243                 while (total_scan >= SHRINK_BATCH) {
244                         long this_scan = SHRINK_BATCH;
245                         int shrink_ret;
246                         int nr_before;
247
248                         nr_before = (*shrinker->shrink)(0, gfp_mask);
249                         shrink_ret = (*shrinker->shrink)(this_scan, gfp_mask);
250                         if (shrink_ret == -1)
251                                 break;
252                         if (shrink_ret < nr_before)
253                                 ret += nr_before - shrink_ret;
254                         count_vm_events(SLABS_SCANNED, this_scan);
255                         total_scan -= this_scan;
256
257                         cond_resched();
258                 }
259
260                 shrinker->nr += total_scan;
261         }
262         up_read(&shrinker_rwsem);
263         return ret;
264 }
265
266 /* Called without lock on whether page is mapped, so answer is unstable */
267 static inline int page_mapping_inuse(struct page *page)
268 {
269         struct address_space *mapping;
270
271         /* Page is in somebody's page tables. */
272         if (page_mapped(page))
273                 return 1;
274
275         /* Be more reluctant to reclaim swapcache than pagecache */
276         if (PageSwapCache(page))
277                 return 1;
278
279         mapping = page_mapping(page);
280         if (!mapping)
281                 return 0;
282
283         /* File is mmap'd by somebody? */
284         return mapping_mapped(mapping);
285 }
286
287 static inline int is_page_cache_freeable(struct page *page)
288 {
289         return page_count(page) - !!page_has_private(page) == 2;
290 }
291
292 static int may_write_to_queue(struct backing_dev_info *bdi)
293 {
294         if (current->flags & PF_SWAPWRITE)
295                 return 1;
296         if (!bdi_write_congested(bdi))
297                 return 1;
298         if (bdi == current->backing_dev_info)
299                 return 1;
300         return 0;
301 }
302
303 /*
304  * We detected a synchronous write error writing a page out.  Probably
305  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
306  * fsync(), msync() or close().
307  *
308  * The tricky part is that after writepage we cannot touch the mapping: nothing
309  * prevents it from being freed up.  But we have a ref on the page and once
310  * that page is locked, the mapping is pinned.
311  *
312  * We're allowed to run sleeping lock_page() here because we know the caller has
313  * __GFP_FS.
314  */
315 static void handle_write_error(struct address_space *mapping,
316                                 struct page *page, int error)
317 {
318         lock_page(page);
319         if (page_mapping(page) == mapping)
320                 mapping_set_error(mapping, error);
321         unlock_page(page);
322 }
323
324 /* Request for sync pageout. */
325 enum pageout_io {
326         PAGEOUT_IO_ASYNC,
327         PAGEOUT_IO_SYNC,
328 };
329
330 /* possible outcome of pageout() */
331 typedef enum {
332         /* failed to write page out, page is locked */
333         PAGE_KEEP,
334         /* move page to the active list, page is locked */
335         PAGE_ACTIVATE,
336         /* page has been sent to the disk successfully, page is unlocked */
337         PAGE_SUCCESS,
338         /* page is clean and locked */
339         PAGE_CLEAN,
340 } pageout_t;
341
342 /*
343  * pageout is called by shrink_page_list() for each dirty page.
344  * Calls ->writepage().
345  */
346 static pageout_t pageout(struct page *page, struct address_space *mapping,
347                                                 enum pageout_io sync_writeback)
348 {
349         /*
350          * If the page is dirty, only perform writeback if that write
351          * will be non-blocking.  To prevent this allocation from being
352          * stalled by pagecache activity.  But note that there may be
353          * stalls if we need to run get_block().  We could test
354          * PagePrivate for that.
355          *
356          * If this process is currently in generic_file_write() against
357          * this page's queue, we can perform writeback even if that
358          * will block.
359          *
360          * If the page is swapcache, write it back even if that would
361          * block, for some throttling. This happens by accident, because
362          * swap_backing_dev_info is bust: it doesn't reflect the
363          * congestion state of the swapdevs.  Easy to fix, if needed.
364          * See swapfile.c:page_queue_congested().
365          */
366         if (!is_page_cache_freeable(page))
367                 return PAGE_KEEP;
368         if (!mapping) {
369                 /*
370                  * Some data journaling orphaned pages can have
371                  * page->mapping == NULL while being dirty with clean buffers.
372                  */
373                 if (page_has_private(page)) {
374                         if (try_to_free_buffers(page)) {
375                                 ClearPageDirty(page);
376                                 printk("%s: orphaned page\n", __func__);
377                                 return PAGE_CLEAN;
378                         }
379                 }
380                 return PAGE_KEEP;
381         }
382         if (mapping->a_ops->writepage == NULL)
383                 return PAGE_ACTIVATE;
384         if (!may_write_to_queue(mapping->backing_dev_info))
385                 return PAGE_KEEP;
386
387         if (clear_page_dirty_for_io(page)) {
388                 int res;
389                 struct writeback_control wbc = {
390                         .sync_mode = WB_SYNC_NONE,
391                         .nr_to_write = SWAP_CLUSTER_MAX,
392                         .range_start = 0,
393                         .range_end = LLONG_MAX,
394                         .nonblocking = 1,
395                         .for_reclaim = 1,
396                 };
397
398                 SetPageReclaim(page);
399                 res = mapping->a_ops->writepage(page, &wbc);
400                 if (res < 0)
401                         handle_write_error(mapping, page, res);
402                 if (res == AOP_WRITEPAGE_ACTIVATE) {
403                         ClearPageReclaim(page);
404                         return PAGE_ACTIVATE;
405                 }
406
407                 /*
408                  * Wait on writeback if requested to. This happens when
409                  * direct reclaiming a large contiguous area and the
410                  * first attempt to free a range of pages fails.
411                  */
412                 if (PageWriteback(page) && sync_writeback == PAGEOUT_IO_SYNC)
413                         wait_on_page_writeback(page);
414
415                 if (!PageWriteback(page)) {
416                         /* synchronous write or broken a_ops? */
417                         ClearPageReclaim(page);
418                 }
419                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
420                 return PAGE_SUCCESS;
421         }
422
423         return PAGE_CLEAN;
424 }
425
426 /*
427  * Same as remove_mapping, but if the page is removed from the mapping, it
428  * gets returned with a refcount of 0.
429  */
430 static int __remove_mapping(struct address_space *mapping, struct page *page)
431 {
432         BUG_ON(!PageLocked(page));
433         BUG_ON(mapping != page_mapping(page));
434
435         spin_lock_irq(&mapping->tree_lock);
436         /*
437          * The non racy check for a busy page.
438          *
439          * Must be careful with the order of the tests. When someone has
440          * a ref to the page, it may be possible that they dirty it then
441          * drop the reference. So if PageDirty is tested before page_count
442          * here, then the following race may occur:
443          *
444          * get_user_pages(&page);
445          * [user mapping goes away]
446          * write_to(page);
447          *                              !PageDirty(page)    [good]
448          * SetPageDirty(page);
449          * put_page(page);
450          *                              !page_count(page)   [good, discard it]
451          *
452          * [oops, our write_to data is lost]
453          *
454          * Reversing the order of the tests ensures such a situation cannot
455          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
456          * load is not satisfied before that of page->_count.
457          *
458          * Note that if SetPageDirty is always performed via set_page_dirty,
459          * and thus under tree_lock, then this ordering is not required.
460          */
461         if (!page_freeze_refs(page, 2))
462                 goto cannot_free;
463         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
464         if (unlikely(PageDirty(page))) {
465                 page_unfreeze_refs(page, 2);
466                 goto cannot_free;
467         }
468
469         if (PageSwapCache(page)) {
470                 swp_entry_t swap = { .val = page_private(page) };
471                 __delete_from_swap_cache(page);
472                 spin_unlock_irq(&mapping->tree_lock);
473                 swapcache_free(swap, page);
474         } else {
475                 __remove_from_page_cache(page);
476                 spin_unlock_irq(&mapping->tree_lock);
477                 mem_cgroup_uncharge_cache_page(page);
478         }
479
480         return 1;
481
482 cannot_free:
483         spin_unlock_irq(&mapping->tree_lock);
484         return 0;
485 }
486
487 /*
488  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
489  * someone else has a ref on the page, abort and return 0.  If it was
490  * successfully detached, return 1.  Assumes the caller has a single ref on
491  * this page.
492  */
493 int remove_mapping(struct address_space *mapping, struct page *page)
494 {
495         if (__remove_mapping(mapping, page)) {
496                 /*
497                  * Unfreezing the refcount with 1 rather than 2 effectively
498                  * drops the pagecache ref for us without requiring another
499                  * atomic operation.
500                  */
501                 page_unfreeze_refs(page, 1);
502                 return 1;
503         }
504         return 0;
505 }
506
507 /**
508  * putback_lru_page - put previously isolated page onto appropriate LRU list
509  * @page: page to be put back to appropriate lru list
510  *
511  * Add previously isolated @page to appropriate LRU list.
512  * Page may still be unevictable for other reasons.
513  *
514  * lru_lock must not be held, interrupts must be enabled.
515  */
516 void putback_lru_page(struct page *page)
517 {
518         int lru;
519         int active = !!TestClearPageActive(page);
520         int was_unevictable = PageUnevictable(page);
521
522         VM_BUG_ON(PageLRU(page));
523
524 redo:
525         ClearPageUnevictable(page);
526
527         if (page_evictable(page, NULL)) {
528                 /*
529                  * For evictable pages, we can use the cache.
530                  * In event of a race, worst case is we end up with an
531                  * unevictable page on [in]active list.
532                  * We know how to handle that.
533                  */
534                 lru = active + page_is_file_cache(page);
535                 lru_cache_add_lru(page, lru);
536         } else {
537                 /*
538                  * Put unevictable pages directly on zone's unevictable
539                  * list.
540                  */
541                 lru = LRU_UNEVICTABLE;
542                 add_page_to_unevictable_list(page);
543         }
544
545         /*
546          * page's status can change while we move it among lru. If an evictable
547          * page is on unevictable list, it never be freed. To avoid that,
548          * check after we added it to the list, again.
549          */
550         if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
551                 if (!isolate_lru_page(page)) {
552                         put_page(page);
553                         goto redo;
554                 }
555                 /* This means someone else dropped this page from LRU
556                  * So, it will be freed or putback to LRU again. There is
557                  * nothing to do here.
558                  */
559         }
560
561         if (was_unevictable && lru != LRU_UNEVICTABLE)
562                 count_vm_event(UNEVICTABLE_PGRESCUED);
563         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
564                 count_vm_event(UNEVICTABLE_PGCULLED);
565
566         put_page(page);         /* drop ref from isolate */
567 }
568
569 /*
570  * shrink_page_list() returns the number of reclaimed pages
571  */
572 static unsigned long shrink_page_list(struct list_head *page_list,
573                                         struct scan_control *sc,
574                                         enum pageout_io sync_writeback)
575 {
576         LIST_HEAD(ret_pages);
577         struct pagevec freed_pvec;
578         int pgactivate = 0;
579         unsigned long nr_reclaimed = 0;
580         unsigned long vm_flags;
581
582         cond_resched();
583
584         pagevec_init(&freed_pvec, 1);
585         while (!list_empty(page_list)) {
586                 struct address_space *mapping;
587                 struct page *page;
588                 int may_enter_fs;
589                 int referenced;
590
591                 cond_resched();
592
593                 page = lru_to_page(page_list);
594                 list_del(&page->lru);
595
596                 if (!trylock_page(page))
597                         goto keep;
598
599                 VM_BUG_ON(PageActive(page));
600
601                 sc->nr_scanned++;
602
603                 if (unlikely(!page_evictable(page, NULL)))
604                         goto cull_mlocked;
605
606                 if (!sc->may_unmap && page_mapped(page))
607                         goto keep_locked;
608
609                 /* Double the slab pressure for mapped and swapcache pages */
610                 if (page_mapped(page) || PageSwapCache(page))
611                         sc->nr_scanned++;
612
613                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
614                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
615
616                 if (PageWriteback(page)) {
617                         /*
618                          * Synchronous reclaim is performed in two passes,
619                          * first an asynchronous pass over the list to
620                          * start parallel writeback, and a second synchronous
621                          * pass to wait for the IO to complete.  Wait here
622                          * for any page for which writeback has already
623                          * started.
624                          */
625                         if (sync_writeback == PAGEOUT_IO_SYNC && may_enter_fs)
626                                 wait_on_page_writeback(page);
627                         else
628                                 goto keep_locked;
629                 }
630
631                 referenced = page_referenced(page, 1,
632                                                 sc->mem_cgroup, &vm_flags);
633                 /* In active use or really unfreeable?  Activate it. */
634                 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER &&
635                                         referenced && page_mapping_inuse(page))
636                         goto activate_locked;
637
638                 /*
639                  * Anonymous process memory has backing store?
640                  * Try to allocate it some swap space here.
641                  */
642                 if (PageAnon(page) && !PageSwapCache(page)) {
643                         if (!(sc->gfp_mask & __GFP_IO))
644                                 goto keep_locked;
645                         if (!add_to_swap(page))
646                                 goto activate_locked;
647                         may_enter_fs = 1;
648                 }
649
650                 mapping = page_mapping(page);
651
652                 /*
653                  * The page is mapped into the page tables of one or more
654                  * processes. Try to unmap it here.
655                  */
656                 if (page_mapped(page) && mapping) {
657                         switch (try_to_unmap(page, 0)) {
658                         case SWAP_FAIL:
659                                 goto activate_locked;
660                         case SWAP_AGAIN:
661                                 goto keep_locked;
662                         case SWAP_MLOCK:
663                                 goto cull_mlocked;
664                         case SWAP_SUCCESS:
665                                 ; /* try to free the page below */
666                         }
667                 }
668
669                 if (PageDirty(page)) {
670                         if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && referenced)
671                                 goto keep_locked;
672                         if (!may_enter_fs)
673                                 goto keep_locked;
674                         if (!sc->may_writepage)
675                                 goto keep_locked;
676
677                         /* Page is dirty, try to write it out here */
678                         switch (pageout(page, mapping, sync_writeback)) {
679                         case PAGE_KEEP:
680                                 goto keep_locked;
681                         case PAGE_ACTIVATE:
682                                 goto activate_locked;
683                         case PAGE_SUCCESS:
684                                 if (PageWriteback(page) || PageDirty(page))
685                                         goto keep;
686                                 /*
687                                  * A synchronous write - probably a ramdisk.  Go
688                                  * ahead and try to reclaim the page.
689                                  */
690                                 if (!trylock_page(page))
691                                         goto keep;
692                                 if (PageDirty(page) || PageWriteback(page))
693                                         goto keep_locked;
694                                 mapping = page_mapping(page);
695                         case PAGE_CLEAN:
696                                 ; /* try to free the page below */
697                         }
698                 }
699
700                 /*
701                  * If the page has buffers, try to free the buffer mappings
702                  * associated with this page. If we succeed we try to free
703                  * the page as well.
704                  *
705                  * We do this even if the page is PageDirty().
706                  * try_to_release_page() does not perform I/O, but it is
707                  * possible for a page to have PageDirty set, but it is actually
708                  * clean (all its buffers are clean).  This happens if the
709                  * buffers were written out directly, with submit_bh(). ext3
710                  * will do this, as well as the blockdev mapping.
711                  * try_to_release_page() will discover that cleanness and will
712                  * drop the buffers and mark the page clean - it can be freed.
713                  *
714                  * Rarely, pages can have buffers and no ->mapping.  These are
715                  * the pages which were not successfully invalidated in
716                  * truncate_complete_page().  We try to drop those buffers here
717                  * and if that worked, and the page is no longer mapped into
718                  * process address space (page_count == 1) it can be freed.
719                  * Otherwise, leave the page on the LRU so it is swappable.
720                  */
721                 if (page_has_private(page)) {
722                         if (!try_to_release_page(page, sc->gfp_mask))
723                                 goto activate_locked;
724                         if (!mapping && page_count(page) == 1) {
725                                 unlock_page(page);
726                                 if (put_page_testzero(page))
727                                         goto free_it;
728                                 else {
729                                         /*
730                                          * rare race with speculative reference.
731                                          * the speculative reference will free
732                                          * this page shortly, so we may
733                                          * increment nr_reclaimed here (and
734                                          * leave it off the LRU).
735                                          */
736                                         nr_reclaimed++;
737                                         continue;
738                                 }
739                         }
740                 }
741
742                 if (!mapping || !__remove_mapping(mapping, page))
743                         goto keep_locked;
744
745                 /*
746                  * At this point, we have no other references and there is
747                  * no way to pick any more up (removed from LRU, removed
748                  * from pagecache). Can use non-atomic bitops now (and
749                  * we obviously don't have to worry about waking up a process
750                  * waiting on the page lock, because there are no references.
751                  */
752                 __clear_page_locked(page);
753 free_it:
754                 nr_reclaimed++;
755                 if (!pagevec_add(&freed_pvec, page)) {
756                         __pagevec_free(&freed_pvec);
757                         pagevec_reinit(&freed_pvec);
758                 }
759                 continue;
760
761 cull_mlocked:
762                 if (PageSwapCache(page))
763                         try_to_free_swap(page);
764                 unlock_page(page);
765                 putback_lru_page(page);
766                 continue;
767
768 activate_locked:
769                 /* Not a candidate for swapping, so reclaim swap space. */
770                 if (PageSwapCache(page) && vm_swap_full())
771                         try_to_free_swap(page);
772                 VM_BUG_ON(PageActive(page));
773                 SetPageActive(page);
774                 pgactivate++;
775 keep_locked:
776                 unlock_page(page);
777 keep:
778                 list_add(&page->lru, &ret_pages);
779                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
780         }
781         list_splice(&ret_pages, page_list);
782         if (pagevec_count(&freed_pvec))
783                 __pagevec_free(&freed_pvec);
784         count_vm_events(PGACTIVATE, pgactivate);
785         return nr_reclaimed;
786 }
787
788 /* LRU Isolation modes. */
789 #define ISOLATE_INACTIVE 0      /* Isolate inactive pages. */
790 #define ISOLATE_ACTIVE 1        /* Isolate active pages. */
791 #define ISOLATE_BOTH 2          /* Isolate both active and inactive pages. */
792
793 /*
794  * Attempt to remove the specified page from its LRU.  Only take this page
795  * if it is of the appropriate PageActive status.  Pages which are being
796  * freed elsewhere are also ignored.
797  *
798  * page:        page to consider
799  * mode:        one of the LRU isolation modes defined above
800  *
801  * returns 0 on success, -ve errno on failure.
802  */
803 int __isolate_lru_page(struct page *page, int mode, int file)
804 {
805         int ret = -EINVAL;
806
807         /* Only take pages on the LRU. */
808         if (!PageLRU(page))
809                 return ret;
810
811         /*
812          * When checking the active state, we need to be sure we are
813          * dealing with comparible boolean values.  Take the logical not
814          * of each.
815          */
816         if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
817                 return ret;
818
819         if (mode != ISOLATE_BOTH && (!page_is_file_cache(page) != !file))
820                 return ret;
821
822         /*
823          * When this function is being called for lumpy reclaim, we
824          * initially look into all LRU pages, active, inactive and
825          * unevictable; only give shrink_page_list evictable pages.
826          */
827         if (PageUnevictable(page))
828                 return ret;
829
830         ret = -EBUSY;
831
832         if (likely(get_page_unless_zero(page))) {
833                 /*
834                  * Be careful not to clear PageLRU until after we're
835                  * sure the page is not being freed elsewhere -- the
836                  * page release code relies on it.
837                  */
838                 ClearPageLRU(page);
839                 ret = 0;
840         }
841
842         return ret;
843 }
844
845 /*
846  * zone->lru_lock is heavily contended.  Some of the functions that
847  * shrink the lists perform better by taking out a batch of pages
848  * and working on them outside the LRU lock.
849  *
850  * For pagecache intensive workloads, this function is the hottest
851  * spot in the kernel (apart from copy_*_user functions).
852  *
853  * Appropriate locks must be held before calling this function.
854  *
855  * @nr_to_scan: The number of pages to look through on the list.
856  * @src:        The LRU list to pull pages off.
857  * @dst:        The temp list to put pages on to.
858  * @scanned:    The number of pages that were scanned.
859  * @order:      The caller's attempted allocation order
860  * @mode:       One of the LRU isolation modes
861  * @file:       True [1] if isolating file [!anon] pages
862  *
863  * returns how many pages were moved onto *@dst.
864  */
865 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
866                 struct list_head *src, struct list_head *dst,
867                 unsigned long *scanned, int order, int mode, int file)
868 {
869         unsigned long nr_taken = 0;
870         unsigned long scan;
871
872         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
873                 struct page *page;
874                 unsigned long pfn;
875                 unsigned long end_pfn;
876                 unsigned long page_pfn;
877                 int zone_id;
878
879                 page = lru_to_page(src);
880                 prefetchw_prev_lru_page(page, src, flags);
881
882                 VM_BUG_ON(!PageLRU(page));
883
884                 switch (__isolate_lru_page(page, mode, file)) {
885                 case 0:
886                         list_move(&page->lru, dst);
887                         mem_cgroup_del_lru(page);
888                         nr_taken++;
889                         break;
890
891                 case -EBUSY:
892                         /* else it is being freed elsewhere */
893                         list_move(&page->lru, src);
894                         mem_cgroup_rotate_lru_list(page, page_lru(page));
895                         continue;
896
897                 default:
898                         BUG();
899                 }
900
901                 if (!order)
902                         continue;
903
904                 /*
905                  * Attempt to take all pages in the order aligned region
906                  * surrounding the tag page.  Only take those pages of
907                  * the same active state as that tag page.  We may safely
908                  * round the target page pfn down to the requested order
909                  * as the mem_map is guarenteed valid out to MAX_ORDER,
910                  * where that page is in a different zone we will detect
911                  * it from its zone id and abort this block scan.
912                  */
913                 zone_id = page_zone_id(page);
914                 page_pfn = page_to_pfn(page);
915                 pfn = page_pfn & ~((1 << order) - 1);
916                 end_pfn = pfn + (1 << order);
917                 for (; pfn < end_pfn; pfn++) {
918                         struct page *cursor_page;
919
920                         /* The target page is in the block, ignore it. */
921                         if (unlikely(pfn == page_pfn))
922                                 continue;
923
924                         /* Avoid holes within the zone. */
925                         if (unlikely(!pfn_valid_within(pfn)))
926                                 break;
927
928                         cursor_page = pfn_to_page(pfn);
929
930                         /* Check that we have not crossed a zone boundary. */
931                         if (unlikely(page_zone_id(cursor_page) != zone_id))
932                                 continue;
933                         if (__isolate_lru_page(cursor_page, mode, file) == 0) {
934                                 list_move(&cursor_page->lru, dst);
935                                 mem_cgroup_del_lru(cursor_page);
936                                 nr_taken++;
937                                 scan++;
938                         }
939                 }
940         }
941
942         *scanned = scan;
943         return nr_taken;
944 }
945
946 static unsigned long isolate_pages_global(unsigned long nr,
947                                         struct list_head *dst,
948                                         unsigned long *scanned, int order,
949                                         int mode, struct zone *z,
950                                         struct mem_cgroup *mem_cont,
951                                         int active, int file)
952 {
953         int lru = LRU_BASE;
954         if (active)
955                 lru += LRU_ACTIVE;
956         if (file)
957                 lru += LRU_FILE;
958         return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
959                                                                 mode, !!file);
960 }
961
962 /*
963  * clear_active_flags() is a helper for shrink_active_list(), clearing
964  * any active bits from the pages in the list.
965  */
966 static unsigned long clear_active_flags(struct list_head *page_list,
967                                         unsigned int *count)
968 {
969         int nr_active = 0;
970         int lru;
971         struct page *page;
972
973         list_for_each_entry(page, page_list, lru) {
974                 lru = page_is_file_cache(page);
975                 if (PageActive(page)) {
976                         lru += LRU_ACTIVE;
977                         ClearPageActive(page);
978                         nr_active++;
979                 }
980                 count[lru]++;
981         }
982
983         return nr_active;
984 }
985
986 /**
987  * isolate_lru_page - tries to isolate a page from its LRU list
988  * @page: page to isolate from its LRU list
989  *
990  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
991  * vmstat statistic corresponding to whatever LRU list the page was on.
992  *
993  * Returns 0 if the page was removed from an LRU list.
994  * Returns -EBUSY if the page was not on an LRU list.
995  *
996  * The returned page will have PageLRU() cleared.  If it was found on
997  * the active list, it will have PageActive set.  If it was found on
998  * the unevictable list, it will have the PageUnevictable bit set. That flag
999  * may need to be cleared by the caller before letting the page go.
1000  *
1001  * The vmstat statistic corresponding to the list on which the page was
1002  * found will be decremented.
1003  *
1004  * Restrictions:
1005  * (1) Must be called with an elevated refcount on the page. This is a
1006  *     fundamentnal difference from isolate_lru_pages (which is called
1007  *     without a stable reference).
1008  * (2) the lru_lock must not be held.
1009  * (3) interrupts must be enabled.
1010  */
1011 int isolate_lru_page(struct page *page)
1012 {
1013         int ret = -EBUSY;
1014
1015         if (PageLRU(page)) {
1016                 struct zone *zone = page_zone(page);
1017
1018                 spin_lock_irq(&zone->lru_lock);
1019                 if (PageLRU(page) && get_page_unless_zero(page)) {
1020                         int lru = page_lru(page);
1021                         ret = 0;
1022                         ClearPageLRU(page);
1023
1024                         del_page_from_lru_list(zone, page, lru);
1025                 }
1026                 spin_unlock_irq(&zone->lru_lock);
1027         }
1028         return ret;
1029 }
1030
1031 /*
1032  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1033  * of reclaimed pages
1034  */
1035 static unsigned long shrink_inactive_list(unsigned long max_scan,
1036                         struct zone *zone, struct scan_control *sc,
1037                         int priority, int file)
1038 {
1039         LIST_HEAD(page_list);
1040         struct pagevec pvec;
1041         unsigned long nr_scanned = 0;
1042         unsigned long nr_reclaimed = 0;
1043         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1044         int lumpy_reclaim = 0;
1045
1046         /*
1047          * If we need a large contiguous chunk of memory, or have
1048          * trouble getting a small set of contiguous pages, we
1049          * will reclaim both active and inactive pages.
1050          *
1051          * We use the same threshold as pageout congestion_wait below.
1052          */
1053         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1054                 lumpy_reclaim = 1;
1055         else if (sc->order && priority < DEF_PRIORITY - 2)
1056                 lumpy_reclaim = 1;
1057
1058         pagevec_init(&pvec, 1);
1059
1060         lru_add_drain();
1061         spin_lock_irq(&zone->lru_lock);
1062         do {
1063                 struct page *page;
1064                 unsigned long nr_taken;
1065                 unsigned long nr_scan;
1066                 unsigned long nr_freed;
1067                 unsigned long nr_active;
1068                 unsigned int count[NR_LRU_LISTS] = { 0, };
1069                 int mode = lumpy_reclaim ? ISOLATE_BOTH : ISOLATE_INACTIVE;
1070
1071                 nr_taken = sc->isolate_pages(sc->swap_cluster_max,
1072                              &page_list, &nr_scan, sc->order, mode,
1073                                 zone, sc->mem_cgroup, 0, file);
1074                 nr_active = clear_active_flags(&page_list, count);
1075                 __count_vm_events(PGDEACTIVATE, nr_active);
1076
1077                 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1078                                                 -count[LRU_ACTIVE_FILE]);
1079                 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1080                                                 -count[LRU_INACTIVE_FILE]);
1081                 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1082                                                 -count[LRU_ACTIVE_ANON]);
1083                 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1084                                                 -count[LRU_INACTIVE_ANON]);
1085
1086                 if (scanning_global_lru(sc))
1087                         zone->pages_scanned += nr_scan;
1088
1089                 reclaim_stat->recent_scanned[0] += count[LRU_INACTIVE_ANON];
1090                 reclaim_stat->recent_scanned[0] += count[LRU_ACTIVE_ANON];
1091                 reclaim_stat->recent_scanned[1] += count[LRU_INACTIVE_FILE];
1092                 reclaim_stat->recent_scanned[1] += count[LRU_ACTIVE_FILE];
1093
1094                 spin_unlock_irq(&zone->lru_lock);
1095
1096                 nr_scanned += nr_scan;
1097                 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
1098
1099                 /*
1100                  * If we are direct reclaiming for contiguous pages and we do
1101                  * not reclaim everything in the list, try again and wait
1102                  * for IO to complete. This will stall high-order allocations
1103                  * but that should be acceptable to the caller
1104                  */
1105                 if (nr_freed < nr_taken && !current_is_kswapd() &&
1106                     lumpy_reclaim) {
1107                         congestion_wait(WRITE, HZ/10);
1108
1109                         /*
1110                          * The attempt at page out may have made some
1111                          * of the pages active, mark them inactive again.
1112                          */
1113                         nr_active = clear_active_flags(&page_list, count);
1114                         count_vm_events(PGDEACTIVATE, nr_active);
1115
1116                         nr_freed += shrink_page_list(&page_list, sc,
1117                                                         PAGEOUT_IO_SYNC);
1118                 }
1119
1120                 nr_reclaimed += nr_freed;
1121                 local_irq_disable();
1122                 if (current_is_kswapd()) {
1123                         __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scan);
1124                         __count_vm_events(KSWAPD_STEAL, nr_freed);
1125                 } else if (scanning_global_lru(sc))
1126                         __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scan);
1127
1128                 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
1129
1130                 if (nr_taken == 0)
1131                         goto done;
1132
1133                 spin_lock(&zone->lru_lock);
1134                 /*
1135                  * Put back any unfreeable pages.
1136                  */
1137                 while (!list_empty(&page_list)) {
1138                         int lru;
1139                         page = lru_to_page(&page_list);
1140                         VM_BUG_ON(PageLRU(page));
1141                         list_del(&page->lru);
1142                         if (unlikely(!page_evictable(page, NULL))) {
1143                                 spin_unlock_irq(&zone->lru_lock);
1144                                 putback_lru_page(page);
1145                                 spin_lock_irq(&zone->lru_lock);
1146                                 continue;
1147                         }
1148                         SetPageLRU(page);
1149                         lru = page_lru(page);
1150                         add_page_to_lru_list(zone, page, lru);
1151                         if (PageActive(page)) {
1152                                 int file = !!page_is_file_cache(page);
1153                                 reclaim_stat->recent_rotated[file]++;
1154                         }
1155                         if (!pagevec_add(&pvec, page)) {
1156                                 spin_unlock_irq(&zone->lru_lock);
1157                                 __pagevec_release(&pvec);
1158                                 spin_lock_irq(&zone->lru_lock);
1159                         }
1160                 }
1161         } while (nr_scanned < max_scan);
1162         spin_unlock(&zone->lru_lock);
1163 done:
1164         local_irq_enable();
1165         pagevec_release(&pvec);
1166         return nr_reclaimed;
1167 }
1168
1169 /*
1170  * We are about to scan this zone at a certain priority level.  If that priority
1171  * level is smaller (ie: more urgent) than the previous priority, then note
1172  * that priority level within the zone.  This is done so that when the next
1173  * process comes in to scan this zone, it will immediately start out at this
1174  * priority level rather than having to build up its own scanning priority.
1175  * Here, this priority affects only the reclaim-mapped threshold.
1176  */
1177 static inline void note_zone_scanning_priority(struct zone *zone, int priority)
1178 {
1179         if (priority < zone->prev_priority)
1180                 zone->prev_priority = priority;
1181 }
1182
1183 /*
1184  * This moves pages from the active list to the inactive list.
1185  *
1186  * We move them the other way if the page is referenced by one or more
1187  * processes, from rmap.
1188  *
1189  * If the pages are mostly unmapped, the processing is fast and it is
1190  * appropriate to hold zone->lru_lock across the whole operation.  But if
1191  * the pages are mapped, the processing is slow (page_referenced()) so we
1192  * should drop zone->lru_lock around each page.  It's impossible to balance
1193  * this, so instead we remove the pages from the LRU while processing them.
1194  * It is safe to rely on PG_active against the non-LRU pages in here because
1195  * nobody will play with that bit on a non-LRU page.
1196  *
1197  * The downside is that we have to touch page->_count against each page.
1198  * But we had to alter page->flags anyway.
1199  */
1200
1201 static void move_active_pages_to_lru(struct zone *zone,
1202                                      struct list_head *list,
1203                                      enum lru_list lru)
1204 {
1205         unsigned long pgmoved = 0;
1206         struct pagevec pvec;
1207         struct page *page;
1208
1209         pagevec_init(&pvec, 1);
1210
1211         while (!list_empty(list)) {
1212                 page = lru_to_page(list);
1213                 prefetchw_prev_lru_page(page, list, flags);
1214
1215                 VM_BUG_ON(PageLRU(page));
1216                 SetPageLRU(page);
1217
1218                 VM_BUG_ON(!PageActive(page));
1219                 if (!is_active_lru(lru))
1220                         ClearPageActive(page);  /* we are de-activating */
1221
1222                 list_move(&page->lru, &zone->lru[lru].list);
1223                 mem_cgroup_add_lru_list(page, lru);
1224                 pgmoved++;
1225
1226                 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1227                         spin_unlock_irq(&zone->lru_lock);
1228                         if (buffer_heads_over_limit)
1229                                 pagevec_strip(&pvec);
1230                         __pagevec_release(&pvec);
1231                         spin_lock_irq(&zone->lru_lock);
1232                 }
1233         }
1234         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1235         if (!is_active_lru(lru))
1236                 __count_vm_events(PGDEACTIVATE, pgmoved);
1237 }
1238
1239 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1240                         struct scan_control *sc, int priority, int file)
1241 {
1242         unsigned long pgmoved;
1243         unsigned long pgscanned;
1244         unsigned long vm_flags;
1245         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1246         LIST_HEAD(l_active);
1247         LIST_HEAD(l_inactive);
1248         struct page *page;
1249         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1250
1251         lru_add_drain();
1252         spin_lock_irq(&zone->lru_lock);
1253         pgmoved = sc->isolate_pages(nr_pages, &l_hold, &pgscanned, sc->order,
1254                                         ISOLATE_ACTIVE, zone,
1255                                         sc->mem_cgroup, 1, file);
1256         /*
1257          * zone->pages_scanned is used for detect zone's oom
1258          * mem_cgroup remembers nr_scan by itself.
1259          */
1260         if (scanning_global_lru(sc)) {
1261                 zone->pages_scanned += pgscanned;
1262         }
1263         reclaim_stat->recent_scanned[!!file] += pgmoved;
1264
1265         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1266         if (file)
1267                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -pgmoved);
1268         else
1269                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -pgmoved);
1270         spin_unlock_irq(&zone->lru_lock);
1271
1272         pgmoved = 0;  /* count referenced (mapping) mapped pages */
1273         while (!list_empty(&l_hold)) {
1274                 cond_resched();
1275                 page = lru_to_page(&l_hold);
1276                 list_del(&page->lru);
1277
1278                 if (unlikely(!page_evictable(page, NULL))) {
1279                         putback_lru_page(page);
1280                         continue;
1281                 }
1282
1283                 /* page_referenced clears PageReferenced */
1284                 if (page_mapping_inuse(page) &&
1285                     page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1286                         pgmoved++;
1287                         /*
1288                          * Identify referenced, file-backed active pages and
1289                          * give them one more trip around the active list. So
1290                          * that executable code get better chances to stay in
1291                          * memory under moderate memory pressure.  Anon pages
1292                          * are not likely to be evicted by use-once streaming
1293                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1294                          * so we ignore them here.
1295                          */
1296                         if ((vm_flags & VM_EXEC) && !PageAnon(page)) {
1297                                 list_add(&page->lru, &l_active);
1298                                 continue;
1299                         }
1300                 }
1301
1302                 list_add(&page->lru, &l_inactive);
1303         }
1304
1305         /*
1306          * Move pages back to the lru list.
1307          */
1308         spin_lock_irq(&zone->lru_lock);
1309         /*
1310          * Count referenced pages from currently used mappings as rotated,
1311          * even though only some of them are actually re-activated.  This
1312          * helps balance scan pressure between file and anonymous pages in
1313          * get_scan_ratio.
1314          */
1315         reclaim_stat->recent_rotated[!!file] += pgmoved;
1316
1317         move_active_pages_to_lru(zone, &l_active,
1318                                                 LRU_ACTIVE + file * LRU_FILE);
1319         move_active_pages_to_lru(zone, &l_inactive,
1320                                                 LRU_BASE   + file * LRU_FILE);
1321
1322         spin_unlock_irq(&zone->lru_lock);
1323 }
1324
1325 static int inactive_anon_is_low_global(struct zone *zone)
1326 {
1327         unsigned long active, inactive;
1328
1329         active = zone_page_state(zone, NR_ACTIVE_ANON);
1330         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1331
1332         if (inactive * zone->inactive_ratio < active)
1333                 return 1;
1334
1335         return 0;
1336 }
1337
1338 /**
1339  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1340  * @zone: zone to check
1341  * @sc:   scan control of this context
1342  *
1343  * Returns true if the zone does not have enough inactive anon pages,
1344  * meaning some active anon pages need to be deactivated.
1345  */
1346 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1347 {
1348         int low;
1349
1350         if (scanning_global_lru(sc))
1351                 low = inactive_anon_is_low_global(zone);
1352         else
1353                 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1354         return low;
1355 }
1356
1357 static int inactive_file_is_low_global(struct zone *zone)
1358 {
1359         unsigned long active, inactive;
1360
1361         active = zone_page_state(zone, NR_ACTIVE_FILE);
1362         inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1363
1364         return (active > inactive);
1365 }
1366
1367 /**
1368  * inactive_file_is_low - check if file pages need to be deactivated
1369  * @zone: zone to check
1370  * @sc:   scan control of this context
1371  *
1372  * When the system is doing streaming IO, memory pressure here
1373  * ensures that active file pages get deactivated, until more
1374  * than half of the file pages are on the inactive list.
1375  *
1376  * Once we get to that situation, protect the system's working
1377  * set from being evicted by disabling active file page aging.
1378  *
1379  * This uses a different ratio than the anonymous pages, because
1380  * the page cache uses a use-once replacement algorithm.
1381  */
1382 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1383 {
1384         int low;
1385
1386         if (scanning_global_lru(sc))
1387                 low = inactive_file_is_low_global(zone);
1388         else
1389                 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1390         return low;
1391 }
1392
1393 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1394         struct zone *zone, struct scan_control *sc, int priority)
1395 {
1396         int file = is_file_lru(lru);
1397
1398         if (lru == LRU_ACTIVE_FILE && inactive_file_is_low(zone, sc)) {
1399                 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1400                 return 0;
1401         }
1402
1403         if (lru == LRU_ACTIVE_ANON && inactive_anon_is_low(zone, sc)) {
1404                 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1405                 return 0;
1406         }
1407         return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1408 }
1409
1410 /*
1411  * Determine how aggressively the anon and file LRU lists should be
1412  * scanned.  The relative value of each set of LRU lists is determined
1413  * by looking at the fraction of the pages scanned we did rotate back
1414  * onto the active list instead of evict.
1415  *
1416  * percent[0] specifies how much pressure to put on ram/swap backed
1417  * memory, while percent[1] determines pressure on the file LRUs.
1418  */
1419 static void get_scan_ratio(struct zone *zone, struct scan_control *sc,
1420                                         unsigned long *percent)
1421 {
1422         unsigned long anon, file, free;
1423         unsigned long anon_prio, file_prio;
1424         unsigned long ap, fp;
1425         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1426
1427         anon  = zone_nr_pages(zone, sc, LRU_ACTIVE_ANON) +
1428                 zone_nr_pages(zone, sc, LRU_INACTIVE_ANON);
1429         file  = zone_nr_pages(zone, sc, LRU_ACTIVE_FILE) +
1430                 zone_nr_pages(zone, sc, LRU_INACTIVE_FILE);
1431
1432         if (scanning_global_lru(sc)) {
1433                 free  = zone_page_state(zone, NR_FREE_PAGES);
1434                 /* If we have very few page cache pages,
1435                    force-scan anon pages. */
1436                 if (unlikely(file + free <= high_wmark_pages(zone))) {
1437                         percent[0] = 100;
1438                         percent[1] = 0;
1439                         return;
1440                 }
1441         }
1442
1443         /*
1444          * OK, so we have swap space and a fair amount of page cache
1445          * pages.  We use the recently rotated / recently scanned
1446          * ratios to determine how valuable each cache is.
1447          *
1448          * Because workloads change over time (and to avoid overflow)
1449          * we keep these statistics as a floating average, which ends
1450          * up weighing recent references more than old ones.
1451          *
1452          * anon in [0], file in [1]
1453          */
1454         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1455                 spin_lock_irq(&zone->lru_lock);
1456                 reclaim_stat->recent_scanned[0] /= 2;
1457                 reclaim_stat->recent_rotated[0] /= 2;
1458                 spin_unlock_irq(&zone->lru_lock);
1459         }
1460
1461         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1462                 spin_lock_irq(&zone->lru_lock);
1463                 reclaim_stat->recent_scanned[1] /= 2;
1464                 reclaim_stat->recent_rotated[1] /= 2;
1465                 spin_unlock_irq(&zone->lru_lock);
1466         }
1467
1468         /*
1469          * With swappiness at 100, anonymous and file have the same priority.
1470          * This scanning priority is essentially the inverse of IO cost.
1471          */
1472         anon_prio = sc->swappiness;
1473         file_prio = 200 - sc->swappiness;
1474
1475         /*
1476          * The amount of pressure on anon vs file pages is inversely
1477          * proportional to the fraction of recently scanned pages on
1478          * each list that were recently referenced and in active use.
1479          */
1480         ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1481         ap /= reclaim_stat->recent_rotated[0] + 1;
1482
1483         fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1484         fp /= reclaim_stat->recent_rotated[1] + 1;
1485
1486         /* Normalize to percentages */
1487         percent[0] = 100 * ap / (ap + fp + 1);
1488         percent[1] = 100 - percent[0];
1489 }
1490
1491 /*
1492  * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1493  * until we collected @swap_cluster_max pages to scan.
1494  */
1495 static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
1496                                        unsigned long *nr_saved_scan,
1497                                        unsigned long swap_cluster_max)
1498 {
1499         unsigned long nr;
1500
1501         *nr_saved_scan += nr_to_scan;
1502         nr = *nr_saved_scan;
1503
1504         if (nr >= swap_cluster_max)
1505                 *nr_saved_scan = 0;
1506         else
1507                 nr = 0;
1508
1509         return nr;
1510 }
1511
1512 /*
1513  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1514  */
1515 static void shrink_zone(int priority, struct zone *zone,
1516                                 struct scan_control *sc)
1517 {
1518         unsigned long nr[NR_LRU_LISTS];
1519         unsigned long nr_to_scan;
1520         unsigned long percent[2];       /* anon @ 0; file @ 1 */
1521         enum lru_list l;
1522         unsigned long nr_reclaimed = sc->nr_reclaimed;
1523         unsigned long swap_cluster_max = sc->swap_cluster_max;
1524         int noswap = 0;
1525
1526         /* If we have no swap space, do not bother scanning anon pages. */
1527         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1528                 noswap = 1;
1529                 percent[0] = 0;
1530                 percent[1] = 100;
1531         } else
1532                 get_scan_ratio(zone, sc, percent);
1533
1534         for_each_evictable_lru(l) {
1535                 int file = is_file_lru(l);
1536                 unsigned long scan;
1537
1538                 scan = zone_nr_pages(zone, sc, l);
1539                 if (priority || noswap) {
1540                         scan >>= priority;
1541                         scan = (scan * percent[file]) / 100;
1542                 }
1543                 if (scanning_global_lru(sc))
1544                         nr[l] = nr_scan_try_batch(scan,
1545                                                   &zone->lru[l].nr_saved_scan,
1546                                                   swap_cluster_max);
1547                 else
1548                         nr[l] = scan;
1549         }
1550
1551         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1552                                         nr[LRU_INACTIVE_FILE]) {
1553                 for_each_evictable_lru(l) {
1554                         if (nr[l]) {
1555                                 nr_to_scan = min(nr[l], swap_cluster_max);
1556                                 nr[l] -= nr_to_scan;
1557
1558                                 nr_reclaimed += shrink_list(l, nr_to_scan,
1559                                                             zone, sc, priority);
1560                         }
1561                 }
1562                 /*
1563                  * On large memory systems, scan >> priority can become
1564                  * really large. This is fine for the starting priority;
1565                  * we want to put equal scanning pressure on each zone.
1566                  * However, if the VM has a harder time of freeing pages,
1567                  * with multiple processes reclaiming pages, the total
1568                  * freeing target can get unreasonably large.
1569                  */
1570                 if (nr_reclaimed > swap_cluster_max &&
1571                         priority < DEF_PRIORITY && !current_is_kswapd())
1572                         break;
1573         }
1574
1575         sc->nr_reclaimed = nr_reclaimed;
1576
1577         /*
1578          * Even if we did not try to evict anon pages at all, we want to
1579          * rebalance the anon lru active/inactive ratio.
1580          */
1581         if (inactive_anon_is_low(zone, sc) && nr_swap_pages > 0)
1582                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1583
1584         throttle_vm_writeout(sc->gfp_mask);
1585 }
1586
1587 /*
1588  * This is the direct reclaim path, for page-allocating processes.  We only
1589  * try to reclaim pages from zones which will satisfy the caller's allocation
1590  * request.
1591  *
1592  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1593  * Because:
1594  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1595  *    allocation or
1596  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1597  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1598  *    zone defense algorithm.
1599  *
1600  * If a zone is deemed to be full of pinned pages then just give it a light
1601  * scan then give up on it.
1602  */
1603 static void shrink_zones(int priority, struct zonelist *zonelist,
1604                                         struct scan_control *sc)
1605 {
1606         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1607         struct zoneref *z;
1608         struct zone *zone;
1609
1610         sc->all_unreclaimable = 1;
1611         for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1612                                         sc->nodemask) {
1613                 if (!populated_zone(zone))
1614                         continue;
1615                 /*
1616                  * Take care memory controller reclaiming has small influence
1617                  * to global LRU.
1618                  */
1619                 if (scanning_global_lru(sc)) {
1620                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1621                                 continue;
1622                         note_zone_scanning_priority(zone, priority);
1623
1624                         if (zone_is_all_unreclaimable(zone) &&
1625                                                 priority != DEF_PRIORITY)
1626                                 continue;       /* Let kswapd poll it */
1627                         sc->all_unreclaimable = 0;
1628                 } else {
1629                         /*
1630                          * Ignore cpuset limitation here. We just want to reduce
1631                          * # of used pages by us regardless of memory shortage.
1632                          */
1633                         sc->all_unreclaimable = 0;
1634                         mem_cgroup_note_reclaim_priority(sc->mem_cgroup,
1635                                                         priority);
1636                 }
1637
1638                 shrink_zone(priority, zone, sc);
1639         }
1640 }
1641
1642 /*
1643  * This is the main entry point to direct page reclaim.
1644  *
1645  * If a full scan of the inactive list fails to free enough memory then we
1646  * are "out of memory" and something needs to be killed.
1647  *
1648  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1649  * high - the zone may be full of dirty or under-writeback pages, which this
1650  * caller can't do much about.  We kick pdflush and take explicit naps in the
1651  * hope that some of these pages can be written.  But if the allocating task
1652  * holds filesystem locks which prevent writeout this might not work, and the
1653  * allocation attempt will fail.
1654  *
1655  * returns:     0, if no pages reclaimed
1656  *              else, the number of pages reclaimed
1657  */
1658 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
1659                                         struct scan_control *sc)
1660 {
1661         int priority;
1662         unsigned long ret = 0;
1663         unsigned long total_scanned = 0;
1664         struct reclaim_state *reclaim_state = current->reclaim_state;
1665         unsigned long lru_pages = 0;
1666         struct zoneref *z;
1667         struct zone *zone;
1668         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1669
1670         delayacct_freepages_start();
1671
1672         if (scanning_global_lru(sc))
1673                 count_vm_event(ALLOCSTALL);
1674         /*
1675          * mem_cgroup will not do shrink_slab.
1676          */
1677         if (scanning_global_lru(sc)) {
1678                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1679
1680                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1681                                 continue;
1682
1683                         lru_pages += zone_lru_pages(zone);
1684                 }
1685         }
1686
1687         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1688                 sc->nr_scanned = 0;
1689                 if (!priority)
1690                         disable_swap_token();
1691                 shrink_zones(priority, zonelist, sc);
1692                 /*
1693                  * Don't shrink slabs when reclaiming memory from
1694                  * over limit cgroups
1695                  */
1696                 if (scanning_global_lru(sc)) {
1697                         shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
1698                         if (reclaim_state) {
1699                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
1700                                 reclaim_state->reclaimed_slab = 0;
1701                         }
1702                 }
1703                 total_scanned += sc->nr_scanned;
1704                 if (sc->nr_reclaimed >= sc->swap_cluster_max) {
1705                         ret = sc->nr_reclaimed;
1706                         goto out;
1707                 }
1708
1709                 /*
1710                  * Try to write back as many pages as we just scanned.  This
1711                  * tends to cause slow streaming writers to write data to the
1712                  * disk smoothly, at the dirtying rate, which is nice.   But
1713                  * that's undesirable in laptop mode, where we *want* lumpy
1714                  * writeout.  So in laptop mode, write out the whole world.
1715                  */
1716                 if (total_scanned > sc->swap_cluster_max +
1717                                         sc->swap_cluster_max / 2) {
1718                         wakeup_pdflush(laptop_mode ? 0 : total_scanned);
1719                         sc->may_writepage = 1;
1720                 }
1721
1722                 /* Take a nap, wait for some writeback to complete */
1723                 if (sc->nr_scanned && priority < DEF_PRIORITY - 2)
1724                         congestion_wait(WRITE, HZ/10);
1725         }
1726         /* top priority shrink_zones still had more to do? don't OOM, then */
1727         if (!sc->all_unreclaimable && scanning_global_lru(sc))
1728                 ret = sc->nr_reclaimed;
1729 out:
1730         /*
1731          * Now that we've scanned all the zones at this priority level, note
1732          * that level within the zone so that the next thread which performs
1733          * scanning of this zone will immediately start out at this priority
1734          * level.  This affects only the decision whether or not to bring
1735          * mapped pages onto the inactive list.
1736          */
1737         if (priority < 0)
1738                 priority = 0;
1739
1740         if (scanning_global_lru(sc)) {
1741                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1742
1743                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1744                                 continue;
1745
1746                         zone->prev_priority = priority;
1747                 }
1748         } else
1749                 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority);
1750
1751         delayacct_freepages_end();
1752
1753         return ret;
1754 }
1755
1756 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
1757                                 gfp_t gfp_mask, nodemask_t *nodemask)
1758 {
1759         struct scan_control sc = {
1760                 .gfp_mask = gfp_mask,
1761                 .may_writepage = !laptop_mode,
1762                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1763                 .may_unmap = 1,
1764                 .may_swap = 1,
1765                 .swappiness = vm_swappiness,
1766                 .order = order,
1767                 .mem_cgroup = NULL,
1768                 .isolate_pages = isolate_pages_global,
1769                 .nodemask = nodemask,
1770         };
1771
1772         return do_try_to_free_pages(zonelist, &sc);
1773 }
1774
1775 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
1776
1777 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
1778                                            gfp_t gfp_mask,
1779                                            bool noswap,
1780                                            unsigned int swappiness)
1781 {
1782         struct scan_control sc = {
1783                 .may_writepage = !laptop_mode,
1784                 .may_unmap = 1,
1785                 .may_swap = !noswap,
1786                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1787                 .swappiness = swappiness,
1788                 .order = 0,
1789                 .mem_cgroup = mem_cont,
1790                 .isolate_pages = mem_cgroup_isolate_pages,
1791                 .nodemask = NULL, /* we don't care the placement */
1792         };
1793         struct zonelist *zonelist;
1794
1795         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1796                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1797         zonelist = NODE_DATA(numa_node_id())->node_zonelists;
1798         return do_try_to_free_pages(zonelist, &sc);
1799 }
1800 #endif
1801
1802 /*
1803  * For kswapd, balance_pgdat() will work across all this node's zones until
1804  * they are all at high_wmark_pages(zone).
1805  *
1806  * Returns the number of pages which were actually freed.
1807  *
1808  * There is special handling here for zones which are full of pinned pages.
1809  * This can happen if the pages are all mlocked, or if they are all used by
1810  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1811  * What we do is to detect the case where all pages in the zone have been
1812  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1813  * dead and from now on, only perform a short scan.  Basically we're polling
1814  * the zone for when the problem goes away.
1815  *
1816  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
1817  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
1818  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
1819  * lower zones regardless of the number of free pages in the lower zones. This
1820  * interoperates with the page allocator fallback scheme to ensure that aging
1821  * of pages is balanced across the zones.
1822  */
1823 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
1824 {
1825         int all_zones_ok;
1826         int priority;
1827         int i;
1828         unsigned long total_scanned;
1829         struct reclaim_state *reclaim_state = current->reclaim_state;
1830         struct scan_control sc = {
1831                 .gfp_mask = GFP_KERNEL,
1832                 .may_unmap = 1,
1833                 .may_swap = 1,
1834                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1835                 .swappiness = vm_swappiness,
1836                 .order = order,
1837                 .mem_cgroup = NULL,
1838                 .isolate_pages = isolate_pages_global,
1839         };
1840         /*
1841          * temp_priority is used to remember the scanning priority at which
1842          * this zone was successfully refilled to
1843          * free_pages == high_wmark_pages(zone).
1844          */
1845         int temp_priority[MAX_NR_ZONES];
1846
1847 loop_again:
1848         total_scanned = 0;
1849         sc.nr_reclaimed = 0;
1850         sc.may_writepage = !laptop_mode;
1851         count_vm_event(PAGEOUTRUN);
1852
1853         for (i = 0; i < pgdat->nr_zones; i++)
1854                 temp_priority[i] = DEF_PRIORITY;
1855
1856         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1857                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1858                 unsigned long lru_pages = 0;
1859
1860                 /* The swap token gets in the way of swapout... */
1861                 if (!priority)
1862                         disable_swap_token();
1863
1864                 all_zones_ok = 1;
1865
1866                 /*
1867                  * Scan in the highmem->dma direction for the highest
1868                  * zone which needs scanning
1869                  */
1870                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1871                         struct zone *zone = pgdat->node_zones + i;
1872
1873                         if (!populated_zone(zone))
1874                                 continue;
1875
1876                         if (zone_is_all_unreclaimable(zone) &&
1877                             priority != DEF_PRIORITY)
1878                                 continue;
1879
1880                         /*
1881                          * Do some background aging of the anon list, to give
1882                          * pages a chance to be referenced before reclaiming.
1883                          */
1884                         if (inactive_anon_is_low(zone, &sc))
1885                                 shrink_active_list(SWAP_CLUSTER_MAX, zone,
1886                                                         &sc, priority, 0);
1887
1888                         if (!zone_watermark_ok(zone, order,
1889                                         high_wmark_pages(zone), 0, 0)) {
1890                                 end_zone = i;
1891                                 break;
1892                         }
1893                 }
1894                 if (i < 0)
1895                         goto out;
1896
1897                 for (i = 0; i <= end_zone; i++) {
1898                         struct zone *zone = pgdat->node_zones + i;
1899
1900                         lru_pages += zone_lru_pages(zone);
1901                 }
1902
1903                 /*
1904                  * Now scan the zone in the dma->highmem direction, stopping
1905                  * at the last zone which needs scanning.
1906                  *
1907                  * We do this because the page allocator works in the opposite
1908                  * direction.  This prevents the page allocator from allocating
1909                  * pages behind kswapd's direction of progress, which would
1910                  * cause too much scanning of the lower zones.
1911                  */
1912                 for (i = 0; i <= end_zone; i++) {
1913                         struct zone *zone = pgdat->node_zones + i;
1914                         int nr_slab;
1915
1916                         if (!populated_zone(zone))
1917                                 continue;
1918
1919                         if (zone_is_all_unreclaimable(zone) &&
1920                                         priority != DEF_PRIORITY)
1921                                 continue;
1922
1923                         if (!zone_watermark_ok(zone, order,
1924                                         high_wmark_pages(zone), end_zone, 0))
1925                                 all_zones_ok = 0;
1926                         temp_priority[i] = priority;
1927                         sc.nr_scanned = 0;
1928                         note_zone_scanning_priority(zone, priority);
1929                         /*
1930                          * We put equal pressure on every zone, unless one
1931                          * zone has way too many pages free already.
1932                          */
1933                         if (!zone_watermark_ok(zone, order,
1934                                         8*high_wmark_pages(zone), end_zone, 0))
1935                                 shrink_zone(priority, zone, &sc);
1936                         reclaim_state->reclaimed_slab = 0;
1937                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1938                                                 lru_pages);
1939                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1940                         total_scanned += sc.nr_scanned;
1941                         if (zone_is_all_unreclaimable(zone))
1942                                 continue;
1943                         if (nr_slab == 0 && zone->pages_scanned >=
1944                                                 (zone_lru_pages(zone) * 6))
1945                                         zone_set_flag(zone,
1946                                                       ZONE_ALL_UNRECLAIMABLE);
1947                         /*
1948                          * If we've done a decent amount of scanning and
1949                          * the reclaim ratio is low, start doing writepage
1950                          * even in laptop mode
1951                          */
1952                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1953                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
1954                                 sc.may_writepage = 1;
1955                 }
1956                 if (all_zones_ok)
1957                         break;          /* kswapd: all done */
1958                 /*
1959                  * OK, kswapd is getting into trouble.  Take a nap, then take
1960                  * another pass across the zones.
1961                  */
1962                 if (total_scanned && priority < DEF_PRIORITY - 2)
1963                         congestion_wait(WRITE, HZ/10);
1964
1965                 /*
1966                  * We do this so kswapd doesn't build up large priorities for
1967                  * example when it is freeing in parallel with allocators. It
1968                  * matches the direct reclaim path behaviour in terms of impact
1969                  * on zone->*_priority.
1970                  */
1971                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
1972                         break;
1973         }
1974 out:
1975         /*
1976          * Note within each zone the priority level at which this zone was
1977          * brought into a happy state.  So that the next thread which scans this
1978          * zone will start out at that priority level.
1979          */
1980         for (i = 0; i < pgdat->nr_zones; i++) {
1981                 struct zone *zone = pgdat->node_zones + i;
1982
1983                 zone->prev_priority = temp_priority[i];
1984         }
1985         if (!all_zones_ok) {
1986                 cond_resched();
1987
1988                 try_to_freeze();
1989
1990                 /*
1991                  * Fragmentation may mean that the system cannot be
1992                  * rebalanced for high-order allocations in all zones.
1993                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
1994                  * it means the zones have been fully scanned and are still
1995                  * not balanced. For high-order allocations, there is
1996                  * little point trying all over again as kswapd may
1997                  * infinite loop.
1998                  *
1999                  * Instead, recheck all watermarks at order-0 as they
2000                  * are the most important. If watermarks are ok, kswapd will go
2001                  * back to sleep. High-order users can still perform direct
2002                  * reclaim if they wish.
2003                  */
2004                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2005                         order = sc.order = 0;
2006
2007                 goto loop_again;
2008         }
2009
2010         return sc.nr_reclaimed;
2011 }
2012
2013 /*
2014  * The background pageout daemon, started as a kernel thread
2015  * from the init process.
2016  *
2017  * This basically trickles out pages so that we have _some_
2018  * free memory available even if there is no other activity
2019  * that frees anything up. This is needed for things like routing
2020  * etc, where we otherwise might have all activity going on in
2021  * asynchronous contexts that cannot page things out.
2022  *
2023  * If there are applications that are active memory-allocators
2024  * (most normal use), this basically shouldn't matter.
2025  */
2026 static int kswapd(void *p)
2027 {
2028         unsigned long order;
2029         pg_data_t *pgdat = (pg_data_t*)p;
2030         struct task_struct *tsk = current;
2031         DEFINE_WAIT(wait);
2032         struct reclaim_state reclaim_state = {
2033                 .reclaimed_slab = 0,
2034         };
2035         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2036
2037         lockdep_set_current_reclaim_state(GFP_KERNEL);
2038
2039         if (!cpumask_empty(cpumask))
2040                 set_cpus_allowed_ptr(tsk, cpumask);
2041         current->reclaim_state = &reclaim_state;
2042
2043         /*
2044          * Tell the memory management that we're a "memory allocator",
2045          * and that if we need more memory we should get access to it
2046          * regardless (see "__alloc_pages()"). "kswapd" should
2047          * never get caught in the normal page freeing logic.
2048          *
2049          * (Kswapd normally doesn't need memory anyway, but sometimes
2050          * you need a small amount of memory in order to be able to
2051          * page out something else, and this flag essentially protects
2052          * us from recursively trying to free more memory as we're
2053          * trying to free the first piece of memory in the first place).
2054          */
2055         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2056         set_freezable();
2057
2058         order = 0;
2059         for ( ; ; ) {
2060                 unsigned long new_order;
2061
2062                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2063                 new_order = pgdat->kswapd_max_order;
2064                 pgdat->kswapd_max_order = 0;
2065                 if (order < new_order) {
2066                         /*
2067                          * Don't sleep if someone wants a larger 'order'
2068                          * allocation
2069                          */
2070                         order = new_order;
2071                 } else {
2072                         if (!freezing(current))
2073                                 schedule();
2074
2075                         order = pgdat->kswapd_max_order;
2076                 }
2077                 finish_wait(&pgdat->kswapd_wait, &wait);
2078
2079                 if (!try_to_freeze()) {
2080                         /* We can speed up thawing tasks if we don't call
2081                          * balance_pgdat after returning from the refrigerator
2082                          */
2083                         balance_pgdat(pgdat, order);
2084                 }
2085         }
2086         return 0;
2087 }
2088
2089 /*
2090  * A zone is low on free memory, so wake its kswapd task to service it.
2091  */
2092 void wakeup_kswapd(struct zone *zone, int order)
2093 {
2094         pg_data_t *pgdat;
2095
2096         if (!populated_zone(zone))
2097                 return;
2098
2099         pgdat = zone->zone_pgdat;
2100         if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
2101                 return;
2102         if (pgdat->kswapd_max_order < order)
2103                 pgdat->kswapd_max_order = order;
2104         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2105                 return;
2106         if (!waitqueue_active(&pgdat->kswapd_wait))
2107                 return;
2108         wake_up_interruptible(&pgdat->kswapd_wait);
2109 }
2110
2111 unsigned long global_lru_pages(void)
2112 {
2113         return global_page_state(NR_ACTIVE_ANON)
2114                 + global_page_state(NR_ACTIVE_FILE)
2115                 + global_page_state(NR_INACTIVE_ANON)
2116                 + global_page_state(NR_INACTIVE_FILE);
2117 }
2118
2119 #ifdef CONFIG_HIBERNATION
2120 /*
2121  * Helper function for shrink_all_memory().  Tries to reclaim 'nr_pages' pages
2122  * from LRU lists system-wide, for given pass and priority.
2123  *
2124  * For pass > 3 we also try to shrink the LRU lists that contain a few pages
2125  */
2126 static void shrink_all_zones(unsigned long nr_pages, int prio,
2127                                       int pass, struct scan_control *sc)
2128 {
2129         struct zone *zone;
2130         unsigned long nr_reclaimed = 0;
2131
2132         for_each_populated_zone(zone) {
2133                 enum lru_list l;
2134
2135                 if (zone_is_all_unreclaimable(zone) && prio != DEF_PRIORITY)
2136                         continue;
2137
2138                 for_each_evictable_lru(l) {
2139                         enum zone_stat_item ls = NR_LRU_BASE + l;
2140                         unsigned long lru_pages = zone_page_state(zone, ls);
2141
2142                         /* For pass = 0, we don't shrink the active list */
2143                         if (pass == 0 && (l == LRU_ACTIVE_ANON ||
2144                                                 l == LRU_ACTIVE_FILE))
2145                                 continue;
2146
2147                         zone->lru[l].nr_saved_scan += (lru_pages >> prio) + 1;
2148                         if (zone->lru[l].nr_saved_scan >= nr_pages || pass > 3) {
2149                                 unsigned long nr_to_scan;
2150
2151                                 zone->lru[l].nr_saved_scan = 0;
2152                                 nr_to_scan = min(nr_pages, lru_pages);
2153                                 nr_reclaimed += shrink_list(l, nr_to_scan, zone,
2154                                                                 sc, prio);
2155                                 if (nr_reclaimed >= nr_pages) {
2156                                         sc->nr_reclaimed += nr_reclaimed;
2157                                         return;
2158                                 }
2159                         }
2160                 }
2161         }
2162         sc->nr_reclaimed += nr_reclaimed;
2163 }
2164
2165 /*
2166  * Try to free `nr_pages' of memory, system-wide, and return the number of
2167  * freed pages.
2168  *
2169  * Rather than trying to age LRUs the aim is to preserve the overall
2170  * LRU order by reclaiming preferentially
2171  * inactive > active > active referenced > active mapped
2172  */
2173 unsigned long shrink_all_memory(unsigned long nr_pages)
2174 {
2175         unsigned long lru_pages, nr_slab;
2176         int pass;
2177         struct reclaim_state reclaim_state;
2178         struct scan_control sc = {
2179                 .gfp_mask = GFP_KERNEL,
2180                 .may_unmap = 0,
2181                 .may_writepage = 1,
2182                 .isolate_pages = isolate_pages_global,
2183                 .nr_reclaimed = 0,
2184         };
2185
2186         current->reclaim_state = &reclaim_state;
2187
2188         lru_pages = global_lru_pages();
2189         nr_slab = global_page_state(NR_SLAB_RECLAIMABLE);
2190         /* If slab caches are huge, it's better to hit them first */
2191         while (nr_slab >= lru_pages) {
2192                 reclaim_state.reclaimed_slab = 0;
2193                 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
2194                 if (!reclaim_state.reclaimed_slab)
2195                         break;
2196
2197                 sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2198                 if (sc.nr_reclaimed >= nr_pages)
2199                         goto out;
2200
2201                 nr_slab -= reclaim_state.reclaimed_slab;
2202         }
2203
2204         /*
2205          * We try to shrink LRUs in 5 passes:
2206          * 0 = Reclaim from inactive_list only
2207          * 1 = Reclaim from active list but don't reclaim mapped
2208          * 2 = 2nd pass of type 1
2209          * 3 = Reclaim mapped (normal reclaim)
2210          * 4 = 2nd pass of type 3
2211          */
2212         for (pass = 0; pass < 5; pass++) {
2213                 int prio;
2214
2215                 /* Force reclaiming mapped pages in the passes #3 and #4 */
2216                 if (pass > 2)
2217                         sc.may_unmap = 1;
2218
2219                 for (prio = DEF_PRIORITY; prio >= 0; prio--) {
2220                         unsigned long nr_to_scan = nr_pages - sc.nr_reclaimed;
2221
2222                         sc.nr_scanned = 0;
2223                         sc.swap_cluster_max = nr_to_scan;
2224                         shrink_all_zones(nr_to_scan, prio, pass, &sc);
2225                         if (sc.nr_reclaimed >= nr_pages)
2226                                 goto out;
2227
2228                         reclaim_state.reclaimed_slab = 0;
2229                         shrink_slab(sc.nr_scanned, sc.gfp_mask,
2230                                         global_lru_pages());
2231                         sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2232                         if (sc.nr_reclaimed >= nr_pages)
2233                                 goto out;
2234
2235                         if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
2236                                 congestion_wait(WRITE, HZ / 10);
2237                 }
2238         }
2239
2240         /*
2241          * If sc.nr_reclaimed = 0, we could not shrink LRUs, but there may be
2242          * something in slab caches
2243          */
2244         if (!sc.nr_reclaimed) {
2245                 do {
2246                         reclaim_state.reclaimed_slab = 0;
2247                         shrink_slab(nr_pages, sc.gfp_mask, global_lru_pages());
2248                         sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2249                 } while (sc.nr_reclaimed < nr_pages &&
2250                                 reclaim_state.reclaimed_slab > 0);
2251         }
2252
2253
2254 out:
2255         current->reclaim_state = NULL;
2256
2257         return sc.nr_reclaimed;
2258 }
2259 #endif /* CONFIG_HIBERNATION */
2260
2261 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2262    not required for correctness.  So if the last cpu in a node goes
2263    away, we get changed to run anywhere: as the first one comes back,
2264    restore their cpu bindings. */
2265 static int __devinit cpu_callback(struct notifier_block *nfb,
2266                                   unsigned long action, void *hcpu)
2267 {
2268         int nid;
2269
2270         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2271                 for_each_node_state(nid, N_HIGH_MEMORY) {
2272                         pg_data_t *pgdat = NODE_DATA(nid);
2273                         const struct cpumask *mask;
2274
2275                         mask = cpumask_of_node(pgdat->node_id);
2276
2277                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2278                                 /* One of our CPUs online: restore mask */
2279                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2280                 }
2281         }
2282         return NOTIFY_OK;
2283 }
2284
2285 /*
2286  * This kswapd start function will be called by init and node-hot-add.
2287  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2288  */
2289 int kswapd_run(int nid)
2290 {
2291         pg_data_t *pgdat = NODE_DATA(nid);
2292         int ret = 0;
2293
2294         if (pgdat->kswapd)
2295                 return 0;
2296
2297         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2298         if (IS_ERR(pgdat->kswapd)) {
2299                 /* failure at boot is fatal */
2300                 BUG_ON(system_state == SYSTEM_BOOTING);
2301                 printk("Failed to start kswapd on node %d\n",nid);
2302                 ret = -1;
2303         }
2304         return ret;
2305 }
2306
2307 static int __init kswapd_init(void)
2308 {
2309         int nid;
2310
2311         swap_setup();
2312         for_each_node_state(nid, N_HIGH_MEMORY)
2313                 kswapd_run(nid);
2314         hotcpu_notifier(cpu_callback, 0);
2315         return 0;
2316 }
2317
2318 module_init(kswapd_init)
2319
2320 #ifdef CONFIG_NUMA
2321 /*
2322  * Zone reclaim mode
2323  *
2324  * If non-zero call zone_reclaim when the number of free pages falls below
2325  * the watermarks.
2326  */
2327 int zone_reclaim_mode __read_mostly;
2328
2329 #define RECLAIM_OFF 0
2330 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
2331 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
2332 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
2333
2334 /*
2335  * Priority for ZONE_RECLAIM. This determines the fraction of pages
2336  * of a node considered for each zone_reclaim. 4 scans 1/16th of
2337  * a zone.
2338  */
2339 #define ZONE_RECLAIM_PRIORITY 4
2340
2341 /*
2342  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2343  * occur.
2344  */
2345 int sysctl_min_unmapped_ratio = 1;
2346
2347 /*
2348  * If the number of slab pages in a zone grows beyond this percentage then
2349  * slab reclaim needs to occur.
2350  */
2351 int sysctl_min_slab_ratio = 5;
2352
2353 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
2354 {
2355         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
2356         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
2357                 zone_page_state(zone, NR_ACTIVE_FILE);
2358
2359         /*
2360          * It's possible for there to be more file mapped pages than
2361          * accounted for by the pages on the file LRU lists because
2362          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2363          */
2364         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
2365 }
2366
2367 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
2368 static long zone_pagecache_reclaimable(struct zone *zone)
2369 {
2370         long nr_pagecache_reclaimable;
2371         long delta = 0;
2372
2373         /*
2374          * If RECLAIM_SWAP is set, then all file pages are considered
2375          * potentially reclaimable. Otherwise, we have to worry about
2376          * pages like swapcache and zone_unmapped_file_pages() provides
2377          * a better estimate
2378          */
2379         if (zone_reclaim_mode & RECLAIM_SWAP)
2380                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
2381         else
2382                 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
2383
2384         /* If we can't clean pages, remove dirty pages from consideration */
2385         if (!(zone_reclaim_mode & RECLAIM_WRITE))
2386                 delta += zone_page_state(zone, NR_FILE_DIRTY);
2387
2388         /* Watch for any possible underflows due to delta */
2389         if (unlikely(delta > nr_pagecache_reclaimable))
2390                 delta = nr_pagecache_reclaimable;
2391
2392         return nr_pagecache_reclaimable - delta;
2393 }
2394
2395 /*
2396  * Try to free up some pages from this zone through reclaim.
2397  */
2398 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2399 {
2400         /* Minimum pages needed in order to stay on node */
2401         const unsigned long nr_pages = 1 << order;
2402         struct task_struct *p = current;
2403         struct reclaim_state reclaim_state;
2404         int priority;
2405         struct scan_control sc = {
2406                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2407                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2408                 .may_swap = 1,
2409                 .swap_cluster_max = max_t(unsigned long, nr_pages,
2410                                         SWAP_CLUSTER_MAX),
2411                 .gfp_mask = gfp_mask,
2412                 .swappiness = vm_swappiness,
2413                 .order = order,
2414                 .isolate_pages = isolate_pages_global,
2415         };
2416         unsigned long slab_reclaimable;
2417
2418         disable_swap_token();
2419         cond_resched();
2420         /*
2421          * We need to be able to allocate from the reserves for RECLAIM_SWAP
2422          * and we also need to be able to write out pages for RECLAIM_WRITE
2423          * and RECLAIM_SWAP.
2424          */
2425         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2426         reclaim_state.reclaimed_slab = 0;
2427         p->reclaim_state = &reclaim_state;
2428
2429         if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
2430                 /*
2431                  * Free memory by calling shrink zone with increasing
2432                  * priorities until we have enough memory freed.
2433                  */
2434                 priority = ZONE_RECLAIM_PRIORITY;
2435                 do {
2436                         note_zone_scanning_priority(zone, priority);
2437                         shrink_zone(priority, zone, &sc);
2438                         priority--;
2439                 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
2440         }
2441
2442         slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2443         if (slab_reclaimable > zone->min_slab_pages) {
2444                 /*
2445                  * shrink_slab() does not currently allow us to determine how
2446                  * many pages were freed in this zone. So we take the current
2447                  * number of slab pages and shake the slab until it is reduced
2448                  * by the same nr_pages that we used for reclaiming unmapped
2449                  * pages.
2450                  *
2451                  * Note that shrink_slab will free memory on all zones and may
2452                  * take a long time.
2453                  */
2454                 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
2455                         zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
2456                                 slab_reclaimable - nr_pages)
2457                         ;
2458
2459                 /*
2460                  * Update nr_reclaimed by the number of slab pages we
2461                  * reclaimed from this zone.
2462                  */
2463                 sc.nr_reclaimed += slab_reclaimable -
2464                         zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2465         }
2466
2467         p->reclaim_state = NULL;
2468         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
2469         return sc.nr_reclaimed >= nr_pages;
2470 }
2471
2472 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2473 {
2474         int node_id;
2475         int ret;
2476
2477         /*
2478          * Zone reclaim reclaims unmapped file backed pages and
2479          * slab pages if we are over the defined limits.
2480          *
2481          * A small portion of unmapped file backed pages is needed for
2482          * file I/O otherwise pages read by file I/O will be immediately
2483          * thrown out if the zone is overallocated. So we do not reclaim
2484          * if less than a specified percentage of the zone is used by
2485          * unmapped file backed pages.
2486          */
2487         if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
2488             zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
2489                 return ZONE_RECLAIM_FULL;
2490
2491         if (zone_is_all_unreclaimable(zone))
2492                 return ZONE_RECLAIM_FULL;
2493
2494         /*
2495          * Do not scan if the allocation should not be delayed.
2496          */
2497         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
2498                 return ZONE_RECLAIM_NOSCAN;
2499
2500         /*
2501          * Only run zone reclaim on the local zone or on zones that do not
2502          * have associated processors. This will favor the local processor
2503          * over remote processors and spread off node memory allocations
2504          * as wide as possible.
2505          */
2506         node_id = zone_to_nid(zone);
2507         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
2508                 return ZONE_RECLAIM_NOSCAN;
2509
2510         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2511                 return ZONE_RECLAIM_NOSCAN;
2512
2513         ret = __zone_reclaim(zone, gfp_mask, order);
2514         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2515
2516         if (!ret)
2517                 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
2518
2519         return ret;
2520 }
2521 #endif
2522
2523 /*
2524  * page_evictable - test whether a page is evictable
2525  * @page: the page to test
2526  * @vma: the VMA in which the page is or will be mapped, may be NULL
2527  *
2528  * Test whether page is evictable--i.e., should be placed on active/inactive
2529  * lists vs unevictable list.  The vma argument is !NULL when called from the
2530  * fault path to determine how to instantate a new page.
2531  *
2532  * Reasons page might not be evictable:
2533  * (1) page's mapping marked unevictable
2534  * (2) page is part of an mlocked VMA
2535  *
2536  */
2537 int page_evictable(struct page *page, struct vm_area_struct *vma)
2538 {
2539
2540         if (mapping_unevictable(page_mapping(page)))
2541                 return 0;
2542
2543         if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2544                 return 0;
2545
2546         return 1;
2547 }
2548
2549 /**
2550  * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2551  * @page: page to check evictability and move to appropriate lru list
2552  * @zone: zone page is in
2553  *
2554  * Checks a page for evictability and moves the page to the appropriate
2555  * zone lru list.
2556  *
2557  * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2558  * have PageUnevictable set.
2559  */
2560 static void check_move_unevictable_page(struct page *page, struct zone *zone)
2561 {
2562         VM_BUG_ON(PageActive(page));
2563
2564 retry:
2565         ClearPageUnevictable(page);
2566         if (page_evictable(page, NULL)) {
2567                 enum lru_list l = LRU_INACTIVE_ANON + page_is_file_cache(page);
2568
2569                 __dec_zone_state(zone, NR_UNEVICTABLE);
2570                 list_move(&page->lru, &zone->lru[l].list);
2571                 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
2572                 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
2573                 __count_vm_event(UNEVICTABLE_PGRESCUED);
2574         } else {
2575                 /*
2576                  * rotate unevictable list
2577                  */
2578                 SetPageUnevictable(page);
2579                 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
2580                 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
2581                 if (page_evictable(page, NULL))
2582                         goto retry;
2583         }
2584 }
2585
2586 /**
2587  * scan_mapping_unevictable_pages - scan an address space for evictable pages
2588  * @mapping: struct address_space to scan for evictable pages
2589  *
2590  * Scan all pages in mapping.  Check unevictable pages for
2591  * evictability and move them to the appropriate zone lru list.
2592  */
2593 void scan_mapping_unevictable_pages(struct address_space *mapping)
2594 {
2595         pgoff_t next = 0;
2596         pgoff_t end   = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
2597                          PAGE_CACHE_SHIFT;
2598         struct zone *zone;
2599         struct pagevec pvec;
2600
2601         if (mapping->nrpages == 0)
2602                 return;
2603
2604         pagevec_init(&pvec, 0);
2605         while (next < end &&
2606                 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
2607                 int i;
2608                 int pg_scanned = 0;
2609
2610                 zone = NULL;
2611
2612                 for (i = 0; i < pagevec_count(&pvec); i++) {
2613                         struct page *page = pvec.pages[i];
2614                         pgoff_t page_index = page->index;
2615                         struct zone *pagezone = page_zone(page);
2616
2617                         pg_scanned++;
2618                         if (page_index > next)
2619                                 next = page_index;
2620                         next++;
2621
2622                         if (pagezone != zone) {
2623                                 if (zone)
2624                                         spin_unlock_irq(&zone->lru_lock);
2625                                 zone = pagezone;
2626                                 spin_lock_irq(&zone->lru_lock);
2627                         }
2628
2629                         if (PageLRU(page) && PageUnevictable(page))
2630                                 check_move_unevictable_page(page, zone);
2631                 }
2632                 if (zone)
2633                         spin_unlock_irq(&zone->lru_lock);
2634                 pagevec_release(&pvec);
2635
2636                 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
2637         }
2638
2639 }
2640
2641 /**
2642  * scan_zone_unevictable_pages - check unevictable list for evictable pages
2643  * @zone - zone of which to scan the unevictable list
2644  *
2645  * Scan @zone's unevictable LRU lists to check for pages that have become
2646  * evictable.  Move those that have to @zone's inactive list where they
2647  * become candidates for reclaim, unless shrink_inactive_zone() decides
2648  * to reactivate them.  Pages that are still unevictable are rotated
2649  * back onto @zone's unevictable list.
2650  */
2651 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
2652 static void scan_zone_unevictable_pages(struct zone *zone)
2653 {
2654         struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
2655         unsigned long scan;
2656         unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
2657
2658         while (nr_to_scan > 0) {
2659                 unsigned long batch_size = min(nr_to_scan,
2660                                                 SCAN_UNEVICTABLE_BATCH_SIZE);
2661
2662                 spin_lock_irq(&zone->lru_lock);
2663                 for (scan = 0;  scan < batch_size; scan++) {
2664                         struct page *page = lru_to_page(l_unevictable);
2665
2666                         if (!trylock_page(page))
2667                                 continue;
2668
2669                         prefetchw_prev_lru_page(page, l_unevictable, flags);
2670
2671                         if (likely(PageLRU(page) && PageUnevictable(page)))
2672                                 check_move_unevictable_page(page, zone);
2673
2674                         unlock_page(page);
2675                 }
2676                 spin_unlock_irq(&zone->lru_lock);
2677
2678                 nr_to_scan -= batch_size;
2679         }
2680 }
2681
2682
2683 /**
2684  * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
2685  *
2686  * A really big hammer:  scan all zones' unevictable LRU lists to check for
2687  * pages that have become evictable.  Move those back to the zones'
2688  * inactive list where they become candidates for reclaim.
2689  * This occurs when, e.g., we have unswappable pages on the unevictable lists,
2690  * and we add swap to the system.  As such, it runs in the context of a task
2691  * that has possibly/probably made some previously unevictable pages
2692  * evictable.
2693  */
2694 static void scan_all_zones_unevictable_pages(void)
2695 {
2696         struct zone *zone;
2697
2698         for_each_zone(zone) {
2699                 scan_zone_unevictable_pages(zone);
2700         }
2701 }
2702
2703 /*
2704  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
2705  * all nodes' unevictable lists for evictable pages
2706  */
2707 unsigned long scan_unevictable_pages;
2708
2709 int scan_unevictable_handler(struct ctl_table *table, int write,
2710                            struct file *file, void __user *buffer,
2711                            size_t *length, loff_t *ppos)
2712 {
2713         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
2714
2715         if (write && *(unsigned long *)table->data)
2716                 scan_all_zones_unevictable_pages();
2717
2718         scan_unevictable_pages = 0;
2719         return 0;
2720 }
2721
2722 /*
2723  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
2724  * a specified node's per zone unevictable lists for evictable pages.
2725  */
2726
2727 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
2728                                           struct sysdev_attribute *attr,
2729                                           char *buf)
2730 {
2731         return sprintf(buf, "0\n");     /* always zero; should fit... */
2732 }
2733
2734 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
2735                                            struct sysdev_attribute *attr,
2736                                         const char *buf, size_t count)
2737 {
2738         struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
2739         struct zone *zone;
2740         unsigned long res;
2741         unsigned long req = strict_strtoul(buf, 10, &res);
2742
2743         if (!req)
2744                 return 1;       /* zero is no-op */
2745
2746         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2747                 if (!populated_zone(zone))
2748                         continue;
2749                 scan_zone_unevictable_pages(zone);
2750         }
2751         return 1;
2752 }
2753
2754
2755 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
2756                         read_scan_unevictable_node,
2757                         write_scan_unevictable_node);
2758
2759 int scan_unevictable_register_node(struct node *node)
2760 {
2761         return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
2762 }
2763
2764 void scan_unevictable_unregister_node(struct node *node)
2765 {
2766         sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
2767 }
2768