slub: Potential stack overflow
[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/gfp.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         /* How many pages shrink_list() should reclaim */
59         unsigned long nr_to_reclaim;
60
61         unsigned long hibernation_mode;
62
63         /* This context's GFP mask */
64         gfp_t gfp_mask;
65
66         int may_writepage;
67
68         /* Can mapped pages be reclaimed? */
69         int may_unmap;
70
71         /* Can pages be swapped as part of reclaim? */
72         int may_swap;
73
74         int swappiness;
75
76         int all_unreclaimable;
77
78         int order;
79
80         /* Which cgroup do we reclaim from */
81         struct mem_cgroup *mem_cgroup;
82
83         /*
84          * Nodemask of nodes allowed by the caller. If NULL, all nodes
85          * are scanned.
86          */
87         nodemask_t      *nodemask;
88
89         /* Pluggable isolate pages callback */
90         unsigned long (*isolate_pages)(unsigned long nr, struct list_head *dst,
91                         unsigned long *scanned, int order, int mode,
92                         struct zone *z, struct mem_cgroup *mem_cont,
93                         int active, int file);
94 };
95
96 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
97
98 #ifdef ARCH_HAS_PREFETCH
99 #define prefetch_prev_lru_page(_page, _base, _field)                    \
100         do {                                                            \
101                 if ((_page)->lru.prev != _base) {                       \
102                         struct page *prev;                              \
103                                                                         \
104                         prev = lru_to_page(&(_page->lru));              \
105                         prefetch(&prev->_field);                        \
106                 }                                                       \
107         } while (0)
108 #else
109 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
110 #endif
111
112 #ifdef ARCH_HAS_PREFETCHW
113 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
114         do {                                                            \
115                 if ((_page)->lru.prev != _base) {                       \
116                         struct page *prev;                              \
117                                                                         \
118                         prev = lru_to_page(&(_page->lru));              \
119                         prefetchw(&prev->_field);                       \
120                 }                                                       \
121         } while (0)
122 #else
123 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
124 #endif
125
126 /*
127  * From 0 .. 100.  Higher means more swappy.
128  */
129 int vm_swappiness = 60;
130 long vm_total_pages;    /* The total number of pages which the VM controls */
131
132 static LIST_HEAD(shrinker_list);
133 static DECLARE_RWSEM(shrinker_rwsem);
134
135 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
136 #define scanning_global_lru(sc) (!(sc)->mem_cgroup)
137 #else
138 #define scanning_global_lru(sc) (1)
139 #endif
140
141 static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
142                                                   struct scan_control *sc)
143 {
144         if (!scanning_global_lru(sc))
145                 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
146
147         return &zone->reclaim_stat;
148 }
149
150 static unsigned long zone_nr_lru_pages(struct zone *zone,
151                                 struct scan_control *sc, enum lru_list lru)
152 {
153         if (!scanning_global_lru(sc))
154                 return mem_cgroup_zone_nr_pages(sc->mem_cgroup, zone, lru);
155
156         return zone_page_state(zone, NR_LRU_BASE + lru);
157 }
158
159
160 /*
161  * Add a shrinker callback to be called from the vm
162  */
163 void register_shrinker(struct shrinker *shrinker)
164 {
165         shrinker->nr = 0;
166         down_write(&shrinker_rwsem);
167         list_add_tail(&shrinker->list, &shrinker_list);
168         up_write(&shrinker_rwsem);
169 }
170 EXPORT_SYMBOL(register_shrinker);
171
172 /*
173  * Remove one
174  */
175 void unregister_shrinker(struct shrinker *shrinker)
176 {
177         down_write(&shrinker_rwsem);
178         list_del(&shrinker->list);
179         up_write(&shrinker_rwsem);
180 }
181 EXPORT_SYMBOL(unregister_shrinker);
182
183 #define SHRINK_BATCH 128
184 /*
185  * Call the shrink functions to age shrinkable caches
186  *
187  * Here we assume it costs one seek to replace a lru page and that it also
188  * takes a seek to recreate a cache object.  With this in mind we age equal
189  * percentages of the lru and ageable caches.  This should balance the seeks
190  * generated by these structures.
191  *
192  * If the vm encountered mapped pages on the LRU it increase the pressure on
193  * slab to avoid swapping.
194  *
195  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
196  *
197  * `lru_pages' represents the number of on-LRU pages in all the zones which
198  * are eligible for the caller's allocation attempt.  It is used for balancing
199  * slab reclaim versus page reclaim.
200  *
201  * Returns the number of slab objects which we shrunk.
202  */
203 unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
204                         unsigned long lru_pages)
205 {
206         struct shrinker *shrinker;
207         unsigned long ret = 0;
208
209         if (scanned == 0)
210                 scanned = SWAP_CLUSTER_MAX;
211
212         if (!down_read_trylock(&shrinker_rwsem))
213                 return 1;       /* Assume we'll be able to shrink next time */
214
215         list_for_each_entry(shrinker, &shrinker_list, list) {
216                 unsigned long long delta;
217                 unsigned long total_scan;
218                 unsigned long max_pass = (*shrinker->shrink)(0, gfp_mask);
219
220                 delta = (4 * scanned) / shrinker->seeks;
221                 delta *= max_pass;
222                 do_div(delta, lru_pages + 1);
223                 shrinker->nr += delta;
224                 if (shrinker->nr < 0) {
225                         printk(KERN_ERR "shrink_slab: %pF negative objects to "
226                                "delete nr=%ld\n",
227                                shrinker->shrink, shrinker->nr);
228                         shrinker->nr = max_pass;
229                 }
230
231                 /*
232                  * Avoid risking looping forever due to too large nr value:
233                  * never try to free more than twice the estimate number of
234                  * freeable entries.
235                  */
236                 if (shrinker->nr > max_pass * 2)
237                         shrinker->nr = max_pass * 2;
238
239                 total_scan = shrinker->nr;
240                 shrinker->nr = 0;
241
242                 while (total_scan >= SHRINK_BATCH) {
243                         long this_scan = SHRINK_BATCH;
244                         int shrink_ret;
245                         int nr_before;
246
247                         nr_before = (*shrinker->shrink)(0, gfp_mask);
248                         shrink_ret = (*shrinker->shrink)(this_scan, gfp_mask);
249                         if (shrink_ret == -1)
250                                 break;
251                         if (shrink_ret < nr_before)
252                                 ret += nr_before - shrink_ret;
253                         count_vm_events(SLABS_SCANNED, this_scan);
254                         total_scan -= this_scan;
255
256                         cond_resched();
257                 }
258
259                 shrinker->nr += total_scan;
260         }
261         up_read(&shrinker_rwsem);
262         return ret;
263 }
264
265 static inline int is_page_cache_freeable(struct page *page)
266 {
267         /*
268          * A freeable page cache page is referenced only by the caller
269          * that isolated the page, the page cache radix tree and
270          * optional buffer heads at page->private.
271          */
272         return page_count(page) - page_has_private(page) == 2;
273 }
274
275 static int may_write_to_queue(struct backing_dev_info *bdi)
276 {
277         if (current->flags & PF_SWAPWRITE)
278                 return 1;
279         if (!bdi_write_congested(bdi))
280                 return 1;
281         if (bdi == current->backing_dev_info)
282                 return 1;
283         return 0;
284 }
285
286 /*
287  * We detected a synchronous write error writing a page out.  Probably
288  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
289  * fsync(), msync() or close().
290  *
291  * The tricky part is that after writepage we cannot touch the mapping: nothing
292  * prevents it from being freed up.  But we have a ref on the page and once
293  * that page is locked, the mapping is pinned.
294  *
295  * We're allowed to run sleeping lock_page() here because we know the caller has
296  * __GFP_FS.
297  */
298 static void handle_write_error(struct address_space *mapping,
299                                 struct page *page, int error)
300 {
301         lock_page(page);
302         if (page_mapping(page) == mapping)
303                 mapping_set_error(mapping, error);
304         unlock_page(page);
305 }
306
307 /* Request for sync pageout. */
308 enum pageout_io {
309         PAGEOUT_IO_ASYNC,
310         PAGEOUT_IO_SYNC,
311 };
312
313 /* possible outcome of pageout() */
314 typedef enum {
315         /* failed to write page out, page is locked */
316         PAGE_KEEP,
317         /* move page to the active list, page is locked */
318         PAGE_ACTIVATE,
319         /* page has been sent to the disk successfully, page is unlocked */
320         PAGE_SUCCESS,
321         /* page is clean and locked */
322         PAGE_CLEAN,
323 } pageout_t;
324
325 /*
326  * pageout is called by shrink_page_list() for each dirty page.
327  * Calls ->writepage().
328  */
329 static pageout_t pageout(struct page *page, struct address_space *mapping,
330                                                 enum pageout_io sync_writeback)
331 {
332         /*
333          * If the page is dirty, only perform writeback if that write
334          * will be non-blocking.  To prevent this allocation from being
335          * stalled by pagecache activity.  But note that there may be
336          * stalls if we need to run get_block().  We could test
337          * PagePrivate for that.
338          *
339          * If this process is currently in __generic_file_aio_write() against
340          * this page's queue, we can perform writeback even if that
341          * will block.
342          *
343          * If the page is swapcache, write it back even if that would
344          * block, for some throttling. This happens by accident, because
345          * swap_backing_dev_info is bust: it doesn't reflect the
346          * congestion state of the swapdevs.  Easy to fix, if needed.
347          */
348         if (!is_page_cache_freeable(page))
349                 return PAGE_KEEP;
350         if (!mapping) {
351                 /*
352                  * Some data journaling orphaned pages can have
353                  * page->mapping == NULL while being dirty with clean buffers.
354                  */
355                 if (page_has_private(page)) {
356                         if (try_to_free_buffers(page)) {
357                                 ClearPageDirty(page);
358                                 printk("%s: orphaned page\n", __func__);
359                                 return PAGE_CLEAN;
360                         }
361                 }
362                 return PAGE_KEEP;
363         }
364         if (mapping->a_ops->writepage == NULL)
365                 return PAGE_ACTIVATE;
366         if (!may_write_to_queue(mapping->backing_dev_info))
367                 return PAGE_KEEP;
368
369         if (clear_page_dirty_for_io(page)) {
370                 int res;
371                 struct writeback_control wbc = {
372                         .sync_mode = WB_SYNC_NONE,
373                         .nr_to_write = SWAP_CLUSTER_MAX,
374                         .range_start = 0,
375                         .range_end = LLONG_MAX,
376                         .nonblocking = 1,
377                         .for_reclaim = 1,
378                 };
379
380                 SetPageReclaim(page);
381                 res = mapping->a_ops->writepage(page, &wbc);
382                 if (res < 0)
383                         handle_write_error(mapping, page, res);
384                 if (res == AOP_WRITEPAGE_ACTIVATE) {
385                         ClearPageReclaim(page);
386                         return PAGE_ACTIVATE;
387                 }
388
389                 /*
390                  * Wait on writeback if requested to. This happens when
391                  * direct reclaiming a large contiguous area and the
392                  * first attempt to free a range of pages fails.
393                  */
394                 if (PageWriteback(page) && sync_writeback == PAGEOUT_IO_SYNC)
395                         wait_on_page_writeback(page);
396
397                 if (!PageWriteback(page)) {
398                         /* synchronous write or broken a_ops? */
399                         ClearPageReclaim(page);
400                 }
401                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
402                 return PAGE_SUCCESS;
403         }
404
405         return PAGE_CLEAN;
406 }
407
408 /*
409  * Same as remove_mapping, but if the page is removed from the mapping, it
410  * gets returned with a refcount of 0.
411  */
412 static int __remove_mapping(struct address_space *mapping, struct page *page)
413 {
414         BUG_ON(!PageLocked(page));
415         BUG_ON(mapping != page_mapping(page));
416
417         spin_lock_irq(&mapping->tree_lock);
418         /*
419          * The non racy check for a busy page.
420          *
421          * Must be careful with the order of the tests. When someone has
422          * a ref to the page, it may be possible that they dirty it then
423          * drop the reference. So if PageDirty is tested before page_count
424          * here, then the following race may occur:
425          *
426          * get_user_pages(&page);
427          * [user mapping goes away]
428          * write_to(page);
429          *                              !PageDirty(page)    [good]
430          * SetPageDirty(page);
431          * put_page(page);
432          *                              !page_count(page)   [good, discard it]
433          *
434          * [oops, our write_to data is lost]
435          *
436          * Reversing the order of the tests ensures such a situation cannot
437          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
438          * load is not satisfied before that of page->_count.
439          *
440          * Note that if SetPageDirty is always performed via set_page_dirty,
441          * and thus under tree_lock, then this ordering is not required.
442          */
443         if (!page_freeze_refs(page, 2))
444                 goto cannot_free;
445         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
446         if (unlikely(PageDirty(page))) {
447                 page_unfreeze_refs(page, 2);
448                 goto cannot_free;
449         }
450
451         if (PageSwapCache(page)) {
452                 swp_entry_t swap = { .val = page_private(page) };
453                 __delete_from_swap_cache(page);
454                 spin_unlock_irq(&mapping->tree_lock);
455                 swapcache_free(swap, page);
456         } else {
457                 __remove_from_page_cache(page);
458                 spin_unlock_irq(&mapping->tree_lock);
459                 mem_cgroup_uncharge_cache_page(page);
460         }
461
462         return 1;
463
464 cannot_free:
465         spin_unlock_irq(&mapping->tree_lock);
466         return 0;
467 }
468
469 /*
470  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
471  * someone else has a ref on the page, abort and return 0.  If it was
472  * successfully detached, return 1.  Assumes the caller has a single ref on
473  * this page.
474  */
475 int remove_mapping(struct address_space *mapping, struct page *page)
476 {
477         if (__remove_mapping(mapping, page)) {
478                 /*
479                  * Unfreezing the refcount with 1 rather than 2 effectively
480                  * drops the pagecache ref for us without requiring another
481                  * atomic operation.
482                  */
483                 page_unfreeze_refs(page, 1);
484                 return 1;
485         }
486         return 0;
487 }
488
489 /**
490  * putback_lru_page - put previously isolated page onto appropriate LRU list
491  * @page: page to be put back to appropriate lru list
492  *
493  * Add previously isolated @page to appropriate LRU list.
494  * Page may still be unevictable for other reasons.
495  *
496  * lru_lock must not be held, interrupts must be enabled.
497  */
498 void putback_lru_page(struct page *page)
499 {
500         int lru;
501         int active = !!TestClearPageActive(page);
502         int was_unevictable = PageUnevictable(page);
503
504         VM_BUG_ON(PageLRU(page));
505
506 redo:
507         ClearPageUnevictable(page);
508
509         if (page_evictable(page, NULL)) {
510                 /*
511                  * For evictable pages, we can use the cache.
512                  * In event of a race, worst case is we end up with an
513                  * unevictable page on [in]active list.
514                  * We know how to handle that.
515                  */
516                 lru = active + page_lru_base_type(page);
517                 lru_cache_add_lru(page, lru);
518         } else {
519                 /*
520                  * Put unevictable pages directly on zone's unevictable
521                  * list.
522                  */
523                 lru = LRU_UNEVICTABLE;
524                 add_page_to_unevictable_list(page);
525                 /*
526                  * When racing with an mlock clearing (page is
527                  * unlocked), make sure that if the other thread does
528                  * not observe our setting of PG_lru and fails
529                  * isolation, we see PG_mlocked cleared below and move
530                  * the page back to the evictable list.
531                  *
532                  * The other side is TestClearPageMlocked().
533                  */
534                 smp_mb();
535         }
536
537         /*
538          * page's status can change while we move it among lru. If an evictable
539          * page is on unevictable list, it never be freed. To avoid that,
540          * check after we added it to the list, again.
541          */
542         if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
543                 if (!isolate_lru_page(page)) {
544                         put_page(page);
545                         goto redo;
546                 }
547                 /* This means someone else dropped this page from LRU
548                  * So, it will be freed or putback to LRU again. There is
549                  * nothing to do here.
550                  */
551         }
552
553         if (was_unevictable && lru != LRU_UNEVICTABLE)
554                 count_vm_event(UNEVICTABLE_PGRESCUED);
555         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
556                 count_vm_event(UNEVICTABLE_PGCULLED);
557
558         put_page(page);         /* drop ref from isolate */
559 }
560
561 enum page_references {
562         PAGEREF_RECLAIM,
563         PAGEREF_RECLAIM_CLEAN,
564         PAGEREF_KEEP,
565         PAGEREF_ACTIVATE,
566 };
567
568 static enum page_references page_check_references(struct page *page,
569                                                   struct scan_control *sc)
570 {
571         int referenced_ptes, referenced_page;
572         unsigned long vm_flags;
573
574         referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
575         referenced_page = TestClearPageReferenced(page);
576
577         /* Lumpy reclaim - ignore references */
578         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
579                 return PAGEREF_RECLAIM;
580
581         /*
582          * Mlock lost the isolation race with us.  Let try_to_unmap()
583          * move the page to the unevictable list.
584          */
585         if (vm_flags & VM_LOCKED)
586                 return PAGEREF_RECLAIM;
587
588         if (referenced_ptes) {
589                 if (PageAnon(page))
590                         return PAGEREF_ACTIVATE;
591                 /*
592                  * All mapped pages start out with page table
593                  * references from the instantiating fault, so we need
594                  * to look twice if a mapped file page is used more
595                  * than once.
596                  *
597                  * Mark it and spare it for another trip around the
598                  * inactive list.  Another page table reference will
599                  * lead to its activation.
600                  *
601                  * Note: the mark is set for activated pages as well
602                  * so that recently deactivated but used pages are
603                  * quickly recovered.
604                  */
605                 SetPageReferenced(page);
606
607                 if (referenced_page)
608                         return PAGEREF_ACTIVATE;
609
610                 return PAGEREF_KEEP;
611         }
612
613         /* Reclaim if clean, defer dirty pages to writeback */
614         if (referenced_page)
615                 return PAGEREF_RECLAIM_CLEAN;
616
617         return PAGEREF_RECLAIM;
618 }
619
620 /*
621  * shrink_page_list() returns the number of reclaimed pages
622  */
623 static unsigned long shrink_page_list(struct list_head *page_list,
624                                         struct scan_control *sc,
625                                         enum pageout_io sync_writeback)
626 {
627         LIST_HEAD(ret_pages);
628         struct pagevec freed_pvec;
629         int pgactivate = 0;
630         unsigned long nr_reclaimed = 0;
631
632         cond_resched();
633
634         pagevec_init(&freed_pvec, 1);
635         while (!list_empty(page_list)) {
636                 enum page_references references;
637                 struct address_space *mapping;
638                 struct page *page;
639                 int may_enter_fs;
640
641                 cond_resched();
642
643                 page = lru_to_page(page_list);
644                 list_del(&page->lru);
645
646                 if (!trylock_page(page))
647                         goto keep;
648
649                 VM_BUG_ON(PageActive(page));
650
651                 sc->nr_scanned++;
652
653                 if (unlikely(!page_evictable(page, NULL)))
654                         goto cull_mlocked;
655
656                 if (!sc->may_unmap && page_mapped(page))
657                         goto keep_locked;
658
659                 /* Double the slab pressure for mapped and swapcache pages */
660                 if (page_mapped(page) || PageSwapCache(page))
661                         sc->nr_scanned++;
662
663                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
664                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
665
666                 if (PageWriteback(page)) {
667                         /*
668                          * Synchronous reclaim is performed in two passes,
669                          * first an asynchronous pass over the list to
670                          * start parallel writeback, and a second synchronous
671                          * pass to wait for the IO to complete.  Wait here
672                          * for any page for which writeback has already
673                          * started.
674                          */
675                         if (sync_writeback == PAGEOUT_IO_SYNC && may_enter_fs)
676                                 wait_on_page_writeback(page);
677                         else
678                                 goto keep_locked;
679                 }
680
681                 references = page_check_references(page, sc);
682                 switch (references) {
683                 case PAGEREF_ACTIVATE:
684                         goto activate_locked;
685                 case PAGEREF_KEEP:
686                         goto keep_locked;
687                 case PAGEREF_RECLAIM:
688                 case PAGEREF_RECLAIM_CLEAN:
689                         ; /* try to reclaim the page below */
690                 }
691
692                 /*
693                  * Anonymous process memory has backing store?
694                  * Try to allocate it some swap space here.
695                  */
696                 if (PageAnon(page) && !PageSwapCache(page)) {
697                         if (!(sc->gfp_mask & __GFP_IO))
698                                 goto keep_locked;
699                         if (!add_to_swap(page))
700                                 goto activate_locked;
701                         may_enter_fs = 1;
702                 }
703
704                 mapping = page_mapping(page);
705
706                 /*
707                  * The page is mapped into the page tables of one or more
708                  * processes. Try to unmap it here.
709                  */
710                 if (page_mapped(page) && mapping) {
711                         switch (try_to_unmap(page, TTU_UNMAP)) {
712                         case SWAP_FAIL:
713                                 goto activate_locked;
714                         case SWAP_AGAIN:
715                                 goto keep_locked;
716                         case SWAP_MLOCK:
717                                 goto cull_mlocked;
718                         case SWAP_SUCCESS:
719                                 ; /* try to free the page below */
720                         }
721                 }
722
723                 if (PageDirty(page)) {
724                         if (references == PAGEREF_RECLAIM_CLEAN)
725                                 goto keep_locked;
726                         if (!may_enter_fs)
727                                 goto keep_locked;
728                         if (!sc->may_writepage)
729                                 goto keep_locked;
730
731                         /* Page is dirty, try to write it out here */
732                         switch (pageout(page, mapping, sync_writeback)) {
733                         case PAGE_KEEP:
734                                 goto keep_locked;
735                         case PAGE_ACTIVATE:
736                                 goto activate_locked;
737                         case PAGE_SUCCESS:
738                                 if (PageWriteback(page) || PageDirty(page))
739                                         goto keep;
740                                 /*
741                                  * A synchronous write - probably a ramdisk.  Go
742                                  * ahead and try to reclaim the page.
743                                  */
744                                 if (!trylock_page(page))
745                                         goto keep;
746                                 if (PageDirty(page) || PageWriteback(page))
747                                         goto keep_locked;
748                                 mapping = page_mapping(page);
749                         case PAGE_CLEAN:
750                                 ; /* try to free the page below */
751                         }
752                 }
753
754                 /*
755                  * If the page has buffers, try to free the buffer mappings
756                  * associated with this page. If we succeed we try to free
757                  * the page as well.
758                  *
759                  * We do this even if the page is PageDirty().
760                  * try_to_release_page() does not perform I/O, but it is
761                  * possible for a page to have PageDirty set, but it is actually
762                  * clean (all its buffers are clean).  This happens if the
763                  * buffers were written out directly, with submit_bh(). ext3
764                  * will do this, as well as the blockdev mapping.
765                  * try_to_release_page() will discover that cleanness and will
766                  * drop the buffers and mark the page clean - it can be freed.
767                  *
768                  * Rarely, pages can have buffers and no ->mapping.  These are
769                  * the pages which were not successfully invalidated in
770                  * truncate_complete_page().  We try to drop those buffers here
771                  * and if that worked, and the page is no longer mapped into
772                  * process address space (page_count == 1) it can be freed.
773                  * Otherwise, leave the page on the LRU so it is swappable.
774                  */
775                 if (page_has_private(page)) {
776                         if (!try_to_release_page(page, sc->gfp_mask))
777                                 goto activate_locked;
778                         if (!mapping && page_count(page) == 1) {
779                                 unlock_page(page);
780                                 if (put_page_testzero(page))
781                                         goto free_it;
782                                 else {
783                                         /*
784                                          * rare race with speculative reference.
785                                          * the speculative reference will free
786                                          * this page shortly, so we may
787                                          * increment nr_reclaimed here (and
788                                          * leave it off the LRU).
789                                          */
790                                         nr_reclaimed++;
791                                         continue;
792                                 }
793                         }
794                 }
795
796                 if (!mapping || !__remove_mapping(mapping, page))
797                         goto keep_locked;
798
799                 /*
800                  * At this point, we have no other references and there is
801                  * no way to pick any more up (removed from LRU, removed
802                  * from pagecache). Can use non-atomic bitops now (and
803                  * we obviously don't have to worry about waking up a process
804                  * waiting on the page lock, because there are no references.
805                  */
806                 __clear_page_locked(page);
807 free_it:
808                 nr_reclaimed++;
809                 if (!pagevec_add(&freed_pvec, page)) {
810                         __pagevec_free(&freed_pvec);
811                         pagevec_reinit(&freed_pvec);
812                 }
813                 continue;
814
815 cull_mlocked:
816                 if (PageSwapCache(page))
817                         try_to_free_swap(page);
818                 unlock_page(page);
819                 putback_lru_page(page);
820                 continue;
821
822 activate_locked:
823                 /* Not a candidate for swapping, so reclaim swap space. */
824                 if (PageSwapCache(page) && vm_swap_full())
825                         try_to_free_swap(page);
826                 VM_BUG_ON(PageActive(page));
827                 SetPageActive(page);
828                 pgactivate++;
829 keep_locked:
830                 unlock_page(page);
831 keep:
832                 list_add(&page->lru, &ret_pages);
833                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
834         }
835         list_splice(&ret_pages, page_list);
836         if (pagevec_count(&freed_pvec))
837                 __pagevec_free(&freed_pvec);
838         count_vm_events(PGACTIVATE, pgactivate);
839         return nr_reclaimed;
840 }
841
842 /* LRU Isolation modes. */
843 #define ISOLATE_INACTIVE 0      /* Isolate inactive pages. */
844 #define ISOLATE_ACTIVE 1        /* Isolate active pages. */
845 #define ISOLATE_BOTH 2          /* Isolate both active and inactive pages. */
846
847 /*
848  * Attempt to remove the specified page from its LRU.  Only take this page
849  * if it is of the appropriate PageActive status.  Pages which are being
850  * freed elsewhere are also ignored.
851  *
852  * page:        page to consider
853  * mode:        one of the LRU isolation modes defined above
854  *
855  * returns 0 on success, -ve errno on failure.
856  */
857 int __isolate_lru_page(struct page *page, int mode, int file)
858 {
859         int ret = -EINVAL;
860
861         /* Only take pages on the LRU. */
862         if (!PageLRU(page))
863                 return ret;
864
865         /*
866          * When checking the active state, we need to be sure we are
867          * dealing with comparible boolean values.  Take the logical not
868          * of each.
869          */
870         if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
871                 return ret;
872
873         if (mode != ISOLATE_BOTH && page_is_file_cache(page) != file)
874                 return ret;
875
876         /*
877          * When this function is being called for lumpy reclaim, we
878          * initially look into all LRU pages, active, inactive and
879          * unevictable; only give shrink_page_list evictable pages.
880          */
881         if (PageUnevictable(page))
882                 return ret;
883
884         ret = -EBUSY;
885
886         if (likely(get_page_unless_zero(page))) {
887                 /*
888                  * Be careful not to clear PageLRU until after we're
889                  * sure the page is not being freed elsewhere -- the
890                  * page release code relies on it.
891                  */
892                 ClearPageLRU(page);
893                 ret = 0;
894         }
895
896         return ret;
897 }
898
899 /*
900  * zone->lru_lock is heavily contended.  Some of the functions that
901  * shrink the lists perform better by taking out a batch of pages
902  * and working on them outside the LRU lock.
903  *
904  * For pagecache intensive workloads, this function is the hottest
905  * spot in the kernel (apart from copy_*_user functions).
906  *
907  * Appropriate locks must be held before calling this function.
908  *
909  * @nr_to_scan: The number of pages to look through on the list.
910  * @src:        The LRU list to pull pages off.
911  * @dst:        The temp list to put pages on to.
912  * @scanned:    The number of pages that were scanned.
913  * @order:      The caller's attempted allocation order
914  * @mode:       One of the LRU isolation modes
915  * @file:       True [1] if isolating file [!anon] pages
916  *
917  * returns how many pages were moved onto *@dst.
918  */
919 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
920                 struct list_head *src, struct list_head *dst,
921                 unsigned long *scanned, int order, int mode, int file)
922 {
923         unsigned long nr_taken = 0;
924         unsigned long scan;
925
926         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
927                 struct page *page;
928                 unsigned long pfn;
929                 unsigned long end_pfn;
930                 unsigned long page_pfn;
931                 int zone_id;
932
933                 page = lru_to_page(src);
934                 prefetchw_prev_lru_page(page, src, flags);
935
936                 VM_BUG_ON(!PageLRU(page));
937
938                 switch (__isolate_lru_page(page, mode, file)) {
939                 case 0:
940                         list_move(&page->lru, dst);
941                         mem_cgroup_del_lru(page);
942                         nr_taken++;
943                         break;
944
945                 case -EBUSY:
946                         /* else it is being freed elsewhere */
947                         list_move(&page->lru, src);
948                         mem_cgroup_rotate_lru_list(page, page_lru(page));
949                         continue;
950
951                 default:
952                         BUG();
953                 }
954
955                 if (!order)
956                         continue;
957
958                 /*
959                  * Attempt to take all pages in the order aligned region
960                  * surrounding the tag page.  Only take those pages of
961                  * the same active state as that tag page.  We may safely
962                  * round the target page pfn down to the requested order
963                  * as the mem_map is guarenteed valid out to MAX_ORDER,
964                  * where that page is in a different zone we will detect
965                  * it from its zone id and abort this block scan.
966                  */
967                 zone_id = page_zone_id(page);
968                 page_pfn = page_to_pfn(page);
969                 pfn = page_pfn & ~((1 << order) - 1);
970                 end_pfn = pfn + (1 << order);
971                 for (; pfn < end_pfn; pfn++) {
972                         struct page *cursor_page;
973
974                         /* The target page is in the block, ignore it. */
975                         if (unlikely(pfn == page_pfn))
976                                 continue;
977
978                         /* Avoid holes within the zone. */
979                         if (unlikely(!pfn_valid_within(pfn)))
980                                 break;
981
982                         cursor_page = pfn_to_page(pfn);
983
984                         /* Check that we have not crossed a zone boundary. */
985                         if (unlikely(page_zone_id(cursor_page) != zone_id))
986                                 continue;
987
988                         /*
989                          * If we don't have enough swap space, reclaiming of
990                          * anon page which don't already have a swap slot is
991                          * pointless.
992                          */
993                         if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
994                                         !PageSwapCache(cursor_page))
995                                 continue;
996
997                         if (__isolate_lru_page(cursor_page, mode, file) == 0) {
998                                 list_move(&cursor_page->lru, dst);
999                                 mem_cgroup_del_lru(cursor_page);
1000                                 nr_taken++;
1001                                 scan++;
1002                         }
1003                 }
1004         }
1005
1006         *scanned = scan;
1007         return nr_taken;
1008 }
1009
1010 static unsigned long isolate_pages_global(unsigned long nr,
1011                                         struct list_head *dst,
1012                                         unsigned long *scanned, int order,
1013                                         int mode, struct zone *z,
1014                                         struct mem_cgroup *mem_cont,
1015                                         int active, int file)
1016 {
1017         int lru = LRU_BASE;
1018         if (active)
1019                 lru += LRU_ACTIVE;
1020         if (file)
1021                 lru += LRU_FILE;
1022         return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1023                                                                 mode, file);
1024 }
1025
1026 /*
1027  * clear_active_flags() is a helper for shrink_active_list(), clearing
1028  * any active bits from the pages in the list.
1029  */
1030 static unsigned long clear_active_flags(struct list_head *page_list,
1031                                         unsigned int *count)
1032 {
1033         int nr_active = 0;
1034         int lru;
1035         struct page *page;
1036
1037         list_for_each_entry(page, page_list, lru) {
1038                 lru = page_lru_base_type(page);
1039                 if (PageActive(page)) {
1040                         lru += LRU_ACTIVE;
1041                         ClearPageActive(page);
1042                         nr_active++;
1043                 }
1044                 count[lru]++;
1045         }
1046
1047         return nr_active;
1048 }
1049
1050 /**
1051  * isolate_lru_page - tries to isolate a page from its LRU list
1052  * @page: page to isolate from its LRU list
1053  *
1054  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1055  * vmstat statistic corresponding to whatever LRU list the page was on.
1056  *
1057  * Returns 0 if the page was removed from an LRU list.
1058  * Returns -EBUSY if the page was not on an LRU list.
1059  *
1060  * The returned page will have PageLRU() cleared.  If it was found on
1061  * the active list, it will have PageActive set.  If it was found on
1062  * the unevictable list, it will have the PageUnevictable bit set. That flag
1063  * may need to be cleared by the caller before letting the page go.
1064  *
1065  * The vmstat statistic corresponding to the list on which the page was
1066  * found will be decremented.
1067  *
1068  * Restrictions:
1069  * (1) Must be called with an elevated refcount on the page. This is a
1070  *     fundamentnal difference from isolate_lru_pages (which is called
1071  *     without a stable reference).
1072  * (2) the lru_lock must not be held.
1073  * (3) interrupts must be enabled.
1074  */
1075 int isolate_lru_page(struct page *page)
1076 {
1077         int ret = -EBUSY;
1078
1079         if (PageLRU(page)) {
1080                 struct zone *zone = page_zone(page);
1081
1082                 spin_lock_irq(&zone->lru_lock);
1083                 if (PageLRU(page) && get_page_unless_zero(page)) {
1084                         int lru = page_lru(page);
1085                         ret = 0;
1086                         ClearPageLRU(page);
1087
1088                         del_page_from_lru_list(zone, page, lru);
1089                 }
1090                 spin_unlock_irq(&zone->lru_lock);
1091         }
1092         return ret;
1093 }
1094
1095 /*
1096  * Are there way too many processes in the direct reclaim path already?
1097  */
1098 static int too_many_isolated(struct zone *zone, int file,
1099                 struct scan_control *sc)
1100 {
1101         unsigned long inactive, isolated;
1102
1103         if (current_is_kswapd())
1104                 return 0;
1105
1106         if (!scanning_global_lru(sc))
1107                 return 0;
1108
1109         if (file) {
1110                 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1111                 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1112         } else {
1113                 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1114                 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1115         }
1116
1117         return isolated > inactive;
1118 }
1119
1120 /*
1121  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1122  * of reclaimed pages
1123  */
1124 static unsigned long shrink_inactive_list(unsigned long max_scan,
1125                         struct zone *zone, struct scan_control *sc,
1126                         int priority, int file)
1127 {
1128         LIST_HEAD(page_list);
1129         struct pagevec pvec;
1130         unsigned long nr_scanned = 0;
1131         unsigned long nr_reclaimed = 0;
1132         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1133         int lumpy_reclaim = 0;
1134
1135         while (unlikely(too_many_isolated(zone, file, sc))) {
1136                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1137
1138                 /* We are about to die and free our memory. Return now. */
1139                 if (fatal_signal_pending(current))
1140                         return SWAP_CLUSTER_MAX;
1141         }
1142
1143         /*
1144          * If we need a large contiguous chunk of memory, or have
1145          * trouble getting a small set of contiguous pages, we
1146          * will reclaim both active and inactive pages.
1147          *
1148          * We use the same threshold as pageout congestion_wait below.
1149          */
1150         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1151                 lumpy_reclaim = 1;
1152         else if (sc->order && priority < DEF_PRIORITY - 2)
1153                 lumpy_reclaim = 1;
1154
1155         pagevec_init(&pvec, 1);
1156
1157         lru_add_drain();
1158         spin_lock_irq(&zone->lru_lock);
1159         do {
1160                 struct page *page;
1161                 unsigned long nr_taken;
1162                 unsigned long nr_scan;
1163                 unsigned long nr_freed;
1164                 unsigned long nr_active;
1165                 unsigned int count[NR_LRU_LISTS] = { 0, };
1166                 int mode = lumpy_reclaim ? ISOLATE_BOTH : ISOLATE_INACTIVE;
1167                 unsigned long nr_anon;
1168                 unsigned long nr_file;
1169
1170                 nr_taken = sc->isolate_pages(SWAP_CLUSTER_MAX,
1171                              &page_list, &nr_scan, sc->order, mode,
1172                                 zone, sc->mem_cgroup, 0, file);
1173
1174                 if (scanning_global_lru(sc)) {
1175                         zone->pages_scanned += nr_scan;
1176                         if (current_is_kswapd())
1177                                 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1178                                                        nr_scan);
1179                         else
1180                                 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1181                                                        nr_scan);
1182                 }
1183
1184                 if (nr_taken == 0)
1185                         goto done;
1186
1187                 nr_active = clear_active_flags(&page_list, count);
1188                 __count_vm_events(PGDEACTIVATE, nr_active);
1189
1190                 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1191                                                 -count[LRU_ACTIVE_FILE]);
1192                 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1193                                                 -count[LRU_INACTIVE_FILE]);
1194                 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1195                                                 -count[LRU_ACTIVE_ANON]);
1196                 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1197                                                 -count[LRU_INACTIVE_ANON]);
1198
1199                 nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1200                 nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1201                 __mod_zone_page_state(zone, NR_ISOLATED_ANON, nr_anon);
1202                 __mod_zone_page_state(zone, NR_ISOLATED_FILE, nr_file);
1203
1204                 reclaim_stat->recent_scanned[0] += nr_anon;
1205                 reclaim_stat->recent_scanned[1] += nr_file;
1206
1207                 spin_unlock_irq(&zone->lru_lock);
1208
1209                 nr_scanned += nr_scan;
1210                 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
1211
1212                 /*
1213                  * If we are direct reclaiming for contiguous pages and we do
1214                  * not reclaim everything in the list, try again and wait
1215                  * for IO to complete. This will stall high-order allocations
1216                  * but that should be acceptable to the caller
1217                  */
1218                 if (nr_freed < nr_taken && !current_is_kswapd() &&
1219                     lumpy_reclaim) {
1220                         congestion_wait(BLK_RW_ASYNC, HZ/10);
1221
1222                         /*
1223                          * The attempt at page out may have made some
1224                          * of the pages active, mark them inactive again.
1225                          */
1226                         nr_active = clear_active_flags(&page_list, count);
1227                         count_vm_events(PGDEACTIVATE, nr_active);
1228
1229                         nr_freed += shrink_page_list(&page_list, sc,
1230                                                         PAGEOUT_IO_SYNC);
1231                 }
1232
1233                 nr_reclaimed += nr_freed;
1234
1235                 local_irq_disable();
1236                 if (current_is_kswapd())
1237                         __count_vm_events(KSWAPD_STEAL, nr_freed);
1238                 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
1239
1240                 spin_lock(&zone->lru_lock);
1241                 /*
1242                  * Put back any unfreeable pages.
1243                  */
1244                 while (!list_empty(&page_list)) {
1245                         int lru;
1246                         page = lru_to_page(&page_list);
1247                         VM_BUG_ON(PageLRU(page));
1248                         list_del(&page->lru);
1249                         if (unlikely(!page_evictable(page, NULL))) {
1250                                 spin_unlock_irq(&zone->lru_lock);
1251                                 putback_lru_page(page);
1252                                 spin_lock_irq(&zone->lru_lock);
1253                                 continue;
1254                         }
1255                         SetPageLRU(page);
1256                         lru = page_lru(page);
1257                         add_page_to_lru_list(zone, page, lru);
1258                         if (is_active_lru(lru)) {
1259                                 int file = is_file_lru(lru);
1260                                 reclaim_stat->recent_rotated[file]++;
1261                         }
1262                         if (!pagevec_add(&pvec, page)) {
1263                                 spin_unlock_irq(&zone->lru_lock);
1264                                 __pagevec_release(&pvec);
1265                                 spin_lock_irq(&zone->lru_lock);
1266                         }
1267                 }
1268                 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1269                 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1270
1271         } while (nr_scanned < max_scan);
1272
1273 done:
1274         spin_unlock_irq(&zone->lru_lock);
1275         pagevec_release(&pvec);
1276         return nr_reclaimed;
1277 }
1278
1279 /*
1280  * We are about to scan this zone at a certain priority level.  If that priority
1281  * level is smaller (ie: more urgent) than the previous priority, then note
1282  * that priority level within the zone.  This is done so that when the next
1283  * process comes in to scan this zone, it will immediately start out at this
1284  * priority level rather than having to build up its own scanning priority.
1285  * Here, this priority affects only the reclaim-mapped threshold.
1286  */
1287 static inline void note_zone_scanning_priority(struct zone *zone, int priority)
1288 {
1289         if (priority < zone->prev_priority)
1290                 zone->prev_priority = priority;
1291 }
1292
1293 /*
1294  * This moves pages from the active list to the inactive list.
1295  *
1296  * We move them the other way if the page is referenced by one or more
1297  * processes, from rmap.
1298  *
1299  * If the pages are mostly unmapped, the processing is fast and it is
1300  * appropriate to hold zone->lru_lock across the whole operation.  But if
1301  * the pages are mapped, the processing is slow (page_referenced()) so we
1302  * should drop zone->lru_lock around each page.  It's impossible to balance
1303  * this, so instead we remove the pages from the LRU while processing them.
1304  * It is safe to rely on PG_active against the non-LRU pages in here because
1305  * nobody will play with that bit on a non-LRU page.
1306  *
1307  * The downside is that we have to touch page->_count against each page.
1308  * But we had to alter page->flags anyway.
1309  */
1310
1311 static void move_active_pages_to_lru(struct zone *zone,
1312                                      struct list_head *list,
1313                                      enum lru_list lru)
1314 {
1315         unsigned long pgmoved = 0;
1316         struct pagevec pvec;
1317         struct page *page;
1318
1319         pagevec_init(&pvec, 1);
1320
1321         while (!list_empty(list)) {
1322                 page = lru_to_page(list);
1323
1324                 VM_BUG_ON(PageLRU(page));
1325                 SetPageLRU(page);
1326
1327                 list_move(&page->lru, &zone->lru[lru].list);
1328                 mem_cgroup_add_lru_list(page, lru);
1329                 pgmoved++;
1330
1331                 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1332                         spin_unlock_irq(&zone->lru_lock);
1333                         if (buffer_heads_over_limit)
1334                                 pagevec_strip(&pvec);
1335                         __pagevec_release(&pvec);
1336                         spin_lock_irq(&zone->lru_lock);
1337                 }
1338         }
1339         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1340         if (!is_active_lru(lru))
1341                 __count_vm_events(PGDEACTIVATE, pgmoved);
1342 }
1343
1344 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1345                         struct scan_control *sc, int priority, int file)
1346 {
1347         unsigned long nr_taken;
1348         unsigned long pgscanned;
1349         unsigned long vm_flags;
1350         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1351         LIST_HEAD(l_active);
1352         LIST_HEAD(l_inactive);
1353         struct page *page;
1354         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1355         unsigned long nr_rotated = 0;
1356
1357         lru_add_drain();
1358         spin_lock_irq(&zone->lru_lock);
1359         nr_taken = sc->isolate_pages(nr_pages, &l_hold, &pgscanned, sc->order,
1360                                         ISOLATE_ACTIVE, zone,
1361                                         sc->mem_cgroup, 1, file);
1362         /*
1363          * zone->pages_scanned is used for detect zone's oom
1364          * mem_cgroup remembers nr_scan by itself.
1365          */
1366         if (scanning_global_lru(sc)) {
1367                 zone->pages_scanned += pgscanned;
1368         }
1369         reclaim_stat->recent_scanned[file] += nr_taken;
1370
1371         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1372         if (file)
1373                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1374         else
1375                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1376         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1377         spin_unlock_irq(&zone->lru_lock);
1378
1379         while (!list_empty(&l_hold)) {
1380                 cond_resched();
1381                 page = lru_to_page(&l_hold);
1382                 list_del(&page->lru);
1383
1384                 if (unlikely(!page_evictable(page, NULL))) {
1385                         putback_lru_page(page);
1386                         continue;
1387                 }
1388
1389                 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1390                         nr_rotated++;
1391                         /*
1392                          * Identify referenced, file-backed active pages and
1393                          * give them one more trip around the active list. So
1394                          * that executable code get better chances to stay in
1395                          * memory under moderate memory pressure.  Anon pages
1396                          * are not likely to be evicted by use-once streaming
1397                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1398                          * so we ignore them here.
1399                          */
1400                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1401                                 list_add(&page->lru, &l_active);
1402                                 continue;
1403                         }
1404                 }
1405
1406                 ClearPageActive(page);  /* we are de-activating */
1407                 list_add(&page->lru, &l_inactive);
1408         }
1409
1410         /*
1411          * Move pages back to the lru list.
1412          */
1413         spin_lock_irq(&zone->lru_lock);
1414         /*
1415          * Count referenced pages from currently used mappings as rotated,
1416          * even though only some of them are actually re-activated.  This
1417          * helps balance scan pressure between file and anonymous pages in
1418          * get_scan_ratio.
1419          */
1420         reclaim_stat->recent_rotated[file] += nr_rotated;
1421
1422         move_active_pages_to_lru(zone, &l_active,
1423                                                 LRU_ACTIVE + file * LRU_FILE);
1424         move_active_pages_to_lru(zone, &l_inactive,
1425                                                 LRU_BASE   + file * LRU_FILE);
1426         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1427         spin_unlock_irq(&zone->lru_lock);
1428 }
1429
1430 static int inactive_anon_is_low_global(struct zone *zone)
1431 {
1432         unsigned long active, inactive;
1433
1434         active = zone_page_state(zone, NR_ACTIVE_ANON);
1435         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1436
1437         if (inactive * zone->inactive_ratio < active)
1438                 return 1;
1439
1440         return 0;
1441 }
1442
1443 /**
1444  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1445  * @zone: zone to check
1446  * @sc:   scan control of this context
1447  *
1448  * Returns true if the zone does not have enough inactive anon pages,
1449  * meaning some active anon pages need to be deactivated.
1450  */
1451 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1452 {
1453         int low;
1454
1455         if (scanning_global_lru(sc))
1456                 low = inactive_anon_is_low_global(zone);
1457         else
1458                 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1459         return low;
1460 }
1461
1462 static int inactive_file_is_low_global(struct zone *zone)
1463 {
1464         unsigned long active, inactive;
1465
1466         active = zone_page_state(zone, NR_ACTIVE_FILE);
1467         inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1468
1469         return (active > inactive);
1470 }
1471
1472 /**
1473  * inactive_file_is_low - check if file pages need to be deactivated
1474  * @zone: zone to check
1475  * @sc:   scan control of this context
1476  *
1477  * When the system is doing streaming IO, memory pressure here
1478  * ensures that active file pages get deactivated, until more
1479  * than half of the file pages are on the inactive list.
1480  *
1481  * Once we get to that situation, protect the system's working
1482  * set from being evicted by disabling active file page aging.
1483  *
1484  * This uses a different ratio than the anonymous pages, because
1485  * the page cache uses a use-once replacement algorithm.
1486  */
1487 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1488 {
1489         int low;
1490
1491         if (scanning_global_lru(sc))
1492                 low = inactive_file_is_low_global(zone);
1493         else
1494                 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1495         return low;
1496 }
1497
1498 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1499                                 int file)
1500 {
1501         if (file)
1502                 return inactive_file_is_low(zone, sc);
1503         else
1504                 return inactive_anon_is_low(zone, sc);
1505 }
1506
1507 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1508         struct zone *zone, struct scan_control *sc, int priority)
1509 {
1510         int file = is_file_lru(lru);
1511
1512         if (is_active_lru(lru)) {
1513                 if (inactive_list_is_low(zone, sc, file))
1514                     shrink_active_list(nr_to_scan, zone, sc, priority, file);
1515                 return 0;
1516         }
1517
1518         return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1519 }
1520
1521 /*
1522  * Determine how aggressively the anon and file LRU lists should be
1523  * scanned.  The relative value of each set of LRU lists is determined
1524  * by looking at the fraction of the pages scanned we did rotate back
1525  * onto the active list instead of evict.
1526  *
1527  * percent[0] specifies how much pressure to put on ram/swap backed
1528  * memory, while percent[1] determines pressure on the file LRUs.
1529  */
1530 static void get_scan_ratio(struct zone *zone, struct scan_control *sc,
1531                                         unsigned long *percent)
1532 {
1533         unsigned long anon, file, free;
1534         unsigned long anon_prio, file_prio;
1535         unsigned long ap, fp;
1536         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1537
1538         anon  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1539                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1540         file  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1541                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1542
1543         if (scanning_global_lru(sc)) {
1544                 free  = zone_page_state(zone, NR_FREE_PAGES);
1545                 /* If we have very few page cache pages,
1546                    force-scan anon pages. */
1547                 if (unlikely(file + free <= high_wmark_pages(zone))) {
1548                         percent[0] = 100;
1549                         percent[1] = 0;
1550                         return;
1551                 }
1552         }
1553
1554         /*
1555          * OK, so we have swap space and a fair amount of page cache
1556          * pages.  We use the recently rotated / recently scanned
1557          * ratios to determine how valuable each cache is.
1558          *
1559          * Because workloads change over time (and to avoid overflow)
1560          * we keep these statistics as a floating average, which ends
1561          * up weighing recent references more than old ones.
1562          *
1563          * anon in [0], file in [1]
1564          */
1565         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1566                 spin_lock_irq(&zone->lru_lock);
1567                 reclaim_stat->recent_scanned[0] /= 2;
1568                 reclaim_stat->recent_rotated[0] /= 2;
1569                 spin_unlock_irq(&zone->lru_lock);
1570         }
1571
1572         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1573                 spin_lock_irq(&zone->lru_lock);
1574                 reclaim_stat->recent_scanned[1] /= 2;
1575                 reclaim_stat->recent_rotated[1] /= 2;
1576                 spin_unlock_irq(&zone->lru_lock);
1577         }
1578
1579         /*
1580          * With swappiness at 100, anonymous and file have the same priority.
1581          * This scanning priority is essentially the inverse of IO cost.
1582          */
1583         anon_prio = sc->swappiness;
1584         file_prio = 200 - sc->swappiness;
1585
1586         /*
1587          * The amount of pressure on anon vs file pages is inversely
1588          * proportional to the fraction of recently scanned pages on
1589          * each list that were recently referenced and in active use.
1590          */
1591         ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1592         ap /= reclaim_stat->recent_rotated[0] + 1;
1593
1594         fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1595         fp /= reclaim_stat->recent_rotated[1] + 1;
1596
1597         /* Normalize to percentages */
1598         percent[0] = 100 * ap / (ap + fp + 1);
1599         percent[1] = 100 - percent[0];
1600 }
1601
1602 /*
1603  * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1604  * until we collected @swap_cluster_max pages to scan.
1605  */
1606 static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
1607                                        unsigned long *nr_saved_scan)
1608 {
1609         unsigned long nr;
1610
1611         *nr_saved_scan += nr_to_scan;
1612         nr = *nr_saved_scan;
1613
1614         if (nr >= SWAP_CLUSTER_MAX)
1615                 *nr_saved_scan = 0;
1616         else
1617                 nr = 0;
1618
1619         return nr;
1620 }
1621
1622 /*
1623  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1624  */
1625 static void shrink_zone(int priority, struct zone *zone,
1626                                 struct scan_control *sc)
1627 {
1628         unsigned long nr[NR_LRU_LISTS];
1629         unsigned long nr_to_scan;
1630         unsigned long percent[2];       /* anon @ 0; file @ 1 */
1631         enum lru_list l;
1632         unsigned long nr_reclaimed = sc->nr_reclaimed;
1633         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1634         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1635         int noswap = 0;
1636
1637         /* If we have no swap space, do not bother scanning anon pages. */
1638         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1639                 noswap = 1;
1640                 percent[0] = 0;
1641                 percent[1] = 100;
1642         } else
1643                 get_scan_ratio(zone, sc, percent);
1644
1645         for_each_evictable_lru(l) {
1646                 int file = is_file_lru(l);
1647                 unsigned long scan;
1648
1649                 scan = zone_nr_lru_pages(zone, sc, l);
1650                 if (priority || noswap) {
1651                         scan >>= priority;
1652                         scan = (scan * percent[file]) / 100;
1653                 }
1654                 nr[l] = nr_scan_try_batch(scan,
1655                                           &reclaim_stat->nr_saved_scan[l]);
1656         }
1657
1658         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1659                                         nr[LRU_INACTIVE_FILE]) {
1660                 for_each_evictable_lru(l) {
1661                         if (nr[l]) {
1662                                 nr_to_scan = min_t(unsigned long,
1663                                                    nr[l], SWAP_CLUSTER_MAX);
1664                                 nr[l] -= nr_to_scan;
1665
1666                                 nr_reclaimed += shrink_list(l, nr_to_scan,
1667                                                             zone, sc, priority);
1668                         }
1669                 }
1670                 /*
1671                  * On large memory systems, scan >> priority can become
1672                  * really large. This is fine for the starting priority;
1673                  * we want to put equal scanning pressure on each zone.
1674                  * However, if the VM has a harder time of freeing pages,
1675                  * with multiple processes reclaiming pages, the total
1676                  * freeing target can get unreasonably large.
1677                  */
1678                 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
1679                         break;
1680         }
1681
1682         sc->nr_reclaimed = nr_reclaimed;
1683
1684         /*
1685          * Even if we did not try to evict anon pages at all, we want to
1686          * rebalance the anon lru active/inactive ratio.
1687          */
1688         if (inactive_anon_is_low(zone, sc) && nr_swap_pages > 0)
1689                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1690
1691         throttle_vm_writeout(sc->gfp_mask);
1692 }
1693
1694 /*
1695  * This is the direct reclaim path, for page-allocating processes.  We only
1696  * try to reclaim pages from zones which will satisfy the caller's allocation
1697  * request.
1698  *
1699  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1700  * Because:
1701  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1702  *    allocation or
1703  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1704  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1705  *    zone defense algorithm.
1706  *
1707  * If a zone is deemed to be full of pinned pages then just give it a light
1708  * scan then give up on it.
1709  */
1710 static void shrink_zones(int priority, struct zonelist *zonelist,
1711                                         struct scan_control *sc)
1712 {
1713         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1714         struct zoneref *z;
1715         struct zone *zone;
1716
1717         sc->all_unreclaimable = 1;
1718         for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1719                                         sc->nodemask) {
1720                 if (!populated_zone(zone))
1721                         continue;
1722                 /*
1723                  * Take care memory controller reclaiming has small influence
1724                  * to global LRU.
1725                  */
1726                 if (scanning_global_lru(sc)) {
1727                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1728                                 continue;
1729                         note_zone_scanning_priority(zone, priority);
1730
1731                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1732                                 continue;       /* Let kswapd poll it */
1733                         sc->all_unreclaimable = 0;
1734                 } else {
1735                         /*
1736                          * Ignore cpuset limitation here. We just want to reduce
1737                          * # of used pages by us regardless of memory shortage.
1738                          */
1739                         sc->all_unreclaimable = 0;
1740                         mem_cgroup_note_reclaim_priority(sc->mem_cgroup,
1741                                                         priority);
1742                 }
1743
1744                 shrink_zone(priority, zone, sc);
1745         }
1746 }
1747
1748 /*
1749  * This is the main entry point to direct page reclaim.
1750  *
1751  * If a full scan of the inactive list fails to free enough memory then we
1752  * are "out of memory" and something needs to be killed.
1753  *
1754  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1755  * high - the zone may be full of dirty or under-writeback pages, which this
1756  * caller can't do much about.  We kick the writeback threads and take explicit
1757  * naps in the hope that some of these pages can be written.  But if the
1758  * allocating task holds filesystem locks which prevent writeout this might not
1759  * work, and the allocation attempt will fail.
1760  *
1761  * returns:     0, if no pages reclaimed
1762  *              else, the number of pages reclaimed
1763  */
1764 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
1765                                         struct scan_control *sc)
1766 {
1767         int priority;
1768         unsigned long ret = 0;
1769         unsigned long total_scanned = 0;
1770         struct reclaim_state *reclaim_state = current->reclaim_state;
1771         unsigned long lru_pages = 0;
1772         struct zoneref *z;
1773         struct zone *zone;
1774         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1775         unsigned long writeback_threshold;
1776
1777         delayacct_freepages_start();
1778
1779         if (scanning_global_lru(sc))
1780                 count_vm_event(ALLOCSTALL);
1781         /*
1782          * mem_cgroup will not do shrink_slab.
1783          */
1784         if (scanning_global_lru(sc)) {
1785                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1786
1787                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1788                                 continue;
1789
1790                         lru_pages += zone_reclaimable_pages(zone);
1791                 }
1792         }
1793
1794         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1795                 sc->nr_scanned = 0;
1796                 if (!priority)
1797                         disable_swap_token();
1798                 shrink_zones(priority, zonelist, sc);
1799                 /*
1800                  * Don't shrink slabs when reclaiming memory from
1801                  * over limit cgroups
1802                  */
1803                 if (scanning_global_lru(sc)) {
1804                         shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
1805                         if (reclaim_state) {
1806                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
1807                                 reclaim_state->reclaimed_slab = 0;
1808                         }
1809                 }
1810                 total_scanned += sc->nr_scanned;
1811                 if (sc->nr_reclaimed >= sc->nr_to_reclaim) {
1812                         ret = sc->nr_reclaimed;
1813                         goto out;
1814                 }
1815
1816                 /*
1817                  * Try to write back as many pages as we just scanned.  This
1818                  * tends to cause slow streaming writers to write data to the
1819                  * disk smoothly, at the dirtying rate, which is nice.   But
1820                  * that's undesirable in laptop mode, where we *want* lumpy
1821                  * writeout.  So in laptop mode, write out the whole world.
1822                  */
1823                 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
1824                 if (total_scanned > writeback_threshold) {
1825                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
1826                         sc->may_writepage = 1;
1827                 }
1828
1829                 /* Take a nap, wait for some writeback to complete */
1830                 if (!sc->hibernation_mode && sc->nr_scanned &&
1831                     priority < DEF_PRIORITY - 2)
1832                         congestion_wait(BLK_RW_ASYNC, HZ/10);
1833         }
1834         /* top priority shrink_zones still had more to do? don't OOM, then */
1835         if (!sc->all_unreclaimable && scanning_global_lru(sc))
1836                 ret = sc->nr_reclaimed;
1837 out:
1838         /*
1839          * Now that we've scanned all the zones at this priority level, note
1840          * that level within the zone so that the next thread which performs
1841          * scanning of this zone will immediately start out at this priority
1842          * level.  This affects only the decision whether or not to bring
1843          * mapped pages onto the inactive list.
1844          */
1845         if (priority < 0)
1846                 priority = 0;
1847
1848         if (scanning_global_lru(sc)) {
1849                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1850
1851                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1852                                 continue;
1853
1854                         zone->prev_priority = priority;
1855                 }
1856         } else
1857                 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority);
1858
1859         delayacct_freepages_end();
1860
1861         return ret;
1862 }
1863
1864 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
1865                                 gfp_t gfp_mask, nodemask_t *nodemask)
1866 {
1867         struct scan_control sc = {
1868                 .gfp_mask = gfp_mask,
1869                 .may_writepage = !laptop_mode,
1870                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
1871                 .may_unmap = 1,
1872                 .may_swap = 1,
1873                 .swappiness = vm_swappiness,
1874                 .order = order,
1875                 .mem_cgroup = NULL,
1876                 .isolate_pages = isolate_pages_global,
1877                 .nodemask = nodemask,
1878         };
1879
1880         return do_try_to_free_pages(zonelist, &sc);
1881 }
1882
1883 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
1884
1885 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
1886                                                 gfp_t gfp_mask, bool noswap,
1887                                                 unsigned int swappiness,
1888                                                 struct zone *zone, int nid)
1889 {
1890         struct scan_control sc = {
1891                 .may_writepage = !laptop_mode,
1892                 .may_unmap = 1,
1893                 .may_swap = !noswap,
1894                 .swappiness = swappiness,
1895                 .order = 0,
1896                 .mem_cgroup = mem,
1897                 .isolate_pages = mem_cgroup_isolate_pages,
1898         };
1899         nodemask_t nm  = nodemask_of_node(nid);
1900
1901         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1902                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1903         sc.nodemask = &nm;
1904         sc.nr_reclaimed = 0;
1905         sc.nr_scanned = 0;
1906         /*
1907          * NOTE: Although we can get the priority field, using it
1908          * here is not a good idea, since it limits the pages we can scan.
1909          * if we don't reclaim here, the shrink_zone from balance_pgdat
1910          * will pick up pages from other mem cgroup's as well. We hack
1911          * the priority and make it zero.
1912          */
1913         shrink_zone(0, zone, &sc);
1914         return sc.nr_reclaimed;
1915 }
1916
1917 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
1918                                            gfp_t gfp_mask,
1919                                            bool noswap,
1920                                            unsigned int swappiness)
1921 {
1922         struct zonelist *zonelist;
1923         struct scan_control sc = {
1924                 .may_writepage = !laptop_mode,
1925                 .may_unmap = 1,
1926                 .may_swap = !noswap,
1927                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
1928                 .swappiness = swappiness,
1929                 .order = 0,
1930                 .mem_cgroup = mem_cont,
1931                 .isolate_pages = mem_cgroup_isolate_pages,
1932                 .nodemask = NULL, /* we don't care the placement */
1933         };
1934
1935         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1936                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1937         zonelist = NODE_DATA(numa_node_id())->node_zonelists;
1938         return do_try_to_free_pages(zonelist, &sc);
1939 }
1940 #endif
1941
1942 /* is kswapd sleeping prematurely? */
1943 static int sleeping_prematurely(pg_data_t *pgdat, int order, long remaining)
1944 {
1945         int i;
1946
1947         /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
1948         if (remaining)
1949                 return 1;
1950
1951         /* If after HZ/10, a zone is below the high mark, it's premature */
1952         for (i = 0; i < pgdat->nr_zones; i++) {
1953                 struct zone *zone = pgdat->node_zones + i;
1954
1955                 if (!populated_zone(zone))
1956                         continue;
1957
1958                 if (zone->all_unreclaimable)
1959                         continue;
1960
1961                 if (!zone_watermark_ok(zone, order, high_wmark_pages(zone),
1962                                                                 0, 0))
1963                         return 1;
1964         }
1965
1966         return 0;
1967 }
1968
1969 /*
1970  * For kswapd, balance_pgdat() will work across all this node's zones until
1971  * they are all at high_wmark_pages(zone).
1972  *
1973  * Returns the number of pages which were actually freed.
1974  *
1975  * There is special handling here for zones which are full of pinned pages.
1976  * This can happen if the pages are all mlocked, or if they are all used by
1977  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1978  * What we do is to detect the case where all pages in the zone have been
1979  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1980  * dead and from now on, only perform a short scan.  Basically we're polling
1981  * the zone for when the problem goes away.
1982  *
1983  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
1984  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
1985  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
1986  * lower zones regardless of the number of free pages in the lower zones. This
1987  * interoperates with the page allocator fallback scheme to ensure that aging
1988  * of pages is balanced across the zones.
1989  */
1990 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
1991 {
1992         int all_zones_ok;
1993         int priority;
1994         int i;
1995         unsigned long total_scanned;
1996         struct reclaim_state *reclaim_state = current->reclaim_state;
1997         struct scan_control sc = {
1998                 .gfp_mask = GFP_KERNEL,
1999                 .may_unmap = 1,
2000                 .may_swap = 1,
2001                 /*
2002                  * kswapd doesn't want to be bailed out while reclaim. because
2003                  * we want to put equal scanning pressure on each zone.
2004                  */
2005                 .nr_to_reclaim = ULONG_MAX,
2006                 .swappiness = vm_swappiness,
2007                 .order = order,
2008                 .mem_cgroup = NULL,
2009                 .isolate_pages = isolate_pages_global,
2010         };
2011         /*
2012          * temp_priority is used to remember the scanning priority at which
2013          * this zone was successfully refilled to
2014          * free_pages == high_wmark_pages(zone).
2015          */
2016         int temp_priority[MAX_NR_ZONES];
2017
2018 loop_again:
2019         total_scanned = 0;
2020         sc.nr_reclaimed = 0;
2021         sc.may_writepage = !laptop_mode;
2022         count_vm_event(PAGEOUTRUN);
2023
2024         for (i = 0; i < pgdat->nr_zones; i++)
2025                 temp_priority[i] = DEF_PRIORITY;
2026
2027         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2028                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
2029                 unsigned long lru_pages = 0;
2030                 int has_under_min_watermark_zone = 0;
2031
2032                 /* The swap token gets in the way of swapout... */
2033                 if (!priority)
2034                         disable_swap_token();
2035
2036                 all_zones_ok = 1;
2037
2038                 /*
2039                  * Scan in the highmem->dma direction for the highest
2040                  * zone which needs scanning
2041                  */
2042                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2043                         struct zone *zone = pgdat->node_zones + i;
2044
2045                         if (!populated_zone(zone))
2046                                 continue;
2047
2048                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2049                                 continue;
2050
2051                         /*
2052                          * Do some background aging of the anon list, to give
2053                          * pages a chance to be referenced before reclaiming.
2054                          */
2055                         if (inactive_anon_is_low(zone, &sc))
2056                                 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2057                                                         &sc, priority, 0);
2058
2059                         if (!zone_watermark_ok(zone, order,
2060                                         high_wmark_pages(zone), 0, 0)) {
2061                                 end_zone = i;
2062                                 break;
2063                         }
2064                 }
2065                 if (i < 0)
2066                         goto out;
2067
2068                 for (i = 0; i <= end_zone; i++) {
2069                         struct zone *zone = pgdat->node_zones + i;
2070
2071                         lru_pages += zone_reclaimable_pages(zone);
2072                 }
2073
2074                 /*
2075                  * Now scan the zone in the dma->highmem direction, stopping
2076                  * at the last zone which needs scanning.
2077                  *
2078                  * We do this because the page allocator works in the opposite
2079                  * direction.  This prevents the page allocator from allocating
2080                  * pages behind kswapd's direction of progress, which would
2081                  * cause too much scanning of the lower zones.
2082                  */
2083                 for (i = 0; i <= end_zone; i++) {
2084                         struct zone *zone = pgdat->node_zones + i;
2085                         int nr_slab;
2086                         int nid, zid;
2087
2088                         if (!populated_zone(zone))
2089                                 continue;
2090
2091                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2092                                 continue;
2093
2094                         temp_priority[i] = priority;
2095                         sc.nr_scanned = 0;
2096                         note_zone_scanning_priority(zone, priority);
2097
2098                         nid = pgdat->node_id;
2099                         zid = zone_idx(zone);
2100                         /*
2101                          * Call soft limit reclaim before calling shrink_zone.
2102                          * For now we ignore the return value
2103                          */
2104                         mem_cgroup_soft_limit_reclaim(zone, order, sc.gfp_mask,
2105                                                         nid, zid);
2106                         /*
2107                          * We put equal pressure on every zone, unless one
2108                          * zone has way too many pages free already.
2109                          */
2110                         if (!zone_watermark_ok(zone, order,
2111                                         8*high_wmark_pages(zone), end_zone, 0))
2112                                 shrink_zone(priority, zone, &sc);
2113                         reclaim_state->reclaimed_slab = 0;
2114                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
2115                                                 lru_pages);
2116                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2117                         total_scanned += sc.nr_scanned;
2118                         if (zone->all_unreclaimable)
2119                                 continue;
2120                         if (nr_slab == 0 &&
2121                             zone->pages_scanned >= (zone_reclaimable_pages(zone) * 6))
2122                                 zone->all_unreclaimable = 1;
2123                         /*
2124                          * If we've done a decent amount of scanning and
2125                          * the reclaim ratio is low, start doing writepage
2126                          * even in laptop mode
2127                          */
2128                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2129                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2130                                 sc.may_writepage = 1;
2131
2132                         if (!zone_watermark_ok(zone, order,
2133                                         high_wmark_pages(zone), end_zone, 0)) {
2134                                 all_zones_ok = 0;
2135                                 /*
2136                                  * We are still under min water mark.  This
2137                                  * means that we have a GFP_ATOMIC allocation
2138                                  * failure risk. Hurry up!
2139                                  */
2140                                 if (!zone_watermark_ok(zone, order,
2141                                             min_wmark_pages(zone), end_zone, 0))
2142                                         has_under_min_watermark_zone = 1;
2143                         }
2144
2145                 }
2146                 if (all_zones_ok)
2147                         break;          /* kswapd: all done */
2148                 /*
2149                  * OK, kswapd is getting into trouble.  Take a nap, then take
2150                  * another pass across the zones.
2151                  */
2152                 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2153                         if (has_under_min_watermark_zone)
2154                                 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2155                         else
2156                                 congestion_wait(BLK_RW_ASYNC, HZ/10);
2157                 }
2158
2159                 /*
2160                  * We do this so kswapd doesn't build up large priorities for
2161                  * example when it is freeing in parallel with allocators. It
2162                  * matches the direct reclaim path behaviour in terms of impact
2163                  * on zone->*_priority.
2164                  */
2165                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2166                         break;
2167         }
2168 out:
2169         /*
2170          * Note within each zone the priority level at which this zone was
2171          * brought into a happy state.  So that the next thread which scans this
2172          * zone will start out at that priority level.
2173          */
2174         for (i = 0; i < pgdat->nr_zones; i++) {
2175                 struct zone *zone = pgdat->node_zones + i;
2176
2177                 zone->prev_priority = temp_priority[i];
2178         }
2179         if (!all_zones_ok) {
2180                 cond_resched();
2181
2182                 try_to_freeze();
2183
2184                 /*
2185                  * Fragmentation may mean that the system cannot be
2186                  * rebalanced for high-order allocations in all zones.
2187                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2188                  * it means the zones have been fully scanned and are still
2189                  * not balanced. For high-order allocations, there is
2190                  * little point trying all over again as kswapd may
2191                  * infinite loop.
2192                  *
2193                  * Instead, recheck all watermarks at order-0 as they
2194                  * are the most important. If watermarks are ok, kswapd will go
2195                  * back to sleep. High-order users can still perform direct
2196                  * reclaim if they wish.
2197                  */
2198                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2199                         order = sc.order = 0;
2200
2201                 goto loop_again;
2202         }
2203
2204         return sc.nr_reclaimed;
2205 }
2206
2207 /*
2208  * The background pageout daemon, started as a kernel thread
2209  * from the init process.
2210  *
2211  * This basically trickles out pages so that we have _some_
2212  * free memory available even if there is no other activity
2213  * that frees anything up. This is needed for things like routing
2214  * etc, where we otherwise might have all activity going on in
2215  * asynchronous contexts that cannot page things out.
2216  *
2217  * If there are applications that are active memory-allocators
2218  * (most normal use), this basically shouldn't matter.
2219  */
2220 static int kswapd(void *p)
2221 {
2222         unsigned long order;
2223         pg_data_t *pgdat = (pg_data_t*)p;
2224         struct task_struct *tsk = current;
2225         DEFINE_WAIT(wait);
2226         struct reclaim_state reclaim_state = {
2227                 .reclaimed_slab = 0,
2228         };
2229         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2230
2231         lockdep_set_current_reclaim_state(GFP_KERNEL);
2232
2233         if (!cpumask_empty(cpumask))
2234                 set_cpus_allowed_ptr(tsk, cpumask);
2235         current->reclaim_state = &reclaim_state;
2236
2237         /*
2238          * Tell the memory management that we're a "memory allocator",
2239          * and that if we need more memory we should get access to it
2240          * regardless (see "__alloc_pages()"). "kswapd" should
2241          * never get caught in the normal page freeing logic.
2242          *
2243          * (Kswapd normally doesn't need memory anyway, but sometimes
2244          * you need a small amount of memory in order to be able to
2245          * page out something else, and this flag essentially protects
2246          * us from recursively trying to free more memory as we're
2247          * trying to free the first piece of memory in the first place).
2248          */
2249         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2250         set_freezable();
2251
2252         order = 0;
2253         for ( ; ; ) {
2254                 unsigned long new_order;
2255                 int ret;
2256
2257                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2258                 new_order = pgdat->kswapd_max_order;
2259                 pgdat->kswapd_max_order = 0;
2260                 if (order < new_order) {
2261                         /*
2262                          * Don't sleep if someone wants a larger 'order'
2263                          * allocation
2264                          */
2265                         order = new_order;
2266                 } else {
2267                         if (!freezing(current) && !kthread_should_stop()) {
2268                                 long remaining = 0;
2269
2270                                 /* Try to sleep for a short interval */
2271                                 if (!sleeping_prematurely(pgdat, order, remaining)) {
2272                                         remaining = schedule_timeout(HZ/10);
2273                                         finish_wait(&pgdat->kswapd_wait, &wait);
2274                                         prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2275                                 }
2276
2277                                 /*
2278                                  * After a short sleep, check if it was a
2279                                  * premature sleep. If not, then go fully
2280                                  * to sleep until explicitly woken up
2281                                  */
2282                                 if (!sleeping_prematurely(pgdat, order, remaining))
2283                                         schedule();
2284                                 else {
2285                                         if (remaining)
2286                                                 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2287                                         else
2288                                                 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2289                                 }
2290                         }
2291
2292                         order = pgdat->kswapd_max_order;
2293                 }
2294                 finish_wait(&pgdat->kswapd_wait, &wait);
2295
2296                 ret = try_to_freeze();
2297                 if (kthread_should_stop())
2298                         break;
2299
2300                 /*
2301                  * We can speed up thawing tasks if we don't call balance_pgdat
2302                  * after returning from the refrigerator
2303                  */
2304                 if (!ret)
2305                         balance_pgdat(pgdat, order);
2306         }
2307         return 0;
2308 }
2309
2310 /*
2311  * A zone is low on free memory, so wake its kswapd task to service it.
2312  */
2313 void wakeup_kswapd(struct zone *zone, int order)
2314 {
2315         pg_data_t *pgdat;
2316
2317         if (!populated_zone(zone))
2318                 return;
2319
2320         pgdat = zone->zone_pgdat;
2321         if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
2322                 return;
2323         if (pgdat->kswapd_max_order < order)
2324                 pgdat->kswapd_max_order = order;
2325         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2326                 return;
2327         if (!waitqueue_active(&pgdat->kswapd_wait))
2328                 return;
2329         wake_up_interruptible(&pgdat->kswapd_wait);
2330 }
2331
2332 /*
2333  * The reclaimable count would be mostly accurate.
2334  * The less reclaimable pages may be
2335  * - mlocked pages, which will be moved to unevictable list when encountered
2336  * - mapped pages, which may require several travels to be reclaimed
2337  * - dirty pages, which is not "instantly" reclaimable
2338  */
2339 unsigned long global_reclaimable_pages(void)
2340 {
2341         int nr;
2342
2343         nr = global_page_state(NR_ACTIVE_FILE) +
2344              global_page_state(NR_INACTIVE_FILE);
2345
2346         if (nr_swap_pages > 0)
2347                 nr += global_page_state(NR_ACTIVE_ANON) +
2348                       global_page_state(NR_INACTIVE_ANON);
2349
2350         return nr;
2351 }
2352
2353 unsigned long zone_reclaimable_pages(struct zone *zone)
2354 {
2355         int nr;
2356
2357         nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2358              zone_page_state(zone, NR_INACTIVE_FILE);
2359
2360         if (nr_swap_pages > 0)
2361                 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2362                       zone_page_state(zone, NR_INACTIVE_ANON);
2363
2364         return nr;
2365 }
2366
2367 #ifdef CONFIG_HIBERNATION
2368 /*
2369  * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2370  * freed pages.
2371  *
2372  * Rather than trying to age LRUs the aim is to preserve the overall
2373  * LRU order by reclaiming preferentially
2374  * inactive > active > active referenced > active mapped
2375  */
2376 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
2377 {
2378         struct reclaim_state reclaim_state;
2379         struct scan_control sc = {
2380                 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2381                 .may_swap = 1,
2382                 .may_unmap = 1,
2383                 .may_writepage = 1,
2384                 .nr_to_reclaim = nr_to_reclaim,
2385                 .hibernation_mode = 1,
2386                 .swappiness = vm_swappiness,
2387                 .order = 0,
2388                 .isolate_pages = isolate_pages_global,
2389         };
2390         struct zonelist * zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2391         struct task_struct *p = current;
2392         unsigned long nr_reclaimed;
2393
2394         p->flags |= PF_MEMALLOC;
2395         lockdep_set_current_reclaim_state(sc.gfp_mask);
2396         reclaim_state.reclaimed_slab = 0;
2397         p->reclaim_state = &reclaim_state;
2398
2399         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2400
2401         p->reclaim_state = NULL;
2402         lockdep_clear_current_reclaim_state();
2403         p->flags &= ~PF_MEMALLOC;
2404
2405         return nr_reclaimed;
2406 }
2407 #endif /* CONFIG_HIBERNATION */
2408
2409 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2410    not required for correctness.  So if the last cpu in a node goes
2411    away, we get changed to run anywhere: as the first one comes back,
2412    restore their cpu bindings. */
2413 static int __devinit cpu_callback(struct notifier_block *nfb,
2414                                   unsigned long action, void *hcpu)
2415 {
2416         int nid;
2417
2418         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2419                 for_each_node_state(nid, N_HIGH_MEMORY) {
2420                         pg_data_t *pgdat = NODE_DATA(nid);
2421                         const struct cpumask *mask;
2422
2423                         mask = cpumask_of_node(pgdat->node_id);
2424
2425                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2426                                 /* One of our CPUs online: restore mask */
2427                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2428                 }
2429         }
2430         return NOTIFY_OK;
2431 }
2432
2433 /*
2434  * This kswapd start function will be called by init and node-hot-add.
2435  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2436  */
2437 int kswapd_run(int nid)
2438 {
2439         pg_data_t *pgdat = NODE_DATA(nid);
2440         int ret = 0;
2441
2442         if (pgdat->kswapd)
2443                 return 0;
2444
2445         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2446         if (IS_ERR(pgdat->kswapd)) {
2447                 /* failure at boot is fatal */
2448                 BUG_ON(system_state == SYSTEM_BOOTING);
2449                 printk("Failed to start kswapd on node %d\n",nid);
2450                 ret = -1;
2451         }
2452         return ret;
2453 }
2454
2455 /*
2456  * Called by memory hotplug when all memory in a node is offlined.
2457  */
2458 void kswapd_stop(int nid)
2459 {
2460         struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
2461
2462         if (kswapd)
2463                 kthread_stop(kswapd);
2464 }
2465
2466 static int __init kswapd_init(void)
2467 {
2468         int nid;
2469
2470         swap_setup();
2471         for_each_node_state(nid, N_HIGH_MEMORY)
2472                 kswapd_run(nid);
2473         hotcpu_notifier(cpu_callback, 0);
2474         return 0;
2475 }
2476
2477 module_init(kswapd_init)
2478
2479 #ifdef CONFIG_NUMA
2480 /*
2481  * Zone reclaim mode
2482  *
2483  * If non-zero call zone_reclaim when the number of free pages falls below
2484  * the watermarks.
2485  */
2486 int zone_reclaim_mode __read_mostly;
2487
2488 #define RECLAIM_OFF 0
2489 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
2490 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
2491 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
2492
2493 /*
2494  * Priority for ZONE_RECLAIM. This determines the fraction of pages
2495  * of a node considered for each zone_reclaim. 4 scans 1/16th of
2496  * a zone.
2497  */
2498 #define ZONE_RECLAIM_PRIORITY 4
2499
2500 /*
2501  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2502  * occur.
2503  */
2504 int sysctl_min_unmapped_ratio = 1;
2505
2506 /*
2507  * If the number of slab pages in a zone grows beyond this percentage then
2508  * slab reclaim needs to occur.
2509  */
2510 int sysctl_min_slab_ratio = 5;
2511
2512 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
2513 {
2514         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
2515         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
2516                 zone_page_state(zone, NR_ACTIVE_FILE);
2517
2518         /*
2519          * It's possible for there to be more file mapped pages than
2520          * accounted for by the pages on the file LRU lists because
2521          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2522          */
2523         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
2524 }
2525
2526 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
2527 static long zone_pagecache_reclaimable(struct zone *zone)
2528 {
2529         long nr_pagecache_reclaimable;
2530         long delta = 0;
2531
2532         /*
2533          * If RECLAIM_SWAP is set, then all file pages are considered
2534          * potentially reclaimable. Otherwise, we have to worry about
2535          * pages like swapcache and zone_unmapped_file_pages() provides
2536          * a better estimate
2537          */
2538         if (zone_reclaim_mode & RECLAIM_SWAP)
2539                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
2540         else
2541                 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
2542
2543         /* If we can't clean pages, remove dirty pages from consideration */
2544         if (!(zone_reclaim_mode & RECLAIM_WRITE))
2545                 delta += zone_page_state(zone, NR_FILE_DIRTY);
2546
2547         /* Watch for any possible underflows due to delta */
2548         if (unlikely(delta > nr_pagecache_reclaimable))
2549                 delta = nr_pagecache_reclaimable;
2550
2551         return nr_pagecache_reclaimable - delta;
2552 }
2553
2554 /*
2555  * Try to free up some pages from this zone through reclaim.
2556  */
2557 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2558 {
2559         /* Minimum pages needed in order to stay on node */
2560         const unsigned long nr_pages = 1 << order;
2561         struct task_struct *p = current;
2562         struct reclaim_state reclaim_state;
2563         int priority;
2564         struct scan_control sc = {
2565                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2566                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2567                 .may_swap = 1,
2568                 .nr_to_reclaim = max_t(unsigned long, nr_pages,
2569                                        SWAP_CLUSTER_MAX),
2570                 .gfp_mask = gfp_mask,
2571                 .swappiness = vm_swappiness,
2572                 .order = order,
2573                 .isolate_pages = isolate_pages_global,
2574         };
2575         unsigned long slab_reclaimable;
2576
2577         disable_swap_token();
2578         cond_resched();
2579         /*
2580          * We need to be able to allocate from the reserves for RECLAIM_SWAP
2581          * and we also need to be able to write out pages for RECLAIM_WRITE
2582          * and RECLAIM_SWAP.
2583          */
2584         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2585         lockdep_set_current_reclaim_state(gfp_mask);
2586         reclaim_state.reclaimed_slab = 0;
2587         p->reclaim_state = &reclaim_state;
2588
2589         if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
2590                 /*
2591                  * Free memory by calling shrink zone with increasing
2592                  * priorities until we have enough memory freed.
2593                  */
2594                 priority = ZONE_RECLAIM_PRIORITY;
2595                 do {
2596                         note_zone_scanning_priority(zone, priority);
2597                         shrink_zone(priority, zone, &sc);
2598                         priority--;
2599                 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
2600         }
2601
2602         slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2603         if (slab_reclaimable > zone->min_slab_pages) {
2604                 /*
2605                  * shrink_slab() does not currently allow us to determine how
2606                  * many pages were freed in this zone. So we take the current
2607                  * number of slab pages and shake the slab until it is reduced
2608                  * by the same nr_pages that we used for reclaiming unmapped
2609                  * pages.
2610                  *
2611                  * Note that shrink_slab will free memory on all zones and may
2612                  * take a long time.
2613                  */
2614                 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
2615                         zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
2616                                 slab_reclaimable - nr_pages)
2617                         ;
2618
2619                 /*
2620                  * Update nr_reclaimed by the number of slab pages we
2621                  * reclaimed from this zone.
2622                  */
2623                 sc.nr_reclaimed += slab_reclaimable -
2624                         zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2625         }
2626
2627         p->reclaim_state = NULL;
2628         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
2629         lockdep_clear_current_reclaim_state();
2630         return sc.nr_reclaimed >= nr_pages;
2631 }
2632
2633 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2634 {
2635         int node_id;
2636         int ret;
2637
2638         /*
2639          * Zone reclaim reclaims unmapped file backed pages and
2640          * slab pages if we are over the defined limits.
2641          *
2642          * A small portion of unmapped file backed pages is needed for
2643          * file I/O otherwise pages read by file I/O will be immediately
2644          * thrown out if the zone is overallocated. So we do not reclaim
2645          * if less than a specified percentage of the zone is used by
2646          * unmapped file backed pages.
2647          */
2648         if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
2649             zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
2650                 return ZONE_RECLAIM_FULL;
2651
2652         if (zone->all_unreclaimable)
2653                 return ZONE_RECLAIM_FULL;
2654
2655         /*
2656          * Do not scan if the allocation should not be delayed.
2657          */
2658         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
2659                 return ZONE_RECLAIM_NOSCAN;
2660
2661         /*
2662          * Only run zone reclaim on the local zone or on zones that do not
2663          * have associated processors. This will favor the local processor
2664          * over remote processors and spread off node memory allocations
2665          * as wide as possible.
2666          */
2667         node_id = zone_to_nid(zone);
2668         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
2669                 return ZONE_RECLAIM_NOSCAN;
2670
2671         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2672                 return ZONE_RECLAIM_NOSCAN;
2673
2674         ret = __zone_reclaim(zone, gfp_mask, order);
2675         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2676
2677         if (!ret)
2678                 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
2679
2680         return ret;
2681 }
2682 #endif
2683
2684 /*
2685  * page_evictable - test whether a page is evictable
2686  * @page: the page to test
2687  * @vma: the VMA in which the page is or will be mapped, may be NULL
2688  *
2689  * Test whether page is evictable--i.e., should be placed on active/inactive
2690  * lists vs unevictable list.  The vma argument is !NULL when called from the
2691  * fault path to determine how to instantate a new page.
2692  *
2693  * Reasons page might not be evictable:
2694  * (1) page's mapping marked unevictable
2695  * (2) page is part of an mlocked VMA
2696  *
2697  */
2698 int page_evictable(struct page *page, struct vm_area_struct *vma)
2699 {
2700
2701         if (mapping_unevictable(page_mapping(page)))
2702                 return 0;
2703
2704         if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2705                 return 0;
2706
2707         return 1;
2708 }
2709
2710 /**
2711  * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2712  * @page: page to check evictability and move to appropriate lru list
2713  * @zone: zone page is in
2714  *
2715  * Checks a page for evictability and moves the page to the appropriate
2716  * zone lru list.
2717  *
2718  * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2719  * have PageUnevictable set.
2720  */
2721 static void check_move_unevictable_page(struct page *page, struct zone *zone)
2722 {
2723         VM_BUG_ON(PageActive(page));
2724
2725 retry:
2726         ClearPageUnevictable(page);
2727         if (page_evictable(page, NULL)) {
2728                 enum lru_list l = page_lru_base_type(page);
2729
2730                 __dec_zone_state(zone, NR_UNEVICTABLE);
2731                 list_move(&page->lru, &zone->lru[l].list);
2732                 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
2733                 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
2734                 __count_vm_event(UNEVICTABLE_PGRESCUED);
2735         } else {
2736                 /*
2737                  * rotate unevictable list
2738                  */
2739                 SetPageUnevictable(page);
2740                 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
2741                 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
2742                 if (page_evictable(page, NULL))
2743                         goto retry;
2744         }
2745 }
2746
2747 /**
2748  * scan_mapping_unevictable_pages - scan an address space for evictable pages
2749  * @mapping: struct address_space to scan for evictable pages
2750  *
2751  * Scan all pages in mapping.  Check unevictable pages for
2752  * evictability and move them to the appropriate zone lru list.
2753  */
2754 void scan_mapping_unevictable_pages(struct address_space *mapping)
2755 {
2756         pgoff_t next = 0;
2757         pgoff_t end   = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
2758                          PAGE_CACHE_SHIFT;
2759         struct zone *zone;
2760         struct pagevec pvec;
2761
2762         if (mapping->nrpages == 0)
2763                 return;
2764
2765         pagevec_init(&pvec, 0);
2766         while (next < end &&
2767                 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
2768                 int i;
2769                 int pg_scanned = 0;
2770
2771                 zone = NULL;
2772
2773                 for (i = 0; i < pagevec_count(&pvec); i++) {
2774                         struct page *page = pvec.pages[i];
2775                         pgoff_t page_index = page->index;
2776                         struct zone *pagezone = page_zone(page);
2777
2778                         pg_scanned++;
2779                         if (page_index > next)
2780                                 next = page_index;
2781                         next++;
2782
2783                         if (pagezone != zone) {
2784                                 if (zone)
2785                                         spin_unlock_irq(&zone->lru_lock);
2786                                 zone = pagezone;
2787                                 spin_lock_irq(&zone->lru_lock);
2788                         }
2789
2790                         if (PageLRU(page) && PageUnevictable(page))
2791                                 check_move_unevictable_page(page, zone);
2792                 }
2793                 if (zone)
2794                         spin_unlock_irq(&zone->lru_lock);
2795                 pagevec_release(&pvec);
2796
2797                 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
2798         }
2799
2800 }
2801
2802 /**
2803  * scan_zone_unevictable_pages - check unevictable list for evictable pages
2804  * @zone - zone of which to scan the unevictable list
2805  *
2806  * Scan @zone's unevictable LRU lists to check for pages that have become
2807  * evictable.  Move those that have to @zone's inactive list where they
2808  * become candidates for reclaim, unless shrink_inactive_zone() decides
2809  * to reactivate them.  Pages that are still unevictable are rotated
2810  * back onto @zone's unevictable list.
2811  */
2812 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
2813 static void scan_zone_unevictable_pages(struct zone *zone)
2814 {
2815         struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
2816         unsigned long scan;
2817         unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
2818
2819         while (nr_to_scan > 0) {
2820                 unsigned long batch_size = min(nr_to_scan,
2821                                                 SCAN_UNEVICTABLE_BATCH_SIZE);
2822
2823                 spin_lock_irq(&zone->lru_lock);
2824                 for (scan = 0;  scan < batch_size; scan++) {
2825                         struct page *page = lru_to_page(l_unevictable);
2826
2827                         if (!trylock_page(page))
2828                                 continue;
2829
2830                         prefetchw_prev_lru_page(page, l_unevictable, flags);
2831
2832                         if (likely(PageLRU(page) && PageUnevictable(page)))
2833                                 check_move_unevictable_page(page, zone);
2834
2835                         unlock_page(page);
2836                 }
2837                 spin_unlock_irq(&zone->lru_lock);
2838
2839                 nr_to_scan -= batch_size;
2840         }
2841 }
2842
2843
2844 /**
2845  * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
2846  *
2847  * A really big hammer:  scan all zones' unevictable LRU lists to check for
2848  * pages that have become evictable.  Move those back to the zones'
2849  * inactive list where they become candidates for reclaim.
2850  * This occurs when, e.g., we have unswappable pages on the unevictable lists,
2851  * and we add swap to the system.  As such, it runs in the context of a task
2852  * that has possibly/probably made some previously unevictable pages
2853  * evictable.
2854  */
2855 static void scan_all_zones_unevictable_pages(void)
2856 {
2857         struct zone *zone;
2858
2859         for_each_zone(zone) {
2860                 scan_zone_unevictable_pages(zone);
2861         }
2862 }
2863
2864 /*
2865  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
2866  * all nodes' unevictable lists for evictable pages
2867  */
2868 unsigned long scan_unevictable_pages;
2869
2870 int scan_unevictable_handler(struct ctl_table *table, int write,
2871                            void __user *buffer,
2872                            size_t *length, loff_t *ppos)
2873 {
2874         proc_doulongvec_minmax(table, write, buffer, length, ppos);
2875
2876         if (write && *(unsigned long *)table->data)
2877                 scan_all_zones_unevictable_pages();
2878
2879         scan_unevictable_pages = 0;
2880         return 0;
2881 }
2882
2883 /*
2884  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
2885  * a specified node's per zone unevictable lists for evictable pages.
2886  */
2887
2888 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
2889                                           struct sysdev_attribute *attr,
2890                                           char *buf)
2891 {
2892         return sprintf(buf, "0\n");     /* always zero; should fit... */
2893 }
2894
2895 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
2896                                            struct sysdev_attribute *attr,
2897                                         const char *buf, size_t count)
2898 {
2899         struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
2900         struct zone *zone;
2901         unsigned long res;
2902         unsigned long req = strict_strtoul(buf, 10, &res);
2903
2904         if (!req)
2905                 return 1;       /* zero is no-op */
2906
2907         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2908                 if (!populated_zone(zone))
2909                         continue;
2910                 scan_zone_unevictable_pages(zone);
2911         }
2912         return 1;
2913 }
2914
2915
2916 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
2917                         read_scan_unevictable_node,
2918                         write_scan_unevictable_node);
2919
2920 int scan_unevictable_register_node(struct node *node)
2921 {
2922         return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
2923 }
2924
2925 void scan_unevictable_unregister_node(struct node *node)
2926 {
2927         sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
2928 }
2929