vmscan: zone_reclaim don't call disable_swap_token()
[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 order;
77
78         /*
79          * Intend to reclaim enough contenious memory rather than to reclaim
80          * enough amount memory. I.e, it's the mode for high order allocation.
81          */
82         bool lumpy_reclaim_mode;
83
84         /* Which cgroup do we reclaim from */
85         struct mem_cgroup *mem_cgroup;
86
87         /*
88          * Nodemask of nodes allowed by the caller. If NULL, all nodes
89          * are scanned.
90          */
91         nodemask_t      *nodemask;
92 };
93
94 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
95
96 #ifdef ARCH_HAS_PREFETCH
97 #define prefetch_prev_lru_page(_page, _base, _field)                    \
98         do {                                                            \
99                 if ((_page)->lru.prev != _base) {                       \
100                         struct page *prev;                              \
101                                                                         \
102                         prev = lru_to_page(&(_page->lru));              \
103                         prefetch(&prev->_field);                        \
104                 }                                                       \
105         } while (0)
106 #else
107 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
108 #endif
109
110 #ifdef ARCH_HAS_PREFETCHW
111 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
112         do {                                                            \
113                 if ((_page)->lru.prev != _base) {                       \
114                         struct page *prev;                              \
115                                                                         \
116                         prev = lru_to_page(&(_page->lru));              \
117                         prefetchw(&prev->_field);                       \
118                 }                                                       \
119         } while (0)
120 #else
121 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
122 #endif
123
124 /*
125  * From 0 .. 100.  Higher means more swappy.
126  */
127 int vm_swappiness = 60;
128 long vm_total_pages;    /* The total number of pages which the VM controls */
129
130 static LIST_HEAD(shrinker_list);
131 static DECLARE_RWSEM(shrinker_rwsem);
132
133 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
134 #define scanning_global_lru(sc) (!(sc)->mem_cgroup)
135 #else
136 #define scanning_global_lru(sc) (1)
137 #endif
138
139 static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
140                                                   struct scan_control *sc)
141 {
142         if (!scanning_global_lru(sc))
143                 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
144
145         return &zone->reclaim_stat;
146 }
147
148 static unsigned long zone_nr_lru_pages(struct zone *zone,
149                                 struct scan_control *sc, enum lru_list lru)
150 {
151         if (!scanning_global_lru(sc))
152                 return mem_cgroup_zone_nr_pages(sc->mem_cgroup, zone, lru);
153
154         return zone_page_state(zone, NR_LRU_BASE + lru);
155 }
156
157
158 /*
159  * Add a shrinker callback to be called from the vm
160  */
161 void register_shrinker(struct shrinker *shrinker)
162 {
163         shrinker->nr = 0;
164         down_write(&shrinker_rwsem);
165         list_add_tail(&shrinker->list, &shrinker_list);
166         up_write(&shrinker_rwsem);
167 }
168 EXPORT_SYMBOL(register_shrinker);
169
170 /*
171  * Remove one
172  */
173 void unregister_shrinker(struct shrinker *shrinker)
174 {
175         down_write(&shrinker_rwsem);
176         list_del(&shrinker->list);
177         up_write(&shrinker_rwsem);
178 }
179 EXPORT_SYMBOL(unregister_shrinker);
180
181 #define SHRINK_BATCH 128
182 /*
183  * Call the shrink functions to age shrinkable caches
184  *
185  * Here we assume it costs one seek to replace a lru page and that it also
186  * takes a seek to recreate a cache object.  With this in mind we age equal
187  * percentages of the lru and ageable caches.  This should balance the seeks
188  * generated by these structures.
189  *
190  * If the vm encountered mapped pages on the LRU it increase the pressure on
191  * slab to avoid swapping.
192  *
193  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
194  *
195  * `lru_pages' represents the number of on-LRU pages in all the zones which
196  * are eligible for the caller's allocation attempt.  It is used for balancing
197  * slab reclaim versus page reclaim.
198  *
199  * Returns the number of slab objects which we shrunk.
200  */
201 unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
202                         unsigned long lru_pages)
203 {
204         struct shrinker *shrinker;
205         unsigned long ret = 0;
206
207         if (scanned == 0)
208                 scanned = SWAP_CLUSTER_MAX;
209
210         if (!down_read_trylock(&shrinker_rwsem))
211                 return 1;       /* Assume we'll be able to shrink next time */
212
213         list_for_each_entry(shrinker, &shrinker_list, list) {
214                 unsigned long long delta;
215                 unsigned long total_scan;
216                 unsigned long max_pass;
217
218                 max_pass = (*shrinker->shrink)(shrinker, 0, gfp_mask);
219                 delta = (4 * scanned) / shrinker->seeks;
220                 delta *= max_pass;
221                 do_div(delta, lru_pages + 1);
222                 shrinker->nr += delta;
223                 if (shrinker->nr < 0) {
224                         printk(KERN_ERR "shrink_slab: %pF negative objects to "
225                                "delete nr=%ld\n",
226                                shrinker->shrink, shrinker->nr);
227                         shrinker->nr = max_pass;
228                 }
229
230                 /*
231                  * Avoid risking looping forever due to too large nr value:
232                  * never try to free more than twice the estimate number of
233                  * freeable entries.
234                  */
235                 if (shrinker->nr > max_pass * 2)
236                         shrinker->nr = max_pass * 2;
237
238                 total_scan = shrinker->nr;
239                 shrinker->nr = 0;
240
241                 while (total_scan >= SHRINK_BATCH) {
242                         long this_scan = SHRINK_BATCH;
243                         int shrink_ret;
244                         int nr_before;
245
246                         nr_before = (*shrinker->shrink)(shrinker, 0, gfp_mask);
247                         shrink_ret = (*shrinker->shrink)(shrinker, this_scan,
248                                                                 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_nosync(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->lumpy_reclaim_mode)
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 /*
843  * Attempt to remove the specified page from its LRU.  Only take this page
844  * if it is of the appropriate PageActive status.  Pages which are being
845  * freed elsewhere are also ignored.
846  *
847  * page:        page to consider
848  * mode:        one of the LRU isolation modes defined above
849  *
850  * returns 0 on success, -ve errno on failure.
851  */
852 int __isolate_lru_page(struct page *page, int mode, int file)
853 {
854         int ret = -EINVAL;
855
856         /* Only take pages on the LRU. */
857         if (!PageLRU(page))
858                 return ret;
859
860         /*
861          * When checking the active state, we need to be sure we are
862          * dealing with comparible boolean values.  Take the logical not
863          * of each.
864          */
865         if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
866                 return ret;
867
868         if (mode != ISOLATE_BOTH && page_is_file_cache(page) != file)
869                 return ret;
870
871         /*
872          * When this function is being called for lumpy reclaim, we
873          * initially look into all LRU pages, active, inactive and
874          * unevictable; only give shrink_page_list evictable pages.
875          */
876         if (PageUnevictable(page))
877                 return ret;
878
879         ret = -EBUSY;
880
881         if (likely(get_page_unless_zero(page))) {
882                 /*
883                  * Be careful not to clear PageLRU until after we're
884                  * sure the page is not being freed elsewhere -- the
885                  * page release code relies on it.
886                  */
887                 ClearPageLRU(page);
888                 ret = 0;
889         }
890
891         return ret;
892 }
893
894 /*
895  * zone->lru_lock is heavily contended.  Some of the functions that
896  * shrink the lists perform better by taking out a batch of pages
897  * and working on them outside the LRU lock.
898  *
899  * For pagecache intensive workloads, this function is the hottest
900  * spot in the kernel (apart from copy_*_user functions).
901  *
902  * Appropriate locks must be held before calling this function.
903  *
904  * @nr_to_scan: The number of pages to look through on the list.
905  * @src:        The LRU list to pull pages off.
906  * @dst:        The temp list to put pages on to.
907  * @scanned:    The number of pages that were scanned.
908  * @order:      The caller's attempted allocation order
909  * @mode:       One of the LRU isolation modes
910  * @file:       True [1] if isolating file [!anon] pages
911  *
912  * returns how many pages were moved onto *@dst.
913  */
914 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
915                 struct list_head *src, struct list_head *dst,
916                 unsigned long *scanned, int order, int mode, int file)
917 {
918         unsigned long nr_taken = 0;
919         unsigned long scan;
920
921         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
922                 struct page *page;
923                 unsigned long pfn;
924                 unsigned long end_pfn;
925                 unsigned long page_pfn;
926                 int zone_id;
927
928                 page = lru_to_page(src);
929                 prefetchw_prev_lru_page(page, src, flags);
930
931                 VM_BUG_ON(!PageLRU(page));
932
933                 switch (__isolate_lru_page(page, mode, file)) {
934                 case 0:
935                         list_move(&page->lru, dst);
936                         mem_cgroup_del_lru(page);
937                         nr_taken++;
938                         break;
939
940                 case -EBUSY:
941                         /* else it is being freed elsewhere */
942                         list_move(&page->lru, src);
943                         mem_cgroup_rotate_lru_list(page, page_lru(page));
944                         continue;
945
946                 default:
947                         BUG();
948                 }
949
950                 if (!order)
951                         continue;
952
953                 /*
954                  * Attempt to take all pages in the order aligned region
955                  * surrounding the tag page.  Only take those pages of
956                  * the same active state as that tag page.  We may safely
957                  * round the target page pfn down to the requested order
958                  * as the mem_map is guarenteed valid out to MAX_ORDER,
959                  * where that page is in a different zone we will detect
960                  * it from its zone id and abort this block scan.
961                  */
962                 zone_id = page_zone_id(page);
963                 page_pfn = page_to_pfn(page);
964                 pfn = page_pfn & ~((1 << order) - 1);
965                 end_pfn = pfn + (1 << order);
966                 for (; pfn < end_pfn; pfn++) {
967                         struct page *cursor_page;
968
969                         /* The target page is in the block, ignore it. */
970                         if (unlikely(pfn == page_pfn))
971                                 continue;
972
973                         /* Avoid holes within the zone. */
974                         if (unlikely(!pfn_valid_within(pfn)))
975                                 break;
976
977                         cursor_page = pfn_to_page(pfn);
978
979                         /* Check that we have not crossed a zone boundary. */
980                         if (unlikely(page_zone_id(cursor_page) != zone_id))
981                                 continue;
982
983                         /*
984                          * If we don't have enough swap space, reclaiming of
985                          * anon page which don't already have a swap slot is
986                          * pointless.
987                          */
988                         if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
989                                         !PageSwapCache(cursor_page))
990                                 continue;
991
992                         if (__isolate_lru_page(cursor_page, mode, file) == 0) {
993                                 list_move(&cursor_page->lru, dst);
994                                 mem_cgroup_del_lru(cursor_page);
995                                 nr_taken++;
996                                 scan++;
997                         }
998                 }
999         }
1000
1001         *scanned = scan;
1002         return nr_taken;
1003 }
1004
1005 static unsigned long isolate_pages_global(unsigned long nr,
1006                                         struct list_head *dst,
1007                                         unsigned long *scanned, int order,
1008                                         int mode, struct zone *z,
1009                                         int active, int file)
1010 {
1011         int lru = LRU_BASE;
1012         if (active)
1013                 lru += LRU_ACTIVE;
1014         if (file)
1015                 lru += LRU_FILE;
1016         return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1017                                                                 mode, file);
1018 }
1019
1020 /*
1021  * clear_active_flags() is a helper for shrink_active_list(), clearing
1022  * any active bits from the pages in the list.
1023  */
1024 static unsigned long clear_active_flags(struct list_head *page_list,
1025                                         unsigned int *count)
1026 {
1027         int nr_active = 0;
1028         int lru;
1029         struct page *page;
1030
1031         list_for_each_entry(page, page_list, lru) {
1032                 lru = page_lru_base_type(page);
1033                 if (PageActive(page)) {
1034                         lru += LRU_ACTIVE;
1035                         ClearPageActive(page);
1036                         nr_active++;
1037                 }
1038                 count[lru]++;
1039         }
1040
1041         return nr_active;
1042 }
1043
1044 /**
1045  * isolate_lru_page - tries to isolate a page from its LRU list
1046  * @page: page to isolate from its LRU list
1047  *
1048  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1049  * vmstat statistic corresponding to whatever LRU list the page was on.
1050  *
1051  * Returns 0 if the page was removed from an LRU list.
1052  * Returns -EBUSY if the page was not on an LRU list.
1053  *
1054  * The returned page will have PageLRU() cleared.  If it was found on
1055  * the active list, it will have PageActive set.  If it was found on
1056  * the unevictable list, it will have the PageUnevictable bit set. That flag
1057  * may need to be cleared by the caller before letting the page go.
1058  *
1059  * The vmstat statistic corresponding to the list on which the page was
1060  * found will be decremented.
1061  *
1062  * Restrictions:
1063  * (1) Must be called with an elevated refcount on the page. This is a
1064  *     fundamentnal difference from isolate_lru_pages (which is called
1065  *     without a stable reference).
1066  * (2) the lru_lock must not be held.
1067  * (3) interrupts must be enabled.
1068  */
1069 int isolate_lru_page(struct page *page)
1070 {
1071         int ret = -EBUSY;
1072
1073         if (PageLRU(page)) {
1074                 struct zone *zone = page_zone(page);
1075
1076                 spin_lock_irq(&zone->lru_lock);
1077                 if (PageLRU(page) && get_page_unless_zero(page)) {
1078                         int lru = page_lru(page);
1079                         ret = 0;
1080                         ClearPageLRU(page);
1081
1082                         del_page_from_lru_list(zone, page, lru);
1083                 }
1084                 spin_unlock_irq(&zone->lru_lock);
1085         }
1086         return ret;
1087 }
1088
1089 /*
1090  * Are there way too many processes in the direct reclaim path already?
1091  */
1092 static int too_many_isolated(struct zone *zone, int file,
1093                 struct scan_control *sc)
1094 {
1095         unsigned long inactive, isolated;
1096
1097         if (current_is_kswapd())
1098                 return 0;
1099
1100         if (!scanning_global_lru(sc))
1101                 return 0;
1102
1103         if (file) {
1104                 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1105                 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1106         } else {
1107                 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1108                 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1109         }
1110
1111         return isolated > inactive;
1112 }
1113
1114 /*
1115  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1116  * of reclaimed pages
1117  */
1118 static unsigned long shrink_inactive_list(unsigned long max_scan,
1119                         struct zone *zone, struct scan_control *sc,
1120                         int priority, int file)
1121 {
1122         LIST_HEAD(page_list);
1123         struct pagevec pvec;
1124         unsigned long nr_scanned = 0;
1125         unsigned long nr_reclaimed = 0;
1126         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1127
1128         while (unlikely(too_many_isolated(zone, file, sc))) {
1129                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1130
1131                 /* We are about to die and free our memory. Return now. */
1132                 if (fatal_signal_pending(current))
1133                         return SWAP_CLUSTER_MAX;
1134         }
1135
1136
1137         pagevec_init(&pvec, 1);
1138
1139         lru_add_drain();
1140         spin_lock_irq(&zone->lru_lock);
1141         do {
1142                 struct page *page;
1143                 unsigned long nr_taken;
1144                 unsigned long nr_scan;
1145                 unsigned long nr_freed;
1146                 unsigned long nr_active;
1147                 unsigned int count[NR_LRU_LISTS] = { 0, };
1148                 int mode = sc->lumpy_reclaim_mode ? ISOLATE_BOTH : ISOLATE_INACTIVE;
1149                 unsigned long nr_anon;
1150                 unsigned long nr_file;
1151
1152                 if (scanning_global_lru(sc)) {
1153                         nr_taken = isolate_pages_global(SWAP_CLUSTER_MAX,
1154                                                         &page_list, &nr_scan,
1155                                                         sc->order, mode,
1156                                                         zone, 0, file);
1157                         zone->pages_scanned += nr_scan;
1158                         if (current_is_kswapd())
1159                                 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1160                                                        nr_scan);
1161                         else
1162                                 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1163                                                        nr_scan);
1164                 } else {
1165                         nr_taken = mem_cgroup_isolate_pages(SWAP_CLUSTER_MAX,
1166                                                         &page_list, &nr_scan,
1167                                                         sc->order, mode,
1168                                                         zone, sc->mem_cgroup,
1169                                                         0, file);
1170                         /*
1171                          * mem_cgroup_isolate_pages() keeps track of
1172                          * scanned pages on its own.
1173                          */
1174                 }
1175
1176                 if (nr_taken == 0)
1177                         goto done;
1178
1179                 nr_active = clear_active_flags(&page_list, count);
1180                 __count_vm_events(PGDEACTIVATE, nr_active);
1181
1182                 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1183                                                 -count[LRU_ACTIVE_FILE]);
1184                 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1185                                                 -count[LRU_INACTIVE_FILE]);
1186                 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1187                                                 -count[LRU_ACTIVE_ANON]);
1188                 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1189                                                 -count[LRU_INACTIVE_ANON]);
1190
1191                 nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1192                 nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1193                 __mod_zone_page_state(zone, NR_ISOLATED_ANON, nr_anon);
1194                 __mod_zone_page_state(zone, NR_ISOLATED_FILE, nr_file);
1195
1196                 reclaim_stat->recent_scanned[0] += nr_anon;
1197                 reclaim_stat->recent_scanned[1] += nr_file;
1198
1199                 spin_unlock_irq(&zone->lru_lock);
1200
1201                 nr_scanned += nr_scan;
1202                 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
1203
1204                 /*
1205                  * If we are direct reclaiming for contiguous pages and we do
1206                  * not reclaim everything in the list, try again and wait
1207                  * for IO to complete. This will stall high-order allocations
1208                  * but that should be acceptable to the caller
1209                  */
1210                 if (nr_freed < nr_taken && !current_is_kswapd() &&
1211                     sc->lumpy_reclaim_mode) {
1212                         congestion_wait(BLK_RW_ASYNC, HZ/10);
1213
1214                         /*
1215                          * The attempt at page out may have made some
1216                          * of the pages active, mark them inactive again.
1217                          */
1218                         nr_active = clear_active_flags(&page_list, count);
1219                         count_vm_events(PGDEACTIVATE, nr_active);
1220
1221                         nr_freed += shrink_page_list(&page_list, sc,
1222                                                         PAGEOUT_IO_SYNC);
1223                 }
1224
1225                 nr_reclaimed += nr_freed;
1226
1227                 local_irq_disable();
1228                 if (current_is_kswapd())
1229                         __count_vm_events(KSWAPD_STEAL, nr_freed);
1230                 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
1231
1232                 spin_lock(&zone->lru_lock);
1233                 /*
1234                  * Put back any unfreeable pages.
1235                  */
1236                 while (!list_empty(&page_list)) {
1237                         int lru;
1238                         page = lru_to_page(&page_list);
1239                         VM_BUG_ON(PageLRU(page));
1240                         list_del(&page->lru);
1241                         if (unlikely(!page_evictable(page, NULL))) {
1242                                 spin_unlock_irq(&zone->lru_lock);
1243                                 putback_lru_page(page);
1244                                 spin_lock_irq(&zone->lru_lock);
1245                                 continue;
1246                         }
1247                         SetPageLRU(page);
1248                         lru = page_lru(page);
1249                         add_page_to_lru_list(zone, page, lru);
1250                         if (is_active_lru(lru)) {
1251                                 int file = is_file_lru(lru);
1252                                 reclaim_stat->recent_rotated[file]++;
1253                         }
1254                         if (!pagevec_add(&pvec, page)) {
1255                                 spin_unlock_irq(&zone->lru_lock);
1256                                 __pagevec_release(&pvec);
1257                                 spin_lock_irq(&zone->lru_lock);
1258                         }
1259                 }
1260                 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1261                 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1262
1263         } while (nr_scanned < max_scan);
1264
1265 done:
1266         spin_unlock_irq(&zone->lru_lock);
1267         pagevec_release(&pvec);
1268         return nr_reclaimed;
1269 }
1270
1271 /*
1272  * We are about to scan this zone at a certain priority level.  If that priority
1273  * level is smaller (ie: more urgent) than the previous priority, then note
1274  * that priority level within the zone.  This is done so that when the next
1275  * process comes in to scan this zone, it will immediately start out at this
1276  * priority level rather than having to build up its own scanning priority.
1277  * Here, this priority affects only the reclaim-mapped threshold.
1278  */
1279 static inline void note_zone_scanning_priority(struct zone *zone, int priority)
1280 {
1281         if (priority < zone->prev_priority)
1282                 zone->prev_priority = priority;
1283 }
1284
1285 /*
1286  * This moves pages from the active list to the inactive list.
1287  *
1288  * We move them the other way if the page is referenced by one or more
1289  * processes, from rmap.
1290  *
1291  * If the pages are mostly unmapped, the processing is fast and it is
1292  * appropriate to hold zone->lru_lock across the whole operation.  But if
1293  * the pages are mapped, the processing is slow (page_referenced()) so we
1294  * should drop zone->lru_lock around each page.  It's impossible to balance
1295  * this, so instead we remove the pages from the LRU while processing them.
1296  * It is safe to rely on PG_active against the non-LRU pages in here because
1297  * nobody will play with that bit on a non-LRU page.
1298  *
1299  * The downside is that we have to touch page->_count against each page.
1300  * But we had to alter page->flags anyway.
1301  */
1302
1303 static void move_active_pages_to_lru(struct zone *zone,
1304                                      struct list_head *list,
1305                                      enum lru_list lru)
1306 {
1307         unsigned long pgmoved = 0;
1308         struct pagevec pvec;
1309         struct page *page;
1310
1311         pagevec_init(&pvec, 1);
1312
1313         while (!list_empty(list)) {
1314                 page = lru_to_page(list);
1315
1316                 VM_BUG_ON(PageLRU(page));
1317                 SetPageLRU(page);
1318
1319                 list_move(&page->lru, &zone->lru[lru].list);
1320                 mem_cgroup_add_lru_list(page, lru);
1321                 pgmoved++;
1322
1323                 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1324                         spin_unlock_irq(&zone->lru_lock);
1325                         if (buffer_heads_over_limit)
1326                                 pagevec_strip(&pvec);
1327                         __pagevec_release(&pvec);
1328                         spin_lock_irq(&zone->lru_lock);
1329                 }
1330         }
1331         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1332         if (!is_active_lru(lru))
1333                 __count_vm_events(PGDEACTIVATE, pgmoved);
1334 }
1335
1336 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1337                         struct scan_control *sc, int priority, int file)
1338 {
1339         unsigned long nr_taken;
1340         unsigned long pgscanned;
1341         unsigned long vm_flags;
1342         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1343         LIST_HEAD(l_active);
1344         LIST_HEAD(l_inactive);
1345         struct page *page;
1346         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1347         unsigned long nr_rotated = 0;
1348
1349         lru_add_drain();
1350         spin_lock_irq(&zone->lru_lock);
1351         if (scanning_global_lru(sc)) {
1352                 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1353                                                 &pgscanned, sc->order,
1354                                                 ISOLATE_ACTIVE, zone,
1355                                                 1, file);
1356                 zone->pages_scanned += pgscanned;
1357         } else {
1358                 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1359                                                 &pgscanned, sc->order,
1360                                                 ISOLATE_ACTIVE, zone,
1361                                                 sc->mem_cgroup, 1, file);
1362                 /*
1363                  * mem_cgroup_isolate_pages() keeps track of
1364                  * scanned pages on its own.
1365                  */
1366         }
1367
1368         reclaim_stat->recent_scanned[file] += nr_taken;
1369
1370         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1371         if (file)
1372                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1373         else
1374                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1375         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1376         spin_unlock_irq(&zone->lru_lock);
1377
1378         while (!list_empty(&l_hold)) {
1379                 cond_resched();
1380                 page = lru_to_page(&l_hold);
1381                 list_del(&page->lru);
1382
1383                 if (unlikely(!page_evictable(page, NULL))) {
1384                         putback_lru_page(page);
1385                         continue;
1386                 }
1387
1388                 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1389                         nr_rotated++;
1390                         /*
1391                          * Identify referenced, file-backed active pages and
1392                          * give them one more trip around the active list. So
1393                          * that executable code get better chances to stay in
1394                          * memory under moderate memory pressure.  Anon pages
1395                          * are not likely to be evicted by use-once streaming
1396                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1397                          * so we ignore them here.
1398                          */
1399                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1400                                 list_add(&page->lru, &l_active);
1401                                 continue;
1402                         }
1403                 }
1404
1405                 ClearPageActive(page);  /* we are de-activating */
1406                 list_add(&page->lru, &l_inactive);
1407         }
1408
1409         /*
1410          * Move pages back to the lru list.
1411          */
1412         spin_lock_irq(&zone->lru_lock);
1413         /*
1414          * Count referenced pages from currently used mappings as rotated,
1415          * even though only some of them are actually re-activated.  This
1416          * helps balance scan pressure between file and anonymous pages in
1417          * get_scan_ratio.
1418          */
1419         reclaim_stat->recent_rotated[file] += nr_rotated;
1420
1421         move_active_pages_to_lru(zone, &l_active,
1422                                                 LRU_ACTIVE + file * LRU_FILE);
1423         move_active_pages_to_lru(zone, &l_inactive,
1424                                                 LRU_BASE   + file * LRU_FILE);
1425         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1426         spin_unlock_irq(&zone->lru_lock);
1427 }
1428
1429 static int inactive_anon_is_low_global(struct zone *zone)
1430 {
1431         unsigned long active, inactive;
1432
1433         active = zone_page_state(zone, NR_ACTIVE_ANON);
1434         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1435
1436         if (inactive * zone->inactive_ratio < active)
1437                 return 1;
1438
1439         return 0;
1440 }
1441
1442 /**
1443  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1444  * @zone: zone to check
1445  * @sc:   scan control of this context
1446  *
1447  * Returns true if the zone does not have enough inactive anon pages,
1448  * meaning some active anon pages need to be deactivated.
1449  */
1450 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1451 {
1452         int low;
1453
1454         if (scanning_global_lru(sc))
1455                 low = inactive_anon_is_low_global(zone);
1456         else
1457                 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1458         return low;
1459 }
1460
1461 static int inactive_file_is_low_global(struct zone *zone)
1462 {
1463         unsigned long active, inactive;
1464
1465         active = zone_page_state(zone, NR_ACTIVE_FILE);
1466         inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1467
1468         return (active > inactive);
1469 }
1470
1471 /**
1472  * inactive_file_is_low - check if file pages need to be deactivated
1473  * @zone: zone to check
1474  * @sc:   scan control of this context
1475  *
1476  * When the system is doing streaming IO, memory pressure here
1477  * ensures that active file pages get deactivated, until more
1478  * than half of the file pages are on the inactive list.
1479  *
1480  * Once we get to that situation, protect the system's working
1481  * set from being evicted by disabling active file page aging.
1482  *
1483  * This uses a different ratio than the anonymous pages, because
1484  * the page cache uses a use-once replacement algorithm.
1485  */
1486 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1487 {
1488         int low;
1489
1490         if (scanning_global_lru(sc))
1491                 low = inactive_file_is_low_global(zone);
1492         else
1493                 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1494         return low;
1495 }
1496
1497 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1498                                 int file)
1499 {
1500         if (file)
1501                 return inactive_file_is_low(zone, sc);
1502         else
1503                 return inactive_anon_is_low(zone, sc);
1504 }
1505
1506 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1507         struct zone *zone, struct scan_control *sc, int priority)
1508 {
1509         int file = is_file_lru(lru);
1510
1511         if (is_active_lru(lru)) {
1512                 if (inactive_list_is_low(zone, sc, file))
1513                     shrink_active_list(nr_to_scan, zone, sc, priority, file);
1514                 return 0;
1515         }
1516
1517         return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1518 }
1519
1520 /*
1521  * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1522  * until we collected @swap_cluster_max pages to scan.
1523  */
1524 static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
1525                                        unsigned long *nr_saved_scan)
1526 {
1527         unsigned long nr;
1528
1529         *nr_saved_scan += nr_to_scan;
1530         nr = *nr_saved_scan;
1531
1532         if (nr >= SWAP_CLUSTER_MAX)
1533                 *nr_saved_scan = 0;
1534         else
1535                 nr = 0;
1536
1537         return nr;
1538 }
1539
1540 /*
1541  * Determine how aggressively the anon and file LRU lists should be
1542  * scanned.  The relative value of each set of LRU lists is determined
1543  * by looking at the fraction of the pages scanned we did rotate back
1544  * onto the active list instead of evict.
1545  *
1546  * nr[0] = anon pages to scan; nr[1] = file pages to scan
1547  */
1548 static void get_scan_count(struct zone *zone, struct scan_control *sc,
1549                                         unsigned long *nr, int priority)
1550 {
1551         unsigned long anon, file, free;
1552         unsigned long anon_prio, file_prio;
1553         unsigned long ap, fp;
1554         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1555         u64 fraction[2], denominator;
1556         enum lru_list l;
1557         int noswap = 0;
1558
1559         /* If we have no swap space, do not bother scanning anon pages. */
1560         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1561                 noswap = 1;
1562                 fraction[0] = 0;
1563                 fraction[1] = 1;
1564                 denominator = 1;
1565                 goto out;
1566         }
1567
1568         anon  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1569                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1570         file  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1571                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1572
1573         if (scanning_global_lru(sc)) {
1574                 free  = zone_page_state(zone, NR_FREE_PAGES);
1575                 /* If we have very few page cache pages,
1576                    force-scan anon pages. */
1577                 if (unlikely(file + free <= high_wmark_pages(zone))) {
1578                         fraction[0] = 1;
1579                         fraction[1] = 0;
1580                         denominator = 1;
1581                         goto out;
1582                 }
1583         }
1584
1585         /*
1586          * OK, so we have swap space and a fair amount of page cache
1587          * pages.  We use the recently rotated / recently scanned
1588          * ratios to determine how valuable each cache is.
1589          *
1590          * Because workloads change over time (and to avoid overflow)
1591          * we keep these statistics as a floating average, which ends
1592          * up weighing recent references more than old ones.
1593          *
1594          * anon in [0], file in [1]
1595          */
1596         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1597                 spin_lock_irq(&zone->lru_lock);
1598                 reclaim_stat->recent_scanned[0] /= 2;
1599                 reclaim_stat->recent_rotated[0] /= 2;
1600                 spin_unlock_irq(&zone->lru_lock);
1601         }
1602
1603         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1604                 spin_lock_irq(&zone->lru_lock);
1605                 reclaim_stat->recent_scanned[1] /= 2;
1606                 reclaim_stat->recent_rotated[1] /= 2;
1607                 spin_unlock_irq(&zone->lru_lock);
1608         }
1609
1610         /*
1611          * With swappiness at 100, anonymous and file have the same priority.
1612          * This scanning priority is essentially the inverse of IO cost.
1613          */
1614         anon_prio = sc->swappiness;
1615         file_prio = 200 - sc->swappiness;
1616
1617         /*
1618          * The amount of pressure on anon vs file pages is inversely
1619          * proportional to the fraction of recently scanned pages on
1620          * each list that were recently referenced and in active use.
1621          */
1622         ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1623         ap /= reclaim_stat->recent_rotated[0] + 1;
1624
1625         fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1626         fp /= reclaim_stat->recent_rotated[1] + 1;
1627
1628         fraction[0] = ap;
1629         fraction[1] = fp;
1630         denominator = ap + fp + 1;
1631 out:
1632         for_each_evictable_lru(l) {
1633                 int file = is_file_lru(l);
1634                 unsigned long scan;
1635
1636                 scan = zone_nr_lru_pages(zone, sc, l);
1637                 if (priority || noswap) {
1638                         scan >>= priority;
1639                         scan = div64_u64(scan * fraction[file], denominator);
1640                 }
1641                 nr[l] = nr_scan_try_batch(scan,
1642                                           &reclaim_stat->nr_saved_scan[l]);
1643         }
1644 }
1645
1646 static void set_lumpy_reclaim_mode(int priority, struct scan_control *sc)
1647 {
1648         /*
1649          * If we need a large contiguous chunk of memory, or have
1650          * trouble getting a small set of contiguous pages, we
1651          * will reclaim both active and inactive pages.
1652          */
1653         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1654                 sc->lumpy_reclaim_mode = 1;
1655         else if (sc->order && priority < DEF_PRIORITY - 2)
1656                 sc->lumpy_reclaim_mode = 1;
1657         else
1658                 sc->lumpy_reclaim_mode = 0;
1659 }
1660
1661 /*
1662  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1663  */
1664 static void shrink_zone(int priority, struct zone *zone,
1665                                 struct scan_control *sc)
1666 {
1667         unsigned long nr[NR_LRU_LISTS];
1668         unsigned long nr_to_scan;
1669         enum lru_list l;
1670         unsigned long nr_reclaimed = sc->nr_reclaimed;
1671         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1672
1673         get_scan_count(zone, sc, nr, priority);
1674
1675         set_lumpy_reclaim_mode(priority, sc);
1676
1677         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1678                                         nr[LRU_INACTIVE_FILE]) {
1679                 for_each_evictable_lru(l) {
1680                         if (nr[l]) {
1681                                 nr_to_scan = min_t(unsigned long,
1682                                                    nr[l], SWAP_CLUSTER_MAX);
1683                                 nr[l] -= nr_to_scan;
1684
1685                                 nr_reclaimed += shrink_list(l, nr_to_scan,
1686                                                             zone, sc, priority);
1687                         }
1688                 }
1689                 /*
1690                  * On large memory systems, scan >> priority can become
1691                  * really large. This is fine for the starting priority;
1692                  * we want to put equal scanning pressure on each zone.
1693                  * However, if the VM has a harder time of freeing pages,
1694                  * with multiple processes reclaiming pages, the total
1695                  * freeing target can get unreasonably large.
1696                  */
1697                 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
1698                         break;
1699         }
1700
1701         sc->nr_reclaimed = nr_reclaimed;
1702
1703         /*
1704          * Even if we did not try to evict anon pages at all, we want to
1705          * rebalance the anon lru active/inactive ratio.
1706          */
1707         if (inactive_anon_is_low(zone, sc) && nr_swap_pages > 0)
1708                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1709
1710         throttle_vm_writeout(sc->gfp_mask);
1711 }
1712
1713 /*
1714  * This is the direct reclaim path, for page-allocating processes.  We only
1715  * try to reclaim pages from zones which will satisfy the caller's allocation
1716  * request.
1717  *
1718  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1719  * Because:
1720  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1721  *    allocation or
1722  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1723  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1724  *    zone defense algorithm.
1725  *
1726  * If a zone is deemed to be full of pinned pages then just give it a light
1727  * scan then give up on it.
1728  */
1729 static bool shrink_zones(int priority, struct zonelist *zonelist,
1730                                         struct scan_control *sc)
1731 {
1732         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1733         struct zoneref *z;
1734         struct zone *zone;
1735         bool all_unreclaimable = true;
1736
1737         for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1738                                         sc->nodemask) {
1739                 if (!populated_zone(zone))
1740                         continue;
1741                 /*
1742                  * Take care memory controller reclaiming has small influence
1743                  * to global LRU.
1744                  */
1745                 if (scanning_global_lru(sc)) {
1746                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1747                                 continue;
1748                         note_zone_scanning_priority(zone, priority);
1749
1750                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1751                                 continue;       /* Let kswapd poll it */
1752                 } else {
1753                         /*
1754                          * Ignore cpuset limitation here. We just want to reduce
1755                          * # of used pages by us regardless of memory shortage.
1756                          */
1757                         mem_cgroup_note_reclaim_priority(sc->mem_cgroup,
1758                                                         priority);
1759                 }
1760
1761                 shrink_zone(priority, zone, sc);
1762                 all_unreclaimable = false;
1763         }
1764         return all_unreclaimable;
1765 }
1766
1767 /*
1768  * This is the main entry point to direct page reclaim.
1769  *
1770  * If a full scan of the inactive list fails to free enough memory then we
1771  * are "out of memory" and something needs to be killed.
1772  *
1773  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1774  * high - the zone may be full of dirty or under-writeback pages, which this
1775  * caller can't do much about.  We kick the writeback threads and take explicit
1776  * naps in the hope that some of these pages can be written.  But if the
1777  * allocating task holds filesystem locks which prevent writeout this might not
1778  * work, and the allocation attempt will fail.
1779  *
1780  * returns:     0, if no pages reclaimed
1781  *              else, the number of pages reclaimed
1782  */
1783 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
1784                                         struct scan_control *sc)
1785 {
1786         int priority;
1787         bool all_unreclaimable;
1788         unsigned long total_scanned = 0;
1789         struct reclaim_state *reclaim_state = current->reclaim_state;
1790         unsigned long lru_pages = 0;
1791         struct zoneref *z;
1792         struct zone *zone;
1793         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1794         unsigned long writeback_threshold;
1795
1796         get_mems_allowed();
1797         delayacct_freepages_start();
1798
1799         if (scanning_global_lru(sc))
1800                 count_vm_event(ALLOCSTALL);
1801         /*
1802          * mem_cgroup will not do shrink_slab.
1803          */
1804         if (scanning_global_lru(sc)) {
1805                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1806
1807                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1808                                 continue;
1809
1810                         lru_pages += zone_reclaimable_pages(zone);
1811                 }
1812         }
1813
1814         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1815                 sc->nr_scanned = 0;
1816                 if (!priority)
1817                         disable_swap_token();
1818                 all_unreclaimable = shrink_zones(priority, zonelist, sc);
1819                 /*
1820                  * Don't shrink slabs when reclaiming memory from
1821                  * over limit cgroups
1822                  */
1823                 if (scanning_global_lru(sc)) {
1824                         shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
1825                         if (reclaim_state) {
1826                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
1827                                 reclaim_state->reclaimed_slab = 0;
1828                         }
1829                 }
1830                 total_scanned += sc->nr_scanned;
1831                 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
1832                         goto out;
1833
1834                 /*
1835                  * Try to write back as many pages as we just scanned.  This
1836                  * tends to cause slow streaming writers to write data to the
1837                  * disk smoothly, at the dirtying rate, which is nice.   But
1838                  * that's undesirable in laptop mode, where we *want* lumpy
1839                  * writeout.  So in laptop mode, write out the whole world.
1840                  */
1841                 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
1842                 if (total_scanned > writeback_threshold) {
1843                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
1844                         sc->may_writepage = 1;
1845                 }
1846
1847                 /* Take a nap, wait for some writeback to complete */
1848                 if (!sc->hibernation_mode && sc->nr_scanned &&
1849                     priority < DEF_PRIORITY - 2)
1850                         congestion_wait(BLK_RW_ASYNC, HZ/10);
1851         }
1852
1853 out:
1854         /*
1855          * Now that we've scanned all the zones at this priority level, note
1856          * that level within the zone so that the next thread which performs
1857          * scanning of this zone will immediately start out at this priority
1858          * level.  This affects only the decision whether or not to bring
1859          * mapped pages onto the inactive list.
1860          */
1861         if (priority < 0)
1862                 priority = 0;
1863
1864         if (scanning_global_lru(sc)) {
1865                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1866
1867                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1868                                 continue;
1869
1870                         zone->prev_priority = priority;
1871                 }
1872         } else
1873                 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority);
1874
1875         delayacct_freepages_end();
1876         put_mems_allowed();
1877
1878         if (sc->nr_reclaimed)
1879                 return sc->nr_reclaimed;
1880
1881         /* top priority shrink_zones still had more to do? don't OOM, then */
1882         if (scanning_global_lru(sc) && !all_unreclaimable)
1883                 return 1;
1884
1885         return 0;
1886 }
1887
1888 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
1889                                 gfp_t gfp_mask, nodemask_t *nodemask)
1890 {
1891         struct scan_control sc = {
1892                 .gfp_mask = gfp_mask,
1893                 .may_writepage = !laptop_mode,
1894                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
1895                 .may_unmap = 1,
1896                 .may_swap = 1,
1897                 .swappiness = vm_swappiness,
1898                 .order = order,
1899                 .mem_cgroup = NULL,
1900                 .nodemask = nodemask,
1901         };
1902
1903         return do_try_to_free_pages(zonelist, &sc);
1904 }
1905
1906 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
1907
1908 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
1909                                                 gfp_t gfp_mask, bool noswap,
1910                                                 unsigned int swappiness,
1911                                                 struct zone *zone, int nid)
1912 {
1913         struct scan_control sc = {
1914                 .may_writepage = !laptop_mode,
1915                 .may_unmap = 1,
1916                 .may_swap = !noswap,
1917                 .swappiness = swappiness,
1918                 .order = 0,
1919                 .mem_cgroup = mem,
1920         };
1921         nodemask_t nm  = nodemask_of_node(nid);
1922
1923         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1924                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1925         sc.nodemask = &nm;
1926         sc.nr_reclaimed = 0;
1927         sc.nr_scanned = 0;
1928         /*
1929          * NOTE: Although we can get the priority field, using it
1930          * here is not a good idea, since it limits the pages we can scan.
1931          * if we don't reclaim here, the shrink_zone from balance_pgdat
1932          * will pick up pages from other mem cgroup's as well. We hack
1933          * the priority and make it zero.
1934          */
1935         shrink_zone(0, zone, &sc);
1936         return sc.nr_reclaimed;
1937 }
1938
1939 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
1940                                            gfp_t gfp_mask,
1941                                            bool noswap,
1942                                            unsigned int swappiness)
1943 {
1944         struct zonelist *zonelist;
1945         struct scan_control sc = {
1946                 .may_writepage = !laptop_mode,
1947                 .may_unmap = 1,
1948                 .may_swap = !noswap,
1949                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
1950                 .swappiness = swappiness,
1951                 .order = 0,
1952                 .mem_cgroup = mem_cont,
1953                 .nodemask = NULL, /* we don't care the placement */
1954         };
1955
1956         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1957                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1958         zonelist = NODE_DATA(numa_node_id())->node_zonelists;
1959         return do_try_to_free_pages(zonelist, &sc);
1960 }
1961 #endif
1962
1963 /* is kswapd sleeping prematurely? */
1964 static int sleeping_prematurely(pg_data_t *pgdat, int order, long remaining)
1965 {
1966         int i;
1967
1968         /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
1969         if (remaining)
1970                 return 1;
1971
1972         /* If after HZ/10, a zone is below the high mark, it's premature */
1973         for (i = 0; i < pgdat->nr_zones; i++) {
1974                 struct zone *zone = pgdat->node_zones + i;
1975
1976                 if (!populated_zone(zone))
1977                         continue;
1978
1979                 if (zone->all_unreclaimable)
1980                         continue;
1981
1982                 if (!zone_watermark_ok(zone, order, high_wmark_pages(zone),
1983                                                                 0, 0))
1984                         return 1;
1985         }
1986
1987         return 0;
1988 }
1989
1990 /*
1991  * For kswapd, balance_pgdat() will work across all this node's zones until
1992  * they are all at high_wmark_pages(zone).
1993  *
1994  * Returns the number of pages which were actually freed.
1995  *
1996  * There is special handling here for zones which are full of pinned pages.
1997  * This can happen if the pages are all mlocked, or if they are all used by
1998  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1999  * What we do is to detect the case where all pages in the zone have been
2000  * scanned twice and there has been zero successful reclaim.  Mark the zone as
2001  * dead and from now on, only perform a short scan.  Basically we're polling
2002  * the zone for when the problem goes away.
2003  *
2004  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
2005  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2006  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2007  * lower zones regardless of the number of free pages in the lower zones. This
2008  * interoperates with the page allocator fallback scheme to ensure that aging
2009  * of pages is balanced across the zones.
2010  */
2011 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
2012 {
2013         int all_zones_ok;
2014         int priority;
2015         int i;
2016         unsigned long total_scanned;
2017         struct reclaim_state *reclaim_state = current->reclaim_state;
2018         struct scan_control sc = {
2019                 .gfp_mask = GFP_KERNEL,
2020                 .may_unmap = 1,
2021                 .may_swap = 1,
2022                 /*
2023                  * kswapd doesn't want to be bailed out while reclaim. because
2024                  * we want to put equal scanning pressure on each zone.
2025                  */
2026                 .nr_to_reclaim = ULONG_MAX,
2027                 .swappiness = vm_swappiness,
2028                 .order = order,
2029                 .mem_cgroup = NULL,
2030         };
2031         /*
2032          * temp_priority is used to remember the scanning priority at which
2033          * this zone was successfully refilled to
2034          * free_pages == high_wmark_pages(zone).
2035          */
2036         int temp_priority[MAX_NR_ZONES];
2037
2038 loop_again:
2039         total_scanned = 0;
2040         sc.nr_reclaimed = 0;
2041         sc.may_writepage = !laptop_mode;
2042         count_vm_event(PAGEOUTRUN);
2043
2044         for (i = 0; i < pgdat->nr_zones; i++)
2045                 temp_priority[i] = DEF_PRIORITY;
2046
2047         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2048                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
2049                 unsigned long lru_pages = 0;
2050                 int has_under_min_watermark_zone = 0;
2051
2052                 /* The swap token gets in the way of swapout... */
2053                 if (!priority)
2054                         disable_swap_token();
2055
2056                 all_zones_ok = 1;
2057
2058                 /*
2059                  * Scan in the highmem->dma direction for the highest
2060                  * zone which needs scanning
2061                  */
2062                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2063                         struct zone *zone = pgdat->node_zones + i;
2064
2065                         if (!populated_zone(zone))
2066                                 continue;
2067
2068                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2069                                 continue;
2070
2071                         /*
2072                          * Do some background aging of the anon list, to give
2073                          * pages a chance to be referenced before reclaiming.
2074                          */
2075                         if (inactive_anon_is_low(zone, &sc))
2076                                 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2077                                                         &sc, priority, 0);
2078
2079                         if (!zone_watermark_ok(zone, order,
2080                                         high_wmark_pages(zone), 0, 0)) {
2081                                 end_zone = i;
2082                                 break;
2083                         }
2084                 }
2085                 if (i < 0)
2086                         goto out;
2087
2088                 for (i = 0; i <= end_zone; i++) {
2089                         struct zone *zone = pgdat->node_zones + i;
2090
2091                         lru_pages += zone_reclaimable_pages(zone);
2092                 }
2093
2094                 /*
2095                  * Now scan the zone in the dma->highmem direction, stopping
2096                  * at the last zone which needs scanning.
2097                  *
2098                  * We do this because the page allocator works in the opposite
2099                  * direction.  This prevents the page allocator from allocating
2100                  * pages behind kswapd's direction of progress, which would
2101                  * cause too much scanning of the lower zones.
2102                  */
2103                 for (i = 0; i <= end_zone; i++) {
2104                         struct zone *zone = pgdat->node_zones + i;
2105                         int nr_slab;
2106                         int nid, zid;
2107
2108                         if (!populated_zone(zone))
2109                                 continue;
2110
2111                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2112                                 continue;
2113
2114                         temp_priority[i] = priority;
2115                         sc.nr_scanned = 0;
2116                         note_zone_scanning_priority(zone, priority);
2117
2118                         nid = pgdat->node_id;
2119                         zid = zone_idx(zone);
2120                         /*
2121                          * Call soft limit reclaim before calling shrink_zone.
2122                          * For now we ignore the return value
2123                          */
2124                         mem_cgroup_soft_limit_reclaim(zone, order, sc.gfp_mask,
2125                                                         nid, zid);
2126                         /*
2127                          * We put equal pressure on every zone, unless one
2128                          * zone has way too many pages free already.
2129                          */
2130                         if (!zone_watermark_ok(zone, order,
2131                                         8*high_wmark_pages(zone), end_zone, 0))
2132                                 shrink_zone(priority, zone, &sc);
2133                         reclaim_state->reclaimed_slab = 0;
2134                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
2135                                                 lru_pages);
2136                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2137                         total_scanned += sc.nr_scanned;
2138                         if (zone->all_unreclaimable)
2139                                 continue;
2140                         if (nr_slab == 0 &&
2141                             zone->pages_scanned >= (zone_reclaimable_pages(zone) * 6))
2142                                 zone->all_unreclaimable = 1;
2143                         /*
2144                          * If we've done a decent amount of scanning and
2145                          * the reclaim ratio is low, start doing writepage
2146                          * even in laptop mode
2147                          */
2148                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2149                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2150                                 sc.may_writepage = 1;
2151
2152                         if (!zone_watermark_ok(zone, order,
2153                                         high_wmark_pages(zone), end_zone, 0)) {
2154                                 all_zones_ok = 0;
2155                                 /*
2156                                  * We are still under min water mark.  This
2157                                  * means that we have a GFP_ATOMIC allocation
2158                                  * failure risk. Hurry up!
2159                                  */
2160                                 if (!zone_watermark_ok(zone, order,
2161                                             min_wmark_pages(zone), end_zone, 0))
2162                                         has_under_min_watermark_zone = 1;
2163                         }
2164
2165                 }
2166                 if (all_zones_ok)
2167                         break;          /* kswapd: all done */
2168                 /*
2169                  * OK, kswapd is getting into trouble.  Take a nap, then take
2170                  * another pass across the zones.
2171                  */
2172                 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2173                         if (has_under_min_watermark_zone)
2174                                 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2175                         else
2176                                 congestion_wait(BLK_RW_ASYNC, HZ/10);
2177                 }
2178
2179                 /*
2180                  * We do this so kswapd doesn't build up large priorities for
2181                  * example when it is freeing in parallel with allocators. It
2182                  * matches the direct reclaim path behaviour in terms of impact
2183                  * on zone->*_priority.
2184                  */
2185                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2186                         break;
2187         }
2188 out:
2189         /*
2190          * Note within each zone the priority level at which this zone was
2191          * brought into a happy state.  So that the next thread which scans this
2192          * zone will start out at that priority level.
2193          */
2194         for (i = 0; i < pgdat->nr_zones; i++) {
2195                 struct zone *zone = pgdat->node_zones + i;
2196
2197                 zone->prev_priority = temp_priority[i];
2198         }
2199         if (!all_zones_ok) {
2200                 cond_resched();
2201
2202                 try_to_freeze();
2203
2204                 /*
2205                  * Fragmentation may mean that the system cannot be
2206                  * rebalanced for high-order allocations in all zones.
2207                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2208                  * it means the zones have been fully scanned and are still
2209                  * not balanced. For high-order allocations, there is
2210                  * little point trying all over again as kswapd may
2211                  * infinite loop.
2212                  *
2213                  * Instead, recheck all watermarks at order-0 as they
2214                  * are the most important. If watermarks are ok, kswapd will go
2215                  * back to sleep. High-order users can still perform direct
2216                  * reclaim if they wish.
2217                  */
2218                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2219                         order = sc.order = 0;
2220
2221                 goto loop_again;
2222         }
2223
2224         return sc.nr_reclaimed;
2225 }
2226
2227 /*
2228  * The background pageout daemon, started as a kernel thread
2229  * from the init process.
2230  *
2231  * This basically trickles out pages so that we have _some_
2232  * free memory available even if there is no other activity
2233  * that frees anything up. This is needed for things like routing
2234  * etc, where we otherwise might have all activity going on in
2235  * asynchronous contexts that cannot page things out.
2236  *
2237  * If there are applications that are active memory-allocators
2238  * (most normal use), this basically shouldn't matter.
2239  */
2240 static int kswapd(void *p)
2241 {
2242         unsigned long order;
2243         pg_data_t *pgdat = (pg_data_t*)p;
2244         struct task_struct *tsk = current;
2245         DEFINE_WAIT(wait);
2246         struct reclaim_state reclaim_state = {
2247                 .reclaimed_slab = 0,
2248         };
2249         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2250
2251         lockdep_set_current_reclaim_state(GFP_KERNEL);
2252
2253         if (!cpumask_empty(cpumask))
2254                 set_cpus_allowed_ptr(tsk, cpumask);
2255         current->reclaim_state = &reclaim_state;
2256
2257         /*
2258          * Tell the memory management that we're a "memory allocator",
2259          * and that if we need more memory we should get access to it
2260          * regardless (see "__alloc_pages()"). "kswapd" should
2261          * never get caught in the normal page freeing logic.
2262          *
2263          * (Kswapd normally doesn't need memory anyway, but sometimes
2264          * you need a small amount of memory in order to be able to
2265          * page out something else, and this flag essentially protects
2266          * us from recursively trying to free more memory as we're
2267          * trying to free the first piece of memory in the first place).
2268          */
2269         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2270         set_freezable();
2271
2272         order = 0;
2273         for ( ; ; ) {
2274                 unsigned long new_order;
2275                 int ret;
2276
2277                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2278                 new_order = pgdat->kswapd_max_order;
2279                 pgdat->kswapd_max_order = 0;
2280                 if (order < new_order) {
2281                         /*
2282                          * Don't sleep if someone wants a larger 'order'
2283                          * allocation
2284                          */
2285                         order = new_order;
2286                 } else {
2287                         if (!freezing(current) && !kthread_should_stop()) {
2288                                 long remaining = 0;
2289
2290                                 /* Try to sleep for a short interval */
2291                                 if (!sleeping_prematurely(pgdat, order, remaining)) {
2292                                         remaining = schedule_timeout(HZ/10);
2293                                         finish_wait(&pgdat->kswapd_wait, &wait);
2294                                         prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2295                                 }
2296
2297                                 /*
2298                                  * After a short sleep, check if it was a
2299                                  * premature sleep. If not, then go fully
2300                                  * to sleep until explicitly woken up
2301                                  */
2302                                 if (!sleeping_prematurely(pgdat, order, remaining))
2303                                         schedule();
2304                                 else {
2305                                         if (remaining)
2306                                                 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2307                                         else
2308                                                 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2309                                 }
2310                         }
2311
2312                         order = pgdat->kswapd_max_order;
2313                 }
2314                 finish_wait(&pgdat->kswapd_wait, &wait);
2315
2316                 ret = try_to_freeze();
2317                 if (kthread_should_stop())
2318                         break;
2319
2320                 /*
2321                  * We can speed up thawing tasks if we don't call balance_pgdat
2322                  * after returning from the refrigerator
2323                  */
2324                 if (!ret)
2325                         balance_pgdat(pgdat, order);
2326         }
2327         return 0;
2328 }
2329
2330 /*
2331  * A zone is low on free memory, so wake its kswapd task to service it.
2332  */
2333 void wakeup_kswapd(struct zone *zone, int order)
2334 {
2335         pg_data_t *pgdat;
2336
2337         if (!populated_zone(zone))
2338                 return;
2339
2340         pgdat = zone->zone_pgdat;
2341         if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
2342                 return;
2343         if (pgdat->kswapd_max_order < order)
2344                 pgdat->kswapd_max_order = order;
2345         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2346                 return;
2347         if (!waitqueue_active(&pgdat->kswapd_wait))
2348                 return;
2349         wake_up_interruptible(&pgdat->kswapd_wait);
2350 }
2351
2352 /*
2353  * The reclaimable count would be mostly accurate.
2354  * The less reclaimable pages may be
2355  * - mlocked pages, which will be moved to unevictable list when encountered
2356  * - mapped pages, which may require several travels to be reclaimed
2357  * - dirty pages, which is not "instantly" reclaimable
2358  */
2359 unsigned long global_reclaimable_pages(void)
2360 {
2361         int nr;
2362
2363         nr = global_page_state(NR_ACTIVE_FILE) +
2364              global_page_state(NR_INACTIVE_FILE);
2365
2366         if (nr_swap_pages > 0)
2367                 nr += global_page_state(NR_ACTIVE_ANON) +
2368                       global_page_state(NR_INACTIVE_ANON);
2369
2370         return nr;
2371 }
2372
2373 unsigned long zone_reclaimable_pages(struct zone *zone)
2374 {
2375         int nr;
2376
2377         nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2378              zone_page_state(zone, NR_INACTIVE_FILE);
2379
2380         if (nr_swap_pages > 0)
2381                 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2382                       zone_page_state(zone, NR_INACTIVE_ANON);
2383
2384         return nr;
2385 }
2386
2387 #ifdef CONFIG_HIBERNATION
2388 /*
2389  * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2390  * freed pages.
2391  *
2392  * Rather than trying to age LRUs the aim is to preserve the overall
2393  * LRU order by reclaiming preferentially
2394  * inactive > active > active referenced > active mapped
2395  */
2396 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
2397 {
2398         struct reclaim_state reclaim_state;
2399         struct scan_control sc = {
2400                 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2401                 .may_swap = 1,
2402                 .may_unmap = 1,
2403                 .may_writepage = 1,
2404                 .nr_to_reclaim = nr_to_reclaim,
2405                 .hibernation_mode = 1,
2406                 .swappiness = vm_swappiness,
2407                 .order = 0,
2408         };
2409         struct zonelist * zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2410         struct task_struct *p = current;
2411         unsigned long nr_reclaimed;
2412
2413         p->flags |= PF_MEMALLOC;
2414         lockdep_set_current_reclaim_state(sc.gfp_mask);
2415         reclaim_state.reclaimed_slab = 0;
2416         p->reclaim_state = &reclaim_state;
2417
2418         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2419
2420         p->reclaim_state = NULL;
2421         lockdep_clear_current_reclaim_state();
2422         p->flags &= ~PF_MEMALLOC;
2423
2424         return nr_reclaimed;
2425 }
2426 #endif /* CONFIG_HIBERNATION */
2427
2428 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2429    not required for correctness.  So if the last cpu in a node goes
2430    away, we get changed to run anywhere: as the first one comes back,
2431    restore their cpu bindings. */
2432 static int __devinit cpu_callback(struct notifier_block *nfb,
2433                                   unsigned long action, void *hcpu)
2434 {
2435         int nid;
2436
2437         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2438                 for_each_node_state(nid, N_HIGH_MEMORY) {
2439                         pg_data_t *pgdat = NODE_DATA(nid);
2440                         const struct cpumask *mask;
2441
2442                         mask = cpumask_of_node(pgdat->node_id);
2443
2444                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2445                                 /* One of our CPUs online: restore mask */
2446                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2447                 }
2448         }
2449         return NOTIFY_OK;
2450 }
2451
2452 /*
2453  * This kswapd start function will be called by init and node-hot-add.
2454  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2455  */
2456 int kswapd_run(int nid)
2457 {
2458         pg_data_t *pgdat = NODE_DATA(nid);
2459         int ret = 0;
2460
2461         if (pgdat->kswapd)
2462                 return 0;
2463
2464         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2465         if (IS_ERR(pgdat->kswapd)) {
2466                 /* failure at boot is fatal */
2467                 BUG_ON(system_state == SYSTEM_BOOTING);
2468                 printk("Failed to start kswapd on node %d\n",nid);
2469                 ret = -1;
2470         }
2471         return ret;
2472 }
2473
2474 /*
2475  * Called by memory hotplug when all memory in a node is offlined.
2476  */
2477 void kswapd_stop(int nid)
2478 {
2479         struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
2480
2481         if (kswapd)
2482                 kthread_stop(kswapd);
2483 }
2484
2485 static int __init kswapd_init(void)
2486 {
2487         int nid;
2488
2489         swap_setup();
2490         for_each_node_state(nid, N_HIGH_MEMORY)
2491                 kswapd_run(nid);
2492         hotcpu_notifier(cpu_callback, 0);
2493         return 0;
2494 }
2495
2496 module_init(kswapd_init)
2497
2498 #ifdef CONFIG_NUMA
2499 /*
2500  * Zone reclaim mode
2501  *
2502  * If non-zero call zone_reclaim when the number of free pages falls below
2503  * the watermarks.
2504  */
2505 int zone_reclaim_mode __read_mostly;
2506
2507 #define RECLAIM_OFF 0
2508 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
2509 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
2510 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
2511
2512 /*
2513  * Priority for ZONE_RECLAIM. This determines the fraction of pages
2514  * of a node considered for each zone_reclaim. 4 scans 1/16th of
2515  * a zone.
2516  */
2517 #define ZONE_RECLAIM_PRIORITY 4
2518
2519 /*
2520  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2521  * occur.
2522  */
2523 int sysctl_min_unmapped_ratio = 1;
2524
2525 /*
2526  * If the number of slab pages in a zone grows beyond this percentage then
2527  * slab reclaim needs to occur.
2528  */
2529 int sysctl_min_slab_ratio = 5;
2530
2531 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
2532 {
2533         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
2534         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
2535                 zone_page_state(zone, NR_ACTIVE_FILE);
2536
2537         /*
2538          * It's possible for there to be more file mapped pages than
2539          * accounted for by the pages on the file LRU lists because
2540          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2541          */
2542         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
2543 }
2544
2545 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
2546 static long zone_pagecache_reclaimable(struct zone *zone)
2547 {
2548         long nr_pagecache_reclaimable;
2549         long delta = 0;
2550
2551         /*
2552          * If RECLAIM_SWAP is set, then all file pages are considered
2553          * potentially reclaimable. Otherwise, we have to worry about
2554          * pages like swapcache and zone_unmapped_file_pages() provides
2555          * a better estimate
2556          */
2557         if (zone_reclaim_mode & RECLAIM_SWAP)
2558                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
2559         else
2560                 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
2561
2562         /* If we can't clean pages, remove dirty pages from consideration */
2563         if (!(zone_reclaim_mode & RECLAIM_WRITE))
2564                 delta += zone_page_state(zone, NR_FILE_DIRTY);
2565
2566         /* Watch for any possible underflows due to delta */
2567         if (unlikely(delta > nr_pagecache_reclaimable))
2568                 delta = nr_pagecache_reclaimable;
2569
2570         return nr_pagecache_reclaimable - delta;
2571 }
2572
2573 /*
2574  * Try to free up some pages from this zone through reclaim.
2575  */
2576 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2577 {
2578         /* Minimum pages needed in order to stay on node */
2579         const unsigned long nr_pages = 1 << order;
2580         struct task_struct *p = current;
2581         struct reclaim_state reclaim_state;
2582         int priority;
2583         struct scan_control sc = {
2584                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2585                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2586                 .may_swap = 1,
2587                 .nr_to_reclaim = max_t(unsigned long, nr_pages,
2588                                        SWAP_CLUSTER_MAX),
2589                 .gfp_mask = gfp_mask,
2590                 .swappiness = vm_swappiness,
2591                 .order = order,
2592         };
2593         unsigned long slab_reclaimable;
2594
2595         cond_resched();
2596         /*
2597          * We need to be able to allocate from the reserves for RECLAIM_SWAP
2598          * and we also need to be able to write out pages for RECLAIM_WRITE
2599          * and RECLAIM_SWAP.
2600          */
2601         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2602         lockdep_set_current_reclaim_state(gfp_mask);
2603         reclaim_state.reclaimed_slab = 0;
2604         p->reclaim_state = &reclaim_state;
2605
2606         if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
2607                 /*
2608                  * Free memory by calling shrink zone with increasing
2609                  * priorities until we have enough memory freed.
2610                  */
2611                 priority = ZONE_RECLAIM_PRIORITY;
2612                 do {
2613                         note_zone_scanning_priority(zone, priority);
2614                         shrink_zone(priority, zone, &sc);
2615                         priority--;
2616                 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
2617         }
2618
2619         slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2620         if (slab_reclaimable > zone->min_slab_pages) {
2621                 /*
2622                  * shrink_slab() does not currently allow us to determine how
2623                  * many pages were freed in this zone. So we take the current
2624                  * number of slab pages and shake the slab until it is reduced
2625                  * by the same nr_pages that we used for reclaiming unmapped
2626                  * pages.
2627                  *
2628                  * Note that shrink_slab will free memory on all zones and may
2629                  * take a long time.
2630                  */
2631                 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
2632                         zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
2633                                 slab_reclaimable - nr_pages)
2634                         ;
2635
2636                 /*
2637                  * Update nr_reclaimed by the number of slab pages we
2638                  * reclaimed from this zone.
2639                  */
2640                 sc.nr_reclaimed += slab_reclaimable -
2641                         zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2642         }
2643
2644         p->reclaim_state = NULL;
2645         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
2646         lockdep_clear_current_reclaim_state();
2647         return sc.nr_reclaimed >= nr_pages;
2648 }
2649
2650 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2651 {
2652         int node_id;
2653         int ret;
2654
2655         /*
2656          * Zone reclaim reclaims unmapped file backed pages and
2657          * slab pages if we are over the defined limits.
2658          *
2659          * A small portion of unmapped file backed pages is needed for
2660          * file I/O otherwise pages read by file I/O will be immediately
2661          * thrown out if the zone is overallocated. So we do not reclaim
2662          * if less than a specified percentage of the zone is used by
2663          * unmapped file backed pages.
2664          */
2665         if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
2666             zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
2667                 return ZONE_RECLAIM_FULL;
2668
2669         if (zone->all_unreclaimable)
2670                 return ZONE_RECLAIM_FULL;
2671
2672         /*
2673          * Do not scan if the allocation should not be delayed.
2674          */
2675         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
2676                 return ZONE_RECLAIM_NOSCAN;
2677
2678         /*
2679          * Only run zone reclaim on the local zone or on zones that do not
2680          * have associated processors. This will favor the local processor
2681          * over remote processors and spread off node memory allocations
2682          * as wide as possible.
2683          */
2684         node_id = zone_to_nid(zone);
2685         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
2686                 return ZONE_RECLAIM_NOSCAN;
2687
2688         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2689                 return ZONE_RECLAIM_NOSCAN;
2690
2691         ret = __zone_reclaim(zone, gfp_mask, order);
2692         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2693
2694         if (!ret)
2695                 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
2696
2697         return ret;
2698 }
2699 #endif
2700
2701 /*
2702  * page_evictable - test whether a page is evictable
2703  * @page: the page to test
2704  * @vma: the VMA in which the page is or will be mapped, may be NULL
2705  *
2706  * Test whether page is evictable--i.e., should be placed on active/inactive
2707  * lists vs unevictable list.  The vma argument is !NULL when called from the
2708  * fault path to determine how to instantate a new page.
2709  *
2710  * Reasons page might not be evictable:
2711  * (1) page's mapping marked unevictable
2712  * (2) page is part of an mlocked VMA
2713  *
2714  */
2715 int page_evictable(struct page *page, struct vm_area_struct *vma)
2716 {
2717
2718         if (mapping_unevictable(page_mapping(page)))
2719                 return 0;
2720
2721         if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2722                 return 0;
2723
2724         return 1;
2725 }
2726
2727 /**
2728  * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2729  * @page: page to check evictability and move to appropriate lru list
2730  * @zone: zone page is in
2731  *
2732  * Checks a page for evictability and moves the page to the appropriate
2733  * zone lru list.
2734  *
2735  * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2736  * have PageUnevictable set.
2737  */
2738 static void check_move_unevictable_page(struct page *page, struct zone *zone)
2739 {
2740         VM_BUG_ON(PageActive(page));
2741
2742 retry:
2743         ClearPageUnevictable(page);
2744         if (page_evictable(page, NULL)) {
2745                 enum lru_list l = page_lru_base_type(page);
2746
2747                 __dec_zone_state(zone, NR_UNEVICTABLE);
2748                 list_move(&page->lru, &zone->lru[l].list);
2749                 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
2750                 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
2751                 __count_vm_event(UNEVICTABLE_PGRESCUED);
2752         } else {
2753                 /*
2754                  * rotate unevictable list
2755                  */
2756                 SetPageUnevictable(page);
2757                 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
2758                 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
2759                 if (page_evictable(page, NULL))
2760                         goto retry;
2761         }
2762 }
2763
2764 /**
2765  * scan_mapping_unevictable_pages - scan an address space for evictable pages
2766  * @mapping: struct address_space to scan for evictable pages
2767  *
2768  * Scan all pages in mapping.  Check unevictable pages for
2769  * evictability and move them to the appropriate zone lru list.
2770  */
2771 void scan_mapping_unevictable_pages(struct address_space *mapping)
2772 {
2773         pgoff_t next = 0;
2774         pgoff_t end   = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
2775                          PAGE_CACHE_SHIFT;
2776         struct zone *zone;
2777         struct pagevec pvec;
2778
2779         if (mapping->nrpages == 0)
2780                 return;
2781
2782         pagevec_init(&pvec, 0);
2783         while (next < end &&
2784                 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
2785                 int i;
2786                 int pg_scanned = 0;
2787
2788                 zone = NULL;
2789
2790                 for (i = 0; i < pagevec_count(&pvec); i++) {
2791                         struct page *page = pvec.pages[i];
2792                         pgoff_t page_index = page->index;
2793                         struct zone *pagezone = page_zone(page);
2794
2795                         pg_scanned++;
2796                         if (page_index > next)
2797                                 next = page_index;
2798                         next++;
2799
2800                         if (pagezone != zone) {
2801                                 if (zone)
2802                                         spin_unlock_irq(&zone->lru_lock);
2803                                 zone = pagezone;
2804                                 spin_lock_irq(&zone->lru_lock);
2805                         }
2806
2807                         if (PageLRU(page) && PageUnevictable(page))
2808                                 check_move_unevictable_page(page, zone);
2809                 }
2810                 if (zone)
2811                         spin_unlock_irq(&zone->lru_lock);
2812                 pagevec_release(&pvec);
2813
2814                 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
2815         }
2816
2817 }
2818
2819 /**
2820  * scan_zone_unevictable_pages - check unevictable list for evictable pages
2821  * @zone - zone of which to scan the unevictable list
2822  *
2823  * Scan @zone's unevictable LRU lists to check for pages that have become
2824  * evictable.  Move those that have to @zone's inactive list where they
2825  * become candidates for reclaim, unless shrink_inactive_zone() decides
2826  * to reactivate them.  Pages that are still unevictable are rotated
2827  * back onto @zone's unevictable list.
2828  */
2829 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
2830 static void scan_zone_unevictable_pages(struct zone *zone)
2831 {
2832         struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
2833         unsigned long scan;
2834         unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
2835
2836         while (nr_to_scan > 0) {
2837                 unsigned long batch_size = min(nr_to_scan,
2838                                                 SCAN_UNEVICTABLE_BATCH_SIZE);
2839
2840                 spin_lock_irq(&zone->lru_lock);
2841                 for (scan = 0;  scan < batch_size; scan++) {
2842                         struct page *page = lru_to_page(l_unevictable);
2843
2844                         if (!trylock_page(page))
2845                                 continue;
2846
2847                         prefetchw_prev_lru_page(page, l_unevictable, flags);
2848
2849                         if (likely(PageLRU(page) && PageUnevictable(page)))
2850                                 check_move_unevictable_page(page, zone);
2851
2852                         unlock_page(page);
2853                 }
2854                 spin_unlock_irq(&zone->lru_lock);
2855
2856                 nr_to_scan -= batch_size;
2857         }
2858 }
2859
2860
2861 /**
2862  * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
2863  *
2864  * A really big hammer:  scan all zones' unevictable LRU lists to check for
2865  * pages that have become evictable.  Move those back to the zones'
2866  * inactive list where they become candidates for reclaim.
2867  * This occurs when, e.g., we have unswappable pages on the unevictable lists,
2868  * and we add swap to the system.  As such, it runs in the context of a task
2869  * that has possibly/probably made some previously unevictable pages
2870  * evictable.
2871  */
2872 static void scan_all_zones_unevictable_pages(void)
2873 {
2874         struct zone *zone;
2875
2876         for_each_zone(zone) {
2877                 scan_zone_unevictable_pages(zone);
2878         }
2879 }
2880
2881 /*
2882  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
2883  * all nodes' unevictable lists for evictable pages
2884  */
2885 unsigned long scan_unevictable_pages;
2886
2887 int scan_unevictable_handler(struct ctl_table *table, int write,
2888                            void __user *buffer,
2889                            size_t *length, loff_t *ppos)
2890 {
2891         proc_doulongvec_minmax(table, write, buffer, length, ppos);
2892
2893         if (write && *(unsigned long *)table->data)
2894                 scan_all_zones_unevictable_pages();
2895
2896         scan_unevictable_pages = 0;
2897         return 0;
2898 }
2899
2900 /*
2901  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
2902  * a specified node's per zone unevictable lists for evictable pages.
2903  */
2904
2905 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
2906                                           struct sysdev_attribute *attr,
2907                                           char *buf)
2908 {
2909         return sprintf(buf, "0\n");     /* always zero; should fit... */
2910 }
2911
2912 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
2913                                            struct sysdev_attribute *attr,
2914                                         const char *buf, size_t count)
2915 {
2916         struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
2917         struct zone *zone;
2918         unsigned long res;
2919         unsigned long req = strict_strtoul(buf, 10, &res);
2920
2921         if (!req)
2922                 return 1;       /* zero is no-op */
2923
2924         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2925                 if (!populated_zone(zone))
2926                         continue;
2927                 scan_zone_unevictable_pages(zone);
2928         }
2929         return 1;
2930 }
2931
2932
2933 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
2934                         read_scan_unevictable_node,
2935                         write_scan_unevictable_node);
2936
2937 int scan_unevictable_register_node(struct node *node)
2938 {
2939         return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
2940 }
2941
2942 void scan_unevictable_unregister_node(struct node *node)
2943 {
2944         sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
2945 }
2946