mm/vmscan: remove page_queue_congested() comment
[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_lru_pages(struct zone *zone,
152                                 struct scan_control *sc, 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         /*
290          * A freeable page cache page is referenced only by the caller
291          * that isolated the page, the page cache radix tree and
292          * optional buffer heads at page->private.
293          */
294         return page_count(page) - page_has_private(page) == 2;
295 }
296
297 static int may_write_to_queue(struct backing_dev_info *bdi)
298 {
299         if (current->flags & PF_SWAPWRITE)
300                 return 1;
301         if (!bdi_write_congested(bdi))
302                 return 1;
303         if (bdi == current->backing_dev_info)
304                 return 1;
305         return 0;
306 }
307
308 /*
309  * We detected a synchronous write error writing a page out.  Probably
310  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
311  * fsync(), msync() or close().
312  *
313  * The tricky part is that after writepage we cannot touch the mapping: nothing
314  * prevents it from being freed up.  But we have a ref on the page and once
315  * that page is locked, the mapping is pinned.
316  *
317  * We're allowed to run sleeping lock_page() here because we know the caller has
318  * __GFP_FS.
319  */
320 static void handle_write_error(struct address_space *mapping,
321                                 struct page *page, int error)
322 {
323         lock_page(page);
324         if (page_mapping(page) == mapping)
325                 mapping_set_error(mapping, error);
326         unlock_page(page);
327 }
328
329 /* Request for sync pageout. */
330 enum pageout_io {
331         PAGEOUT_IO_ASYNC,
332         PAGEOUT_IO_SYNC,
333 };
334
335 /* possible outcome of pageout() */
336 typedef enum {
337         /* failed to write page out, page is locked */
338         PAGE_KEEP,
339         /* move page to the active list, page is locked */
340         PAGE_ACTIVATE,
341         /* page has been sent to the disk successfully, page is unlocked */
342         PAGE_SUCCESS,
343         /* page is clean and locked */
344         PAGE_CLEAN,
345 } pageout_t;
346
347 /*
348  * pageout is called by shrink_page_list() for each dirty page.
349  * Calls ->writepage().
350  */
351 static pageout_t pageout(struct page *page, struct address_space *mapping,
352                                                 enum pageout_io sync_writeback)
353 {
354         /*
355          * If the page is dirty, only perform writeback if that write
356          * will be non-blocking.  To prevent this allocation from being
357          * stalled by pagecache activity.  But note that there may be
358          * stalls if we need to run get_block().  We could test
359          * PagePrivate for that.
360          *
361          * If this process is currently in generic_file_write() against
362          * this page's queue, we can perform writeback even if that
363          * will block.
364          *
365          * If the page is swapcache, write it back even if that would
366          * block, for some throttling. This happens by accident, because
367          * swap_backing_dev_info is bust: it doesn't reflect the
368          * congestion state of the swapdevs.  Easy to fix, if needed.
369          */
370         if (!is_page_cache_freeable(page))
371                 return PAGE_KEEP;
372         if (!mapping) {
373                 /*
374                  * Some data journaling orphaned pages can have
375                  * page->mapping == NULL while being dirty with clean buffers.
376                  */
377                 if (page_has_private(page)) {
378                         if (try_to_free_buffers(page)) {
379                                 ClearPageDirty(page);
380                                 printk("%s: orphaned page\n", __func__);
381                                 return PAGE_CLEAN;
382                         }
383                 }
384                 return PAGE_KEEP;
385         }
386         if (mapping->a_ops->writepage == NULL)
387                 return PAGE_ACTIVATE;
388         if (!may_write_to_queue(mapping->backing_dev_info))
389                 return PAGE_KEEP;
390
391         if (clear_page_dirty_for_io(page)) {
392                 int res;
393                 struct writeback_control wbc = {
394                         .sync_mode = WB_SYNC_NONE,
395                         .nr_to_write = SWAP_CLUSTER_MAX,
396                         .range_start = 0,
397                         .range_end = LLONG_MAX,
398                         .nonblocking = 1,
399                         .for_reclaim = 1,
400                 };
401
402                 SetPageReclaim(page);
403                 res = mapping->a_ops->writepage(page, &wbc);
404                 if (res < 0)
405                         handle_write_error(mapping, page, res);
406                 if (res == AOP_WRITEPAGE_ACTIVATE) {
407                         ClearPageReclaim(page);
408                         return PAGE_ACTIVATE;
409                 }
410
411                 /*
412                  * Wait on writeback if requested to. This happens when
413                  * direct reclaiming a large contiguous area and the
414                  * first attempt to free a range of pages fails.
415                  */
416                 if (PageWriteback(page) && sync_writeback == PAGEOUT_IO_SYNC)
417                         wait_on_page_writeback(page);
418
419                 if (!PageWriteback(page)) {
420                         /* synchronous write or broken a_ops? */
421                         ClearPageReclaim(page);
422                 }
423                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
424                 return PAGE_SUCCESS;
425         }
426
427         return PAGE_CLEAN;
428 }
429
430 /*
431  * Same as remove_mapping, but if the page is removed from the mapping, it
432  * gets returned with a refcount of 0.
433  */
434 static int __remove_mapping(struct address_space *mapping, struct page *page)
435 {
436         BUG_ON(!PageLocked(page));
437         BUG_ON(mapping != page_mapping(page));
438
439         spin_lock_irq(&mapping->tree_lock);
440         /*
441          * The non racy check for a busy page.
442          *
443          * Must be careful with the order of the tests. When someone has
444          * a ref to the page, it may be possible that they dirty it then
445          * drop the reference. So if PageDirty is tested before page_count
446          * here, then the following race may occur:
447          *
448          * get_user_pages(&page);
449          * [user mapping goes away]
450          * write_to(page);
451          *                              !PageDirty(page)    [good]
452          * SetPageDirty(page);
453          * put_page(page);
454          *                              !page_count(page)   [good, discard it]
455          *
456          * [oops, our write_to data is lost]
457          *
458          * Reversing the order of the tests ensures such a situation cannot
459          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
460          * load is not satisfied before that of page->_count.
461          *
462          * Note that if SetPageDirty is always performed via set_page_dirty,
463          * and thus under tree_lock, then this ordering is not required.
464          */
465         if (!page_freeze_refs(page, 2))
466                 goto cannot_free;
467         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
468         if (unlikely(PageDirty(page))) {
469                 page_unfreeze_refs(page, 2);
470                 goto cannot_free;
471         }
472
473         if (PageSwapCache(page)) {
474                 swp_entry_t swap = { .val = page_private(page) };
475                 __delete_from_swap_cache(page);
476                 spin_unlock_irq(&mapping->tree_lock);
477                 swapcache_free(swap, page);
478         } else {
479                 __remove_from_page_cache(page);
480                 spin_unlock_irq(&mapping->tree_lock);
481                 mem_cgroup_uncharge_cache_page(page);
482         }
483
484         return 1;
485
486 cannot_free:
487         spin_unlock_irq(&mapping->tree_lock);
488         return 0;
489 }
490
491 /*
492  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
493  * someone else has a ref on the page, abort and return 0.  If it was
494  * successfully detached, return 1.  Assumes the caller has a single ref on
495  * this page.
496  */
497 int remove_mapping(struct address_space *mapping, struct page *page)
498 {
499         if (__remove_mapping(mapping, page)) {
500                 /*
501                  * Unfreezing the refcount with 1 rather than 2 effectively
502                  * drops the pagecache ref for us without requiring another
503                  * atomic operation.
504                  */
505                 page_unfreeze_refs(page, 1);
506                 return 1;
507         }
508         return 0;
509 }
510
511 /**
512  * putback_lru_page - put previously isolated page onto appropriate LRU list
513  * @page: page to be put back to appropriate lru list
514  *
515  * Add previously isolated @page to appropriate LRU list.
516  * Page may still be unevictable for other reasons.
517  *
518  * lru_lock must not be held, interrupts must be enabled.
519  */
520 void putback_lru_page(struct page *page)
521 {
522         int lru;
523         int active = !!TestClearPageActive(page);
524         int was_unevictable = PageUnevictable(page);
525
526         VM_BUG_ON(PageLRU(page));
527
528 redo:
529         ClearPageUnevictable(page);
530
531         if (page_evictable(page, NULL)) {
532                 /*
533                  * For evictable pages, we can use the cache.
534                  * In event of a race, worst case is we end up with an
535                  * unevictable page on [in]active list.
536                  * We know how to handle that.
537                  */
538                 lru = active + page_lru_base_type(page);
539                 lru_cache_add_lru(page, lru);
540         } else {
541                 /*
542                  * Put unevictable pages directly on zone's unevictable
543                  * list.
544                  */
545                 lru = LRU_UNEVICTABLE;
546                 add_page_to_unevictable_list(page);
547         }
548
549         /*
550          * page's status can change while we move it among lru. If an evictable
551          * page is on unevictable list, it never be freed. To avoid that,
552          * check after we added it to the list, again.
553          */
554         if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
555                 if (!isolate_lru_page(page)) {
556                         put_page(page);
557                         goto redo;
558                 }
559                 /* This means someone else dropped this page from LRU
560                  * So, it will be freed or putback to LRU again. There is
561                  * nothing to do here.
562                  */
563         }
564
565         if (was_unevictable && lru != LRU_UNEVICTABLE)
566                 count_vm_event(UNEVICTABLE_PGRESCUED);
567         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
568                 count_vm_event(UNEVICTABLE_PGCULLED);
569
570         put_page(page);         /* drop ref from isolate */
571 }
572
573 /*
574  * shrink_page_list() returns the number of reclaimed pages
575  */
576 static unsigned long shrink_page_list(struct list_head *page_list,
577                                         struct scan_control *sc,
578                                         enum pageout_io sync_writeback)
579 {
580         LIST_HEAD(ret_pages);
581         struct pagevec freed_pvec;
582         int pgactivate = 0;
583         unsigned long nr_reclaimed = 0;
584         unsigned long vm_flags;
585
586         cond_resched();
587
588         pagevec_init(&freed_pvec, 1);
589         while (!list_empty(page_list)) {
590                 struct address_space *mapping;
591                 struct page *page;
592                 int may_enter_fs;
593                 int referenced;
594
595                 cond_resched();
596
597                 page = lru_to_page(page_list);
598                 list_del(&page->lru);
599
600                 if (!trylock_page(page))
601                         goto keep;
602
603                 VM_BUG_ON(PageActive(page));
604
605                 sc->nr_scanned++;
606
607                 if (unlikely(!page_evictable(page, NULL)))
608                         goto cull_mlocked;
609
610                 if (!sc->may_unmap && page_mapped(page))
611                         goto keep_locked;
612
613                 /* Double the slab pressure for mapped and swapcache pages */
614                 if (page_mapped(page) || PageSwapCache(page))
615                         sc->nr_scanned++;
616
617                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
618                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
619
620                 if (PageWriteback(page)) {
621                         /*
622                          * Synchronous reclaim is performed in two passes,
623                          * first an asynchronous pass over the list to
624                          * start parallel writeback, and a second synchronous
625                          * pass to wait for the IO to complete.  Wait here
626                          * for any page for which writeback has already
627                          * started.
628                          */
629                         if (sync_writeback == PAGEOUT_IO_SYNC && may_enter_fs)
630                                 wait_on_page_writeback(page);
631                         else
632                                 goto keep_locked;
633                 }
634
635                 referenced = page_referenced(page, 1,
636                                                 sc->mem_cgroup, &vm_flags);
637                 /*
638                  * In active use or really unfreeable?  Activate it.
639                  * If page which have PG_mlocked lost isoltation race,
640                  * try_to_unmap moves it to unevictable list
641                  */
642                 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER &&
643                                         referenced && page_mapping_inuse(page)
644                                         && !(vm_flags & VM_LOCKED))
645                         goto activate_locked;
646
647                 /*
648                  * Anonymous process memory has backing store?
649                  * Try to allocate it some swap space here.
650                  */
651                 if (PageAnon(page) && !PageSwapCache(page)) {
652                         if (!(sc->gfp_mask & __GFP_IO))
653                                 goto keep_locked;
654                         if (!add_to_swap(page))
655                                 goto activate_locked;
656                         may_enter_fs = 1;
657                 }
658
659                 mapping = page_mapping(page);
660
661                 /*
662                  * The page is mapped into the page tables of one or more
663                  * processes. Try to unmap it here.
664                  */
665                 if (page_mapped(page) && mapping) {
666                         switch (try_to_unmap(page, 0)) {
667                         case SWAP_FAIL:
668                                 goto activate_locked;
669                         case SWAP_AGAIN:
670                                 goto keep_locked;
671                         case SWAP_MLOCK:
672                                 goto cull_mlocked;
673                         case SWAP_SUCCESS:
674                                 ; /* try to free the page below */
675                         }
676                 }
677
678                 if (PageDirty(page)) {
679                         if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && referenced)
680                                 goto keep_locked;
681                         if (!may_enter_fs)
682                                 goto keep_locked;
683                         if (!sc->may_writepage)
684                                 goto keep_locked;
685
686                         /* Page is dirty, try to write it out here */
687                         switch (pageout(page, mapping, sync_writeback)) {
688                         case PAGE_KEEP:
689                                 goto keep_locked;
690                         case PAGE_ACTIVATE:
691                                 goto activate_locked;
692                         case PAGE_SUCCESS:
693                                 if (PageWriteback(page) || PageDirty(page))
694                                         goto keep;
695                                 /*
696                                  * A synchronous write - probably a ramdisk.  Go
697                                  * ahead and try to reclaim the page.
698                                  */
699                                 if (!trylock_page(page))
700                                         goto keep;
701                                 if (PageDirty(page) || PageWriteback(page))
702                                         goto keep_locked;
703                                 mapping = page_mapping(page);
704                         case PAGE_CLEAN:
705                                 ; /* try to free the page below */
706                         }
707                 }
708
709                 /*
710                  * If the page has buffers, try to free the buffer mappings
711                  * associated with this page. If we succeed we try to free
712                  * the page as well.
713                  *
714                  * We do this even if the page is PageDirty().
715                  * try_to_release_page() does not perform I/O, but it is
716                  * possible for a page to have PageDirty set, but it is actually
717                  * clean (all its buffers are clean).  This happens if the
718                  * buffers were written out directly, with submit_bh(). ext3
719                  * will do this, as well as the blockdev mapping.
720                  * try_to_release_page() will discover that cleanness and will
721                  * drop the buffers and mark the page clean - it can be freed.
722                  *
723                  * Rarely, pages can have buffers and no ->mapping.  These are
724                  * the pages which were not successfully invalidated in
725                  * truncate_complete_page().  We try to drop those buffers here
726                  * and if that worked, and the page is no longer mapped into
727                  * process address space (page_count == 1) it can be freed.
728                  * Otherwise, leave the page on the LRU so it is swappable.
729                  */
730                 if (page_has_private(page)) {
731                         if (!try_to_release_page(page, sc->gfp_mask))
732                                 goto activate_locked;
733                         if (!mapping && page_count(page) == 1) {
734                                 unlock_page(page);
735                                 if (put_page_testzero(page))
736                                         goto free_it;
737                                 else {
738                                         /*
739                                          * rare race with speculative reference.
740                                          * the speculative reference will free
741                                          * this page shortly, so we may
742                                          * increment nr_reclaimed here (and
743                                          * leave it off the LRU).
744                                          */
745                                         nr_reclaimed++;
746                                         continue;
747                                 }
748                         }
749                 }
750
751                 if (!mapping || !__remove_mapping(mapping, page))
752                         goto keep_locked;
753
754                 /*
755                  * At this point, we have no other references and there is
756                  * no way to pick any more up (removed from LRU, removed
757                  * from pagecache). Can use non-atomic bitops now (and
758                  * we obviously don't have to worry about waking up a process
759                  * waiting on the page lock, because there are no references.
760                  */
761                 __clear_page_locked(page);
762 free_it:
763                 nr_reclaimed++;
764                 if (!pagevec_add(&freed_pvec, page)) {
765                         __pagevec_free(&freed_pvec);
766                         pagevec_reinit(&freed_pvec);
767                 }
768                 continue;
769
770 cull_mlocked:
771                 if (PageSwapCache(page))
772                         try_to_free_swap(page);
773                 unlock_page(page);
774                 putback_lru_page(page);
775                 continue;
776
777 activate_locked:
778                 /* Not a candidate for swapping, so reclaim swap space. */
779                 if (PageSwapCache(page) && vm_swap_full())
780                         try_to_free_swap(page);
781                 VM_BUG_ON(PageActive(page));
782                 SetPageActive(page);
783                 pgactivate++;
784 keep_locked:
785                 unlock_page(page);
786 keep:
787                 list_add(&page->lru, &ret_pages);
788                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
789         }
790         list_splice(&ret_pages, page_list);
791         if (pagevec_count(&freed_pvec))
792                 __pagevec_free(&freed_pvec);
793         count_vm_events(PGACTIVATE, pgactivate);
794         return nr_reclaimed;
795 }
796
797 /* LRU Isolation modes. */
798 #define ISOLATE_INACTIVE 0      /* Isolate inactive pages. */
799 #define ISOLATE_ACTIVE 1        /* Isolate active pages. */
800 #define ISOLATE_BOTH 2          /* Isolate both active and inactive pages. */
801
802 /*
803  * Attempt to remove the specified page from its LRU.  Only take this page
804  * if it is of the appropriate PageActive status.  Pages which are being
805  * freed elsewhere are also ignored.
806  *
807  * page:        page to consider
808  * mode:        one of the LRU isolation modes defined above
809  *
810  * returns 0 on success, -ve errno on failure.
811  */
812 int __isolate_lru_page(struct page *page, int mode, int file)
813 {
814         int ret = -EINVAL;
815
816         /* Only take pages on the LRU. */
817         if (!PageLRU(page))
818                 return ret;
819
820         /*
821          * When checking the active state, we need to be sure we are
822          * dealing with comparible boolean values.  Take the logical not
823          * of each.
824          */
825         if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
826                 return ret;
827
828         if (mode != ISOLATE_BOTH && page_is_file_cache(page) != file)
829                 return ret;
830
831         /*
832          * When this function is being called for lumpy reclaim, we
833          * initially look into all LRU pages, active, inactive and
834          * unevictable; only give shrink_page_list evictable pages.
835          */
836         if (PageUnevictable(page))
837                 return ret;
838
839         ret = -EBUSY;
840
841         if (likely(get_page_unless_zero(page))) {
842                 /*
843                  * Be careful not to clear PageLRU until after we're
844                  * sure the page is not being freed elsewhere -- the
845                  * page release code relies on it.
846                  */
847                 ClearPageLRU(page);
848                 ret = 0;
849         }
850
851         return ret;
852 }
853
854 /*
855  * zone->lru_lock is heavily contended.  Some of the functions that
856  * shrink the lists perform better by taking out a batch of pages
857  * and working on them outside the LRU lock.
858  *
859  * For pagecache intensive workloads, this function is the hottest
860  * spot in the kernel (apart from copy_*_user functions).
861  *
862  * Appropriate locks must be held before calling this function.
863  *
864  * @nr_to_scan: The number of pages to look through on the list.
865  * @src:        The LRU list to pull pages off.
866  * @dst:        The temp list to put pages on to.
867  * @scanned:    The number of pages that were scanned.
868  * @order:      The caller's attempted allocation order
869  * @mode:       One of the LRU isolation modes
870  * @file:       True [1] if isolating file [!anon] pages
871  *
872  * returns how many pages were moved onto *@dst.
873  */
874 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
875                 struct list_head *src, struct list_head *dst,
876                 unsigned long *scanned, int order, int mode, int file)
877 {
878         unsigned long nr_taken = 0;
879         unsigned long scan;
880
881         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
882                 struct page *page;
883                 unsigned long pfn;
884                 unsigned long end_pfn;
885                 unsigned long page_pfn;
886                 int zone_id;
887
888                 page = lru_to_page(src);
889                 prefetchw_prev_lru_page(page, src, flags);
890
891                 VM_BUG_ON(!PageLRU(page));
892
893                 switch (__isolate_lru_page(page, mode, file)) {
894                 case 0:
895                         list_move(&page->lru, dst);
896                         mem_cgroup_del_lru(page);
897                         nr_taken++;
898                         break;
899
900                 case -EBUSY:
901                         /* else it is being freed elsewhere */
902                         list_move(&page->lru, src);
903                         mem_cgroup_rotate_lru_list(page, page_lru(page));
904                         continue;
905
906                 default:
907                         BUG();
908                 }
909
910                 if (!order)
911                         continue;
912
913                 /*
914                  * Attempt to take all pages in the order aligned region
915                  * surrounding the tag page.  Only take those pages of
916                  * the same active state as that tag page.  We may safely
917                  * round the target page pfn down to the requested order
918                  * as the mem_map is guarenteed valid out to MAX_ORDER,
919                  * where that page is in a different zone we will detect
920                  * it from its zone id and abort this block scan.
921                  */
922                 zone_id = page_zone_id(page);
923                 page_pfn = page_to_pfn(page);
924                 pfn = page_pfn & ~((1 << order) - 1);
925                 end_pfn = pfn + (1 << order);
926                 for (; pfn < end_pfn; pfn++) {
927                         struct page *cursor_page;
928
929                         /* The target page is in the block, ignore it. */
930                         if (unlikely(pfn == page_pfn))
931                                 continue;
932
933                         /* Avoid holes within the zone. */
934                         if (unlikely(!pfn_valid_within(pfn)))
935                                 break;
936
937                         cursor_page = pfn_to_page(pfn);
938
939                         /* Check that we have not crossed a zone boundary. */
940                         if (unlikely(page_zone_id(cursor_page) != zone_id))
941                                 continue;
942
943                         /*
944                          * If we don't have enough swap space, reclaiming of
945                          * anon page which don't already have a swap slot is
946                          * pointless.
947                          */
948                         if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
949                                         !PageSwapCache(cursor_page))
950                                 continue;
951
952                         if (__isolate_lru_page(cursor_page, mode, file) == 0) {
953                                 list_move(&cursor_page->lru, dst);
954                                 mem_cgroup_del_lru(cursor_page);
955                                 nr_taken++;
956                                 scan++;
957                         }
958                 }
959         }
960
961         *scanned = scan;
962         return nr_taken;
963 }
964
965 static unsigned long isolate_pages_global(unsigned long nr,
966                                         struct list_head *dst,
967                                         unsigned long *scanned, int order,
968                                         int mode, struct zone *z,
969                                         struct mem_cgroup *mem_cont,
970                                         int active, int file)
971 {
972         int lru = LRU_BASE;
973         if (active)
974                 lru += LRU_ACTIVE;
975         if (file)
976                 lru += LRU_FILE;
977         return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
978                                                                 mode, file);
979 }
980
981 /*
982  * clear_active_flags() is a helper for shrink_active_list(), clearing
983  * any active bits from the pages in the list.
984  */
985 static unsigned long clear_active_flags(struct list_head *page_list,
986                                         unsigned int *count)
987 {
988         int nr_active = 0;
989         int lru;
990         struct page *page;
991
992         list_for_each_entry(page, page_list, lru) {
993                 lru = page_lru_base_type(page);
994                 if (PageActive(page)) {
995                         lru += LRU_ACTIVE;
996                         ClearPageActive(page);
997                         nr_active++;
998                 }
999                 count[lru]++;
1000         }
1001
1002         return nr_active;
1003 }
1004
1005 /**
1006  * isolate_lru_page - tries to isolate a page from its LRU list
1007  * @page: page to isolate from its LRU list
1008  *
1009  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1010  * vmstat statistic corresponding to whatever LRU list the page was on.
1011  *
1012  * Returns 0 if the page was removed from an LRU list.
1013  * Returns -EBUSY if the page was not on an LRU list.
1014  *
1015  * The returned page will have PageLRU() cleared.  If it was found on
1016  * the active list, it will have PageActive set.  If it was found on
1017  * the unevictable list, it will have the PageUnevictable bit set. That flag
1018  * may need to be cleared by the caller before letting the page go.
1019  *
1020  * The vmstat statistic corresponding to the list on which the page was
1021  * found will be decremented.
1022  *
1023  * Restrictions:
1024  * (1) Must be called with an elevated refcount on the page. This is a
1025  *     fundamentnal difference from isolate_lru_pages (which is called
1026  *     without a stable reference).
1027  * (2) the lru_lock must not be held.
1028  * (3) interrupts must be enabled.
1029  */
1030 int isolate_lru_page(struct page *page)
1031 {
1032         int ret = -EBUSY;
1033
1034         if (PageLRU(page)) {
1035                 struct zone *zone = page_zone(page);
1036
1037                 spin_lock_irq(&zone->lru_lock);
1038                 if (PageLRU(page) && get_page_unless_zero(page)) {
1039                         int lru = page_lru(page);
1040                         ret = 0;
1041                         ClearPageLRU(page);
1042
1043                         del_page_from_lru_list(zone, page, lru);
1044                 }
1045                 spin_unlock_irq(&zone->lru_lock);
1046         }
1047         return ret;
1048 }
1049
1050 /*
1051  * Are there way too many processes in the direct reclaim path already?
1052  */
1053 static int too_many_isolated(struct zone *zone, int file,
1054                 struct scan_control *sc)
1055 {
1056         unsigned long inactive, isolated;
1057
1058         if (current_is_kswapd())
1059                 return 0;
1060
1061         if (!scanning_global_lru(sc))
1062                 return 0;
1063
1064         if (file) {
1065                 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1066                 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1067         } else {
1068                 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1069                 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1070         }
1071
1072         return isolated > inactive;
1073 }
1074
1075 /*
1076  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1077  * of reclaimed pages
1078  */
1079 static unsigned long shrink_inactive_list(unsigned long max_scan,
1080                         struct zone *zone, struct scan_control *sc,
1081                         int priority, int file)
1082 {
1083         LIST_HEAD(page_list);
1084         struct pagevec pvec;
1085         unsigned long nr_scanned = 0;
1086         unsigned long nr_reclaimed = 0;
1087         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1088         int lumpy_reclaim = 0;
1089
1090         while (unlikely(too_many_isolated(zone, file, sc))) {
1091                 congestion_wait(WRITE, HZ/10);
1092
1093                 /* We are about to die and free our memory. Return now. */
1094                 if (fatal_signal_pending(current))
1095                         return SWAP_CLUSTER_MAX;
1096         }
1097
1098         /*
1099          * If we need a large contiguous chunk of memory, or have
1100          * trouble getting a small set of contiguous pages, we
1101          * will reclaim both active and inactive pages.
1102          *
1103          * We use the same threshold as pageout congestion_wait below.
1104          */
1105         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1106                 lumpy_reclaim = 1;
1107         else if (sc->order && priority < DEF_PRIORITY - 2)
1108                 lumpy_reclaim = 1;
1109
1110         pagevec_init(&pvec, 1);
1111
1112         lru_add_drain();
1113         spin_lock_irq(&zone->lru_lock);
1114         do {
1115                 struct page *page;
1116                 unsigned long nr_taken;
1117                 unsigned long nr_scan;
1118                 unsigned long nr_freed;
1119                 unsigned long nr_active;
1120                 unsigned int count[NR_LRU_LISTS] = { 0, };
1121                 int mode = lumpy_reclaim ? ISOLATE_BOTH : ISOLATE_INACTIVE;
1122                 unsigned long nr_anon;
1123                 unsigned long nr_file;
1124
1125                 nr_taken = sc->isolate_pages(sc->swap_cluster_max,
1126                              &page_list, &nr_scan, sc->order, mode,
1127                                 zone, sc->mem_cgroup, 0, file);
1128
1129                 if (scanning_global_lru(sc)) {
1130                         zone->pages_scanned += nr_scan;
1131                         if (current_is_kswapd())
1132                                 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1133                                                        nr_scan);
1134                         else
1135                                 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1136                                                        nr_scan);
1137                 }
1138
1139                 if (nr_taken == 0)
1140                         goto done;
1141
1142                 nr_active = clear_active_flags(&page_list, count);
1143                 __count_vm_events(PGDEACTIVATE, nr_active);
1144
1145                 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1146                                                 -count[LRU_ACTIVE_FILE]);
1147                 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1148                                                 -count[LRU_INACTIVE_FILE]);
1149                 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1150                                                 -count[LRU_ACTIVE_ANON]);
1151                 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1152                                                 -count[LRU_INACTIVE_ANON]);
1153
1154                 nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1155                 nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1156                 __mod_zone_page_state(zone, NR_ISOLATED_ANON, nr_anon);
1157                 __mod_zone_page_state(zone, NR_ISOLATED_FILE, nr_file);
1158
1159                 reclaim_stat->recent_scanned[0] += count[LRU_INACTIVE_ANON];
1160                 reclaim_stat->recent_scanned[0] += count[LRU_ACTIVE_ANON];
1161                 reclaim_stat->recent_scanned[1] += count[LRU_INACTIVE_FILE];
1162                 reclaim_stat->recent_scanned[1] += count[LRU_ACTIVE_FILE];
1163
1164                 spin_unlock_irq(&zone->lru_lock);
1165
1166                 nr_scanned += nr_scan;
1167                 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
1168
1169                 /*
1170                  * If we are direct reclaiming for contiguous pages and we do
1171                  * not reclaim everything in the list, try again and wait
1172                  * for IO to complete. This will stall high-order allocations
1173                  * but that should be acceptable to the caller
1174                  */
1175                 if (nr_freed < nr_taken && !current_is_kswapd() &&
1176                     lumpy_reclaim) {
1177                         congestion_wait(BLK_RW_ASYNC, HZ/10);
1178
1179                         /*
1180                          * The attempt at page out may have made some
1181                          * of the pages active, mark them inactive again.
1182                          */
1183                         nr_active = clear_active_flags(&page_list, count);
1184                         count_vm_events(PGDEACTIVATE, nr_active);
1185
1186                         nr_freed += shrink_page_list(&page_list, sc,
1187                                                         PAGEOUT_IO_SYNC);
1188                 }
1189
1190                 nr_reclaimed += nr_freed;
1191
1192                 local_irq_disable();
1193                 if (current_is_kswapd())
1194                         __count_vm_events(KSWAPD_STEAL, nr_freed);
1195                 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
1196
1197                 spin_lock(&zone->lru_lock);
1198                 /*
1199                  * Put back any unfreeable pages.
1200                  */
1201                 while (!list_empty(&page_list)) {
1202                         int lru;
1203                         page = lru_to_page(&page_list);
1204                         VM_BUG_ON(PageLRU(page));
1205                         list_del(&page->lru);
1206                         if (unlikely(!page_evictable(page, NULL))) {
1207                                 spin_unlock_irq(&zone->lru_lock);
1208                                 putback_lru_page(page);
1209                                 spin_lock_irq(&zone->lru_lock);
1210                                 continue;
1211                         }
1212                         SetPageLRU(page);
1213                         lru = page_lru(page);
1214                         add_page_to_lru_list(zone, page, lru);
1215                         if (is_active_lru(lru)) {
1216                                 int file = is_file_lru(lru);
1217                                 reclaim_stat->recent_rotated[file]++;
1218                         }
1219                         if (!pagevec_add(&pvec, page)) {
1220                                 spin_unlock_irq(&zone->lru_lock);
1221                                 __pagevec_release(&pvec);
1222                                 spin_lock_irq(&zone->lru_lock);
1223                         }
1224                 }
1225                 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1226                 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1227
1228         } while (nr_scanned < max_scan);
1229
1230 done:
1231         spin_unlock_irq(&zone->lru_lock);
1232         pagevec_release(&pvec);
1233         return nr_reclaimed;
1234 }
1235
1236 /*
1237  * We are about to scan this zone at a certain priority level.  If that priority
1238  * level is smaller (ie: more urgent) than the previous priority, then note
1239  * that priority level within the zone.  This is done so that when the next
1240  * process comes in to scan this zone, it will immediately start out at this
1241  * priority level rather than having to build up its own scanning priority.
1242  * Here, this priority affects only the reclaim-mapped threshold.
1243  */
1244 static inline void note_zone_scanning_priority(struct zone *zone, int priority)
1245 {
1246         if (priority < zone->prev_priority)
1247                 zone->prev_priority = priority;
1248 }
1249
1250 /*
1251  * This moves pages from the active list to the inactive list.
1252  *
1253  * We move them the other way if the page is referenced by one or more
1254  * processes, from rmap.
1255  *
1256  * If the pages are mostly unmapped, the processing is fast and it is
1257  * appropriate to hold zone->lru_lock across the whole operation.  But if
1258  * the pages are mapped, the processing is slow (page_referenced()) so we
1259  * should drop zone->lru_lock around each page.  It's impossible to balance
1260  * this, so instead we remove the pages from the LRU while processing them.
1261  * It is safe to rely on PG_active against the non-LRU pages in here because
1262  * nobody will play with that bit on a non-LRU page.
1263  *
1264  * The downside is that we have to touch page->_count against each page.
1265  * But we had to alter page->flags anyway.
1266  */
1267
1268 static void move_active_pages_to_lru(struct zone *zone,
1269                                      struct list_head *list,
1270                                      enum lru_list lru)
1271 {
1272         unsigned long pgmoved = 0;
1273         struct pagevec pvec;
1274         struct page *page;
1275
1276         pagevec_init(&pvec, 1);
1277
1278         while (!list_empty(list)) {
1279                 page = lru_to_page(list);
1280
1281                 VM_BUG_ON(PageLRU(page));
1282                 SetPageLRU(page);
1283
1284                 list_move(&page->lru, &zone->lru[lru].list);
1285                 mem_cgroup_add_lru_list(page, lru);
1286                 pgmoved++;
1287
1288                 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1289                         spin_unlock_irq(&zone->lru_lock);
1290                         if (buffer_heads_over_limit)
1291                                 pagevec_strip(&pvec);
1292                         __pagevec_release(&pvec);
1293                         spin_lock_irq(&zone->lru_lock);
1294                 }
1295         }
1296         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1297         if (!is_active_lru(lru))
1298                 __count_vm_events(PGDEACTIVATE, pgmoved);
1299 }
1300
1301 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1302                         struct scan_control *sc, int priority, int file)
1303 {
1304         unsigned long nr_taken;
1305         unsigned long pgscanned;
1306         unsigned long vm_flags;
1307         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1308         LIST_HEAD(l_active);
1309         LIST_HEAD(l_inactive);
1310         struct page *page;
1311         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1312         unsigned long nr_rotated = 0;
1313
1314         lru_add_drain();
1315         spin_lock_irq(&zone->lru_lock);
1316         nr_taken = sc->isolate_pages(nr_pages, &l_hold, &pgscanned, sc->order,
1317                                         ISOLATE_ACTIVE, zone,
1318                                         sc->mem_cgroup, 1, file);
1319         /*
1320          * zone->pages_scanned is used for detect zone's oom
1321          * mem_cgroup remembers nr_scan by itself.
1322          */
1323         if (scanning_global_lru(sc)) {
1324                 zone->pages_scanned += pgscanned;
1325         }
1326         reclaim_stat->recent_scanned[file] += nr_taken;
1327
1328         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1329         if (file)
1330                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1331         else
1332                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1333         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1334         spin_unlock_irq(&zone->lru_lock);
1335
1336         while (!list_empty(&l_hold)) {
1337                 cond_resched();
1338                 page = lru_to_page(&l_hold);
1339                 list_del(&page->lru);
1340
1341                 if (unlikely(!page_evictable(page, NULL))) {
1342                         putback_lru_page(page);
1343                         continue;
1344                 }
1345
1346                 /* page_referenced clears PageReferenced */
1347                 if (page_mapping_inuse(page) &&
1348                     page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1349                         nr_rotated++;
1350                         /*
1351                          * Identify referenced, file-backed active pages and
1352                          * give them one more trip around the active list. So
1353                          * that executable code get better chances to stay in
1354                          * memory under moderate memory pressure.  Anon pages
1355                          * are not likely to be evicted by use-once streaming
1356                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1357                          * so we ignore them here.
1358                          */
1359                         if ((vm_flags & VM_EXEC) && !PageAnon(page)) {
1360                                 list_add(&page->lru, &l_active);
1361                                 continue;
1362                         }
1363                 }
1364
1365                 ClearPageActive(page);  /* we are de-activating */
1366                 list_add(&page->lru, &l_inactive);
1367         }
1368
1369         /*
1370          * Move pages back to the lru list.
1371          */
1372         spin_lock_irq(&zone->lru_lock);
1373         /*
1374          * Count referenced pages from currently used mappings as rotated,
1375          * even though only some of them are actually re-activated.  This
1376          * helps balance scan pressure between file and anonymous pages in
1377          * get_scan_ratio.
1378          */
1379         reclaim_stat->recent_rotated[file] += nr_rotated;
1380
1381         move_active_pages_to_lru(zone, &l_active,
1382                                                 LRU_ACTIVE + file * LRU_FILE);
1383         move_active_pages_to_lru(zone, &l_inactive,
1384                                                 LRU_BASE   + file * LRU_FILE);
1385         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1386         spin_unlock_irq(&zone->lru_lock);
1387 }
1388
1389 static int inactive_anon_is_low_global(struct zone *zone)
1390 {
1391         unsigned long active, inactive;
1392
1393         active = zone_page_state(zone, NR_ACTIVE_ANON);
1394         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1395
1396         if (inactive * zone->inactive_ratio < active)
1397                 return 1;
1398
1399         return 0;
1400 }
1401
1402 /**
1403  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1404  * @zone: zone to check
1405  * @sc:   scan control of this context
1406  *
1407  * Returns true if the zone does not have enough inactive anon pages,
1408  * meaning some active anon pages need to be deactivated.
1409  */
1410 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1411 {
1412         int low;
1413
1414         if (scanning_global_lru(sc))
1415                 low = inactive_anon_is_low_global(zone);
1416         else
1417                 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1418         return low;
1419 }
1420
1421 static int inactive_file_is_low_global(struct zone *zone)
1422 {
1423         unsigned long active, inactive;
1424
1425         active = zone_page_state(zone, NR_ACTIVE_FILE);
1426         inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1427
1428         return (active > inactive);
1429 }
1430
1431 /**
1432  * inactive_file_is_low - check if file pages need to be deactivated
1433  * @zone: zone to check
1434  * @sc:   scan control of this context
1435  *
1436  * When the system is doing streaming IO, memory pressure here
1437  * ensures that active file pages get deactivated, until more
1438  * than half of the file pages are on the inactive list.
1439  *
1440  * Once we get to that situation, protect the system's working
1441  * set from being evicted by disabling active file page aging.
1442  *
1443  * This uses a different ratio than the anonymous pages, because
1444  * the page cache uses a use-once replacement algorithm.
1445  */
1446 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1447 {
1448         int low;
1449
1450         if (scanning_global_lru(sc))
1451                 low = inactive_file_is_low_global(zone);
1452         else
1453                 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1454         return low;
1455 }
1456
1457 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1458         struct zone *zone, struct scan_control *sc, int priority)
1459 {
1460         int file = is_file_lru(lru);
1461
1462         if (lru == LRU_ACTIVE_FILE && inactive_file_is_low(zone, sc)) {
1463                 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1464                 return 0;
1465         }
1466
1467         if (lru == LRU_ACTIVE_ANON && inactive_anon_is_low(zone, sc)) {
1468                 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1469                 return 0;
1470         }
1471         return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1472 }
1473
1474 /*
1475  * Determine how aggressively the anon and file LRU lists should be
1476  * scanned.  The relative value of each set of LRU lists is determined
1477  * by looking at the fraction of the pages scanned we did rotate back
1478  * onto the active list instead of evict.
1479  *
1480  * percent[0] specifies how much pressure to put on ram/swap backed
1481  * memory, while percent[1] determines pressure on the file LRUs.
1482  */
1483 static void get_scan_ratio(struct zone *zone, struct scan_control *sc,
1484                                         unsigned long *percent)
1485 {
1486         unsigned long anon, file, free;
1487         unsigned long anon_prio, file_prio;
1488         unsigned long ap, fp;
1489         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1490
1491         anon  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1492                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1493         file  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1494                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1495
1496         if (scanning_global_lru(sc)) {
1497                 free  = zone_page_state(zone, NR_FREE_PAGES);
1498                 /* If we have very few page cache pages,
1499                    force-scan anon pages. */
1500                 if (unlikely(file + free <= high_wmark_pages(zone))) {
1501                         percent[0] = 100;
1502                         percent[1] = 0;
1503                         return;
1504                 }
1505         }
1506
1507         /*
1508          * OK, so we have swap space and a fair amount of page cache
1509          * pages.  We use the recently rotated / recently scanned
1510          * ratios to determine how valuable each cache is.
1511          *
1512          * Because workloads change over time (and to avoid overflow)
1513          * we keep these statistics as a floating average, which ends
1514          * up weighing recent references more than old ones.
1515          *
1516          * anon in [0], file in [1]
1517          */
1518         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1519                 spin_lock_irq(&zone->lru_lock);
1520                 reclaim_stat->recent_scanned[0] /= 2;
1521                 reclaim_stat->recent_rotated[0] /= 2;
1522                 spin_unlock_irq(&zone->lru_lock);
1523         }
1524
1525         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1526                 spin_lock_irq(&zone->lru_lock);
1527                 reclaim_stat->recent_scanned[1] /= 2;
1528                 reclaim_stat->recent_rotated[1] /= 2;
1529                 spin_unlock_irq(&zone->lru_lock);
1530         }
1531
1532         /*
1533          * With swappiness at 100, anonymous and file have the same priority.
1534          * This scanning priority is essentially the inverse of IO cost.
1535          */
1536         anon_prio = sc->swappiness;
1537         file_prio = 200 - sc->swappiness;
1538
1539         /*
1540          * The amount of pressure on anon vs file pages is inversely
1541          * proportional to the fraction of recently scanned pages on
1542          * each list that were recently referenced and in active use.
1543          */
1544         ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1545         ap /= reclaim_stat->recent_rotated[0] + 1;
1546
1547         fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1548         fp /= reclaim_stat->recent_rotated[1] + 1;
1549
1550         /* Normalize to percentages */
1551         percent[0] = 100 * ap / (ap + fp + 1);
1552         percent[1] = 100 - percent[0];
1553 }
1554
1555 /*
1556  * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1557  * until we collected @swap_cluster_max pages to scan.
1558  */
1559 static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
1560                                        unsigned long *nr_saved_scan,
1561                                        unsigned long swap_cluster_max)
1562 {
1563         unsigned long nr;
1564
1565         *nr_saved_scan += nr_to_scan;
1566         nr = *nr_saved_scan;
1567
1568         if (nr >= swap_cluster_max)
1569                 *nr_saved_scan = 0;
1570         else
1571                 nr = 0;
1572
1573         return nr;
1574 }
1575
1576 /*
1577  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1578  */
1579 static void shrink_zone(int priority, struct zone *zone,
1580                                 struct scan_control *sc)
1581 {
1582         unsigned long nr[NR_LRU_LISTS];
1583         unsigned long nr_to_scan;
1584         unsigned long percent[2];       /* anon @ 0; file @ 1 */
1585         enum lru_list l;
1586         unsigned long nr_reclaimed = sc->nr_reclaimed;
1587         unsigned long swap_cluster_max = sc->swap_cluster_max;
1588         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1589         int noswap = 0;
1590
1591         /* If we have no swap space, do not bother scanning anon pages. */
1592         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1593                 noswap = 1;
1594                 percent[0] = 0;
1595                 percent[1] = 100;
1596         } else
1597                 get_scan_ratio(zone, sc, percent);
1598
1599         for_each_evictable_lru(l) {
1600                 int file = is_file_lru(l);
1601                 unsigned long scan;
1602
1603                 scan = zone_nr_lru_pages(zone, sc, l);
1604                 if (priority || noswap) {
1605                         scan >>= priority;
1606                         scan = (scan * percent[file]) / 100;
1607                 }
1608                 nr[l] = nr_scan_try_batch(scan,
1609                                           &reclaim_stat->nr_saved_scan[l],
1610                                           swap_cluster_max);
1611         }
1612
1613         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1614                                         nr[LRU_INACTIVE_FILE]) {
1615                 for_each_evictable_lru(l) {
1616                         if (nr[l]) {
1617                                 nr_to_scan = min(nr[l], swap_cluster_max);
1618                                 nr[l] -= nr_to_scan;
1619
1620                                 nr_reclaimed += shrink_list(l, nr_to_scan,
1621                                                             zone, sc, priority);
1622                         }
1623                 }
1624                 /*
1625                  * On large memory systems, scan >> priority can become
1626                  * really large. This is fine for the starting priority;
1627                  * we want to put equal scanning pressure on each zone.
1628                  * However, if the VM has a harder time of freeing pages,
1629                  * with multiple processes reclaiming pages, the total
1630                  * freeing target can get unreasonably large.
1631                  */
1632                 if (nr_reclaimed > swap_cluster_max &&
1633                         priority < DEF_PRIORITY && !current_is_kswapd())
1634                         break;
1635         }
1636
1637         sc->nr_reclaimed = nr_reclaimed;
1638
1639         /*
1640          * Even if we did not try to evict anon pages at all, we want to
1641          * rebalance the anon lru active/inactive ratio.
1642          */
1643         if (inactive_anon_is_low(zone, sc) && nr_swap_pages > 0)
1644                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1645
1646         throttle_vm_writeout(sc->gfp_mask);
1647 }
1648
1649 /*
1650  * This is the direct reclaim path, for page-allocating processes.  We only
1651  * try to reclaim pages from zones which will satisfy the caller's allocation
1652  * request.
1653  *
1654  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1655  * Because:
1656  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1657  *    allocation or
1658  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1659  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1660  *    zone defense algorithm.
1661  *
1662  * If a zone is deemed to be full of pinned pages then just give it a light
1663  * scan then give up on it.
1664  */
1665 static void shrink_zones(int priority, struct zonelist *zonelist,
1666                                         struct scan_control *sc)
1667 {
1668         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1669         struct zoneref *z;
1670         struct zone *zone;
1671
1672         sc->all_unreclaimable = 1;
1673         for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1674                                         sc->nodemask) {
1675                 if (!populated_zone(zone))
1676                         continue;
1677                 /*
1678                  * Take care memory controller reclaiming has small influence
1679                  * to global LRU.
1680                  */
1681                 if (scanning_global_lru(sc)) {
1682                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1683                                 continue;
1684                         note_zone_scanning_priority(zone, priority);
1685
1686                         if (zone_is_all_unreclaimable(zone) &&
1687                                                 priority != DEF_PRIORITY)
1688                                 continue;       /* Let kswapd poll it */
1689                         sc->all_unreclaimable = 0;
1690                 } else {
1691                         /*
1692                          * Ignore cpuset limitation here. We just want to reduce
1693                          * # of used pages by us regardless of memory shortage.
1694                          */
1695                         sc->all_unreclaimable = 0;
1696                         mem_cgroup_note_reclaim_priority(sc->mem_cgroup,
1697                                                         priority);
1698                 }
1699
1700                 shrink_zone(priority, zone, sc);
1701         }
1702 }
1703
1704 /*
1705  * This is the main entry point to direct page reclaim.
1706  *
1707  * If a full scan of the inactive list fails to free enough memory then we
1708  * are "out of memory" and something needs to be killed.
1709  *
1710  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1711  * high - the zone may be full of dirty or under-writeback pages, which this
1712  * caller can't do much about.  We kick pdflush and take explicit naps in the
1713  * hope that some of these pages can be written.  But if the allocating task
1714  * holds filesystem locks which prevent writeout this might not work, and the
1715  * allocation attempt will fail.
1716  *
1717  * returns:     0, if no pages reclaimed
1718  *              else, the number of pages reclaimed
1719  */
1720 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
1721                                         struct scan_control *sc)
1722 {
1723         int priority;
1724         unsigned long ret = 0;
1725         unsigned long total_scanned = 0;
1726         struct reclaim_state *reclaim_state = current->reclaim_state;
1727         unsigned long lru_pages = 0;
1728         struct zoneref *z;
1729         struct zone *zone;
1730         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1731
1732         delayacct_freepages_start();
1733
1734         if (scanning_global_lru(sc))
1735                 count_vm_event(ALLOCSTALL);
1736         /*
1737          * mem_cgroup will not do shrink_slab.
1738          */
1739         if (scanning_global_lru(sc)) {
1740                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1741
1742                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1743                                 continue;
1744
1745                         lru_pages += zone_reclaimable_pages(zone);
1746                 }
1747         }
1748
1749         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1750                 sc->nr_scanned = 0;
1751                 if (!priority)
1752                         disable_swap_token();
1753                 shrink_zones(priority, zonelist, sc);
1754                 /*
1755                  * Don't shrink slabs when reclaiming memory from
1756                  * over limit cgroups
1757                  */
1758                 if (scanning_global_lru(sc)) {
1759                         shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
1760                         if (reclaim_state) {
1761                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
1762                                 reclaim_state->reclaimed_slab = 0;
1763                         }
1764                 }
1765                 total_scanned += sc->nr_scanned;
1766                 if (sc->nr_reclaimed >= sc->swap_cluster_max) {
1767                         ret = sc->nr_reclaimed;
1768                         goto out;
1769                 }
1770
1771                 /*
1772                  * Try to write back as many pages as we just scanned.  This
1773                  * tends to cause slow streaming writers to write data to the
1774                  * disk smoothly, at the dirtying rate, which is nice.   But
1775                  * that's undesirable in laptop mode, where we *want* lumpy
1776                  * writeout.  So in laptop mode, write out the whole world.
1777                  */
1778                 if (total_scanned > sc->swap_cluster_max +
1779                                         sc->swap_cluster_max / 2) {
1780                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
1781                         sc->may_writepage = 1;
1782                 }
1783
1784                 /* Take a nap, wait for some writeback to complete */
1785                 if (sc->nr_scanned && priority < DEF_PRIORITY - 2)
1786                         congestion_wait(BLK_RW_ASYNC, HZ/10);
1787         }
1788         /* top priority shrink_zones still had more to do? don't OOM, then */
1789         if (!sc->all_unreclaimable && scanning_global_lru(sc))
1790                 ret = sc->nr_reclaimed;
1791 out:
1792         /*
1793          * Now that we've scanned all the zones at this priority level, note
1794          * that level within the zone so that the next thread which performs
1795          * scanning of this zone will immediately start out at this priority
1796          * level.  This affects only the decision whether or not to bring
1797          * mapped pages onto the inactive list.
1798          */
1799         if (priority < 0)
1800                 priority = 0;
1801
1802         if (scanning_global_lru(sc)) {
1803                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1804
1805                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1806                                 continue;
1807
1808                         zone->prev_priority = priority;
1809                 }
1810         } else
1811                 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority);
1812
1813         delayacct_freepages_end();
1814
1815         return ret;
1816 }
1817
1818 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
1819                                 gfp_t gfp_mask, nodemask_t *nodemask)
1820 {
1821         struct scan_control sc = {
1822                 .gfp_mask = gfp_mask,
1823                 .may_writepage = !laptop_mode,
1824                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1825                 .may_unmap = 1,
1826                 .may_swap = 1,
1827                 .swappiness = vm_swappiness,
1828                 .order = order,
1829                 .mem_cgroup = NULL,
1830                 .isolate_pages = isolate_pages_global,
1831                 .nodemask = nodemask,
1832         };
1833
1834         return do_try_to_free_pages(zonelist, &sc);
1835 }
1836
1837 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
1838
1839 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
1840                                            gfp_t gfp_mask,
1841                                            bool noswap,
1842                                            unsigned int swappiness)
1843 {
1844         struct scan_control sc = {
1845                 .may_writepage = !laptop_mode,
1846                 .may_unmap = 1,
1847                 .may_swap = !noswap,
1848                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1849                 .swappiness = swappiness,
1850                 .order = 0,
1851                 .mem_cgroup = mem_cont,
1852                 .isolate_pages = mem_cgroup_isolate_pages,
1853                 .nodemask = NULL, /* we don't care the placement */
1854         };
1855         struct zonelist *zonelist;
1856
1857         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1858                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1859         zonelist = NODE_DATA(numa_node_id())->node_zonelists;
1860         return do_try_to_free_pages(zonelist, &sc);
1861 }
1862 #endif
1863
1864 /*
1865  * For kswapd, balance_pgdat() will work across all this node's zones until
1866  * they are all at high_wmark_pages(zone).
1867  *
1868  * Returns the number of pages which were actually freed.
1869  *
1870  * There is special handling here for zones which are full of pinned pages.
1871  * This can happen if the pages are all mlocked, or if they are all used by
1872  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1873  * What we do is to detect the case where all pages in the zone have been
1874  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1875  * dead and from now on, only perform a short scan.  Basically we're polling
1876  * the zone for when the problem goes away.
1877  *
1878  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
1879  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
1880  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
1881  * lower zones regardless of the number of free pages in the lower zones. This
1882  * interoperates with the page allocator fallback scheme to ensure that aging
1883  * of pages is balanced across the zones.
1884  */
1885 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
1886 {
1887         int all_zones_ok;
1888         int priority;
1889         int i;
1890         unsigned long total_scanned;
1891         struct reclaim_state *reclaim_state = current->reclaim_state;
1892         struct scan_control sc = {
1893                 .gfp_mask = GFP_KERNEL,
1894                 .may_unmap = 1,
1895                 .may_swap = 1,
1896                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1897                 .swappiness = vm_swappiness,
1898                 .order = order,
1899                 .mem_cgroup = NULL,
1900                 .isolate_pages = isolate_pages_global,
1901         };
1902         /*
1903          * temp_priority is used to remember the scanning priority at which
1904          * this zone was successfully refilled to
1905          * free_pages == high_wmark_pages(zone).
1906          */
1907         int temp_priority[MAX_NR_ZONES];
1908
1909 loop_again:
1910         total_scanned = 0;
1911         sc.nr_reclaimed = 0;
1912         sc.may_writepage = !laptop_mode;
1913         count_vm_event(PAGEOUTRUN);
1914
1915         for (i = 0; i < pgdat->nr_zones; i++)
1916                 temp_priority[i] = DEF_PRIORITY;
1917
1918         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1919                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1920                 unsigned long lru_pages = 0;
1921
1922                 /* The swap token gets in the way of swapout... */
1923                 if (!priority)
1924                         disable_swap_token();
1925
1926                 all_zones_ok = 1;
1927
1928                 /*
1929                  * Scan in the highmem->dma direction for the highest
1930                  * zone which needs scanning
1931                  */
1932                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1933                         struct zone *zone = pgdat->node_zones + i;
1934
1935                         if (!populated_zone(zone))
1936                                 continue;
1937
1938                         if (zone_is_all_unreclaimable(zone) &&
1939                             priority != DEF_PRIORITY)
1940                                 continue;
1941
1942                         /*
1943                          * Do some background aging of the anon list, to give
1944                          * pages a chance to be referenced before reclaiming.
1945                          */
1946                         if (inactive_anon_is_low(zone, &sc))
1947                                 shrink_active_list(SWAP_CLUSTER_MAX, zone,
1948                                                         &sc, priority, 0);
1949
1950                         if (!zone_watermark_ok(zone, order,
1951                                         high_wmark_pages(zone), 0, 0)) {
1952                                 end_zone = i;
1953                                 break;
1954                         }
1955                 }
1956                 if (i < 0)
1957                         goto out;
1958
1959                 for (i = 0; i <= end_zone; i++) {
1960                         struct zone *zone = pgdat->node_zones + i;
1961
1962                         lru_pages += zone_reclaimable_pages(zone);
1963                 }
1964
1965                 /*
1966                  * Now scan the zone in the dma->highmem direction, stopping
1967                  * at the last zone which needs scanning.
1968                  *
1969                  * We do this because the page allocator works in the opposite
1970                  * direction.  This prevents the page allocator from allocating
1971                  * pages behind kswapd's direction of progress, which would
1972                  * cause too much scanning of the lower zones.
1973                  */
1974                 for (i = 0; i <= end_zone; i++) {
1975                         struct zone *zone = pgdat->node_zones + i;
1976                         int nr_slab;
1977
1978                         if (!populated_zone(zone))
1979                                 continue;
1980
1981                         if (zone_is_all_unreclaimable(zone) &&
1982                                         priority != DEF_PRIORITY)
1983                                 continue;
1984
1985                         if (!zone_watermark_ok(zone, order,
1986                                         high_wmark_pages(zone), end_zone, 0))
1987                                 all_zones_ok = 0;
1988                         temp_priority[i] = priority;
1989                         sc.nr_scanned = 0;
1990                         note_zone_scanning_priority(zone, priority);
1991                         /*
1992                          * We put equal pressure on every zone, unless one
1993                          * zone has way too many pages free already.
1994                          */
1995                         if (!zone_watermark_ok(zone, order,
1996                                         8*high_wmark_pages(zone), end_zone, 0))
1997                                 shrink_zone(priority, zone, &sc);
1998                         reclaim_state->reclaimed_slab = 0;
1999                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
2000                                                 lru_pages);
2001                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2002                         total_scanned += sc.nr_scanned;
2003                         if (zone_is_all_unreclaimable(zone))
2004                                 continue;
2005                         if (nr_slab == 0 && zone->pages_scanned >=
2006                                         (zone_reclaimable_pages(zone) * 6))
2007                                         zone_set_flag(zone,
2008                                                       ZONE_ALL_UNRECLAIMABLE);
2009                         /*
2010                          * If we've done a decent amount of scanning and
2011                          * the reclaim ratio is low, start doing writepage
2012                          * even in laptop mode
2013                          */
2014                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2015                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2016                                 sc.may_writepage = 1;
2017                 }
2018                 if (all_zones_ok)
2019                         break;          /* kswapd: all done */
2020                 /*
2021                  * OK, kswapd is getting into trouble.  Take a nap, then take
2022                  * another pass across the zones.
2023                  */
2024                 if (total_scanned && priority < DEF_PRIORITY - 2)
2025                         congestion_wait(BLK_RW_ASYNC, HZ/10);
2026
2027                 /*
2028                  * We do this so kswapd doesn't build up large priorities for
2029                  * example when it is freeing in parallel with allocators. It
2030                  * matches the direct reclaim path behaviour in terms of impact
2031                  * on zone->*_priority.
2032                  */
2033                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2034                         break;
2035         }
2036 out:
2037         /*
2038          * Note within each zone the priority level at which this zone was
2039          * brought into a happy state.  So that the next thread which scans this
2040          * zone will start out at that priority level.
2041          */
2042         for (i = 0; i < pgdat->nr_zones; i++) {
2043                 struct zone *zone = pgdat->node_zones + i;
2044
2045                 zone->prev_priority = temp_priority[i];
2046         }
2047         if (!all_zones_ok) {
2048                 cond_resched();
2049
2050                 try_to_freeze();
2051
2052                 /*
2053                  * Fragmentation may mean that the system cannot be
2054                  * rebalanced for high-order allocations in all zones.
2055                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2056                  * it means the zones have been fully scanned and are still
2057                  * not balanced. For high-order allocations, there is
2058                  * little point trying all over again as kswapd may
2059                  * infinite loop.
2060                  *
2061                  * Instead, recheck all watermarks at order-0 as they
2062                  * are the most important. If watermarks are ok, kswapd will go
2063                  * back to sleep. High-order users can still perform direct
2064                  * reclaim if they wish.
2065                  */
2066                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2067                         order = sc.order = 0;
2068
2069                 goto loop_again;
2070         }
2071
2072         return sc.nr_reclaimed;
2073 }
2074
2075 /*
2076  * The background pageout daemon, started as a kernel thread
2077  * from the init process.
2078  *
2079  * This basically trickles out pages so that we have _some_
2080  * free memory available even if there is no other activity
2081  * that frees anything up. This is needed for things like routing
2082  * etc, where we otherwise might have all activity going on in
2083  * asynchronous contexts that cannot page things out.
2084  *
2085  * If there are applications that are active memory-allocators
2086  * (most normal use), this basically shouldn't matter.
2087  */
2088 static int kswapd(void *p)
2089 {
2090         unsigned long order;
2091         pg_data_t *pgdat = (pg_data_t*)p;
2092         struct task_struct *tsk = current;
2093         DEFINE_WAIT(wait);
2094         struct reclaim_state reclaim_state = {
2095                 .reclaimed_slab = 0,
2096         };
2097         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2098
2099         lockdep_set_current_reclaim_state(GFP_KERNEL);
2100
2101         if (!cpumask_empty(cpumask))
2102                 set_cpus_allowed_ptr(tsk, cpumask);
2103         current->reclaim_state = &reclaim_state;
2104
2105         /*
2106          * Tell the memory management that we're a "memory allocator",
2107          * and that if we need more memory we should get access to it
2108          * regardless (see "__alloc_pages()"). "kswapd" should
2109          * never get caught in the normal page freeing logic.
2110          *
2111          * (Kswapd normally doesn't need memory anyway, but sometimes
2112          * you need a small amount of memory in order to be able to
2113          * page out something else, and this flag essentially protects
2114          * us from recursively trying to free more memory as we're
2115          * trying to free the first piece of memory in the first place).
2116          */
2117         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2118         set_freezable();
2119
2120         order = 0;
2121         for ( ; ; ) {
2122                 unsigned long new_order;
2123
2124                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2125                 new_order = pgdat->kswapd_max_order;
2126                 pgdat->kswapd_max_order = 0;
2127                 if (order < new_order) {
2128                         /*
2129                          * Don't sleep if someone wants a larger 'order'
2130                          * allocation
2131                          */
2132                         order = new_order;
2133                 } else {
2134                         if (!freezing(current))
2135                                 schedule();
2136
2137                         order = pgdat->kswapd_max_order;
2138                 }
2139                 finish_wait(&pgdat->kswapd_wait, &wait);
2140
2141                 if (!try_to_freeze()) {
2142                         /* We can speed up thawing tasks if we don't call
2143                          * balance_pgdat after returning from the refrigerator
2144                          */
2145                         balance_pgdat(pgdat, order);
2146                 }
2147         }
2148         return 0;
2149 }
2150
2151 /*
2152  * A zone is low on free memory, so wake its kswapd task to service it.
2153  */
2154 void wakeup_kswapd(struct zone *zone, int order)
2155 {
2156         pg_data_t *pgdat;
2157
2158         if (!populated_zone(zone))
2159                 return;
2160
2161         pgdat = zone->zone_pgdat;
2162         if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
2163                 return;
2164         if (pgdat->kswapd_max_order < order)
2165                 pgdat->kswapd_max_order = order;
2166         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2167                 return;
2168         if (!waitqueue_active(&pgdat->kswapd_wait))
2169                 return;
2170         wake_up_interruptible(&pgdat->kswapd_wait);
2171 }
2172
2173 /*
2174  * The reclaimable count would be mostly accurate.
2175  * The less reclaimable pages may be
2176  * - mlocked pages, which will be moved to unevictable list when encountered
2177  * - mapped pages, which may require several travels to be reclaimed
2178  * - dirty pages, which is not "instantly" reclaimable
2179  */
2180 unsigned long global_reclaimable_pages(void)
2181 {
2182         int nr;
2183
2184         nr = global_page_state(NR_ACTIVE_FILE) +
2185              global_page_state(NR_INACTIVE_FILE);
2186
2187         if (nr_swap_pages > 0)
2188                 nr += global_page_state(NR_ACTIVE_ANON) +
2189                       global_page_state(NR_INACTIVE_ANON);
2190
2191         return nr;
2192 }
2193
2194 unsigned long zone_reclaimable_pages(struct zone *zone)
2195 {
2196         int nr;
2197
2198         nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2199              zone_page_state(zone, NR_INACTIVE_FILE);
2200
2201         if (nr_swap_pages > 0)
2202                 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2203                       zone_page_state(zone, NR_INACTIVE_ANON);
2204
2205         return nr;
2206 }
2207
2208 #ifdef CONFIG_HIBERNATION
2209 /*
2210  * Helper function for shrink_all_memory().  Tries to reclaim 'nr_pages' pages
2211  * from LRU lists system-wide, for given pass and priority.
2212  *
2213  * For pass > 3 we also try to shrink the LRU lists that contain a few pages
2214  */
2215 static void shrink_all_zones(unsigned long nr_pages, int prio,
2216                                       int pass, struct scan_control *sc)
2217 {
2218         struct zone *zone;
2219         unsigned long nr_reclaimed = 0;
2220         struct zone_reclaim_stat *reclaim_stat;
2221
2222         for_each_populated_zone(zone) {
2223                 enum lru_list l;
2224
2225                 if (zone_is_all_unreclaimable(zone) && prio != DEF_PRIORITY)
2226                         continue;
2227
2228                 for_each_evictable_lru(l) {
2229                         enum zone_stat_item ls = NR_LRU_BASE + l;
2230                         unsigned long lru_pages = zone_page_state(zone, ls);
2231
2232                         /* For pass = 0, we don't shrink the active list */
2233                         if (pass == 0 && (l == LRU_ACTIVE_ANON ||
2234                                                 l == LRU_ACTIVE_FILE))
2235                                 continue;
2236
2237                         reclaim_stat = get_reclaim_stat(zone, sc);
2238                         reclaim_stat->nr_saved_scan[l] +=
2239                                                 (lru_pages >> prio) + 1;
2240                         if (reclaim_stat->nr_saved_scan[l]
2241                                                 >= nr_pages || pass > 3) {
2242                                 unsigned long nr_to_scan;
2243
2244                                 reclaim_stat->nr_saved_scan[l] = 0;
2245                                 nr_to_scan = min(nr_pages, lru_pages);
2246                                 nr_reclaimed += shrink_list(l, nr_to_scan, zone,
2247                                                                 sc, prio);
2248                                 if (nr_reclaimed >= nr_pages) {
2249                                         sc->nr_reclaimed += nr_reclaimed;
2250                                         return;
2251                                 }
2252                         }
2253                 }
2254         }
2255         sc->nr_reclaimed += nr_reclaimed;
2256 }
2257
2258 /*
2259  * Try to free `nr_pages' of memory, system-wide, and return the number of
2260  * freed pages.
2261  *
2262  * Rather than trying to age LRUs the aim is to preserve the overall
2263  * LRU order by reclaiming preferentially
2264  * inactive > active > active referenced > active mapped
2265  */
2266 unsigned long shrink_all_memory(unsigned long nr_pages)
2267 {
2268         unsigned long lru_pages, nr_slab;
2269         int pass;
2270         struct reclaim_state reclaim_state;
2271         struct scan_control sc = {
2272                 .gfp_mask = GFP_KERNEL,
2273                 .may_unmap = 0,
2274                 .may_writepage = 1,
2275                 .isolate_pages = isolate_pages_global,
2276                 .nr_reclaimed = 0,
2277         };
2278
2279         current->reclaim_state = &reclaim_state;
2280
2281         lru_pages = global_reclaimable_pages();
2282         nr_slab = global_page_state(NR_SLAB_RECLAIMABLE);
2283         /* If slab caches are huge, it's better to hit them first */
2284         while (nr_slab >= lru_pages) {
2285                 reclaim_state.reclaimed_slab = 0;
2286                 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
2287                 if (!reclaim_state.reclaimed_slab)
2288                         break;
2289
2290                 sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2291                 if (sc.nr_reclaimed >= nr_pages)
2292                         goto out;
2293
2294                 nr_slab -= reclaim_state.reclaimed_slab;
2295         }
2296
2297         /*
2298          * We try to shrink LRUs in 5 passes:
2299          * 0 = Reclaim from inactive_list only
2300          * 1 = Reclaim from active list but don't reclaim mapped
2301          * 2 = 2nd pass of type 1
2302          * 3 = Reclaim mapped (normal reclaim)
2303          * 4 = 2nd pass of type 3
2304          */
2305         for (pass = 0; pass < 5; pass++) {
2306                 int prio;
2307
2308                 /* Force reclaiming mapped pages in the passes #3 and #4 */
2309                 if (pass > 2)
2310                         sc.may_unmap = 1;
2311
2312                 for (prio = DEF_PRIORITY; prio >= 0; prio--) {
2313                         unsigned long nr_to_scan = nr_pages - sc.nr_reclaimed;
2314
2315                         sc.nr_scanned = 0;
2316                         sc.swap_cluster_max = nr_to_scan;
2317                         shrink_all_zones(nr_to_scan, prio, pass, &sc);
2318                         if (sc.nr_reclaimed >= nr_pages)
2319                                 goto out;
2320
2321                         reclaim_state.reclaimed_slab = 0;
2322                         shrink_slab(sc.nr_scanned, sc.gfp_mask,
2323                                     global_reclaimable_pages());
2324                         sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2325                         if (sc.nr_reclaimed >= nr_pages)
2326                                 goto out;
2327
2328                         if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
2329                                 congestion_wait(BLK_RW_ASYNC, HZ / 10);
2330                 }
2331         }
2332
2333         /*
2334          * If sc.nr_reclaimed = 0, we could not shrink LRUs, but there may be
2335          * something in slab caches
2336          */
2337         if (!sc.nr_reclaimed) {
2338                 do {
2339                         reclaim_state.reclaimed_slab = 0;
2340                         shrink_slab(nr_pages, sc.gfp_mask,
2341                                     global_reclaimable_pages());
2342                         sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2343                 } while (sc.nr_reclaimed < nr_pages &&
2344                                 reclaim_state.reclaimed_slab > 0);
2345         }
2346
2347
2348 out:
2349         current->reclaim_state = NULL;
2350
2351         return sc.nr_reclaimed;
2352 }
2353 #endif /* CONFIG_HIBERNATION */
2354
2355 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2356    not required for correctness.  So if the last cpu in a node goes
2357    away, we get changed to run anywhere: as the first one comes back,
2358    restore their cpu bindings. */
2359 static int __devinit cpu_callback(struct notifier_block *nfb,
2360                                   unsigned long action, void *hcpu)
2361 {
2362         int nid;
2363
2364         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2365                 for_each_node_state(nid, N_HIGH_MEMORY) {
2366                         pg_data_t *pgdat = NODE_DATA(nid);
2367                         const struct cpumask *mask;
2368
2369                         mask = cpumask_of_node(pgdat->node_id);
2370
2371                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2372                                 /* One of our CPUs online: restore mask */
2373                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2374                 }
2375         }
2376         return NOTIFY_OK;
2377 }
2378
2379 /*
2380  * This kswapd start function will be called by init and node-hot-add.
2381  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2382  */
2383 int kswapd_run(int nid)
2384 {
2385         pg_data_t *pgdat = NODE_DATA(nid);
2386         int ret = 0;
2387
2388         if (pgdat->kswapd)
2389                 return 0;
2390
2391         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2392         if (IS_ERR(pgdat->kswapd)) {
2393                 /* failure at boot is fatal */
2394                 BUG_ON(system_state == SYSTEM_BOOTING);
2395                 printk("Failed to start kswapd on node %d\n",nid);
2396                 ret = -1;
2397         }
2398         return ret;
2399 }
2400
2401 static int __init kswapd_init(void)
2402 {
2403         int nid;
2404
2405         swap_setup();
2406         for_each_node_state(nid, N_HIGH_MEMORY)
2407                 kswapd_run(nid);
2408         hotcpu_notifier(cpu_callback, 0);
2409         return 0;
2410 }
2411
2412 module_init(kswapd_init)
2413
2414 #ifdef CONFIG_NUMA
2415 /*
2416  * Zone reclaim mode
2417  *
2418  * If non-zero call zone_reclaim when the number of free pages falls below
2419  * the watermarks.
2420  */
2421 int zone_reclaim_mode __read_mostly;
2422
2423 #define RECLAIM_OFF 0
2424 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
2425 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
2426 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
2427
2428 /*
2429  * Priority for ZONE_RECLAIM. This determines the fraction of pages
2430  * of a node considered for each zone_reclaim. 4 scans 1/16th of
2431  * a zone.
2432  */
2433 #define ZONE_RECLAIM_PRIORITY 4
2434
2435 /*
2436  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2437  * occur.
2438  */
2439 int sysctl_min_unmapped_ratio = 1;
2440
2441 /*
2442  * If the number of slab pages in a zone grows beyond this percentage then
2443  * slab reclaim needs to occur.
2444  */
2445 int sysctl_min_slab_ratio = 5;
2446
2447 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
2448 {
2449         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
2450         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
2451                 zone_page_state(zone, NR_ACTIVE_FILE);
2452
2453         /*
2454          * It's possible for there to be more file mapped pages than
2455          * accounted for by the pages on the file LRU lists because
2456          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2457          */
2458         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
2459 }
2460
2461 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
2462 static long zone_pagecache_reclaimable(struct zone *zone)
2463 {
2464         long nr_pagecache_reclaimable;
2465         long delta = 0;
2466
2467         /*
2468          * If RECLAIM_SWAP is set, then all file pages are considered
2469          * potentially reclaimable. Otherwise, we have to worry about
2470          * pages like swapcache and zone_unmapped_file_pages() provides
2471          * a better estimate
2472          */
2473         if (zone_reclaim_mode & RECLAIM_SWAP)
2474                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
2475         else
2476                 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
2477
2478         /* If we can't clean pages, remove dirty pages from consideration */
2479         if (!(zone_reclaim_mode & RECLAIM_WRITE))
2480                 delta += zone_page_state(zone, NR_FILE_DIRTY);
2481
2482         /* Watch for any possible underflows due to delta */
2483         if (unlikely(delta > nr_pagecache_reclaimable))
2484                 delta = nr_pagecache_reclaimable;
2485
2486         return nr_pagecache_reclaimable - delta;
2487 }
2488
2489 /*
2490  * Try to free up some pages from this zone through reclaim.
2491  */
2492 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2493 {
2494         /* Minimum pages needed in order to stay on node */
2495         const unsigned long nr_pages = 1 << order;
2496         struct task_struct *p = current;
2497         struct reclaim_state reclaim_state;
2498         int priority;
2499         struct scan_control sc = {
2500                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2501                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2502                 .may_swap = 1,
2503                 .swap_cluster_max = max_t(unsigned long, nr_pages,
2504                                         SWAP_CLUSTER_MAX),
2505                 .gfp_mask = gfp_mask,
2506                 .swappiness = vm_swappiness,
2507                 .order = order,
2508                 .isolate_pages = isolate_pages_global,
2509         };
2510         unsigned long slab_reclaimable;
2511
2512         disable_swap_token();
2513         cond_resched();
2514         /*
2515          * We need to be able to allocate from the reserves for RECLAIM_SWAP
2516          * and we also need to be able to write out pages for RECLAIM_WRITE
2517          * and RECLAIM_SWAP.
2518          */
2519         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2520         reclaim_state.reclaimed_slab = 0;
2521         p->reclaim_state = &reclaim_state;
2522
2523         if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
2524                 /*
2525                  * Free memory by calling shrink zone with increasing
2526                  * priorities until we have enough memory freed.
2527                  */
2528                 priority = ZONE_RECLAIM_PRIORITY;
2529                 do {
2530                         note_zone_scanning_priority(zone, priority);
2531                         shrink_zone(priority, zone, &sc);
2532                         priority--;
2533                 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
2534         }
2535
2536         slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2537         if (slab_reclaimable > zone->min_slab_pages) {
2538                 /*
2539                  * shrink_slab() does not currently allow us to determine how
2540                  * many pages were freed in this zone. So we take the current
2541                  * number of slab pages and shake the slab until it is reduced
2542                  * by the same nr_pages that we used for reclaiming unmapped
2543                  * pages.
2544                  *
2545                  * Note that shrink_slab will free memory on all zones and may
2546                  * take a long time.
2547                  */
2548                 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
2549                         zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
2550                                 slab_reclaimable - nr_pages)
2551                         ;
2552
2553                 /*
2554                  * Update nr_reclaimed by the number of slab pages we
2555                  * reclaimed from this zone.
2556                  */
2557                 sc.nr_reclaimed += slab_reclaimable -
2558                         zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2559         }
2560
2561         p->reclaim_state = NULL;
2562         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
2563         return sc.nr_reclaimed >= nr_pages;
2564 }
2565
2566 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2567 {
2568         int node_id;
2569         int ret;
2570
2571         /*
2572          * Zone reclaim reclaims unmapped file backed pages and
2573          * slab pages if we are over the defined limits.
2574          *
2575          * A small portion of unmapped file backed pages is needed for
2576          * file I/O otherwise pages read by file I/O will be immediately
2577          * thrown out if the zone is overallocated. So we do not reclaim
2578          * if less than a specified percentage of the zone is used by
2579          * unmapped file backed pages.
2580          */
2581         if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
2582             zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
2583                 return ZONE_RECLAIM_FULL;
2584
2585         if (zone_is_all_unreclaimable(zone))
2586                 return ZONE_RECLAIM_FULL;
2587
2588         /*
2589          * Do not scan if the allocation should not be delayed.
2590          */
2591         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
2592                 return ZONE_RECLAIM_NOSCAN;
2593
2594         /*
2595          * Only run zone reclaim on the local zone or on zones that do not
2596          * have associated processors. This will favor the local processor
2597          * over remote processors and spread off node memory allocations
2598          * as wide as possible.
2599          */
2600         node_id = zone_to_nid(zone);
2601         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
2602                 return ZONE_RECLAIM_NOSCAN;
2603
2604         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2605                 return ZONE_RECLAIM_NOSCAN;
2606
2607         ret = __zone_reclaim(zone, gfp_mask, order);
2608         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2609
2610         if (!ret)
2611                 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
2612
2613         return ret;
2614 }
2615 #endif
2616
2617 /*
2618  * page_evictable - test whether a page is evictable
2619  * @page: the page to test
2620  * @vma: the VMA in which the page is or will be mapped, may be NULL
2621  *
2622  * Test whether page is evictable--i.e., should be placed on active/inactive
2623  * lists vs unevictable list.  The vma argument is !NULL when called from the
2624  * fault path to determine how to instantate a new page.
2625  *
2626  * Reasons page might not be evictable:
2627  * (1) page's mapping marked unevictable
2628  * (2) page is part of an mlocked VMA
2629  *
2630  */
2631 int page_evictable(struct page *page, struct vm_area_struct *vma)
2632 {
2633
2634         if (mapping_unevictable(page_mapping(page)))
2635                 return 0;
2636
2637         if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2638                 return 0;
2639
2640         return 1;
2641 }
2642
2643 /**
2644  * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2645  * @page: page to check evictability and move to appropriate lru list
2646  * @zone: zone page is in
2647  *
2648  * Checks a page for evictability and moves the page to the appropriate
2649  * zone lru list.
2650  *
2651  * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2652  * have PageUnevictable set.
2653  */
2654 static void check_move_unevictable_page(struct page *page, struct zone *zone)
2655 {
2656         VM_BUG_ON(PageActive(page));
2657
2658 retry:
2659         ClearPageUnevictable(page);
2660         if (page_evictable(page, NULL)) {
2661                 enum lru_list l = page_lru_base_type(page);
2662
2663                 __dec_zone_state(zone, NR_UNEVICTABLE);
2664                 list_move(&page->lru, &zone->lru[l].list);
2665                 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
2666                 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
2667                 __count_vm_event(UNEVICTABLE_PGRESCUED);
2668         } else {
2669                 /*
2670                  * rotate unevictable list
2671                  */
2672                 SetPageUnevictable(page);
2673                 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
2674                 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
2675                 if (page_evictable(page, NULL))
2676                         goto retry;
2677         }
2678 }
2679
2680 /**
2681  * scan_mapping_unevictable_pages - scan an address space for evictable pages
2682  * @mapping: struct address_space to scan for evictable pages
2683  *
2684  * Scan all pages in mapping.  Check unevictable pages for
2685  * evictability and move them to the appropriate zone lru list.
2686  */
2687 void scan_mapping_unevictable_pages(struct address_space *mapping)
2688 {
2689         pgoff_t next = 0;
2690         pgoff_t end   = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
2691                          PAGE_CACHE_SHIFT;
2692         struct zone *zone;
2693         struct pagevec pvec;
2694
2695         if (mapping->nrpages == 0)
2696                 return;
2697
2698         pagevec_init(&pvec, 0);
2699         while (next < end &&
2700                 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
2701                 int i;
2702                 int pg_scanned = 0;
2703
2704                 zone = NULL;
2705
2706                 for (i = 0; i < pagevec_count(&pvec); i++) {
2707                         struct page *page = pvec.pages[i];
2708                         pgoff_t page_index = page->index;
2709                         struct zone *pagezone = page_zone(page);
2710
2711                         pg_scanned++;
2712                         if (page_index > next)
2713                                 next = page_index;
2714                         next++;
2715
2716                         if (pagezone != zone) {
2717                                 if (zone)
2718                                         spin_unlock_irq(&zone->lru_lock);
2719                                 zone = pagezone;
2720                                 spin_lock_irq(&zone->lru_lock);
2721                         }
2722
2723                         if (PageLRU(page) && PageUnevictable(page))
2724                                 check_move_unevictable_page(page, zone);
2725                 }
2726                 if (zone)
2727                         spin_unlock_irq(&zone->lru_lock);
2728                 pagevec_release(&pvec);
2729
2730                 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
2731         }
2732
2733 }
2734
2735 /**
2736  * scan_zone_unevictable_pages - check unevictable list for evictable pages
2737  * @zone - zone of which to scan the unevictable list
2738  *
2739  * Scan @zone's unevictable LRU lists to check for pages that have become
2740  * evictable.  Move those that have to @zone's inactive list where they
2741  * become candidates for reclaim, unless shrink_inactive_zone() decides
2742  * to reactivate them.  Pages that are still unevictable are rotated
2743  * back onto @zone's unevictable list.
2744  */
2745 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
2746 static void scan_zone_unevictable_pages(struct zone *zone)
2747 {
2748         struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
2749         unsigned long scan;
2750         unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
2751
2752         while (nr_to_scan > 0) {
2753                 unsigned long batch_size = min(nr_to_scan,
2754                                                 SCAN_UNEVICTABLE_BATCH_SIZE);
2755
2756                 spin_lock_irq(&zone->lru_lock);
2757                 for (scan = 0;  scan < batch_size; scan++) {
2758                         struct page *page = lru_to_page(l_unevictable);
2759
2760                         if (!trylock_page(page))
2761                                 continue;
2762
2763                         prefetchw_prev_lru_page(page, l_unevictable, flags);
2764
2765                         if (likely(PageLRU(page) && PageUnevictable(page)))
2766                                 check_move_unevictable_page(page, zone);
2767
2768                         unlock_page(page);
2769                 }
2770                 spin_unlock_irq(&zone->lru_lock);
2771
2772                 nr_to_scan -= batch_size;
2773         }
2774 }
2775
2776
2777 /**
2778  * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
2779  *
2780  * A really big hammer:  scan all zones' unevictable LRU lists to check for
2781  * pages that have become evictable.  Move those back to the zones'
2782  * inactive list where they become candidates for reclaim.
2783  * This occurs when, e.g., we have unswappable pages on the unevictable lists,
2784  * and we add swap to the system.  As such, it runs in the context of a task
2785  * that has possibly/probably made some previously unevictable pages
2786  * evictable.
2787  */
2788 static void scan_all_zones_unevictable_pages(void)
2789 {
2790         struct zone *zone;
2791
2792         for_each_zone(zone) {
2793                 scan_zone_unevictable_pages(zone);
2794         }
2795 }
2796
2797 /*
2798  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
2799  * all nodes' unevictable lists for evictable pages
2800  */
2801 unsigned long scan_unevictable_pages;
2802
2803 int scan_unevictable_handler(struct ctl_table *table, int write,
2804                            struct file *file, void __user *buffer,
2805                            size_t *length, loff_t *ppos)
2806 {
2807         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
2808
2809         if (write && *(unsigned long *)table->data)
2810                 scan_all_zones_unevictable_pages();
2811
2812         scan_unevictable_pages = 0;
2813         return 0;
2814 }
2815
2816 /*
2817  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
2818  * a specified node's per zone unevictable lists for evictable pages.
2819  */
2820
2821 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
2822                                           struct sysdev_attribute *attr,
2823                                           char *buf)
2824 {
2825         return sprintf(buf, "0\n");     /* always zero; should fit... */
2826 }
2827
2828 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
2829                                            struct sysdev_attribute *attr,
2830                                         const char *buf, size_t count)
2831 {
2832         struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
2833         struct zone *zone;
2834         unsigned long res;
2835         unsigned long req = strict_strtoul(buf, 10, &res);
2836
2837         if (!req)
2838                 return 1;       /* zero is no-op */
2839
2840         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2841                 if (!populated_zone(zone))
2842                         continue;
2843                 scan_zone_unevictable_pages(zone);
2844         }
2845         return 1;
2846 }
2847
2848
2849 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
2850                         read_scan_unevictable_node,
2851                         write_scan_unevictable_node);
2852
2853 int scan_unevictable_register_node(struct node *node)
2854 {
2855         return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
2856 }
2857
2858 void scan_unevictable_unregister_node(struct node *node)
2859 {
2860         sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
2861 }
2862