vmscan: properly account for the number of page cache pages zone_reclaim() can reclaim
[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                 mem_cgroup_del_lru(page);
841         }
842
843         return ret;
844 }
845
846 /*
847  * zone->lru_lock is heavily contended.  Some of the functions that
848  * shrink the lists perform better by taking out a batch of pages
849  * and working on them outside the LRU lock.
850  *
851  * For pagecache intensive workloads, this function is the hottest
852  * spot in the kernel (apart from copy_*_user functions).
853  *
854  * Appropriate locks must be held before calling this function.
855  *
856  * @nr_to_scan: The number of pages to look through on the list.
857  * @src:        The LRU list to pull pages off.
858  * @dst:        The temp list to put pages on to.
859  * @scanned:    The number of pages that were scanned.
860  * @order:      The caller's attempted allocation order
861  * @mode:       One of the LRU isolation modes
862  * @file:       True [1] if isolating file [!anon] pages
863  *
864  * returns how many pages were moved onto *@dst.
865  */
866 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
867                 struct list_head *src, struct list_head *dst,
868                 unsigned long *scanned, int order, int mode, int file)
869 {
870         unsigned long nr_taken = 0;
871         unsigned long scan;
872
873         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
874                 struct page *page;
875                 unsigned long pfn;
876                 unsigned long end_pfn;
877                 unsigned long page_pfn;
878                 int zone_id;
879
880                 page = lru_to_page(src);
881                 prefetchw_prev_lru_page(page, src, flags);
882
883                 VM_BUG_ON(!PageLRU(page));
884
885                 switch (__isolate_lru_page(page, mode, file)) {
886                 case 0:
887                         list_move(&page->lru, dst);
888                         nr_taken++;
889                         break;
890
891                 case -EBUSY:
892                         /* else it is being freed elsewhere */
893                         list_move(&page->lru, src);
894                         continue;
895
896                 default:
897                         BUG();
898                 }
899
900                 if (!order)
901                         continue;
902
903                 /*
904                  * Attempt to take all pages in the order aligned region
905                  * surrounding the tag page.  Only take those pages of
906                  * the same active state as that tag page.  We may safely
907                  * round the target page pfn down to the requested order
908                  * as the mem_map is guarenteed valid out to MAX_ORDER,
909                  * where that page is in a different zone we will detect
910                  * it from its zone id and abort this block scan.
911                  */
912                 zone_id = page_zone_id(page);
913                 page_pfn = page_to_pfn(page);
914                 pfn = page_pfn & ~((1 << order) - 1);
915                 end_pfn = pfn + (1 << order);
916                 for (; pfn < end_pfn; pfn++) {
917                         struct page *cursor_page;
918
919                         /* The target page is in the block, ignore it. */
920                         if (unlikely(pfn == page_pfn))
921                                 continue;
922
923                         /* Avoid holes within the zone. */
924                         if (unlikely(!pfn_valid_within(pfn)))
925                                 break;
926
927                         cursor_page = pfn_to_page(pfn);
928
929                         /* Check that we have not crossed a zone boundary. */
930                         if (unlikely(page_zone_id(cursor_page) != zone_id))
931                                 continue;
932                         switch (__isolate_lru_page(cursor_page, mode, file)) {
933                         case 0:
934                                 list_move(&cursor_page->lru, dst);
935                                 nr_taken++;
936                                 scan++;
937                                 break;
938
939                         case -EBUSY:
940                                 /* else it is being freed elsewhere */
941                                 list_move(&cursor_page->lru, src);
942                         default:
943                                 break;  /* ! on LRU or wrong list */
944                         }
945                 }
946         }
947
948         *scanned = scan;
949         return nr_taken;
950 }
951
952 static unsigned long isolate_pages_global(unsigned long nr,
953                                         struct list_head *dst,
954                                         unsigned long *scanned, int order,
955                                         int mode, struct zone *z,
956                                         struct mem_cgroup *mem_cont,
957                                         int active, int file)
958 {
959         int lru = LRU_BASE;
960         if (active)
961                 lru += LRU_ACTIVE;
962         if (file)
963                 lru += LRU_FILE;
964         return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
965                                                                 mode, !!file);
966 }
967
968 /*
969  * clear_active_flags() is a helper for shrink_active_list(), clearing
970  * any active bits from the pages in the list.
971  */
972 static unsigned long clear_active_flags(struct list_head *page_list,
973                                         unsigned int *count)
974 {
975         int nr_active = 0;
976         int lru;
977         struct page *page;
978
979         list_for_each_entry(page, page_list, lru) {
980                 lru = page_is_file_cache(page);
981                 if (PageActive(page)) {
982                         lru += LRU_ACTIVE;
983                         ClearPageActive(page);
984                         nr_active++;
985                 }
986                 count[lru]++;
987         }
988
989         return nr_active;
990 }
991
992 /**
993  * isolate_lru_page - tries to isolate a page from its LRU list
994  * @page: page to isolate from its LRU list
995  *
996  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
997  * vmstat statistic corresponding to whatever LRU list the page was on.
998  *
999  * Returns 0 if the page was removed from an LRU list.
1000  * Returns -EBUSY if the page was not on an LRU list.
1001  *
1002  * The returned page will have PageLRU() cleared.  If it was found on
1003  * the active list, it will have PageActive set.  If it was found on
1004  * the unevictable list, it will have the PageUnevictable bit set. That flag
1005  * may need to be cleared by the caller before letting the page go.
1006  *
1007  * The vmstat statistic corresponding to the list on which the page was
1008  * found will be decremented.
1009  *
1010  * Restrictions:
1011  * (1) Must be called with an elevated refcount on the page. This is a
1012  *     fundamentnal difference from isolate_lru_pages (which is called
1013  *     without a stable reference).
1014  * (2) the lru_lock must not be held.
1015  * (3) interrupts must be enabled.
1016  */
1017 int isolate_lru_page(struct page *page)
1018 {
1019         int ret = -EBUSY;
1020
1021         if (PageLRU(page)) {
1022                 struct zone *zone = page_zone(page);
1023
1024                 spin_lock_irq(&zone->lru_lock);
1025                 if (PageLRU(page) && get_page_unless_zero(page)) {
1026                         int lru = page_lru(page);
1027                         ret = 0;
1028                         ClearPageLRU(page);
1029
1030                         del_page_from_lru_list(zone, page, lru);
1031                 }
1032                 spin_unlock_irq(&zone->lru_lock);
1033         }
1034         return ret;
1035 }
1036
1037 /*
1038  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1039  * of reclaimed pages
1040  */
1041 static unsigned long shrink_inactive_list(unsigned long max_scan,
1042                         struct zone *zone, struct scan_control *sc,
1043                         int priority, int file)
1044 {
1045         LIST_HEAD(page_list);
1046         struct pagevec pvec;
1047         unsigned long nr_scanned = 0;
1048         unsigned long nr_reclaimed = 0;
1049         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1050         int lumpy_reclaim = 0;
1051
1052         /*
1053          * If we need a large contiguous chunk of memory, or have
1054          * trouble getting a small set of contiguous pages, we
1055          * will reclaim both active and inactive pages.
1056          *
1057          * We use the same threshold as pageout congestion_wait below.
1058          */
1059         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1060                 lumpy_reclaim = 1;
1061         else if (sc->order && priority < DEF_PRIORITY - 2)
1062                 lumpy_reclaim = 1;
1063
1064         pagevec_init(&pvec, 1);
1065
1066         lru_add_drain();
1067         spin_lock_irq(&zone->lru_lock);
1068         do {
1069                 struct page *page;
1070                 unsigned long nr_taken;
1071                 unsigned long nr_scan;
1072                 unsigned long nr_freed;
1073                 unsigned long nr_active;
1074                 unsigned int count[NR_LRU_LISTS] = { 0, };
1075                 int mode = lumpy_reclaim ? ISOLATE_BOTH : ISOLATE_INACTIVE;
1076
1077                 nr_taken = sc->isolate_pages(sc->swap_cluster_max,
1078                              &page_list, &nr_scan, sc->order, mode,
1079                                 zone, sc->mem_cgroup, 0, file);
1080                 nr_active = clear_active_flags(&page_list, count);
1081                 __count_vm_events(PGDEACTIVATE, nr_active);
1082
1083                 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1084                                                 -count[LRU_ACTIVE_FILE]);
1085                 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1086                                                 -count[LRU_INACTIVE_FILE]);
1087                 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1088                                                 -count[LRU_ACTIVE_ANON]);
1089                 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1090                                                 -count[LRU_INACTIVE_ANON]);
1091
1092                 if (scanning_global_lru(sc))
1093                         zone->pages_scanned += nr_scan;
1094
1095                 reclaim_stat->recent_scanned[0] += count[LRU_INACTIVE_ANON];
1096                 reclaim_stat->recent_scanned[0] += count[LRU_ACTIVE_ANON];
1097                 reclaim_stat->recent_scanned[1] += count[LRU_INACTIVE_FILE];
1098                 reclaim_stat->recent_scanned[1] += count[LRU_ACTIVE_FILE];
1099
1100                 spin_unlock_irq(&zone->lru_lock);
1101
1102                 nr_scanned += nr_scan;
1103                 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
1104
1105                 /*
1106                  * If we are direct reclaiming for contiguous pages and we do
1107                  * not reclaim everything in the list, try again and wait
1108                  * for IO to complete. This will stall high-order allocations
1109                  * but that should be acceptable to the caller
1110                  */
1111                 if (nr_freed < nr_taken && !current_is_kswapd() &&
1112                     lumpy_reclaim) {
1113                         congestion_wait(WRITE, HZ/10);
1114
1115                         /*
1116                          * The attempt at page out may have made some
1117                          * of the pages active, mark them inactive again.
1118                          */
1119                         nr_active = clear_active_flags(&page_list, count);
1120                         count_vm_events(PGDEACTIVATE, nr_active);
1121
1122                         nr_freed += shrink_page_list(&page_list, sc,
1123                                                         PAGEOUT_IO_SYNC);
1124                 }
1125
1126                 nr_reclaimed += nr_freed;
1127                 local_irq_disable();
1128                 if (current_is_kswapd()) {
1129                         __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scan);
1130                         __count_vm_events(KSWAPD_STEAL, nr_freed);
1131                 } else if (scanning_global_lru(sc))
1132                         __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scan);
1133
1134                 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
1135
1136                 if (nr_taken == 0)
1137                         goto done;
1138
1139                 spin_lock(&zone->lru_lock);
1140                 /*
1141                  * Put back any unfreeable pages.
1142                  */
1143                 while (!list_empty(&page_list)) {
1144                         int lru;
1145                         page = lru_to_page(&page_list);
1146                         VM_BUG_ON(PageLRU(page));
1147                         list_del(&page->lru);
1148                         if (unlikely(!page_evictable(page, NULL))) {
1149                                 spin_unlock_irq(&zone->lru_lock);
1150                                 putback_lru_page(page);
1151                                 spin_lock_irq(&zone->lru_lock);
1152                                 continue;
1153                         }
1154                         SetPageLRU(page);
1155                         lru = page_lru(page);
1156                         add_page_to_lru_list(zone, page, lru);
1157                         if (PageActive(page)) {
1158                                 int file = !!page_is_file_cache(page);
1159                                 reclaim_stat->recent_rotated[file]++;
1160                         }
1161                         if (!pagevec_add(&pvec, page)) {
1162                                 spin_unlock_irq(&zone->lru_lock);
1163                                 __pagevec_release(&pvec);
1164                                 spin_lock_irq(&zone->lru_lock);
1165                         }
1166                 }
1167         } while (nr_scanned < max_scan);
1168         spin_unlock(&zone->lru_lock);
1169 done:
1170         local_irq_enable();
1171         pagevec_release(&pvec);
1172         return nr_reclaimed;
1173 }
1174
1175 /*
1176  * We are about to scan this zone at a certain priority level.  If that priority
1177  * level is smaller (ie: more urgent) than the previous priority, then note
1178  * that priority level within the zone.  This is done so that when the next
1179  * process comes in to scan this zone, it will immediately start out at this
1180  * priority level rather than having to build up its own scanning priority.
1181  * Here, this priority affects only the reclaim-mapped threshold.
1182  */
1183 static inline void note_zone_scanning_priority(struct zone *zone, int priority)
1184 {
1185         if (priority < zone->prev_priority)
1186                 zone->prev_priority = priority;
1187 }
1188
1189 /*
1190  * This moves pages from the active list to the inactive list.
1191  *
1192  * We move them the other way if the page is referenced by one or more
1193  * processes, from rmap.
1194  *
1195  * If the pages are mostly unmapped, the processing is fast and it is
1196  * appropriate to hold zone->lru_lock across the whole operation.  But if
1197  * the pages are mapped, the processing is slow (page_referenced()) so we
1198  * should drop zone->lru_lock around each page.  It's impossible to balance
1199  * this, so instead we remove the pages from the LRU while processing them.
1200  * It is safe to rely on PG_active against the non-LRU pages in here because
1201  * nobody will play with that bit on a non-LRU page.
1202  *
1203  * The downside is that we have to touch page->_count against each page.
1204  * But we had to alter page->flags anyway.
1205  */
1206
1207 static void move_active_pages_to_lru(struct zone *zone,
1208                                      struct list_head *list,
1209                                      enum lru_list lru)
1210 {
1211         unsigned long pgmoved = 0;
1212         struct pagevec pvec;
1213         struct page *page;
1214
1215         pagevec_init(&pvec, 1);
1216
1217         while (!list_empty(list)) {
1218                 page = lru_to_page(list);
1219                 prefetchw_prev_lru_page(page, list, flags);
1220
1221                 VM_BUG_ON(PageLRU(page));
1222                 SetPageLRU(page);
1223
1224                 VM_BUG_ON(!PageActive(page));
1225                 if (!is_active_lru(lru))
1226                         ClearPageActive(page);  /* we are de-activating */
1227
1228                 list_move(&page->lru, &zone->lru[lru].list);
1229                 mem_cgroup_add_lru_list(page, lru);
1230                 pgmoved++;
1231
1232                 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1233                         spin_unlock_irq(&zone->lru_lock);
1234                         if (buffer_heads_over_limit)
1235                                 pagevec_strip(&pvec);
1236                         __pagevec_release(&pvec);
1237                         spin_lock_irq(&zone->lru_lock);
1238                 }
1239         }
1240         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1241         if (!is_active_lru(lru))
1242                 __count_vm_events(PGDEACTIVATE, pgmoved);
1243 }
1244
1245 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1246                         struct scan_control *sc, int priority, int file)
1247 {
1248         unsigned long pgmoved;
1249         unsigned long pgscanned;
1250         unsigned long vm_flags;
1251         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1252         LIST_HEAD(l_active);
1253         LIST_HEAD(l_inactive);
1254         struct page *page;
1255         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1256
1257         lru_add_drain();
1258         spin_lock_irq(&zone->lru_lock);
1259         pgmoved = sc->isolate_pages(nr_pages, &l_hold, &pgscanned, sc->order,
1260                                         ISOLATE_ACTIVE, zone,
1261                                         sc->mem_cgroup, 1, file);
1262         /*
1263          * zone->pages_scanned is used for detect zone's oom
1264          * mem_cgroup remembers nr_scan by itself.
1265          */
1266         if (scanning_global_lru(sc)) {
1267                 zone->pages_scanned += pgscanned;
1268         }
1269         reclaim_stat->recent_scanned[!!file] += pgmoved;
1270
1271         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1272         if (file)
1273                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -pgmoved);
1274         else
1275                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -pgmoved);
1276         spin_unlock_irq(&zone->lru_lock);
1277
1278         pgmoved = 0;  /* count referenced (mapping) mapped pages */
1279         while (!list_empty(&l_hold)) {
1280                 cond_resched();
1281                 page = lru_to_page(&l_hold);
1282                 list_del(&page->lru);
1283
1284                 if (unlikely(!page_evictable(page, NULL))) {
1285                         putback_lru_page(page);
1286                         continue;
1287                 }
1288
1289                 /* page_referenced clears PageReferenced */
1290                 if (page_mapping_inuse(page) &&
1291                     page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1292                         pgmoved++;
1293                         /*
1294                          * Identify referenced, file-backed active pages and
1295                          * give them one more trip around the active list. So
1296                          * that executable code get better chances to stay in
1297                          * memory under moderate memory pressure.  Anon pages
1298                          * are not likely to be evicted by use-once streaming
1299                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1300                          * so we ignore them here.
1301                          */
1302                         if ((vm_flags & VM_EXEC) && !PageAnon(page)) {
1303                                 list_add(&page->lru, &l_active);
1304                                 continue;
1305                         }
1306                 }
1307
1308                 list_add(&page->lru, &l_inactive);
1309         }
1310
1311         /*
1312          * Move pages back to the lru list.
1313          */
1314         spin_lock_irq(&zone->lru_lock);
1315         /*
1316          * Count referenced pages from currently used mappings as rotated,
1317          * even though only some of them are actually re-activated.  This
1318          * helps balance scan pressure between file and anonymous pages in
1319          * get_scan_ratio.
1320          */
1321         reclaim_stat->recent_rotated[!!file] += pgmoved;
1322
1323         move_active_pages_to_lru(zone, &l_active,
1324                                                 LRU_ACTIVE + file * LRU_FILE);
1325         move_active_pages_to_lru(zone, &l_inactive,
1326                                                 LRU_BASE   + file * LRU_FILE);
1327
1328         spin_unlock_irq(&zone->lru_lock);
1329 }
1330
1331 static int inactive_anon_is_low_global(struct zone *zone)
1332 {
1333         unsigned long active, inactive;
1334
1335         active = zone_page_state(zone, NR_ACTIVE_ANON);
1336         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1337
1338         if (inactive * zone->inactive_ratio < active)
1339                 return 1;
1340
1341         return 0;
1342 }
1343
1344 /**
1345  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1346  * @zone: zone to check
1347  * @sc:   scan control of this context
1348  *
1349  * Returns true if the zone does not have enough inactive anon pages,
1350  * meaning some active anon pages need to be deactivated.
1351  */
1352 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1353 {
1354         int low;
1355
1356         if (scanning_global_lru(sc))
1357                 low = inactive_anon_is_low_global(zone);
1358         else
1359                 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1360         return low;
1361 }
1362
1363 static int inactive_file_is_low_global(struct zone *zone)
1364 {
1365         unsigned long active, inactive;
1366
1367         active = zone_page_state(zone, NR_ACTIVE_FILE);
1368         inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1369
1370         return (active > inactive);
1371 }
1372
1373 /**
1374  * inactive_file_is_low - check if file pages need to be deactivated
1375  * @zone: zone to check
1376  * @sc:   scan control of this context
1377  *
1378  * When the system is doing streaming IO, memory pressure here
1379  * ensures that active file pages get deactivated, until more
1380  * than half of the file pages are on the inactive list.
1381  *
1382  * Once we get to that situation, protect the system's working
1383  * set from being evicted by disabling active file page aging.
1384  *
1385  * This uses a different ratio than the anonymous pages, because
1386  * the page cache uses a use-once replacement algorithm.
1387  */
1388 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1389 {
1390         int low;
1391
1392         if (scanning_global_lru(sc))
1393                 low = inactive_file_is_low_global(zone);
1394         else
1395                 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1396         return low;
1397 }
1398
1399 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1400         struct zone *zone, struct scan_control *sc, int priority)
1401 {
1402         int file = is_file_lru(lru);
1403
1404         if (lru == LRU_ACTIVE_FILE && inactive_file_is_low(zone, sc)) {
1405                 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1406                 return 0;
1407         }
1408
1409         if (lru == LRU_ACTIVE_ANON && inactive_anon_is_low(zone, sc)) {
1410                 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1411                 return 0;
1412         }
1413         return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1414 }
1415
1416 /*
1417  * Determine how aggressively the anon and file LRU lists should be
1418  * scanned.  The relative value of each set of LRU lists is determined
1419  * by looking at the fraction of the pages scanned we did rotate back
1420  * onto the active list instead of evict.
1421  *
1422  * percent[0] specifies how much pressure to put on ram/swap backed
1423  * memory, while percent[1] determines pressure on the file LRUs.
1424  */
1425 static void get_scan_ratio(struct zone *zone, struct scan_control *sc,
1426                                         unsigned long *percent)
1427 {
1428         unsigned long anon, file, free;
1429         unsigned long anon_prio, file_prio;
1430         unsigned long ap, fp;
1431         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1432
1433         anon  = zone_nr_pages(zone, sc, LRU_ACTIVE_ANON) +
1434                 zone_nr_pages(zone, sc, LRU_INACTIVE_ANON);
1435         file  = zone_nr_pages(zone, sc, LRU_ACTIVE_FILE) +
1436                 zone_nr_pages(zone, sc, LRU_INACTIVE_FILE);
1437
1438         if (scanning_global_lru(sc)) {
1439                 free  = zone_page_state(zone, NR_FREE_PAGES);
1440                 /* If we have very few page cache pages,
1441                    force-scan anon pages. */
1442                 if (unlikely(file + free <= high_wmark_pages(zone))) {
1443                         percent[0] = 100;
1444                         percent[1] = 0;
1445                         return;
1446                 }
1447         }
1448
1449         /*
1450          * OK, so we have swap space and a fair amount of page cache
1451          * pages.  We use the recently rotated / recently scanned
1452          * ratios to determine how valuable each cache is.
1453          *
1454          * Because workloads change over time (and to avoid overflow)
1455          * we keep these statistics as a floating average, which ends
1456          * up weighing recent references more than old ones.
1457          *
1458          * anon in [0], file in [1]
1459          */
1460         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1461                 spin_lock_irq(&zone->lru_lock);
1462                 reclaim_stat->recent_scanned[0] /= 2;
1463                 reclaim_stat->recent_rotated[0] /= 2;
1464                 spin_unlock_irq(&zone->lru_lock);
1465         }
1466
1467         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1468                 spin_lock_irq(&zone->lru_lock);
1469                 reclaim_stat->recent_scanned[1] /= 2;
1470                 reclaim_stat->recent_rotated[1] /= 2;
1471                 spin_unlock_irq(&zone->lru_lock);
1472         }
1473
1474         /*
1475          * With swappiness at 100, anonymous and file have the same priority.
1476          * This scanning priority is essentially the inverse of IO cost.
1477          */
1478         anon_prio = sc->swappiness;
1479         file_prio = 200 - sc->swappiness;
1480
1481         /*
1482          * The amount of pressure on anon vs file pages is inversely
1483          * proportional to the fraction of recently scanned pages on
1484          * each list that were recently referenced and in active use.
1485          */
1486         ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1487         ap /= reclaim_stat->recent_rotated[0] + 1;
1488
1489         fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1490         fp /= reclaim_stat->recent_rotated[1] + 1;
1491
1492         /* Normalize to percentages */
1493         percent[0] = 100 * ap / (ap + fp + 1);
1494         percent[1] = 100 - percent[0];
1495 }
1496
1497 /*
1498  * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1499  * until we collected @swap_cluster_max pages to scan.
1500  */
1501 static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
1502                                        unsigned long *nr_saved_scan,
1503                                        unsigned long swap_cluster_max)
1504 {
1505         unsigned long nr;
1506
1507         *nr_saved_scan += nr_to_scan;
1508         nr = *nr_saved_scan;
1509
1510         if (nr >= swap_cluster_max)
1511                 *nr_saved_scan = 0;
1512         else
1513                 nr = 0;
1514
1515         return nr;
1516 }
1517
1518 /*
1519  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1520  */
1521 static void shrink_zone(int priority, struct zone *zone,
1522                                 struct scan_control *sc)
1523 {
1524         unsigned long nr[NR_LRU_LISTS];
1525         unsigned long nr_to_scan;
1526         unsigned long percent[2];       /* anon @ 0; file @ 1 */
1527         enum lru_list l;
1528         unsigned long nr_reclaimed = sc->nr_reclaimed;
1529         unsigned long swap_cluster_max = sc->swap_cluster_max;
1530         int noswap = 0;
1531
1532         /* If we have no swap space, do not bother scanning anon pages. */
1533         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1534                 noswap = 1;
1535                 percent[0] = 0;
1536                 percent[1] = 100;
1537         } else
1538                 get_scan_ratio(zone, sc, percent);
1539
1540         for_each_evictable_lru(l) {
1541                 int file = is_file_lru(l);
1542                 unsigned long scan;
1543
1544                 scan = zone_nr_pages(zone, sc, l);
1545                 if (priority || noswap) {
1546                         scan >>= priority;
1547                         scan = (scan * percent[file]) / 100;
1548                 }
1549                 if (scanning_global_lru(sc))
1550                         nr[l] = nr_scan_try_batch(scan,
1551                                                   &zone->lru[l].nr_saved_scan,
1552                                                   swap_cluster_max);
1553                 else
1554                         nr[l] = scan;
1555         }
1556
1557         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1558                                         nr[LRU_INACTIVE_FILE]) {
1559                 for_each_evictable_lru(l) {
1560                         if (nr[l]) {
1561                                 nr_to_scan = min(nr[l], swap_cluster_max);
1562                                 nr[l] -= nr_to_scan;
1563
1564                                 nr_reclaimed += shrink_list(l, nr_to_scan,
1565                                                             zone, sc, priority);
1566                         }
1567                 }
1568                 /*
1569                  * On large memory systems, scan >> priority can become
1570                  * really large. This is fine for the starting priority;
1571                  * we want to put equal scanning pressure on each zone.
1572                  * However, if the VM has a harder time of freeing pages,
1573                  * with multiple processes reclaiming pages, the total
1574                  * freeing target can get unreasonably large.
1575                  */
1576                 if (nr_reclaimed > swap_cluster_max &&
1577                         priority < DEF_PRIORITY && !current_is_kswapd())
1578                         break;
1579         }
1580
1581         sc->nr_reclaimed = nr_reclaimed;
1582
1583         /*
1584          * Even if we did not try to evict anon pages at all, we want to
1585          * rebalance the anon lru active/inactive ratio.
1586          */
1587         if (inactive_anon_is_low(zone, sc) && nr_swap_pages > 0)
1588                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1589
1590         throttle_vm_writeout(sc->gfp_mask);
1591 }
1592
1593 /*
1594  * This is the direct reclaim path, for page-allocating processes.  We only
1595  * try to reclaim pages from zones which will satisfy the caller's allocation
1596  * request.
1597  *
1598  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1599  * Because:
1600  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1601  *    allocation or
1602  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1603  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1604  *    zone defense algorithm.
1605  *
1606  * If a zone is deemed to be full of pinned pages then just give it a light
1607  * scan then give up on it.
1608  */
1609 static void shrink_zones(int priority, struct zonelist *zonelist,
1610                                         struct scan_control *sc)
1611 {
1612         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1613         struct zoneref *z;
1614         struct zone *zone;
1615
1616         sc->all_unreclaimable = 1;
1617         for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1618                                         sc->nodemask) {
1619                 if (!populated_zone(zone))
1620                         continue;
1621                 /*
1622                  * Take care memory controller reclaiming has small influence
1623                  * to global LRU.
1624                  */
1625                 if (scanning_global_lru(sc)) {
1626                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1627                                 continue;
1628                         note_zone_scanning_priority(zone, priority);
1629
1630                         if (zone_is_all_unreclaimable(zone) &&
1631                                                 priority != DEF_PRIORITY)
1632                                 continue;       /* Let kswapd poll it */
1633                         sc->all_unreclaimable = 0;
1634                 } else {
1635                         /*
1636                          * Ignore cpuset limitation here. We just want to reduce
1637                          * # of used pages by us regardless of memory shortage.
1638                          */
1639                         sc->all_unreclaimable = 0;
1640                         mem_cgroup_note_reclaim_priority(sc->mem_cgroup,
1641                                                         priority);
1642                 }
1643
1644                 shrink_zone(priority, zone, sc);
1645         }
1646 }
1647
1648 /*
1649  * This is the main entry point to direct page reclaim.
1650  *
1651  * If a full scan of the inactive list fails to free enough memory then we
1652  * are "out of memory" and something needs to be killed.
1653  *
1654  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1655  * high - the zone may be full of dirty or under-writeback pages, which this
1656  * caller can't do much about.  We kick pdflush and take explicit naps in the
1657  * hope that some of these pages can be written.  But if the allocating task
1658  * holds filesystem locks which prevent writeout this might not work, and the
1659  * allocation attempt will fail.
1660  *
1661  * returns:     0, if no pages reclaimed
1662  *              else, the number of pages reclaimed
1663  */
1664 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
1665                                         struct scan_control *sc)
1666 {
1667         int priority;
1668         unsigned long ret = 0;
1669         unsigned long total_scanned = 0;
1670         struct reclaim_state *reclaim_state = current->reclaim_state;
1671         unsigned long lru_pages = 0;
1672         struct zoneref *z;
1673         struct zone *zone;
1674         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1675
1676         delayacct_freepages_start();
1677
1678         if (scanning_global_lru(sc))
1679                 count_vm_event(ALLOCSTALL);
1680         /*
1681          * mem_cgroup will not do shrink_slab.
1682          */
1683         if (scanning_global_lru(sc)) {
1684                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1685
1686                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1687                                 continue;
1688
1689                         lru_pages += zone_lru_pages(zone);
1690                 }
1691         }
1692
1693         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1694                 sc->nr_scanned = 0;
1695                 if (!priority)
1696                         disable_swap_token();
1697                 shrink_zones(priority, zonelist, sc);
1698                 /*
1699                  * Don't shrink slabs when reclaiming memory from
1700                  * over limit cgroups
1701                  */
1702                 if (scanning_global_lru(sc)) {
1703                         shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
1704                         if (reclaim_state) {
1705                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
1706                                 reclaim_state->reclaimed_slab = 0;
1707                         }
1708                 }
1709                 total_scanned += sc->nr_scanned;
1710                 if (sc->nr_reclaimed >= sc->swap_cluster_max) {
1711                         ret = sc->nr_reclaimed;
1712                         goto out;
1713                 }
1714
1715                 /*
1716                  * Try to write back as many pages as we just scanned.  This
1717                  * tends to cause slow streaming writers to write data to the
1718                  * disk smoothly, at the dirtying rate, which is nice.   But
1719                  * that's undesirable in laptop mode, where we *want* lumpy
1720                  * writeout.  So in laptop mode, write out the whole world.
1721                  */
1722                 if (total_scanned > sc->swap_cluster_max +
1723                                         sc->swap_cluster_max / 2) {
1724                         wakeup_pdflush(laptop_mode ? 0 : total_scanned);
1725                         sc->may_writepage = 1;
1726                 }
1727
1728                 /* Take a nap, wait for some writeback to complete */
1729                 if (sc->nr_scanned && priority < DEF_PRIORITY - 2)
1730                         congestion_wait(WRITE, HZ/10);
1731         }
1732         /* top priority shrink_zones still had more to do? don't OOM, then */
1733         if (!sc->all_unreclaimable && scanning_global_lru(sc))
1734                 ret = sc->nr_reclaimed;
1735 out:
1736         /*
1737          * Now that we've scanned all the zones at this priority level, note
1738          * that level within the zone so that the next thread which performs
1739          * scanning of this zone will immediately start out at this priority
1740          * level.  This affects only the decision whether or not to bring
1741          * mapped pages onto the inactive list.
1742          */
1743         if (priority < 0)
1744                 priority = 0;
1745
1746         if (scanning_global_lru(sc)) {
1747                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1748
1749                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1750                                 continue;
1751
1752                         zone->prev_priority = priority;
1753                 }
1754         } else
1755                 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority);
1756
1757         delayacct_freepages_end();
1758
1759         return ret;
1760 }
1761
1762 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
1763                                 gfp_t gfp_mask, nodemask_t *nodemask)
1764 {
1765         struct scan_control sc = {
1766                 .gfp_mask = gfp_mask,
1767                 .may_writepage = !laptop_mode,
1768                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1769                 .may_unmap = 1,
1770                 .may_swap = 1,
1771                 .swappiness = vm_swappiness,
1772                 .order = order,
1773                 .mem_cgroup = NULL,
1774                 .isolate_pages = isolate_pages_global,
1775                 .nodemask = nodemask,
1776         };
1777
1778         return do_try_to_free_pages(zonelist, &sc);
1779 }
1780
1781 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
1782
1783 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
1784                                            gfp_t gfp_mask,
1785                                            bool noswap,
1786                                            unsigned int swappiness)
1787 {
1788         struct scan_control sc = {
1789                 .may_writepage = !laptop_mode,
1790                 .may_unmap = 1,
1791                 .may_swap = !noswap,
1792                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1793                 .swappiness = swappiness,
1794                 .order = 0,
1795                 .mem_cgroup = mem_cont,
1796                 .isolate_pages = mem_cgroup_isolate_pages,
1797                 .nodemask = NULL, /* we don't care the placement */
1798         };
1799         struct zonelist *zonelist;
1800
1801         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1802                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1803         zonelist = NODE_DATA(numa_node_id())->node_zonelists;
1804         return do_try_to_free_pages(zonelist, &sc);
1805 }
1806 #endif
1807
1808 /*
1809  * For kswapd, balance_pgdat() will work across all this node's zones until
1810  * they are all at high_wmark_pages(zone).
1811  *
1812  * Returns the number of pages which were actually freed.
1813  *
1814  * There is special handling here for zones which are full of pinned pages.
1815  * This can happen if the pages are all mlocked, or if they are all used by
1816  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1817  * What we do is to detect the case where all pages in the zone have been
1818  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1819  * dead and from now on, only perform a short scan.  Basically we're polling
1820  * the zone for when the problem goes away.
1821  *
1822  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
1823  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
1824  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
1825  * lower zones regardless of the number of free pages in the lower zones. This
1826  * interoperates with the page allocator fallback scheme to ensure that aging
1827  * of pages is balanced across the zones.
1828  */
1829 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
1830 {
1831         int all_zones_ok;
1832         int priority;
1833         int i;
1834         unsigned long total_scanned;
1835         struct reclaim_state *reclaim_state = current->reclaim_state;
1836         struct scan_control sc = {
1837                 .gfp_mask = GFP_KERNEL,
1838                 .may_unmap = 1,
1839                 .may_swap = 1,
1840                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1841                 .swappiness = vm_swappiness,
1842                 .order = order,
1843                 .mem_cgroup = NULL,
1844                 .isolate_pages = isolate_pages_global,
1845         };
1846         /*
1847          * temp_priority is used to remember the scanning priority at which
1848          * this zone was successfully refilled to
1849          * free_pages == high_wmark_pages(zone).
1850          */
1851         int temp_priority[MAX_NR_ZONES];
1852
1853 loop_again:
1854         total_scanned = 0;
1855         sc.nr_reclaimed = 0;
1856         sc.may_writepage = !laptop_mode;
1857         count_vm_event(PAGEOUTRUN);
1858
1859         for (i = 0; i < pgdat->nr_zones; i++)
1860                 temp_priority[i] = DEF_PRIORITY;
1861
1862         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1863                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1864                 unsigned long lru_pages = 0;
1865
1866                 /* The swap token gets in the way of swapout... */
1867                 if (!priority)
1868                         disable_swap_token();
1869
1870                 all_zones_ok = 1;
1871
1872                 /*
1873                  * Scan in the highmem->dma direction for the highest
1874                  * zone which needs scanning
1875                  */
1876                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1877                         struct zone *zone = pgdat->node_zones + i;
1878
1879                         if (!populated_zone(zone))
1880                                 continue;
1881
1882                         if (zone_is_all_unreclaimable(zone) &&
1883                             priority != DEF_PRIORITY)
1884                                 continue;
1885
1886                         /*
1887                          * Do some background aging of the anon list, to give
1888                          * pages a chance to be referenced before reclaiming.
1889                          */
1890                         if (inactive_anon_is_low(zone, &sc))
1891                                 shrink_active_list(SWAP_CLUSTER_MAX, zone,
1892                                                         &sc, priority, 0);
1893
1894                         if (!zone_watermark_ok(zone, order,
1895                                         high_wmark_pages(zone), 0, 0)) {
1896                                 end_zone = i;
1897                                 break;
1898                         }
1899                 }
1900                 if (i < 0)
1901                         goto out;
1902
1903                 for (i = 0; i <= end_zone; i++) {
1904                         struct zone *zone = pgdat->node_zones + i;
1905
1906                         lru_pages += zone_lru_pages(zone);
1907                 }
1908
1909                 /*
1910                  * Now scan the zone in the dma->highmem direction, stopping
1911                  * at the last zone which needs scanning.
1912                  *
1913                  * We do this because the page allocator works in the opposite
1914                  * direction.  This prevents the page allocator from allocating
1915                  * pages behind kswapd's direction of progress, which would
1916                  * cause too much scanning of the lower zones.
1917                  */
1918                 for (i = 0; i <= end_zone; i++) {
1919                         struct zone *zone = pgdat->node_zones + i;
1920                         int nr_slab;
1921
1922                         if (!populated_zone(zone))
1923                                 continue;
1924
1925                         if (zone_is_all_unreclaimable(zone) &&
1926                                         priority != DEF_PRIORITY)
1927                                 continue;
1928
1929                         if (!zone_watermark_ok(zone, order,
1930                                         high_wmark_pages(zone), end_zone, 0))
1931                                 all_zones_ok = 0;
1932                         temp_priority[i] = priority;
1933                         sc.nr_scanned = 0;
1934                         note_zone_scanning_priority(zone, priority);
1935                         /*
1936                          * We put equal pressure on every zone, unless one
1937                          * zone has way too many pages free already.
1938                          */
1939                         if (!zone_watermark_ok(zone, order,
1940                                         8*high_wmark_pages(zone), end_zone, 0))
1941                                 shrink_zone(priority, zone, &sc);
1942                         reclaim_state->reclaimed_slab = 0;
1943                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1944                                                 lru_pages);
1945                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1946                         total_scanned += sc.nr_scanned;
1947                         if (zone_is_all_unreclaimable(zone))
1948                                 continue;
1949                         if (nr_slab == 0 && zone->pages_scanned >=
1950                                                 (zone_lru_pages(zone) * 6))
1951                                         zone_set_flag(zone,
1952                                                       ZONE_ALL_UNRECLAIMABLE);
1953                         /*
1954                          * If we've done a decent amount of scanning and
1955                          * the reclaim ratio is low, start doing writepage
1956                          * even in laptop mode
1957                          */
1958                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1959                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
1960                                 sc.may_writepage = 1;
1961                 }
1962                 if (all_zones_ok)
1963                         break;          /* kswapd: all done */
1964                 /*
1965                  * OK, kswapd is getting into trouble.  Take a nap, then take
1966                  * another pass across the zones.
1967                  */
1968                 if (total_scanned && priority < DEF_PRIORITY - 2)
1969                         congestion_wait(WRITE, HZ/10);
1970
1971                 /*
1972                  * We do this so kswapd doesn't build up large priorities for
1973                  * example when it is freeing in parallel with allocators. It
1974                  * matches the direct reclaim path behaviour in terms of impact
1975                  * on zone->*_priority.
1976                  */
1977                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
1978                         break;
1979         }
1980 out:
1981         /*
1982          * Note within each zone the priority level at which this zone was
1983          * brought into a happy state.  So that the next thread which scans this
1984          * zone will start out at that priority level.
1985          */
1986         for (i = 0; i < pgdat->nr_zones; i++) {
1987                 struct zone *zone = pgdat->node_zones + i;
1988
1989                 zone->prev_priority = temp_priority[i];
1990         }
1991         if (!all_zones_ok) {
1992                 cond_resched();
1993
1994                 try_to_freeze();
1995
1996                 /*
1997                  * Fragmentation may mean that the system cannot be
1998                  * rebalanced for high-order allocations in all zones.
1999                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2000                  * it means the zones have been fully scanned and are still
2001                  * not balanced. For high-order allocations, there is
2002                  * little point trying all over again as kswapd may
2003                  * infinite loop.
2004                  *
2005                  * Instead, recheck all watermarks at order-0 as they
2006                  * are the most important. If watermarks are ok, kswapd will go
2007                  * back to sleep. High-order users can still perform direct
2008                  * reclaim if they wish.
2009                  */
2010                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2011                         order = sc.order = 0;
2012
2013                 goto loop_again;
2014         }
2015
2016         return sc.nr_reclaimed;
2017 }
2018
2019 /*
2020  * The background pageout daemon, started as a kernel thread
2021  * from the init process.
2022  *
2023  * This basically trickles out pages so that we have _some_
2024  * free memory available even if there is no other activity
2025  * that frees anything up. This is needed for things like routing
2026  * etc, where we otherwise might have all activity going on in
2027  * asynchronous contexts that cannot page things out.
2028  *
2029  * If there are applications that are active memory-allocators
2030  * (most normal use), this basically shouldn't matter.
2031  */
2032 static int kswapd(void *p)
2033 {
2034         unsigned long order;
2035         pg_data_t *pgdat = (pg_data_t*)p;
2036         struct task_struct *tsk = current;
2037         DEFINE_WAIT(wait);
2038         struct reclaim_state reclaim_state = {
2039                 .reclaimed_slab = 0,
2040         };
2041         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2042
2043         lockdep_set_current_reclaim_state(GFP_KERNEL);
2044
2045         if (!cpumask_empty(cpumask))
2046                 set_cpus_allowed_ptr(tsk, cpumask);
2047         current->reclaim_state = &reclaim_state;
2048
2049         /*
2050          * Tell the memory management that we're a "memory allocator",
2051          * and that if we need more memory we should get access to it
2052          * regardless (see "__alloc_pages()"). "kswapd" should
2053          * never get caught in the normal page freeing logic.
2054          *
2055          * (Kswapd normally doesn't need memory anyway, but sometimes
2056          * you need a small amount of memory in order to be able to
2057          * page out something else, and this flag essentially protects
2058          * us from recursively trying to free more memory as we're
2059          * trying to free the first piece of memory in the first place).
2060          */
2061         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2062         set_freezable();
2063
2064         order = 0;
2065         for ( ; ; ) {
2066                 unsigned long new_order;
2067
2068                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2069                 new_order = pgdat->kswapd_max_order;
2070                 pgdat->kswapd_max_order = 0;
2071                 if (order < new_order) {
2072                         /*
2073                          * Don't sleep if someone wants a larger 'order'
2074                          * allocation
2075                          */
2076                         order = new_order;
2077                 } else {
2078                         if (!freezing(current))
2079                                 schedule();
2080
2081                         order = pgdat->kswapd_max_order;
2082                 }
2083                 finish_wait(&pgdat->kswapd_wait, &wait);
2084
2085                 if (!try_to_freeze()) {
2086                         /* We can speed up thawing tasks if we don't call
2087                          * balance_pgdat after returning from the refrigerator
2088                          */
2089                         balance_pgdat(pgdat, order);
2090                 }
2091         }
2092         return 0;
2093 }
2094
2095 /*
2096  * A zone is low on free memory, so wake its kswapd task to service it.
2097  */
2098 void wakeup_kswapd(struct zone *zone, int order)
2099 {
2100         pg_data_t *pgdat;
2101
2102         if (!populated_zone(zone))
2103                 return;
2104
2105         pgdat = zone->zone_pgdat;
2106         if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
2107                 return;
2108         if (pgdat->kswapd_max_order < order)
2109                 pgdat->kswapd_max_order = order;
2110         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2111                 return;
2112         if (!waitqueue_active(&pgdat->kswapd_wait))
2113                 return;
2114         wake_up_interruptible(&pgdat->kswapd_wait);
2115 }
2116
2117 unsigned long global_lru_pages(void)
2118 {
2119         return global_page_state(NR_ACTIVE_ANON)
2120                 + global_page_state(NR_ACTIVE_FILE)
2121                 + global_page_state(NR_INACTIVE_ANON)
2122                 + global_page_state(NR_INACTIVE_FILE);
2123 }
2124
2125 #ifdef CONFIG_HIBERNATION
2126 /*
2127  * Helper function for shrink_all_memory().  Tries to reclaim 'nr_pages' pages
2128  * from LRU lists system-wide, for given pass and priority.
2129  *
2130  * For pass > 3 we also try to shrink the LRU lists that contain a few pages
2131  */
2132 static void shrink_all_zones(unsigned long nr_pages, int prio,
2133                                       int pass, struct scan_control *sc)
2134 {
2135         struct zone *zone;
2136         unsigned long nr_reclaimed = 0;
2137
2138         for_each_populated_zone(zone) {
2139                 enum lru_list l;
2140
2141                 if (zone_is_all_unreclaimable(zone) && prio != DEF_PRIORITY)
2142                         continue;
2143
2144                 for_each_evictable_lru(l) {
2145                         enum zone_stat_item ls = NR_LRU_BASE + l;
2146                         unsigned long lru_pages = zone_page_state(zone, ls);
2147
2148                         /* For pass = 0, we don't shrink the active list */
2149                         if (pass == 0 && (l == LRU_ACTIVE_ANON ||
2150                                                 l == LRU_ACTIVE_FILE))
2151                                 continue;
2152
2153                         zone->lru[l].nr_saved_scan += (lru_pages >> prio) + 1;
2154                         if (zone->lru[l].nr_saved_scan >= nr_pages || pass > 3) {
2155                                 unsigned long nr_to_scan;
2156
2157                                 zone->lru[l].nr_saved_scan = 0;
2158                                 nr_to_scan = min(nr_pages, lru_pages);
2159                                 nr_reclaimed += shrink_list(l, nr_to_scan, zone,
2160                                                                 sc, prio);
2161                                 if (nr_reclaimed >= nr_pages) {
2162                                         sc->nr_reclaimed += nr_reclaimed;
2163                                         return;
2164                                 }
2165                         }
2166                 }
2167         }
2168         sc->nr_reclaimed += nr_reclaimed;
2169 }
2170
2171 /*
2172  * Try to free `nr_pages' of memory, system-wide, and return the number of
2173  * freed pages.
2174  *
2175  * Rather than trying to age LRUs the aim is to preserve the overall
2176  * LRU order by reclaiming preferentially
2177  * inactive > active > active referenced > active mapped
2178  */
2179 unsigned long shrink_all_memory(unsigned long nr_pages)
2180 {
2181         unsigned long lru_pages, nr_slab;
2182         int pass;
2183         struct reclaim_state reclaim_state;
2184         struct scan_control sc = {
2185                 .gfp_mask = GFP_KERNEL,
2186                 .may_unmap = 0,
2187                 .may_writepage = 1,
2188                 .isolate_pages = isolate_pages_global,
2189                 .nr_reclaimed = 0,
2190         };
2191
2192         current->reclaim_state = &reclaim_state;
2193
2194         lru_pages = global_lru_pages();
2195         nr_slab = global_page_state(NR_SLAB_RECLAIMABLE);
2196         /* If slab caches are huge, it's better to hit them first */
2197         while (nr_slab >= lru_pages) {
2198                 reclaim_state.reclaimed_slab = 0;
2199                 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
2200                 if (!reclaim_state.reclaimed_slab)
2201                         break;
2202
2203                 sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2204                 if (sc.nr_reclaimed >= nr_pages)
2205                         goto out;
2206
2207                 nr_slab -= reclaim_state.reclaimed_slab;
2208         }
2209
2210         /*
2211          * We try to shrink LRUs in 5 passes:
2212          * 0 = Reclaim from inactive_list only
2213          * 1 = Reclaim from active list but don't reclaim mapped
2214          * 2 = 2nd pass of type 1
2215          * 3 = Reclaim mapped (normal reclaim)
2216          * 4 = 2nd pass of type 3
2217          */
2218         for (pass = 0; pass < 5; pass++) {
2219                 int prio;
2220
2221                 /* Force reclaiming mapped pages in the passes #3 and #4 */
2222                 if (pass > 2)
2223                         sc.may_unmap = 1;
2224
2225                 for (prio = DEF_PRIORITY; prio >= 0; prio--) {
2226                         unsigned long nr_to_scan = nr_pages - sc.nr_reclaimed;
2227
2228                         sc.nr_scanned = 0;
2229                         sc.swap_cluster_max = nr_to_scan;
2230                         shrink_all_zones(nr_to_scan, prio, pass, &sc);
2231                         if (sc.nr_reclaimed >= nr_pages)
2232                                 goto out;
2233
2234                         reclaim_state.reclaimed_slab = 0;
2235                         shrink_slab(sc.nr_scanned, sc.gfp_mask,
2236                                         global_lru_pages());
2237                         sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2238                         if (sc.nr_reclaimed >= nr_pages)
2239                                 goto out;
2240
2241                         if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
2242                                 congestion_wait(WRITE, HZ / 10);
2243                 }
2244         }
2245
2246         /*
2247          * If sc.nr_reclaimed = 0, we could not shrink LRUs, but there may be
2248          * something in slab caches
2249          */
2250         if (!sc.nr_reclaimed) {
2251                 do {
2252                         reclaim_state.reclaimed_slab = 0;
2253                         shrink_slab(nr_pages, sc.gfp_mask, global_lru_pages());
2254                         sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2255                 } while (sc.nr_reclaimed < nr_pages &&
2256                                 reclaim_state.reclaimed_slab > 0);
2257         }
2258
2259
2260 out:
2261         current->reclaim_state = NULL;
2262
2263         return sc.nr_reclaimed;
2264 }
2265 #endif /* CONFIG_HIBERNATION */
2266
2267 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2268    not required for correctness.  So if the last cpu in a node goes
2269    away, we get changed to run anywhere: as the first one comes back,
2270    restore their cpu bindings. */
2271 static int __devinit cpu_callback(struct notifier_block *nfb,
2272                                   unsigned long action, void *hcpu)
2273 {
2274         int nid;
2275
2276         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2277                 for_each_node_state(nid, N_HIGH_MEMORY) {
2278                         pg_data_t *pgdat = NODE_DATA(nid);
2279                         const struct cpumask *mask;
2280
2281                         mask = cpumask_of_node(pgdat->node_id);
2282
2283                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2284                                 /* One of our CPUs online: restore mask */
2285                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2286                 }
2287         }
2288         return NOTIFY_OK;
2289 }
2290
2291 /*
2292  * This kswapd start function will be called by init and node-hot-add.
2293  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2294  */
2295 int kswapd_run(int nid)
2296 {
2297         pg_data_t *pgdat = NODE_DATA(nid);
2298         int ret = 0;
2299
2300         if (pgdat->kswapd)
2301                 return 0;
2302
2303         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2304         if (IS_ERR(pgdat->kswapd)) {
2305                 /* failure at boot is fatal */
2306                 BUG_ON(system_state == SYSTEM_BOOTING);
2307                 printk("Failed to start kswapd on node %d\n",nid);
2308                 ret = -1;
2309         }
2310         return ret;
2311 }
2312
2313 static int __init kswapd_init(void)
2314 {
2315         int nid;
2316
2317         swap_setup();
2318         for_each_node_state(nid, N_HIGH_MEMORY)
2319                 kswapd_run(nid);
2320         hotcpu_notifier(cpu_callback, 0);
2321         return 0;
2322 }
2323
2324 module_init(kswapd_init)
2325
2326 #ifdef CONFIG_NUMA
2327 /*
2328  * Zone reclaim mode
2329  *
2330  * If non-zero call zone_reclaim when the number of free pages falls below
2331  * the watermarks.
2332  */
2333 int zone_reclaim_mode __read_mostly;
2334
2335 #define RECLAIM_OFF 0
2336 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
2337 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
2338 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
2339
2340 /*
2341  * Priority for ZONE_RECLAIM. This determines the fraction of pages
2342  * of a node considered for each zone_reclaim. 4 scans 1/16th of
2343  * a zone.
2344  */
2345 #define ZONE_RECLAIM_PRIORITY 4
2346
2347 /*
2348  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2349  * occur.
2350  */
2351 int sysctl_min_unmapped_ratio = 1;
2352
2353 /*
2354  * If the number of slab pages in a zone grows beyond this percentage then
2355  * slab reclaim needs to occur.
2356  */
2357 int sysctl_min_slab_ratio = 5;
2358
2359 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
2360 {
2361         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
2362         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
2363                 zone_page_state(zone, NR_ACTIVE_FILE);
2364
2365         /*
2366          * It's possible for there to be more file mapped pages than
2367          * accounted for by the pages on the file LRU lists because
2368          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2369          */
2370         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
2371 }
2372
2373 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
2374 static long zone_pagecache_reclaimable(struct zone *zone)
2375 {
2376         long nr_pagecache_reclaimable;
2377         long delta = 0;
2378
2379         /*
2380          * If RECLAIM_SWAP is set, then all file pages are considered
2381          * potentially reclaimable. Otherwise, we have to worry about
2382          * pages like swapcache and zone_unmapped_file_pages() provides
2383          * a better estimate
2384          */
2385         if (zone_reclaim_mode & RECLAIM_SWAP)
2386                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
2387         else
2388                 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
2389
2390         /* If we can't clean pages, remove dirty pages from consideration */
2391         if (!(zone_reclaim_mode & RECLAIM_WRITE))
2392                 delta += zone_page_state(zone, NR_FILE_DIRTY);
2393
2394         /* Watch for any possible underflows due to delta */
2395         if (unlikely(delta > nr_pagecache_reclaimable))
2396                 delta = nr_pagecache_reclaimable;
2397
2398         return nr_pagecache_reclaimable - delta;
2399 }
2400
2401 /*
2402  * Try to free up some pages from this zone through reclaim.
2403  */
2404 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2405 {
2406         /* Minimum pages needed in order to stay on node */
2407         const unsigned long nr_pages = 1 << order;
2408         struct task_struct *p = current;
2409         struct reclaim_state reclaim_state;
2410         int priority;
2411         struct scan_control sc = {
2412                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2413                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2414                 .may_swap = 1,
2415                 .swap_cluster_max = max_t(unsigned long, nr_pages,
2416                                         SWAP_CLUSTER_MAX),
2417                 .gfp_mask = gfp_mask,
2418                 .swappiness = vm_swappiness,
2419                 .order = order,
2420                 .isolate_pages = isolate_pages_global,
2421         };
2422         unsigned long slab_reclaimable;
2423
2424         disable_swap_token();
2425         cond_resched();
2426         /*
2427          * We need to be able to allocate from the reserves for RECLAIM_SWAP
2428          * and we also need to be able to write out pages for RECLAIM_WRITE
2429          * and RECLAIM_SWAP.
2430          */
2431         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2432         reclaim_state.reclaimed_slab = 0;
2433         p->reclaim_state = &reclaim_state;
2434
2435         if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
2436                 /*
2437                  * Free memory by calling shrink zone with increasing
2438                  * priorities until we have enough memory freed.
2439                  */
2440                 priority = ZONE_RECLAIM_PRIORITY;
2441                 do {
2442                         note_zone_scanning_priority(zone, priority);
2443                         shrink_zone(priority, zone, &sc);
2444                         priority--;
2445                 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
2446         }
2447
2448         slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2449         if (slab_reclaimable > zone->min_slab_pages) {
2450                 /*
2451                  * shrink_slab() does not currently allow us to determine how
2452                  * many pages were freed in this zone. So we take the current
2453                  * number of slab pages and shake the slab until it is reduced
2454                  * by the same nr_pages that we used for reclaiming unmapped
2455                  * pages.
2456                  *
2457                  * Note that shrink_slab will free memory on all zones and may
2458                  * take a long time.
2459                  */
2460                 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
2461                         zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
2462                                 slab_reclaimable - nr_pages)
2463                         ;
2464
2465                 /*
2466                  * Update nr_reclaimed by the number of slab pages we
2467                  * reclaimed from this zone.
2468                  */
2469                 sc.nr_reclaimed += slab_reclaimable -
2470                         zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2471         }
2472
2473         p->reclaim_state = NULL;
2474         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
2475         return sc.nr_reclaimed >= nr_pages;
2476 }
2477
2478 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2479 {
2480         int node_id;
2481         int ret;
2482
2483         /*
2484          * Zone reclaim reclaims unmapped file backed pages and
2485          * slab pages if we are over the defined limits.
2486          *
2487          * A small portion of unmapped file backed pages is needed for
2488          * file I/O otherwise pages read by file I/O will be immediately
2489          * thrown out if the zone is overallocated. So we do not reclaim
2490          * if less than a specified percentage of the zone is used by
2491          * unmapped file backed pages.
2492          */
2493         if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
2494             zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
2495                 return 0;
2496
2497         if (zone_is_all_unreclaimable(zone))
2498                 return 0;
2499
2500         /*
2501          * Do not scan if the allocation should not be delayed.
2502          */
2503         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
2504                         return 0;
2505
2506         /*
2507          * Only run zone reclaim on the local zone or on zones that do not
2508          * have associated processors. This will favor the local processor
2509          * over remote processors and spread off node memory allocations
2510          * as wide as possible.
2511          */
2512         node_id = zone_to_nid(zone);
2513         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
2514                 return 0;
2515
2516         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2517                 return 0;
2518         ret = __zone_reclaim(zone, gfp_mask, order);
2519         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2520
2521         return ret;
2522 }
2523 #endif
2524
2525 /*
2526  * page_evictable - test whether a page is evictable
2527  * @page: the page to test
2528  * @vma: the VMA in which the page is or will be mapped, may be NULL
2529  *
2530  * Test whether page is evictable--i.e., should be placed on active/inactive
2531  * lists vs unevictable list.  The vma argument is !NULL when called from the
2532  * fault path to determine how to instantate a new page.
2533  *
2534  * Reasons page might not be evictable:
2535  * (1) page's mapping marked unevictable
2536  * (2) page is part of an mlocked VMA
2537  *
2538  */
2539 int page_evictable(struct page *page, struct vm_area_struct *vma)
2540 {
2541
2542         if (mapping_unevictable(page_mapping(page)))
2543                 return 0;
2544
2545         if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2546                 return 0;
2547
2548         return 1;
2549 }
2550
2551 /**
2552  * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2553  * @page: page to check evictability and move to appropriate lru list
2554  * @zone: zone page is in
2555  *
2556  * Checks a page for evictability and moves the page to the appropriate
2557  * zone lru list.
2558  *
2559  * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2560  * have PageUnevictable set.
2561  */
2562 static void check_move_unevictable_page(struct page *page, struct zone *zone)
2563 {
2564         VM_BUG_ON(PageActive(page));
2565
2566 retry:
2567         ClearPageUnevictable(page);
2568         if (page_evictable(page, NULL)) {
2569                 enum lru_list l = LRU_INACTIVE_ANON + page_is_file_cache(page);
2570
2571                 __dec_zone_state(zone, NR_UNEVICTABLE);
2572                 list_move(&page->lru, &zone->lru[l].list);
2573                 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
2574                 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
2575                 __count_vm_event(UNEVICTABLE_PGRESCUED);
2576         } else {
2577                 /*
2578                  * rotate unevictable list
2579                  */
2580                 SetPageUnevictable(page);
2581                 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
2582                 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
2583                 if (page_evictable(page, NULL))
2584                         goto retry;
2585         }
2586 }
2587
2588 /**
2589  * scan_mapping_unevictable_pages - scan an address space for evictable pages
2590  * @mapping: struct address_space to scan for evictable pages
2591  *
2592  * Scan all pages in mapping.  Check unevictable pages for
2593  * evictability and move them to the appropriate zone lru list.
2594  */
2595 void scan_mapping_unevictable_pages(struct address_space *mapping)
2596 {
2597         pgoff_t next = 0;
2598         pgoff_t end   = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
2599                          PAGE_CACHE_SHIFT;
2600         struct zone *zone;
2601         struct pagevec pvec;
2602
2603         if (mapping->nrpages == 0)
2604                 return;
2605
2606         pagevec_init(&pvec, 0);
2607         while (next < end &&
2608                 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
2609                 int i;
2610                 int pg_scanned = 0;
2611
2612                 zone = NULL;
2613
2614                 for (i = 0; i < pagevec_count(&pvec); i++) {
2615                         struct page *page = pvec.pages[i];
2616                         pgoff_t page_index = page->index;
2617                         struct zone *pagezone = page_zone(page);
2618
2619                         pg_scanned++;
2620                         if (page_index > next)
2621                                 next = page_index;
2622                         next++;
2623
2624                         if (pagezone != zone) {
2625                                 if (zone)
2626                                         spin_unlock_irq(&zone->lru_lock);
2627                                 zone = pagezone;
2628                                 spin_lock_irq(&zone->lru_lock);
2629                         }
2630
2631                         if (PageLRU(page) && PageUnevictable(page))
2632                                 check_move_unevictable_page(page, zone);
2633                 }
2634                 if (zone)
2635                         spin_unlock_irq(&zone->lru_lock);
2636                 pagevec_release(&pvec);
2637
2638                 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
2639         }
2640
2641 }
2642
2643 /**
2644  * scan_zone_unevictable_pages - check unevictable list for evictable pages
2645  * @zone - zone of which to scan the unevictable list
2646  *
2647  * Scan @zone's unevictable LRU lists to check for pages that have become
2648  * evictable.  Move those that have to @zone's inactive list where they
2649  * become candidates for reclaim, unless shrink_inactive_zone() decides
2650  * to reactivate them.  Pages that are still unevictable are rotated
2651  * back onto @zone's unevictable list.
2652  */
2653 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
2654 static void scan_zone_unevictable_pages(struct zone *zone)
2655 {
2656         struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
2657         unsigned long scan;
2658         unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
2659
2660         while (nr_to_scan > 0) {
2661                 unsigned long batch_size = min(nr_to_scan,
2662                                                 SCAN_UNEVICTABLE_BATCH_SIZE);
2663
2664                 spin_lock_irq(&zone->lru_lock);
2665                 for (scan = 0;  scan < batch_size; scan++) {
2666                         struct page *page = lru_to_page(l_unevictable);
2667
2668                         if (!trylock_page(page))
2669                                 continue;
2670
2671                         prefetchw_prev_lru_page(page, l_unevictable, flags);
2672
2673                         if (likely(PageLRU(page) && PageUnevictable(page)))
2674                                 check_move_unevictable_page(page, zone);
2675
2676                         unlock_page(page);
2677                 }
2678                 spin_unlock_irq(&zone->lru_lock);
2679
2680                 nr_to_scan -= batch_size;
2681         }
2682 }
2683
2684
2685 /**
2686  * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
2687  *
2688  * A really big hammer:  scan all zones' unevictable LRU lists to check for
2689  * pages that have become evictable.  Move those back to the zones'
2690  * inactive list where they become candidates for reclaim.
2691  * This occurs when, e.g., we have unswappable pages on the unevictable lists,
2692  * and we add swap to the system.  As such, it runs in the context of a task
2693  * that has possibly/probably made some previously unevictable pages
2694  * evictable.
2695  */
2696 static void scan_all_zones_unevictable_pages(void)
2697 {
2698         struct zone *zone;
2699
2700         for_each_zone(zone) {
2701                 scan_zone_unevictable_pages(zone);
2702         }
2703 }
2704
2705 /*
2706  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
2707  * all nodes' unevictable lists for evictable pages
2708  */
2709 unsigned long scan_unevictable_pages;
2710
2711 int scan_unevictable_handler(struct ctl_table *table, int write,
2712                            struct file *file, void __user *buffer,
2713                            size_t *length, loff_t *ppos)
2714 {
2715         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
2716
2717         if (write && *(unsigned long *)table->data)
2718                 scan_all_zones_unevictable_pages();
2719
2720         scan_unevictable_pages = 0;
2721         return 0;
2722 }
2723
2724 /*
2725  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
2726  * a specified node's per zone unevictable lists for evictable pages.
2727  */
2728
2729 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
2730                                           struct sysdev_attribute *attr,
2731                                           char *buf)
2732 {
2733         return sprintf(buf, "0\n");     /* always zero; should fit... */
2734 }
2735
2736 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
2737                                            struct sysdev_attribute *attr,
2738                                         const char *buf, size_t count)
2739 {
2740         struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
2741         struct zone *zone;
2742         unsigned long res;
2743         unsigned long req = strict_strtoul(buf, 10, &res);
2744
2745         if (!req)
2746                 return 1;       /* zero is no-op */
2747
2748         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2749                 if (!populated_zone(zone))
2750                         continue;
2751                 scan_zone_unevictable_pages(zone);
2752         }
2753         return 1;
2754 }
2755
2756
2757 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
2758                         read_scan_unevictable_node,
2759                         write_scan_unevictable_node);
2760
2761 int scan_unevictable_register_node(struct node *node)
2762 {
2763         return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
2764 }
2765
2766 void scan_unevictable_unregister_node(struct node *node)
2767 {
2768         sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
2769 }
2770