memcg, vmscan: add memcg reclaim tracepoint
[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 #define CREATE_TRACE_POINTS
52 #include <trace/events/vmscan.h>
53
54 struct scan_control {
55         /* Incremented by the number of inactive pages that were scanned */
56         unsigned long nr_scanned;
57
58         /* Number of pages freed so far during a call to shrink_zones() */
59         unsigned long nr_reclaimed;
60
61         /* How many pages shrink_list() should reclaim */
62         unsigned long nr_to_reclaim;
63
64         unsigned long hibernation_mode;
65
66         /* This context's GFP mask */
67         gfp_t gfp_mask;
68
69         int may_writepage;
70
71         /* Can mapped pages be reclaimed? */
72         int may_unmap;
73
74         /* Can pages be swapped as part of reclaim? */
75         int may_swap;
76
77         int swappiness;
78
79         int order;
80
81         /*
82          * Intend to reclaim enough contenious memory rather than to reclaim
83          * enough amount memory. I.e, it's the mode for high order allocation.
84          */
85         bool lumpy_reclaim_mode;
86
87         /* Which cgroup do we reclaim from */
88         struct mem_cgroup *mem_cgroup;
89
90         /*
91          * Nodemask of nodes allowed by the caller. If NULL, all nodes
92          * are scanned.
93          */
94         nodemask_t      *nodemask;
95 };
96
97 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
98
99 #ifdef ARCH_HAS_PREFETCH
100 #define prefetch_prev_lru_page(_page, _base, _field)                    \
101         do {                                                            \
102                 if ((_page)->lru.prev != _base) {                       \
103                         struct page *prev;                              \
104                                                                         \
105                         prev = lru_to_page(&(_page->lru));              \
106                         prefetch(&prev->_field);                        \
107                 }                                                       \
108         } while (0)
109 #else
110 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
111 #endif
112
113 #ifdef ARCH_HAS_PREFETCHW
114 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
115         do {                                                            \
116                 if ((_page)->lru.prev != _base) {                       \
117                         struct page *prev;                              \
118                                                                         \
119                         prev = lru_to_page(&(_page->lru));              \
120                         prefetchw(&prev->_field);                       \
121                 }                                                       \
122         } while (0)
123 #else
124 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
125 #endif
126
127 /*
128  * From 0 .. 100.  Higher means more swappy.
129  */
130 int vm_swappiness = 60;
131 long vm_total_pages;    /* The total number of pages which the VM controls */
132
133 static LIST_HEAD(shrinker_list);
134 static DECLARE_RWSEM(shrinker_rwsem);
135
136 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
137 #define scanning_global_lru(sc) (!(sc)->mem_cgroup)
138 #else
139 #define scanning_global_lru(sc) (1)
140 #endif
141
142 static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
143                                                   struct scan_control *sc)
144 {
145         if (!scanning_global_lru(sc))
146                 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
147
148         return &zone->reclaim_stat;
149 }
150
151 static unsigned long zone_nr_lru_pages(struct zone *zone,
152                                 struct scan_control *sc, enum lru_list lru)
153 {
154         if (!scanning_global_lru(sc))
155                 return mem_cgroup_zone_nr_pages(sc->mem_cgroup, zone, lru);
156
157         return zone_page_state(zone, NR_LRU_BASE + lru);
158 }
159
160
161 /*
162  * Add a shrinker callback to be called from the vm
163  */
164 void register_shrinker(struct shrinker *shrinker)
165 {
166         shrinker->nr = 0;
167         down_write(&shrinker_rwsem);
168         list_add_tail(&shrinker->list, &shrinker_list);
169         up_write(&shrinker_rwsem);
170 }
171 EXPORT_SYMBOL(register_shrinker);
172
173 /*
174  * Remove one
175  */
176 void unregister_shrinker(struct shrinker *shrinker)
177 {
178         down_write(&shrinker_rwsem);
179         list_del(&shrinker->list);
180         up_write(&shrinker_rwsem);
181 }
182 EXPORT_SYMBOL(unregister_shrinker);
183
184 #define SHRINK_BATCH 128
185 /*
186  * Call the shrink functions to age shrinkable caches
187  *
188  * Here we assume it costs one seek to replace a lru page and that it also
189  * takes a seek to recreate a cache object.  With this in mind we age equal
190  * percentages of the lru and ageable caches.  This should balance the seeks
191  * generated by these structures.
192  *
193  * If the vm encountered mapped pages on the LRU it increase the pressure on
194  * slab to avoid swapping.
195  *
196  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
197  *
198  * `lru_pages' represents the number of on-LRU pages in all the zones which
199  * are eligible for the caller's allocation attempt.  It is used for balancing
200  * slab reclaim versus page reclaim.
201  *
202  * Returns the number of slab objects which we shrunk.
203  */
204 unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
205                         unsigned long lru_pages)
206 {
207         struct shrinker *shrinker;
208         unsigned long ret = 0;
209
210         if (scanned == 0)
211                 scanned = SWAP_CLUSTER_MAX;
212
213         if (!down_read_trylock(&shrinker_rwsem))
214                 return 1;       /* Assume we'll be able to shrink next time */
215
216         list_for_each_entry(shrinker, &shrinker_list, list) {
217                 unsigned long long delta;
218                 unsigned long total_scan;
219                 unsigned long max_pass;
220
221                 max_pass = (*shrinker->shrink)(shrinker, 0, gfp_mask);
222                 delta = (4 * scanned) / shrinker->seeks;
223                 delta *= max_pass;
224                 do_div(delta, lru_pages + 1);
225                 shrinker->nr += delta;
226                 if (shrinker->nr < 0) {
227                         printk(KERN_ERR "shrink_slab: %pF negative objects to "
228                                "delete nr=%ld\n",
229                                shrinker->shrink, shrinker->nr);
230                         shrinker->nr = max_pass;
231                 }
232
233                 /*
234                  * Avoid risking looping forever due to too large nr value:
235                  * never try to free more than twice the estimate number of
236                  * freeable entries.
237                  */
238                 if (shrinker->nr > max_pass * 2)
239                         shrinker->nr = max_pass * 2;
240
241                 total_scan = shrinker->nr;
242                 shrinker->nr = 0;
243
244                 while (total_scan >= SHRINK_BATCH) {
245                         long this_scan = SHRINK_BATCH;
246                         int shrink_ret;
247                         int nr_before;
248
249                         nr_before = (*shrinker->shrink)(shrinker, 0, gfp_mask);
250                         shrink_ret = (*shrinker->shrink)(shrinker, this_scan,
251                                                                 gfp_mask);
252                         if (shrink_ret == -1)
253                                 break;
254                         if (shrink_ret < nr_before)
255                                 ret += nr_before - shrink_ret;
256                         count_vm_events(SLABS_SCANNED, this_scan);
257                         total_scan -= this_scan;
258
259                         cond_resched();
260                 }
261
262                 shrinker->nr += total_scan;
263         }
264         up_read(&shrinker_rwsem);
265         return ret;
266 }
267
268 static inline int is_page_cache_freeable(struct page *page)
269 {
270         /*
271          * A freeable page cache page is referenced only by the caller
272          * that isolated the page, the page cache radix tree and
273          * optional buffer heads at page->private.
274          */
275         return page_count(page) - page_has_private(page) == 2;
276 }
277
278 static int may_write_to_queue(struct backing_dev_info *bdi)
279 {
280         if (current->flags & PF_SWAPWRITE)
281                 return 1;
282         if (!bdi_write_congested(bdi))
283                 return 1;
284         if (bdi == current->backing_dev_info)
285                 return 1;
286         return 0;
287 }
288
289 /*
290  * We detected a synchronous write error writing a page out.  Probably
291  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
292  * fsync(), msync() or close().
293  *
294  * The tricky part is that after writepage we cannot touch the mapping: nothing
295  * prevents it from being freed up.  But we have a ref on the page and once
296  * that page is locked, the mapping is pinned.
297  *
298  * We're allowed to run sleeping lock_page() here because we know the caller has
299  * __GFP_FS.
300  */
301 static void handle_write_error(struct address_space *mapping,
302                                 struct page *page, int error)
303 {
304         lock_page_nosync(page);
305         if (page_mapping(page) == mapping)
306                 mapping_set_error(mapping, error);
307         unlock_page(page);
308 }
309
310 /* Request for sync pageout. */
311 enum pageout_io {
312         PAGEOUT_IO_ASYNC,
313         PAGEOUT_IO_SYNC,
314 };
315
316 /* possible outcome of pageout() */
317 typedef enum {
318         /* failed to write page out, page is locked */
319         PAGE_KEEP,
320         /* move page to the active list, page is locked */
321         PAGE_ACTIVATE,
322         /* page has been sent to the disk successfully, page is unlocked */
323         PAGE_SUCCESS,
324         /* page is clean and locked */
325         PAGE_CLEAN,
326 } pageout_t;
327
328 /*
329  * pageout is called by shrink_page_list() for each dirty page.
330  * Calls ->writepage().
331  */
332 static pageout_t pageout(struct page *page, struct address_space *mapping,
333                                                 enum pageout_io sync_writeback)
334 {
335         /*
336          * If the page is dirty, only perform writeback if that write
337          * will be non-blocking.  To prevent this allocation from being
338          * stalled by pagecache activity.  But note that there may be
339          * stalls if we need to run get_block().  We could test
340          * PagePrivate for that.
341          *
342          * If this process is currently in __generic_file_aio_write() against
343          * this page's queue, we can perform writeback even if that
344          * will block.
345          *
346          * If the page is swapcache, write it back even if that would
347          * block, for some throttling. This happens by accident, because
348          * swap_backing_dev_info is bust: it doesn't reflect the
349          * congestion state of the swapdevs.  Easy to fix, if needed.
350          */
351         if (!is_page_cache_freeable(page))
352                 return PAGE_KEEP;
353         if (!mapping) {
354                 /*
355                  * Some data journaling orphaned pages can have
356                  * page->mapping == NULL while being dirty with clean buffers.
357                  */
358                 if (page_has_private(page)) {
359                         if (try_to_free_buffers(page)) {
360                                 ClearPageDirty(page);
361                                 printk("%s: orphaned page\n", __func__);
362                                 return PAGE_CLEAN;
363                         }
364                 }
365                 return PAGE_KEEP;
366         }
367         if (mapping->a_ops->writepage == NULL)
368                 return PAGE_ACTIVATE;
369         if (!may_write_to_queue(mapping->backing_dev_info))
370                 return PAGE_KEEP;
371
372         if (clear_page_dirty_for_io(page)) {
373                 int res;
374                 struct writeback_control wbc = {
375                         .sync_mode = WB_SYNC_NONE,
376                         .nr_to_write = SWAP_CLUSTER_MAX,
377                         .range_start = 0,
378                         .range_end = LLONG_MAX,
379                         .nonblocking = 1,
380                         .for_reclaim = 1,
381                 };
382
383                 SetPageReclaim(page);
384                 res = mapping->a_ops->writepage(page, &wbc);
385                 if (res < 0)
386                         handle_write_error(mapping, page, res);
387                 if (res == AOP_WRITEPAGE_ACTIVATE) {
388                         ClearPageReclaim(page);
389                         return PAGE_ACTIVATE;
390                 }
391
392                 /*
393                  * Wait on writeback if requested to. This happens when
394                  * direct reclaiming a large contiguous area and the
395                  * first attempt to free a range of pages fails.
396                  */
397                 if (PageWriteback(page) && sync_writeback == PAGEOUT_IO_SYNC)
398                         wait_on_page_writeback(page);
399
400                 if (!PageWriteback(page)) {
401                         /* synchronous write or broken a_ops? */
402                         ClearPageReclaim(page);
403                 }
404                 trace_mm_vmscan_writepage(page,
405                         trace_reclaim_flags(page, sync_writeback));
406                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
407                 return PAGE_SUCCESS;
408         }
409
410         return PAGE_CLEAN;
411 }
412
413 /*
414  * Same as remove_mapping, but if the page is removed from the mapping, it
415  * gets returned with a refcount of 0.
416  */
417 static int __remove_mapping(struct address_space *mapping, struct page *page)
418 {
419         BUG_ON(!PageLocked(page));
420         BUG_ON(mapping != page_mapping(page));
421
422         spin_lock_irq(&mapping->tree_lock);
423         /*
424          * The non racy check for a busy page.
425          *
426          * Must be careful with the order of the tests. When someone has
427          * a ref to the page, it may be possible that they dirty it then
428          * drop the reference. So if PageDirty is tested before page_count
429          * here, then the following race may occur:
430          *
431          * get_user_pages(&page);
432          * [user mapping goes away]
433          * write_to(page);
434          *                              !PageDirty(page)    [good]
435          * SetPageDirty(page);
436          * put_page(page);
437          *                              !page_count(page)   [good, discard it]
438          *
439          * [oops, our write_to data is lost]
440          *
441          * Reversing the order of the tests ensures such a situation cannot
442          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
443          * load is not satisfied before that of page->_count.
444          *
445          * Note that if SetPageDirty is always performed via set_page_dirty,
446          * and thus under tree_lock, then this ordering is not required.
447          */
448         if (!page_freeze_refs(page, 2))
449                 goto cannot_free;
450         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
451         if (unlikely(PageDirty(page))) {
452                 page_unfreeze_refs(page, 2);
453                 goto cannot_free;
454         }
455
456         if (PageSwapCache(page)) {
457                 swp_entry_t swap = { .val = page_private(page) };
458                 __delete_from_swap_cache(page);
459                 spin_unlock_irq(&mapping->tree_lock);
460                 swapcache_free(swap, page);
461         } else {
462                 __remove_from_page_cache(page);
463                 spin_unlock_irq(&mapping->tree_lock);
464                 mem_cgroup_uncharge_cache_page(page);
465         }
466
467         return 1;
468
469 cannot_free:
470         spin_unlock_irq(&mapping->tree_lock);
471         return 0;
472 }
473
474 /*
475  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
476  * someone else has a ref on the page, abort and return 0.  If it was
477  * successfully detached, return 1.  Assumes the caller has a single ref on
478  * this page.
479  */
480 int remove_mapping(struct address_space *mapping, struct page *page)
481 {
482         if (__remove_mapping(mapping, page)) {
483                 /*
484                  * Unfreezing the refcount with 1 rather than 2 effectively
485                  * drops the pagecache ref for us without requiring another
486                  * atomic operation.
487                  */
488                 page_unfreeze_refs(page, 1);
489                 return 1;
490         }
491         return 0;
492 }
493
494 /**
495  * putback_lru_page - put previously isolated page onto appropriate LRU list
496  * @page: page to be put back to appropriate lru list
497  *
498  * Add previously isolated @page to appropriate LRU list.
499  * Page may still be unevictable for other reasons.
500  *
501  * lru_lock must not be held, interrupts must be enabled.
502  */
503 void putback_lru_page(struct page *page)
504 {
505         int lru;
506         int active = !!TestClearPageActive(page);
507         int was_unevictable = PageUnevictable(page);
508
509         VM_BUG_ON(PageLRU(page));
510
511 redo:
512         ClearPageUnevictable(page);
513
514         if (page_evictable(page, NULL)) {
515                 /*
516                  * For evictable pages, we can use the cache.
517                  * In event of a race, worst case is we end up with an
518                  * unevictable page on [in]active list.
519                  * We know how to handle that.
520                  */
521                 lru = active + page_lru_base_type(page);
522                 lru_cache_add_lru(page, lru);
523         } else {
524                 /*
525                  * Put unevictable pages directly on zone's unevictable
526                  * list.
527                  */
528                 lru = LRU_UNEVICTABLE;
529                 add_page_to_unevictable_list(page);
530                 /*
531                  * When racing with an mlock clearing (page is
532                  * unlocked), make sure that if the other thread does
533                  * not observe our setting of PG_lru and fails
534                  * isolation, we see PG_mlocked cleared below and move
535                  * the page back to the evictable list.
536                  *
537                  * The other side is TestClearPageMlocked().
538                  */
539                 smp_mb();
540         }
541
542         /*
543          * page's status can change while we move it among lru. If an evictable
544          * page is on unevictable list, it never be freed. To avoid that,
545          * check after we added it to the list, again.
546          */
547         if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
548                 if (!isolate_lru_page(page)) {
549                         put_page(page);
550                         goto redo;
551                 }
552                 /* This means someone else dropped this page from LRU
553                  * So, it will be freed or putback to LRU again. There is
554                  * nothing to do here.
555                  */
556         }
557
558         if (was_unevictable && lru != LRU_UNEVICTABLE)
559                 count_vm_event(UNEVICTABLE_PGRESCUED);
560         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
561                 count_vm_event(UNEVICTABLE_PGCULLED);
562
563         put_page(page);         /* drop ref from isolate */
564 }
565
566 enum page_references {
567         PAGEREF_RECLAIM,
568         PAGEREF_RECLAIM_CLEAN,
569         PAGEREF_KEEP,
570         PAGEREF_ACTIVATE,
571 };
572
573 static enum page_references page_check_references(struct page *page,
574                                                   struct scan_control *sc)
575 {
576         int referenced_ptes, referenced_page;
577         unsigned long vm_flags;
578
579         referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
580         referenced_page = TestClearPageReferenced(page);
581
582         /* Lumpy reclaim - ignore references */
583         if (sc->lumpy_reclaim_mode)
584                 return PAGEREF_RECLAIM;
585
586         /*
587          * Mlock lost the isolation race with us.  Let try_to_unmap()
588          * move the page to the unevictable list.
589          */
590         if (vm_flags & VM_LOCKED)
591                 return PAGEREF_RECLAIM;
592
593         if (referenced_ptes) {
594                 if (PageAnon(page))
595                         return PAGEREF_ACTIVATE;
596                 /*
597                  * All mapped pages start out with page table
598                  * references from the instantiating fault, so we need
599                  * to look twice if a mapped file page is used more
600                  * than once.
601                  *
602                  * Mark it and spare it for another trip around the
603                  * inactive list.  Another page table reference will
604                  * lead to its activation.
605                  *
606                  * Note: the mark is set for activated pages as well
607                  * so that recently deactivated but used pages are
608                  * quickly recovered.
609                  */
610                 SetPageReferenced(page);
611
612                 if (referenced_page)
613                         return PAGEREF_ACTIVATE;
614
615                 return PAGEREF_KEEP;
616         }
617
618         /* Reclaim if clean, defer dirty pages to writeback */
619         if (referenced_page)
620                 return PAGEREF_RECLAIM_CLEAN;
621
622         return PAGEREF_RECLAIM;
623 }
624
625 static noinline_for_stack void free_page_list(struct list_head *free_pages)
626 {
627         struct pagevec freed_pvec;
628         struct page *page, *tmp;
629
630         pagevec_init(&freed_pvec, 1);
631
632         list_for_each_entry_safe(page, tmp, free_pages, lru) {
633                 list_del(&page->lru);
634                 if (!pagevec_add(&freed_pvec, page)) {
635                         __pagevec_free(&freed_pvec);
636                         pagevec_reinit(&freed_pvec);
637                 }
638         }
639
640         pagevec_free(&freed_pvec);
641 }
642
643 /*
644  * shrink_page_list() returns the number of reclaimed pages
645  */
646 static unsigned long shrink_page_list(struct list_head *page_list,
647                                         struct scan_control *sc,
648                                         enum pageout_io sync_writeback)
649 {
650         LIST_HEAD(ret_pages);
651         LIST_HEAD(free_pages);
652         int pgactivate = 0;
653         unsigned long nr_reclaimed = 0;
654
655         cond_resched();
656
657         while (!list_empty(page_list)) {
658                 enum page_references references;
659                 struct address_space *mapping;
660                 struct page *page;
661                 int may_enter_fs;
662
663                 cond_resched();
664
665                 page = lru_to_page(page_list);
666                 list_del(&page->lru);
667
668                 if (!trylock_page(page))
669                         goto keep;
670
671                 VM_BUG_ON(PageActive(page));
672
673                 sc->nr_scanned++;
674
675                 if (unlikely(!page_evictable(page, NULL)))
676                         goto cull_mlocked;
677
678                 if (!sc->may_unmap && page_mapped(page))
679                         goto keep_locked;
680
681                 /* Double the slab pressure for mapped and swapcache pages */
682                 if (page_mapped(page) || PageSwapCache(page))
683                         sc->nr_scanned++;
684
685                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
686                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
687
688                 if (PageWriteback(page)) {
689                         /*
690                          * Synchronous reclaim is performed in two passes,
691                          * first an asynchronous pass over the list to
692                          * start parallel writeback, and a second synchronous
693                          * pass to wait for the IO to complete.  Wait here
694                          * for any page for which writeback has already
695                          * started.
696                          */
697                         if (sync_writeback == PAGEOUT_IO_SYNC && may_enter_fs)
698                                 wait_on_page_writeback(page);
699                         else
700                                 goto keep_locked;
701                 }
702
703                 references = page_check_references(page, sc);
704                 switch (references) {
705                 case PAGEREF_ACTIVATE:
706                         goto activate_locked;
707                 case PAGEREF_KEEP:
708                         goto keep_locked;
709                 case PAGEREF_RECLAIM:
710                 case PAGEREF_RECLAIM_CLEAN:
711                         ; /* try to reclaim the page below */
712                 }
713
714                 /*
715                  * Anonymous process memory has backing store?
716                  * Try to allocate it some swap space here.
717                  */
718                 if (PageAnon(page) && !PageSwapCache(page)) {
719                         if (!(sc->gfp_mask & __GFP_IO))
720                                 goto keep_locked;
721                         if (!add_to_swap(page))
722                                 goto activate_locked;
723                         may_enter_fs = 1;
724                 }
725
726                 mapping = page_mapping(page);
727
728                 /*
729                  * The page is mapped into the page tables of one or more
730                  * processes. Try to unmap it here.
731                  */
732                 if (page_mapped(page) && mapping) {
733                         switch (try_to_unmap(page, TTU_UNMAP)) {
734                         case SWAP_FAIL:
735                                 goto activate_locked;
736                         case SWAP_AGAIN:
737                                 goto keep_locked;
738                         case SWAP_MLOCK:
739                                 goto cull_mlocked;
740                         case SWAP_SUCCESS:
741                                 ; /* try to free the page below */
742                         }
743                 }
744
745                 if (PageDirty(page)) {
746                         if (references == PAGEREF_RECLAIM_CLEAN)
747                                 goto keep_locked;
748                         if (!may_enter_fs)
749                                 goto keep_locked;
750                         if (!sc->may_writepage)
751                                 goto keep_locked;
752
753                         /* Page is dirty, try to write it out here */
754                         switch (pageout(page, mapping, sync_writeback)) {
755                         case PAGE_KEEP:
756                                 goto keep_locked;
757                         case PAGE_ACTIVATE:
758                                 goto activate_locked;
759                         case PAGE_SUCCESS:
760                                 if (PageWriteback(page) || PageDirty(page))
761                                         goto keep;
762                                 /*
763                                  * A synchronous write - probably a ramdisk.  Go
764                                  * ahead and try to reclaim the page.
765                                  */
766                                 if (!trylock_page(page))
767                                         goto keep;
768                                 if (PageDirty(page) || PageWriteback(page))
769                                         goto keep_locked;
770                                 mapping = page_mapping(page);
771                         case PAGE_CLEAN:
772                                 ; /* try to free the page below */
773                         }
774                 }
775
776                 /*
777                  * If the page has buffers, try to free the buffer mappings
778                  * associated with this page. If we succeed we try to free
779                  * the page as well.
780                  *
781                  * We do this even if the page is PageDirty().
782                  * try_to_release_page() does not perform I/O, but it is
783                  * possible for a page to have PageDirty set, but it is actually
784                  * clean (all its buffers are clean).  This happens if the
785                  * buffers were written out directly, with submit_bh(). ext3
786                  * will do this, as well as the blockdev mapping.
787                  * try_to_release_page() will discover that cleanness and will
788                  * drop the buffers and mark the page clean - it can be freed.
789                  *
790                  * Rarely, pages can have buffers and no ->mapping.  These are
791                  * the pages which were not successfully invalidated in
792                  * truncate_complete_page().  We try to drop those buffers here
793                  * and if that worked, and the page is no longer mapped into
794                  * process address space (page_count == 1) it can be freed.
795                  * Otherwise, leave the page on the LRU so it is swappable.
796                  */
797                 if (page_has_private(page)) {
798                         if (!try_to_release_page(page, sc->gfp_mask))
799                                 goto activate_locked;
800                         if (!mapping && page_count(page) == 1) {
801                                 unlock_page(page);
802                                 if (put_page_testzero(page))
803                                         goto free_it;
804                                 else {
805                                         /*
806                                          * rare race with speculative reference.
807                                          * the speculative reference will free
808                                          * this page shortly, so we may
809                                          * increment nr_reclaimed here (and
810                                          * leave it off the LRU).
811                                          */
812                                         nr_reclaimed++;
813                                         continue;
814                                 }
815                         }
816                 }
817
818                 if (!mapping || !__remove_mapping(mapping, page))
819                         goto keep_locked;
820
821                 /*
822                  * At this point, we have no other references and there is
823                  * no way to pick any more up (removed from LRU, removed
824                  * from pagecache). Can use non-atomic bitops now (and
825                  * we obviously don't have to worry about waking up a process
826                  * waiting on the page lock, because there are no references.
827                  */
828                 __clear_page_locked(page);
829 free_it:
830                 nr_reclaimed++;
831
832                 /*
833                  * Is there need to periodically free_page_list? It would
834                  * appear not as the counts should be low
835                  */
836                 list_add(&page->lru, &free_pages);
837                 continue;
838
839 cull_mlocked:
840                 if (PageSwapCache(page))
841                         try_to_free_swap(page);
842                 unlock_page(page);
843                 putback_lru_page(page);
844                 continue;
845
846 activate_locked:
847                 /* Not a candidate for swapping, so reclaim swap space. */
848                 if (PageSwapCache(page) && vm_swap_full())
849                         try_to_free_swap(page);
850                 VM_BUG_ON(PageActive(page));
851                 SetPageActive(page);
852                 pgactivate++;
853 keep_locked:
854                 unlock_page(page);
855 keep:
856                 list_add(&page->lru, &ret_pages);
857                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
858         }
859
860         free_page_list(&free_pages);
861
862         list_splice(&ret_pages, page_list);
863         count_vm_events(PGACTIVATE, pgactivate);
864         return nr_reclaimed;
865 }
866
867 /*
868  * Attempt to remove the specified page from its LRU.  Only take this page
869  * if it is of the appropriate PageActive status.  Pages which are being
870  * freed elsewhere are also ignored.
871  *
872  * page:        page to consider
873  * mode:        one of the LRU isolation modes defined above
874  *
875  * returns 0 on success, -ve errno on failure.
876  */
877 int __isolate_lru_page(struct page *page, int mode, int file)
878 {
879         int ret = -EINVAL;
880
881         /* Only take pages on the LRU. */
882         if (!PageLRU(page))
883                 return ret;
884
885         /*
886          * When checking the active state, we need to be sure we are
887          * dealing with comparible boolean values.  Take the logical not
888          * of each.
889          */
890         if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
891                 return ret;
892
893         if (mode != ISOLATE_BOTH && page_is_file_cache(page) != file)
894                 return ret;
895
896         /*
897          * When this function is being called for lumpy reclaim, we
898          * initially look into all LRU pages, active, inactive and
899          * unevictable; only give shrink_page_list evictable pages.
900          */
901         if (PageUnevictable(page))
902                 return ret;
903
904         ret = -EBUSY;
905
906         if (likely(get_page_unless_zero(page))) {
907                 /*
908                  * Be careful not to clear PageLRU until after we're
909                  * sure the page is not being freed elsewhere -- the
910                  * page release code relies on it.
911                  */
912                 ClearPageLRU(page);
913                 ret = 0;
914         }
915
916         return ret;
917 }
918
919 /*
920  * zone->lru_lock is heavily contended.  Some of the functions that
921  * shrink the lists perform better by taking out a batch of pages
922  * and working on them outside the LRU lock.
923  *
924  * For pagecache intensive workloads, this function is the hottest
925  * spot in the kernel (apart from copy_*_user functions).
926  *
927  * Appropriate locks must be held before calling this function.
928  *
929  * @nr_to_scan: The number of pages to look through on the list.
930  * @src:        The LRU list to pull pages off.
931  * @dst:        The temp list to put pages on to.
932  * @scanned:    The number of pages that were scanned.
933  * @order:      The caller's attempted allocation order
934  * @mode:       One of the LRU isolation modes
935  * @file:       True [1] if isolating file [!anon] pages
936  *
937  * returns how many pages were moved onto *@dst.
938  */
939 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
940                 struct list_head *src, struct list_head *dst,
941                 unsigned long *scanned, int order, int mode, int file)
942 {
943         unsigned long nr_taken = 0;
944         unsigned long nr_lumpy_taken = 0;
945         unsigned long nr_lumpy_dirty = 0;
946         unsigned long nr_lumpy_failed = 0;
947         unsigned long scan;
948
949         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
950                 struct page *page;
951                 unsigned long pfn;
952                 unsigned long end_pfn;
953                 unsigned long page_pfn;
954                 int zone_id;
955
956                 page = lru_to_page(src);
957                 prefetchw_prev_lru_page(page, src, flags);
958
959                 VM_BUG_ON(!PageLRU(page));
960
961                 switch (__isolate_lru_page(page, mode, file)) {
962                 case 0:
963                         list_move(&page->lru, dst);
964                         mem_cgroup_del_lru(page);
965                         nr_taken++;
966                         break;
967
968                 case -EBUSY:
969                         /* else it is being freed elsewhere */
970                         list_move(&page->lru, src);
971                         mem_cgroup_rotate_lru_list(page, page_lru(page));
972                         continue;
973
974                 default:
975                         BUG();
976                 }
977
978                 if (!order)
979                         continue;
980
981                 /*
982                  * Attempt to take all pages in the order aligned region
983                  * surrounding the tag page.  Only take those pages of
984                  * the same active state as that tag page.  We may safely
985                  * round the target page pfn down to the requested order
986                  * as the mem_map is guarenteed valid out to MAX_ORDER,
987                  * where that page is in a different zone we will detect
988                  * it from its zone id and abort this block scan.
989                  */
990                 zone_id = page_zone_id(page);
991                 page_pfn = page_to_pfn(page);
992                 pfn = page_pfn & ~((1 << order) - 1);
993                 end_pfn = pfn + (1 << order);
994                 for (; pfn < end_pfn; pfn++) {
995                         struct page *cursor_page;
996
997                         /* The target page is in the block, ignore it. */
998                         if (unlikely(pfn == page_pfn))
999                                 continue;
1000
1001                         /* Avoid holes within the zone. */
1002                         if (unlikely(!pfn_valid_within(pfn)))
1003                                 break;
1004
1005                         cursor_page = pfn_to_page(pfn);
1006
1007                         /* Check that we have not crossed a zone boundary. */
1008                         if (unlikely(page_zone_id(cursor_page) != zone_id))
1009                                 continue;
1010
1011                         /*
1012                          * If we don't have enough swap space, reclaiming of
1013                          * anon page which don't already have a swap slot is
1014                          * pointless.
1015                          */
1016                         if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
1017                                         !PageSwapCache(cursor_page))
1018                                 continue;
1019
1020                         if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1021                                 list_move(&cursor_page->lru, dst);
1022                                 mem_cgroup_del_lru(cursor_page);
1023                                 nr_taken++;
1024                                 nr_lumpy_taken++;
1025                                 if (PageDirty(cursor_page))
1026                                         nr_lumpy_dirty++;
1027                                 scan++;
1028                         } else {
1029                                 if (mode == ISOLATE_BOTH &&
1030                                                 page_count(cursor_page))
1031                                         nr_lumpy_failed++;
1032                         }
1033                 }
1034         }
1035
1036         *scanned = scan;
1037
1038         trace_mm_vmscan_lru_isolate(order,
1039                         nr_to_scan, scan,
1040                         nr_taken,
1041                         nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1042                         mode);
1043         return nr_taken;
1044 }
1045
1046 static unsigned long isolate_pages_global(unsigned long nr,
1047                                         struct list_head *dst,
1048                                         unsigned long *scanned, int order,
1049                                         int mode, struct zone *z,
1050                                         int active, int file)
1051 {
1052         int lru = LRU_BASE;
1053         if (active)
1054                 lru += LRU_ACTIVE;
1055         if (file)
1056                 lru += LRU_FILE;
1057         return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1058                                                                 mode, file);
1059 }
1060
1061 /*
1062  * clear_active_flags() is a helper for shrink_active_list(), clearing
1063  * any active bits from the pages in the list.
1064  */
1065 static unsigned long clear_active_flags(struct list_head *page_list,
1066                                         unsigned int *count)
1067 {
1068         int nr_active = 0;
1069         int lru;
1070         struct page *page;
1071
1072         list_for_each_entry(page, page_list, lru) {
1073                 lru = page_lru_base_type(page);
1074                 if (PageActive(page)) {
1075                         lru += LRU_ACTIVE;
1076                         ClearPageActive(page);
1077                         nr_active++;
1078                 }
1079                 if (count)
1080                         count[lru]++;
1081         }
1082
1083         return nr_active;
1084 }
1085
1086 /**
1087  * isolate_lru_page - tries to isolate a page from its LRU list
1088  * @page: page to isolate from its LRU list
1089  *
1090  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1091  * vmstat statistic corresponding to whatever LRU list the page was on.
1092  *
1093  * Returns 0 if the page was removed from an LRU list.
1094  * Returns -EBUSY if the page was not on an LRU list.
1095  *
1096  * The returned page will have PageLRU() cleared.  If it was found on
1097  * the active list, it will have PageActive set.  If it was found on
1098  * the unevictable list, it will have the PageUnevictable bit set. That flag
1099  * may need to be cleared by the caller before letting the page go.
1100  *
1101  * The vmstat statistic corresponding to the list on which the page was
1102  * found will be decremented.
1103  *
1104  * Restrictions:
1105  * (1) Must be called with an elevated refcount on the page. This is a
1106  *     fundamentnal difference from isolate_lru_pages (which is called
1107  *     without a stable reference).
1108  * (2) the lru_lock must not be held.
1109  * (3) interrupts must be enabled.
1110  */
1111 int isolate_lru_page(struct page *page)
1112 {
1113         int ret = -EBUSY;
1114
1115         if (PageLRU(page)) {
1116                 struct zone *zone = page_zone(page);
1117
1118                 spin_lock_irq(&zone->lru_lock);
1119                 if (PageLRU(page) && get_page_unless_zero(page)) {
1120                         int lru = page_lru(page);
1121                         ret = 0;
1122                         ClearPageLRU(page);
1123
1124                         del_page_from_lru_list(zone, page, lru);
1125                 }
1126                 spin_unlock_irq(&zone->lru_lock);
1127         }
1128         return ret;
1129 }
1130
1131 /*
1132  * Are there way too many processes in the direct reclaim path already?
1133  */
1134 static int too_many_isolated(struct zone *zone, int file,
1135                 struct scan_control *sc)
1136 {
1137         unsigned long inactive, isolated;
1138
1139         if (current_is_kswapd())
1140                 return 0;
1141
1142         if (!scanning_global_lru(sc))
1143                 return 0;
1144
1145         if (file) {
1146                 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1147                 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1148         } else {
1149                 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1150                 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1151         }
1152
1153         return isolated > inactive;
1154 }
1155
1156 /*
1157  * TODO: Try merging with migrations version of putback_lru_pages
1158  */
1159 static noinline_for_stack void
1160 putback_lru_pages(struct zone *zone, struct scan_control *sc,
1161                                 unsigned long nr_anon, unsigned long nr_file,
1162                                 struct list_head *page_list)
1163 {
1164         struct page *page;
1165         struct pagevec pvec;
1166         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1167
1168         pagevec_init(&pvec, 1);
1169
1170         /*
1171          * Put back any unfreeable pages.
1172          */
1173         spin_lock(&zone->lru_lock);
1174         while (!list_empty(page_list)) {
1175                 int lru;
1176                 page = lru_to_page(page_list);
1177                 VM_BUG_ON(PageLRU(page));
1178                 list_del(&page->lru);
1179                 if (unlikely(!page_evictable(page, NULL))) {
1180                         spin_unlock_irq(&zone->lru_lock);
1181                         putback_lru_page(page);
1182                         spin_lock_irq(&zone->lru_lock);
1183                         continue;
1184                 }
1185                 SetPageLRU(page);
1186                 lru = page_lru(page);
1187                 add_page_to_lru_list(zone, page, lru);
1188                 if (is_active_lru(lru)) {
1189                         int file = is_file_lru(lru);
1190                         reclaim_stat->recent_rotated[file]++;
1191                 }
1192                 if (!pagevec_add(&pvec, page)) {
1193                         spin_unlock_irq(&zone->lru_lock);
1194                         __pagevec_release(&pvec);
1195                         spin_lock_irq(&zone->lru_lock);
1196                 }
1197         }
1198         __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1199         __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1200
1201         spin_unlock_irq(&zone->lru_lock);
1202         pagevec_release(&pvec);
1203 }
1204
1205 static noinline_for_stack void update_isolated_counts(struct zone *zone,
1206                                         struct scan_control *sc,
1207                                         unsigned long *nr_anon,
1208                                         unsigned long *nr_file,
1209                                         struct list_head *isolated_list)
1210 {
1211         unsigned long nr_active;
1212         unsigned int count[NR_LRU_LISTS] = { 0, };
1213         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1214
1215         nr_active = clear_active_flags(isolated_list, count);
1216         __count_vm_events(PGDEACTIVATE, nr_active);
1217
1218         __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1219                               -count[LRU_ACTIVE_FILE]);
1220         __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1221                               -count[LRU_INACTIVE_FILE]);
1222         __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1223                               -count[LRU_ACTIVE_ANON]);
1224         __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1225                               -count[LRU_INACTIVE_ANON]);
1226
1227         *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1228         *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1229         __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1230         __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1231
1232         reclaim_stat->recent_scanned[0] += *nr_anon;
1233         reclaim_stat->recent_scanned[1] += *nr_file;
1234 }
1235
1236 /*
1237  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1238  * of reclaimed pages
1239  */
1240 static noinline_for_stack unsigned long
1241 shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1242                         struct scan_control *sc, int priority, int file)
1243 {
1244         LIST_HEAD(page_list);
1245         unsigned long nr_scanned;
1246         unsigned long nr_reclaimed = 0;
1247         unsigned long nr_taken;
1248         unsigned long nr_active;
1249         unsigned long nr_anon;
1250         unsigned long nr_file;
1251
1252         while (unlikely(too_many_isolated(zone, file, sc))) {
1253                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1254
1255                 /* We are about to die and free our memory. Return now. */
1256                 if (fatal_signal_pending(current))
1257                         return SWAP_CLUSTER_MAX;
1258         }
1259
1260
1261         lru_add_drain();
1262         spin_lock_irq(&zone->lru_lock);
1263
1264         if (scanning_global_lru(sc)) {
1265                 nr_taken = isolate_pages_global(nr_to_scan,
1266                         &page_list, &nr_scanned, sc->order,
1267                         sc->lumpy_reclaim_mode ?
1268                                 ISOLATE_BOTH : ISOLATE_INACTIVE,
1269                         zone, 0, file);
1270                 zone->pages_scanned += nr_scanned;
1271                 if (current_is_kswapd())
1272                         __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1273                                                nr_scanned);
1274                 else
1275                         __count_zone_vm_events(PGSCAN_DIRECT, zone,
1276                                                nr_scanned);
1277         } else {
1278                 nr_taken = mem_cgroup_isolate_pages(nr_to_scan,
1279                         &page_list, &nr_scanned, sc->order,
1280                         sc->lumpy_reclaim_mode ?
1281                                 ISOLATE_BOTH : ISOLATE_INACTIVE,
1282                         zone, sc->mem_cgroup,
1283                         0, file);
1284                 /*
1285                  * mem_cgroup_isolate_pages() keeps track of
1286                  * scanned pages on its own.
1287                  */
1288         }
1289
1290         if (nr_taken == 0) {
1291                 spin_unlock_irq(&zone->lru_lock);
1292                 return 0;
1293         }
1294
1295         update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
1296
1297         spin_unlock_irq(&zone->lru_lock);
1298
1299         nr_reclaimed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
1300
1301         /*
1302          * If we are direct reclaiming for contiguous pages and we do
1303          * not reclaim everything in the list, try again and wait
1304          * for IO to complete. This will stall high-order allocations
1305          * but that should be acceptable to the caller
1306          */
1307         if (nr_reclaimed < nr_taken && !current_is_kswapd() &&
1308                         sc->lumpy_reclaim_mode) {
1309                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1310
1311                 /*
1312                  * The attempt at page out may have made some
1313                  * of the pages active, mark them inactive again.
1314                  */
1315                 nr_active = clear_active_flags(&page_list, NULL);
1316                 count_vm_events(PGDEACTIVATE, nr_active);
1317
1318                 nr_reclaimed += shrink_page_list(&page_list, sc, PAGEOUT_IO_SYNC);
1319         }
1320
1321         local_irq_disable();
1322         if (current_is_kswapd())
1323                 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1324         __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1325
1326         putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
1327         return nr_reclaimed;
1328 }
1329
1330 /*
1331  * This moves pages from the active list to the inactive list.
1332  *
1333  * We move them the other way if the page is referenced by one or more
1334  * processes, from rmap.
1335  *
1336  * If the pages are mostly unmapped, the processing is fast and it is
1337  * appropriate to hold zone->lru_lock across the whole operation.  But if
1338  * the pages are mapped, the processing is slow (page_referenced()) so we
1339  * should drop zone->lru_lock around each page.  It's impossible to balance
1340  * this, so instead we remove the pages from the LRU while processing them.
1341  * It is safe to rely on PG_active against the non-LRU pages in here because
1342  * nobody will play with that bit on a non-LRU page.
1343  *
1344  * The downside is that we have to touch page->_count against each page.
1345  * But we had to alter page->flags anyway.
1346  */
1347
1348 static void move_active_pages_to_lru(struct zone *zone,
1349                                      struct list_head *list,
1350                                      enum lru_list lru)
1351 {
1352         unsigned long pgmoved = 0;
1353         struct pagevec pvec;
1354         struct page *page;
1355
1356         pagevec_init(&pvec, 1);
1357
1358         while (!list_empty(list)) {
1359                 page = lru_to_page(list);
1360
1361                 VM_BUG_ON(PageLRU(page));
1362                 SetPageLRU(page);
1363
1364                 list_move(&page->lru, &zone->lru[lru].list);
1365                 mem_cgroup_add_lru_list(page, lru);
1366                 pgmoved++;
1367
1368                 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1369                         spin_unlock_irq(&zone->lru_lock);
1370                         if (buffer_heads_over_limit)
1371                                 pagevec_strip(&pvec);
1372                         __pagevec_release(&pvec);
1373                         spin_lock_irq(&zone->lru_lock);
1374                 }
1375         }
1376         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1377         if (!is_active_lru(lru))
1378                 __count_vm_events(PGDEACTIVATE, pgmoved);
1379 }
1380
1381 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1382                         struct scan_control *sc, int priority, int file)
1383 {
1384         unsigned long nr_taken;
1385         unsigned long pgscanned;
1386         unsigned long vm_flags;
1387         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1388         LIST_HEAD(l_active);
1389         LIST_HEAD(l_inactive);
1390         struct page *page;
1391         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1392         unsigned long nr_rotated = 0;
1393
1394         lru_add_drain();
1395         spin_lock_irq(&zone->lru_lock);
1396         if (scanning_global_lru(sc)) {
1397                 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1398                                                 &pgscanned, sc->order,
1399                                                 ISOLATE_ACTIVE, zone,
1400                                                 1, file);
1401                 zone->pages_scanned += pgscanned;
1402         } else {
1403                 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1404                                                 &pgscanned, sc->order,
1405                                                 ISOLATE_ACTIVE, zone,
1406                                                 sc->mem_cgroup, 1, file);
1407                 /*
1408                  * mem_cgroup_isolate_pages() keeps track of
1409                  * scanned pages on its own.
1410                  */
1411         }
1412
1413         reclaim_stat->recent_scanned[file] += nr_taken;
1414
1415         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1416         if (file)
1417                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1418         else
1419                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1420         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1421         spin_unlock_irq(&zone->lru_lock);
1422
1423         while (!list_empty(&l_hold)) {
1424                 cond_resched();
1425                 page = lru_to_page(&l_hold);
1426                 list_del(&page->lru);
1427
1428                 if (unlikely(!page_evictable(page, NULL))) {
1429                         putback_lru_page(page);
1430                         continue;
1431                 }
1432
1433                 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1434                         nr_rotated++;
1435                         /*
1436                          * Identify referenced, file-backed active pages and
1437                          * give them one more trip around the active list. So
1438                          * that executable code get better chances to stay in
1439                          * memory under moderate memory pressure.  Anon pages
1440                          * are not likely to be evicted by use-once streaming
1441                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1442                          * so we ignore them here.
1443                          */
1444                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1445                                 list_add(&page->lru, &l_active);
1446                                 continue;
1447                         }
1448                 }
1449
1450                 ClearPageActive(page);  /* we are de-activating */
1451                 list_add(&page->lru, &l_inactive);
1452         }
1453
1454         /*
1455          * Move pages back to the lru list.
1456          */
1457         spin_lock_irq(&zone->lru_lock);
1458         /*
1459          * Count referenced pages from currently used mappings as rotated,
1460          * even though only some of them are actually re-activated.  This
1461          * helps balance scan pressure between file and anonymous pages in
1462          * get_scan_ratio.
1463          */
1464         reclaim_stat->recent_rotated[file] += nr_rotated;
1465
1466         move_active_pages_to_lru(zone, &l_active,
1467                                                 LRU_ACTIVE + file * LRU_FILE);
1468         move_active_pages_to_lru(zone, &l_inactive,
1469                                                 LRU_BASE   + file * LRU_FILE);
1470         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1471         spin_unlock_irq(&zone->lru_lock);
1472 }
1473
1474 static int inactive_anon_is_low_global(struct zone *zone)
1475 {
1476         unsigned long active, inactive;
1477
1478         active = zone_page_state(zone, NR_ACTIVE_ANON);
1479         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1480
1481         if (inactive * zone->inactive_ratio < active)
1482                 return 1;
1483
1484         return 0;
1485 }
1486
1487 /**
1488  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1489  * @zone: zone to check
1490  * @sc:   scan control of this context
1491  *
1492  * Returns true if the zone does not have enough inactive anon pages,
1493  * meaning some active anon pages need to be deactivated.
1494  */
1495 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1496 {
1497         int low;
1498
1499         if (scanning_global_lru(sc))
1500                 low = inactive_anon_is_low_global(zone);
1501         else
1502                 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1503         return low;
1504 }
1505
1506 static int inactive_file_is_low_global(struct zone *zone)
1507 {
1508         unsigned long active, inactive;
1509
1510         active = zone_page_state(zone, NR_ACTIVE_FILE);
1511         inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1512
1513         return (active > inactive);
1514 }
1515
1516 /**
1517  * inactive_file_is_low - check if file pages need to be deactivated
1518  * @zone: zone to check
1519  * @sc:   scan control of this context
1520  *
1521  * When the system is doing streaming IO, memory pressure here
1522  * ensures that active file pages get deactivated, until more
1523  * than half of the file pages are on the inactive list.
1524  *
1525  * Once we get to that situation, protect the system's working
1526  * set from being evicted by disabling active file page aging.
1527  *
1528  * This uses a different ratio than the anonymous pages, because
1529  * the page cache uses a use-once replacement algorithm.
1530  */
1531 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1532 {
1533         int low;
1534
1535         if (scanning_global_lru(sc))
1536                 low = inactive_file_is_low_global(zone);
1537         else
1538                 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1539         return low;
1540 }
1541
1542 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1543                                 int file)
1544 {
1545         if (file)
1546                 return inactive_file_is_low(zone, sc);
1547         else
1548                 return inactive_anon_is_low(zone, sc);
1549 }
1550
1551 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1552         struct zone *zone, struct scan_control *sc, int priority)
1553 {
1554         int file = is_file_lru(lru);
1555
1556         if (is_active_lru(lru)) {
1557                 if (inactive_list_is_low(zone, sc, file))
1558                     shrink_active_list(nr_to_scan, zone, sc, priority, file);
1559                 return 0;
1560         }
1561
1562         return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1563 }
1564
1565 /*
1566  * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1567  * until we collected @swap_cluster_max pages to scan.
1568  */
1569 static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
1570                                        unsigned long *nr_saved_scan)
1571 {
1572         unsigned long nr;
1573
1574         *nr_saved_scan += nr_to_scan;
1575         nr = *nr_saved_scan;
1576
1577         if (nr >= SWAP_CLUSTER_MAX)
1578                 *nr_saved_scan = 0;
1579         else
1580                 nr = 0;
1581
1582         return nr;
1583 }
1584
1585 /*
1586  * Determine how aggressively the anon and file LRU lists should be
1587  * scanned.  The relative value of each set of LRU lists is determined
1588  * by looking at the fraction of the pages scanned we did rotate back
1589  * onto the active list instead of evict.
1590  *
1591  * nr[0] = anon pages to scan; nr[1] = file pages to scan
1592  */
1593 static void get_scan_count(struct zone *zone, struct scan_control *sc,
1594                                         unsigned long *nr, int priority)
1595 {
1596         unsigned long anon, file, free;
1597         unsigned long anon_prio, file_prio;
1598         unsigned long ap, fp;
1599         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1600         u64 fraction[2], denominator;
1601         enum lru_list l;
1602         int noswap = 0;
1603
1604         /* If we have no swap space, do not bother scanning anon pages. */
1605         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1606                 noswap = 1;
1607                 fraction[0] = 0;
1608                 fraction[1] = 1;
1609                 denominator = 1;
1610                 goto out;
1611         }
1612
1613         anon  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1614                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1615         file  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1616                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1617
1618         if (scanning_global_lru(sc)) {
1619                 free  = zone_page_state(zone, NR_FREE_PAGES);
1620                 /* If we have very few page cache pages,
1621                    force-scan anon pages. */
1622                 if (unlikely(file + free <= high_wmark_pages(zone))) {
1623                         fraction[0] = 1;
1624                         fraction[1] = 0;
1625                         denominator = 1;
1626                         goto out;
1627                 }
1628         }
1629
1630         /*
1631          * With swappiness at 100, anonymous and file have the same priority.
1632          * This scanning priority is essentially the inverse of IO cost.
1633          */
1634         anon_prio = sc->swappiness;
1635         file_prio = 200 - sc->swappiness;
1636
1637         /*
1638          * OK, so we have swap space and a fair amount of page cache
1639          * pages.  We use the recently rotated / recently scanned
1640          * ratios to determine how valuable each cache is.
1641          *
1642          * Because workloads change over time (and to avoid overflow)
1643          * we keep these statistics as a floating average, which ends
1644          * up weighing recent references more than old ones.
1645          *
1646          * anon in [0], file in [1]
1647          */
1648         spin_lock_irq(&zone->lru_lock);
1649         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1650                 reclaim_stat->recent_scanned[0] /= 2;
1651                 reclaim_stat->recent_rotated[0] /= 2;
1652         }
1653
1654         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1655                 reclaim_stat->recent_scanned[1] /= 2;
1656                 reclaim_stat->recent_rotated[1] /= 2;
1657         }
1658
1659         /*
1660          * The amount of pressure on anon vs file pages is inversely
1661          * proportional to the fraction of recently scanned pages on
1662          * each list that were recently referenced and in active use.
1663          */
1664         ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1665         ap /= reclaim_stat->recent_rotated[0] + 1;
1666
1667         fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1668         fp /= reclaim_stat->recent_rotated[1] + 1;
1669         spin_unlock_irq(&zone->lru_lock);
1670
1671         fraction[0] = ap;
1672         fraction[1] = fp;
1673         denominator = ap + fp + 1;
1674 out:
1675         for_each_evictable_lru(l) {
1676                 int file = is_file_lru(l);
1677                 unsigned long scan;
1678
1679                 scan = zone_nr_lru_pages(zone, sc, l);
1680                 if (priority || noswap) {
1681                         scan >>= priority;
1682                         scan = div64_u64(scan * fraction[file], denominator);
1683                 }
1684                 nr[l] = nr_scan_try_batch(scan,
1685                                           &reclaim_stat->nr_saved_scan[l]);
1686         }
1687 }
1688
1689 static void set_lumpy_reclaim_mode(int priority, struct scan_control *sc)
1690 {
1691         /*
1692          * If we need a large contiguous chunk of memory, or have
1693          * trouble getting a small set of contiguous pages, we
1694          * will reclaim both active and inactive pages.
1695          */
1696         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1697                 sc->lumpy_reclaim_mode = 1;
1698         else if (sc->order && priority < DEF_PRIORITY - 2)
1699                 sc->lumpy_reclaim_mode = 1;
1700         else
1701                 sc->lumpy_reclaim_mode = 0;
1702 }
1703
1704 /*
1705  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1706  */
1707 static void shrink_zone(int priority, struct zone *zone,
1708                                 struct scan_control *sc)
1709 {
1710         unsigned long nr[NR_LRU_LISTS];
1711         unsigned long nr_to_scan;
1712         enum lru_list l;
1713         unsigned long nr_reclaimed = sc->nr_reclaimed;
1714         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1715
1716         get_scan_count(zone, sc, nr, priority);
1717
1718         set_lumpy_reclaim_mode(priority, sc);
1719
1720         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1721                                         nr[LRU_INACTIVE_FILE]) {
1722                 for_each_evictable_lru(l) {
1723                         if (nr[l]) {
1724                                 nr_to_scan = min_t(unsigned long,
1725                                                    nr[l], SWAP_CLUSTER_MAX);
1726                                 nr[l] -= nr_to_scan;
1727
1728                                 nr_reclaimed += shrink_list(l, nr_to_scan,
1729                                                             zone, sc, priority);
1730                         }
1731                 }
1732                 /*
1733                  * On large memory systems, scan >> priority can become
1734                  * really large. This is fine for the starting priority;
1735                  * we want to put equal scanning pressure on each zone.
1736                  * However, if the VM has a harder time of freeing pages,
1737                  * with multiple processes reclaiming pages, the total
1738                  * freeing target can get unreasonably large.
1739                  */
1740                 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
1741                         break;
1742         }
1743
1744         sc->nr_reclaimed = nr_reclaimed;
1745
1746         /*
1747          * Even if we did not try to evict anon pages at all, we want to
1748          * rebalance the anon lru active/inactive ratio.
1749          */
1750         if (inactive_anon_is_low(zone, sc) && nr_swap_pages > 0)
1751                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1752
1753         throttle_vm_writeout(sc->gfp_mask);
1754 }
1755
1756 /*
1757  * This is the direct reclaim path, for page-allocating processes.  We only
1758  * try to reclaim pages from zones which will satisfy the caller's allocation
1759  * request.
1760  *
1761  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1762  * Because:
1763  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1764  *    allocation or
1765  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1766  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1767  *    zone defense algorithm.
1768  *
1769  * If a zone is deemed to be full of pinned pages then just give it a light
1770  * scan then give up on it.
1771  */
1772 static bool shrink_zones(int priority, struct zonelist *zonelist,
1773                                         struct scan_control *sc)
1774 {
1775         struct zoneref *z;
1776         struct zone *zone;
1777         bool all_unreclaimable = true;
1778
1779         for_each_zone_zonelist_nodemask(zone, z, zonelist,
1780                                         gfp_zone(sc->gfp_mask), sc->nodemask) {
1781                 if (!populated_zone(zone))
1782                         continue;
1783                 /*
1784                  * Take care memory controller reclaiming has small influence
1785                  * to global LRU.
1786                  */
1787                 if (scanning_global_lru(sc)) {
1788                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1789                                 continue;
1790                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1791                                 continue;       /* Let kswapd poll it */
1792                 }
1793
1794                 shrink_zone(priority, zone, sc);
1795                 all_unreclaimable = false;
1796         }
1797         return all_unreclaimable;
1798 }
1799
1800 /*
1801  * This is the main entry point to direct page reclaim.
1802  *
1803  * If a full scan of the inactive list fails to free enough memory then we
1804  * are "out of memory" and something needs to be killed.
1805  *
1806  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1807  * high - the zone may be full of dirty or under-writeback pages, which this
1808  * caller can't do much about.  We kick the writeback threads and take explicit
1809  * naps in the hope that some of these pages can be written.  But if the
1810  * allocating task holds filesystem locks which prevent writeout this might not
1811  * work, and the allocation attempt will fail.
1812  *
1813  * returns:     0, if no pages reclaimed
1814  *              else, the number of pages reclaimed
1815  */
1816 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
1817                                         struct scan_control *sc)
1818 {
1819         int priority;
1820         bool all_unreclaimable;
1821         unsigned long total_scanned = 0;
1822         struct reclaim_state *reclaim_state = current->reclaim_state;
1823         struct zoneref *z;
1824         struct zone *zone;
1825         unsigned long writeback_threshold;
1826
1827         get_mems_allowed();
1828         delayacct_freepages_start();
1829
1830         if (scanning_global_lru(sc))
1831                 count_vm_event(ALLOCSTALL);
1832
1833         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1834                 sc->nr_scanned = 0;
1835                 if (!priority)
1836                         disable_swap_token();
1837                 all_unreclaimable = shrink_zones(priority, zonelist, sc);
1838                 /*
1839                  * Don't shrink slabs when reclaiming memory from
1840                  * over limit cgroups
1841                  */
1842                 if (scanning_global_lru(sc)) {
1843                         unsigned long lru_pages = 0;
1844                         for_each_zone_zonelist(zone, z, zonelist,
1845                                         gfp_zone(sc->gfp_mask)) {
1846                                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1847                                         continue;
1848
1849                                 lru_pages += zone_reclaimable_pages(zone);
1850                         }
1851
1852                         shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
1853                         if (reclaim_state) {
1854                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
1855                                 reclaim_state->reclaimed_slab = 0;
1856                         }
1857                 }
1858                 total_scanned += sc->nr_scanned;
1859                 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
1860                         goto out;
1861
1862                 /*
1863                  * Try to write back as many pages as we just scanned.  This
1864                  * tends to cause slow streaming writers to write data to the
1865                  * disk smoothly, at the dirtying rate, which is nice.   But
1866                  * that's undesirable in laptop mode, where we *want* lumpy
1867                  * writeout.  So in laptop mode, write out the whole world.
1868                  */
1869                 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
1870                 if (total_scanned > writeback_threshold) {
1871                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
1872                         sc->may_writepage = 1;
1873                 }
1874
1875                 /* Take a nap, wait for some writeback to complete */
1876                 if (!sc->hibernation_mode && sc->nr_scanned &&
1877                     priority < DEF_PRIORITY - 2)
1878                         congestion_wait(BLK_RW_ASYNC, HZ/10);
1879         }
1880
1881 out:
1882         /*
1883          * Now that we've scanned all the zones at this priority level, note
1884          * that level within the zone so that the next thread which performs
1885          * scanning of this zone will immediately start out at this priority
1886          * level.  This affects only the decision whether or not to bring
1887          * mapped pages onto the inactive list.
1888          */
1889         if (priority < 0)
1890                 priority = 0;
1891
1892         delayacct_freepages_end();
1893         put_mems_allowed();
1894
1895         if (sc->nr_reclaimed)
1896                 return sc->nr_reclaimed;
1897
1898         /* top priority shrink_zones still had more to do? don't OOM, then */
1899         if (scanning_global_lru(sc) && !all_unreclaimable)
1900                 return 1;
1901
1902         return 0;
1903 }
1904
1905 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
1906                                 gfp_t gfp_mask, nodemask_t *nodemask)
1907 {
1908         unsigned long nr_reclaimed;
1909         struct scan_control sc = {
1910                 .gfp_mask = gfp_mask,
1911                 .may_writepage = !laptop_mode,
1912                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
1913                 .may_unmap = 1,
1914                 .may_swap = 1,
1915                 .swappiness = vm_swappiness,
1916                 .order = order,
1917                 .mem_cgroup = NULL,
1918                 .nodemask = nodemask,
1919         };
1920
1921         trace_mm_vmscan_direct_reclaim_begin(order,
1922                                 sc.may_writepage,
1923                                 gfp_mask);
1924
1925         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
1926
1927         trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
1928
1929         return nr_reclaimed;
1930 }
1931
1932 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
1933
1934 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
1935                                                 gfp_t gfp_mask, bool noswap,
1936                                                 unsigned int swappiness,
1937                                                 struct zone *zone, int nid)
1938 {
1939         struct scan_control sc = {
1940                 .may_writepage = !laptop_mode,
1941                 .may_unmap = 1,
1942                 .may_swap = !noswap,
1943                 .swappiness = swappiness,
1944                 .order = 0,
1945                 .mem_cgroup = mem,
1946         };
1947         nodemask_t nm  = nodemask_of_node(nid);
1948
1949         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1950                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1951         sc.nodemask = &nm;
1952         sc.nr_reclaimed = 0;
1953         sc.nr_scanned = 0;
1954
1955         trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
1956                                                       sc.may_writepage,
1957                                                       sc.gfp_mask);
1958
1959         /*
1960          * NOTE: Although we can get the priority field, using it
1961          * here is not a good idea, since it limits the pages we can scan.
1962          * if we don't reclaim here, the shrink_zone from balance_pgdat
1963          * will pick up pages from other mem cgroup's as well. We hack
1964          * the priority and make it zero.
1965          */
1966         shrink_zone(0, zone, &sc);
1967
1968         trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
1969
1970         return sc.nr_reclaimed;
1971 }
1972
1973 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
1974                                            gfp_t gfp_mask,
1975                                            bool noswap,
1976                                            unsigned int swappiness)
1977 {
1978         struct zonelist *zonelist;
1979         unsigned long nr_reclaimed;
1980         struct scan_control sc = {
1981                 .may_writepage = !laptop_mode,
1982                 .may_unmap = 1,
1983                 .may_swap = !noswap,
1984                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
1985                 .swappiness = swappiness,
1986                 .order = 0,
1987                 .mem_cgroup = mem_cont,
1988                 .nodemask = NULL, /* we don't care the placement */
1989         };
1990
1991         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1992                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1993         zonelist = NODE_DATA(numa_node_id())->node_zonelists;
1994
1995         trace_mm_vmscan_memcg_reclaim_begin(0,
1996                                             sc.may_writepage,
1997                                             sc.gfp_mask);
1998
1999         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2000
2001         trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2002
2003         return nr_reclaimed;
2004 }
2005 #endif
2006
2007 /* is kswapd sleeping prematurely? */
2008 static int sleeping_prematurely(pg_data_t *pgdat, int order, long remaining)
2009 {
2010         int i;
2011
2012         /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2013         if (remaining)
2014                 return 1;
2015
2016         /* If after HZ/10, a zone is below the high mark, it's premature */
2017         for (i = 0; i < pgdat->nr_zones; i++) {
2018                 struct zone *zone = pgdat->node_zones + i;
2019
2020                 if (!populated_zone(zone))
2021                         continue;
2022
2023                 if (zone->all_unreclaimable)
2024                         continue;
2025
2026                 if (!zone_watermark_ok(zone, order, high_wmark_pages(zone),
2027                                                                 0, 0))
2028                         return 1;
2029         }
2030
2031         return 0;
2032 }
2033
2034 /*
2035  * For kswapd, balance_pgdat() will work across all this node's zones until
2036  * they are all at high_wmark_pages(zone).
2037  *
2038  * Returns the number of pages which were actually freed.
2039  *
2040  * There is special handling here for zones which are full of pinned pages.
2041  * This can happen if the pages are all mlocked, or if they are all used by
2042  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
2043  * What we do is to detect the case where all pages in the zone have been
2044  * scanned twice and there has been zero successful reclaim.  Mark the zone as
2045  * dead and from now on, only perform a short scan.  Basically we're polling
2046  * the zone for when the problem goes away.
2047  *
2048  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
2049  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2050  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2051  * lower zones regardless of the number of free pages in the lower zones. This
2052  * interoperates with the page allocator fallback scheme to ensure that aging
2053  * of pages is balanced across the zones.
2054  */
2055 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
2056 {
2057         int all_zones_ok;
2058         int priority;
2059         int i;
2060         unsigned long total_scanned;
2061         struct reclaim_state *reclaim_state = current->reclaim_state;
2062         struct scan_control sc = {
2063                 .gfp_mask = GFP_KERNEL,
2064                 .may_unmap = 1,
2065                 .may_swap = 1,
2066                 /*
2067                  * kswapd doesn't want to be bailed out while reclaim. because
2068                  * we want to put equal scanning pressure on each zone.
2069                  */
2070                 .nr_to_reclaim = ULONG_MAX,
2071                 .swappiness = vm_swappiness,
2072                 .order = order,
2073                 .mem_cgroup = NULL,
2074         };
2075 loop_again:
2076         total_scanned = 0;
2077         sc.nr_reclaimed = 0;
2078         sc.may_writepage = !laptop_mode;
2079         count_vm_event(PAGEOUTRUN);
2080
2081         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2082                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
2083                 unsigned long lru_pages = 0;
2084                 int has_under_min_watermark_zone = 0;
2085
2086                 /* The swap token gets in the way of swapout... */
2087                 if (!priority)
2088                         disable_swap_token();
2089
2090                 all_zones_ok = 1;
2091
2092                 /*
2093                  * Scan in the highmem->dma direction for the highest
2094                  * zone which needs scanning
2095                  */
2096                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2097                         struct zone *zone = pgdat->node_zones + i;
2098
2099                         if (!populated_zone(zone))
2100                                 continue;
2101
2102                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2103                                 continue;
2104
2105                         /*
2106                          * Do some background aging of the anon list, to give
2107                          * pages a chance to be referenced before reclaiming.
2108                          */
2109                         if (inactive_anon_is_low(zone, &sc))
2110                                 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2111                                                         &sc, priority, 0);
2112
2113                         if (!zone_watermark_ok(zone, order,
2114                                         high_wmark_pages(zone), 0, 0)) {
2115                                 end_zone = i;
2116                                 break;
2117                         }
2118                 }
2119                 if (i < 0)
2120                         goto out;
2121
2122                 for (i = 0; i <= end_zone; i++) {
2123                         struct zone *zone = pgdat->node_zones + i;
2124
2125                         lru_pages += zone_reclaimable_pages(zone);
2126                 }
2127
2128                 /*
2129                  * Now scan the zone in the dma->highmem direction, stopping
2130                  * at the last zone which needs scanning.
2131                  *
2132                  * We do this because the page allocator works in the opposite
2133                  * direction.  This prevents the page allocator from allocating
2134                  * pages behind kswapd's direction of progress, which would
2135                  * cause too much scanning of the lower zones.
2136                  */
2137                 for (i = 0; i <= end_zone; i++) {
2138                         struct zone *zone = pgdat->node_zones + i;
2139                         int nr_slab;
2140                         int nid, zid;
2141
2142                         if (!populated_zone(zone))
2143                                 continue;
2144
2145                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2146                                 continue;
2147
2148                         sc.nr_scanned = 0;
2149
2150                         nid = pgdat->node_id;
2151                         zid = zone_idx(zone);
2152                         /*
2153                          * Call soft limit reclaim before calling shrink_zone.
2154                          * For now we ignore the return value
2155                          */
2156                         mem_cgroup_soft_limit_reclaim(zone, order, sc.gfp_mask,
2157                                                         nid, zid);
2158                         /*
2159                          * We put equal pressure on every zone, unless one
2160                          * zone has way too many pages free already.
2161                          */
2162                         if (!zone_watermark_ok(zone, order,
2163                                         8*high_wmark_pages(zone), end_zone, 0))
2164                                 shrink_zone(priority, zone, &sc);
2165                         reclaim_state->reclaimed_slab = 0;
2166                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
2167                                                 lru_pages);
2168                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2169                         total_scanned += sc.nr_scanned;
2170                         if (zone->all_unreclaimable)
2171                                 continue;
2172                         if (nr_slab == 0 &&
2173                             zone->pages_scanned >= (zone_reclaimable_pages(zone) * 6))
2174                                 zone->all_unreclaimable = 1;
2175                         /*
2176                          * If we've done a decent amount of scanning and
2177                          * the reclaim ratio is low, start doing writepage
2178                          * even in laptop mode
2179                          */
2180                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2181                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2182                                 sc.may_writepage = 1;
2183
2184                         if (!zone_watermark_ok(zone, order,
2185                                         high_wmark_pages(zone), end_zone, 0)) {
2186                                 all_zones_ok = 0;
2187                                 /*
2188                                  * We are still under min water mark.  This
2189                                  * means that we have a GFP_ATOMIC allocation
2190                                  * failure risk. Hurry up!
2191                                  */
2192                                 if (!zone_watermark_ok(zone, order,
2193                                             min_wmark_pages(zone), end_zone, 0))
2194                                         has_under_min_watermark_zone = 1;
2195                         }
2196
2197                 }
2198                 if (all_zones_ok)
2199                         break;          /* kswapd: all done */
2200                 /*
2201                  * OK, kswapd is getting into trouble.  Take a nap, then take
2202                  * another pass across the zones.
2203                  */
2204                 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2205                         if (has_under_min_watermark_zone)
2206                                 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2207                         else
2208                                 congestion_wait(BLK_RW_ASYNC, HZ/10);
2209                 }
2210
2211                 /*
2212                  * We do this so kswapd doesn't build up large priorities for
2213                  * example when it is freeing in parallel with allocators. It
2214                  * matches the direct reclaim path behaviour in terms of impact
2215                  * on zone->*_priority.
2216                  */
2217                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2218                         break;
2219         }
2220 out:
2221         if (!all_zones_ok) {
2222                 cond_resched();
2223
2224                 try_to_freeze();
2225
2226                 /*
2227                  * Fragmentation may mean that the system cannot be
2228                  * rebalanced for high-order allocations in all zones.
2229                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2230                  * it means the zones have been fully scanned and are still
2231                  * not balanced. For high-order allocations, there is
2232                  * little point trying all over again as kswapd may
2233                  * infinite loop.
2234                  *
2235                  * Instead, recheck all watermarks at order-0 as they
2236                  * are the most important. If watermarks are ok, kswapd will go
2237                  * back to sleep. High-order users can still perform direct
2238                  * reclaim if they wish.
2239                  */
2240                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2241                         order = sc.order = 0;
2242
2243                 goto loop_again;
2244         }
2245
2246         return sc.nr_reclaimed;
2247 }
2248
2249 /*
2250  * The background pageout daemon, started as a kernel thread
2251  * from the init process.
2252  *
2253  * This basically trickles out pages so that we have _some_
2254  * free memory available even if there is no other activity
2255  * that frees anything up. This is needed for things like routing
2256  * etc, where we otherwise might have all activity going on in
2257  * asynchronous contexts that cannot page things out.
2258  *
2259  * If there are applications that are active memory-allocators
2260  * (most normal use), this basically shouldn't matter.
2261  */
2262 static int kswapd(void *p)
2263 {
2264         unsigned long order;
2265         pg_data_t *pgdat = (pg_data_t*)p;
2266         struct task_struct *tsk = current;
2267         DEFINE_WAIT(wait);
2268         struct reclaim_state reclaim_state = {
2269                 .reclaimed_slab = 0,
2270         };
2271         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2272
2273         lockdep_set_current_reclaim_state(GFP_KERNEL);
2274
2275         if (!cpumask_empty(cpumask))
2276                 set_cpus_allowed_ptr(tsk, cpumask);
2277         current->reclaim_state = &reclaim_state;
2278
2279         /*
2280          * Tell the memory management that we're a "memory allocator",
2281          * and that if we need more memory we should get access to it
2282          * regardless (see "__alloc_pages()"). "kswapd" should
2283          * never get caught in the normal page freeing logic.
2284          *
2285          * (Kswapd normally doesn't need memory anyway, but sometimes
2286          * you need a small amount of memory in order to be able to
2287          * page out something else, and this flag essentially protects
2288          * us from recursively trying to free more memory as we're
2289          * trying to free the first piece of memory in the first place).
2290          */
2291         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2292         set_freezable();
2293
2294         order = 0;
2295         for ( ; ; ) {
2296                 unsigned long new_order;
2297                 int ret;
2298
2299                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2300                 new_order = pgdat->kswapd_max_order;
2301                 pgdat->kswapd_max_order = 0;
2302                 if (order < new_order) {
2303                         /*
2304                          * Don't sleep if someone wants a larger 'order'
2305                          * allocation
2306                          */
2307                         order = new_order;
2308                 } else {
2309                         if (!freezing(current) && !kthread_should_stop()) {
2310                                 long remaining = 0;
2311
2312                                 /* Try to sleep for a short interval */
2313                                 if (!sleeping_prematurely(pgdat, order, remaining)) {
2314                                         remaining = schedule_timeout(HZ/10);
2315                                         finish_wait(&pgdat->kswapd_wait, &wait);
2316                                         prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2317                                 }
2318
2319                                 /*
2320                                  * After a short sleep, check if it was a
2321                                  * premature sleep. If not, then go fully
2322                                  * to sleep until explicitly woken up
2323                                  */
2324                                 if (!sleeping_prematurely(pgdat, order, remaining)) {
2325                                         trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2326                                         schedule();
2327                                 } else {
2328                                         if (remaining)
2329                                                 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2330                                         else
2331                                                 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2332                                 }
2333                         }
2334
2335                         order = pgdat->kswapd_max_order;
2336                 }
2337                 finish_wait(&pgdat->kswapd_wait, &wait);
2338
2339                 ret = try_to_freeze();
2340                 if (kthread_should_stop())
2341                         break;
2342
2343                 /*
2344                  * We can speed up thawing tasks if we don't call balance_pgdat
2345                  * after returning from the refrigerator
2346                  */
2347                 if (!ret) {
2348                         trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2349                         balance_pgdat(pgdat, order);
2350                 }
2351         }
2352         return 0;
2353 }
2354
2355 /*
2356  * A zone is low on free memory, so wake its kswapd task to service it.
2357  */
2358 void wakeup_kswapd(struct zone *zone, int order)
2359 {
2360         pg_data_t *pgdat;
2361
2362         if (!populated_zone(zone))
2363                 return;
2364
2365         pgdat = zone->zone_pgdat;
2366         if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
2367                 return;
2368         if (pgdat->kswapd_max_order < order)
2369                 pgdat->kswapd_max_order = order;
2370         trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
2371         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2372                 return;
2373         if (!waitqueue_active(&pgdat->kswapd_wait))
2374                 return;
2375         wake_up_interruptible(&pgdat->kswapd_wait);
2376 }
2377
2378 /*
2379  * The reclaimable count would be mostly accurate.
2380  * The less reclaimable pages may be
2381  * - mlocked pages, which will be moved to unevictable list when encountered
2382  * - mapped pages, which may require several travels to be reclaimed
2383  * - dirty pages, which is not "instantly" reclaimable
2384  */
2385 unsigned long global_reclaimable_pages(void)
2386 {
2387         int nr;
2388
2389         nr = global_page_state(NR_ACTIVE_FILE) +
2390              global_page_state(NR_INACTIVE_FILE);
2391
2392         if (nr_swap_pages > 0)
2393                 nr += global_page_state(NR_ACTIVE_ANON) +
2394                       global_page_state(NR_INACTIVE_ANON);
2395
2396         return nr;
2397 }
2398
2399 unsigned long zone_reclaimable_pages(struct zone *zone)
2400 {
2401         int nr;
2402
2403         nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2404              zone_page_state(zone, NR_INACTIVE_FILE);
2405
2406         if (nr_swap_pages > 0)
2407                 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2408                       zone_page_state(zone, NR_INACTIVE_ANON);
2409
2410         return nr;
2411 }
2412
2413 #ifdef CONFIG_HIBERNATION
2414 /*
2415  * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2416  * freed pages.
2417  *
2418  * Rather than trying to age LRUs the aim is to preserve the overall
2419  * LRU order by reclaiming preferentially
2420  * inactive > active > active referenced > active mapped
2421  */
2422 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
2423 {
2424         struct reclaim_state reclaim_state;
2425         struct scan_control sc = {
2426                 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2427                 .may_swap = 1,
2428                 .may_unmap = 1,
2429                 .may_writepage = 1,
2430                 .nr_to_reclaim = nr_to_reclaim,
2431                 .hibernation_mode = 1,
2432                 .swappiness = vm_swappiness,
2433                 .order = 0,
2434         };
2435         struct zonelist * zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2436         struct task_struct *p = current;
2437         unsigned long nr_reclaimed;
2438
2439         p->flags |= PF_MEMALLOC;
2440         lockdep_set_current_reclaim_state(sc.gfp_mask);
2441         reclaim_state.reclaimed_slab = 0;
2442         p->reclaim_state = &reclaim_state;
2443
2444         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2445
2446         p->reclaim_state = NULL;
2447         lockdep_clear_current_reclaim_state();
2448         p->flags &= ~PF_MEMALLOC;
2449
2450         return nr_reclaimed;
2451 }
2452 #endif /* CONFIG_HIBERNATION */
2453
2454 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2455    not required for correctness.  So if the last cpu in a node goes
2456    away, we get changed to run anywhere: as the first one comes back,
2457    restore their cpu bindings. */
2458 static int __devinit cpu_callback(struct notifier_block *nfb,
2459                                   unsigned long action, void *hcpu)
2460 {
2461         int nid;
2462
2463         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2464                 for_each_node_state(nid, N_HIGH_MEMORY) {
2465                         pg_data_t *pgdat = NODE_DATA(nid);
2466                         const struct cpumask *mask;
2467
2468                         mask = cpumask_of_node(pgdat->node_id);
2469
2470                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2471                                 /* One of our CPUs online: restore mask */
2472                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2473                 }
2474         }
2475         return NOTIFY_OK;
2476 }
2477
2478 /*
2479  * This kswapd start function will be called by init and node-hot-add.
2480  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2481  */
2482 int kswapd_run(int nid)
2483 {
2484         pg_data_t *pgdat = NODE_DATA(nid);
2485         int ret = 0;
2486
2487         if (pgdat->kswapd)
2488                 return 0;
2489
2490         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2491         if (IS_ERR(pgdat->kswapd)) {
2492                 /* failure at boot is fatal */
2493                 BUG_ON(system_state == SYSTEM_BOOTING);
2494                 printk("Failed to start kswapd on node %d\n",nid);
2495                 ret = -1;
2496         }
2497         return ret;
2498 }
2499
2500 /*
2501  * Called by memory hotplug when all memory in a node is offlined.
2502  */
2503 void kswapd_stop(int nid)
2504 {
2505         struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
2506
2507         if (kswapd)
2508                 kthread_stop(kswapd);
2509 }
2510
2511 static int __init kswapd_init(void)
2512 {
2513         int nid;
2514
2515         swap_setup();
2516         for_each_node_state(nid, N_HIGH_MEMORY)
2517                 kswapd_run(nid);
2518         hotcpu_notifier(cpu_callback, 0);
2519         return 0;
2520 }
2521
2522 module_init(kswapd_init)
2523
2524 #ifdef CONFIG_NUMA
2525 /*
2526  * Zone reclaim mode
2527  *
2528  * If non-zero call zone_reclaim when the number of free pages falls below
2529  * the watermarks.
2530  */
2531 int zone_reclaim_mode __read_mostly;
2532
2533 #define RECLAIM_OFF 0
2534 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
2535 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
2536 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
2537
2538 /*
2539  * Priority for ZONE_RECLAIM. This determines the fraction of pages
2540  * of a node considered for each zone_reclaim. 4 scans 1/16th of
2541  * a zone.
2542  */
2543 #define ZONE_RECLAIM_PRIORITY 4
2544
2545 /*
2546  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2547  * occur.
2548  */
2549 int sysctl_min_unmapped_ratio = 1;
2550
2551 /*
2552  * If the number of slab pages in a zone grows beyond this percentage then
2553  * slab reclaim needs to occur.
2554  */
2555 int sysctl_min_slab_ratio = 5;
2556
2557 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
2558 {
2559         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
2560         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
2561                 zone_page_state(zone, NR_ACTIVE_FILE);
2562
2563         /*
2564          * It's possible for there to be more file mapped pages than
2565          * accounted for by the pages on the file LRU lists because
2566          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2567          */
2568         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
2569 }
2570
2571 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
2572 static long zone_pagecache_reclaimable(struct zone *zone)
2573 {
2574         long nr_pagecache_reclaimable;
2575         long delta = 0;
2576
2577         /*
2578          * If RECLAIM_SWAP is set, then all file pages are considered
2579          * potentially reclaimable. Otherwise, we have to worry about
2580          * pages like swapcache and zone_unmapped_file_pages() provides
2581          * a better estimate
2582          */
2583         if (zone_reclaim_mode & RECLAIM_SWAP)
2584                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
2585         else
2586                 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
2587
2588         /* If we can't clean pages, remove dirty pages from consideration */
2589         if (!(zone_reclaim_mode & RECLAIM_WRITE))
2590                 delta += zone_page_state(zone, NR_FILE_DIRTY);
2591
2592         /* Watch for any possible underflows due to delta */
2593         if (unlikely(delta > nr_pagecache_reclaimable))
2594                 delta = nr_pagecache_reclaimable;
2595
2596         return nr_pagecache_reclaimable - delta;
2597 }
2598
2599 /*
2600  * Try to free up some pages from this zone through reclaim.
2601  */
2602 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2603 {
2604         /* Minimum pages needed in order to stay on node */
2605         const unsigned long nr_pages = 1 << order;
2606         struct task_struct *p = current;
2607         struct reclaim_state reclaim_state;
2608         int priority;
2609         struct scan_control sc = {
2610                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2611                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2612                 .may_swap = 1,
2613                 .nr_to_reclaim = max_t(unsigned long, nr_pages,
2614                                        SWAP_CLUSTER_MAX),
2615                 .gfp_mask = gfp_mask,
2616                 .swappiness = vm_swappiness,
2617                 .order = order,
2618         };
2619         unsigned long nr_slab_pages0, nr_slab_pages1;
2620
2621         cond_resched();
2622         /*
2623          * We need to be able to allocate from the reserves for RECLAIM_SWAP
2624          * and we also need to be able to write out pages for RECLAIM_WRITE
2625          * and RECLAIM_SWAP.
2626          */
2627         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2628         lockdep_set_current_reclaim_state(gfp_mask);
2629         reclaim_state.reclaimed_slab = 0;
2630         p->reclaim_state = &reclaim_state;
2631
2632         if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
2633                 /*
2634                  * Free memory by calling shrink zone with increasing
2635                  * priorities until we have enough memory freed.
2636                  */
2637                 priority = ZONE_RECLAIM_PRIORITY;
2638                 do {
2639                         shrink_zone(priority, zone, &sc);
2640                         priority--;
2641                 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
2642         }
2643
2644         nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2645         if (nr_slab_pages0 > zone->min_slab_pages) {
2646                 /*
2647                  * shrink_slab() does not currently allow us to determine how
2648                  * many pages were freed in this zone. So we take the current
2649                  * number of slab pages and shake the slab until it is reduced
2650                  * by the same nr_pages that we used for reclaiming unmapped
2651                  * pages.
2652                  *
2653                  * Note that shrink_slab will free memory on all zones and may
2654                  * take a long time.
2655                  */
2656                 for (;;) {
2657                         unsigned long lru_pages = zone_reclaimable_pages(zone);
2658
2659                         /* No reclaimable slab or very low memory pressure */
2660                         if (!shrink_slab(sc.nr_scanned, gfp_mask, lru_pages))
2661                                 break;
2662
2663                         /* Freed enough memory */
2664                         nr_slab_pages1 = zone_page_state(zone,
2665                                                         NR_SLAB_RECLAIMABLE);
2666                         if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
2667                                 break;
2668                 }
2669
2670                 /*
2671                  * Update nr_reclaimed by the number of slab pages we
2672                  * reclaimed from this zone.
2673                  */
2674                 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2675                 if (nr_slab_pages1 < nr_slab_pages0)
2676                         sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
2677         }
2678
2679         p->reclaim_state = NULL;
2680         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
2681         lockdep_clear_current_reclaim_state();
2682         return sc.nr_reclaimed >= nr_pages;
2683 }
2684
2685 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2686 {
2687         int node_id;
2688         int ret;
2689
2690         /*
2691          * Zone reclaim reclaims unmapped file backed pages and
2692          * slab pages if we are over the defined limits.
2693          *
2694          * A small portion of unmapped file backed pages is needed for
2695          * file I/O otherwise pages read by file I/O will be immediately
2696          * thrown out if the zone is overallocated. So we do not reclaim
2697          * if less than a specified percentage of the zone is used by
2698          * unmapped file backed pages.
2699          */
2700         if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
2701             zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
2702                 return ZONE_RECLAIM_FULL;
2703
2704         if (zone->all_unreclaimable)
2705                 return ZONE_RECLAIM_FULL;
2706
2707         /*
2708          * Do not scan if the allocation should not be delayed.
2709          */
2710         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
2711                 return ZONE_RECLAIM_NOSCAN;
2712
2713         /*
2714          * Only run zone reclaim on the local zone or on zones that do not
2715          * have associated processors. This will favor the local processor
2716          * over remote processors and spread off node memory allocations
2717          * as wide as possible.
2718          */
2719         node_id = zone_to_nid(zone);
2720         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
2721                 return ZONE_RECLAIM_NOSCAN;
2722
2723         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2724                 return ZONE_RECLAIM_NOSCAN;
2725
2726         ret = __zone_reclaim(zone, gfp_mask, order);
2727         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2728
2729         if (!ret)
2730                 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
2731
2732         return ret;
2733 }
2734 #endif
2735
2736 /*
2737  * page_evictable - test whether a page is evictable
2738  * @page: the page to test
2739  * @vma: the VMA in which the page is or will be mapped, may be NULL
2740  *
2741  * Test whether page is evictable--i.e., should be placed on active/inactive
2742  * lists vs unevictable list.  The vma argument is !NULL when called from the
2743  * fault path to determine how to instantate a new page.
2744  *
2745  * Reasons page might not be evictable:
2746  * (1) page's mapping marked unevictable
2747  * (2) page is part of an mlocked VMA
2748  *
2749  */
2750 int page_evictable(struct page *page, struct vm_area_struct *vma)
2751 {
2752
2753         if (mapping_unevictable(page_mapping(page)))
2754                 return 0;
2755
2756         if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2757                 return 0;
2758
2759         return 1;
2760 }
2761
2762 /**
2763  * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2764  * @page: page to check evictability and move to appropriate lru list
2765  * @zone: zone page is in
2766  *
2767  * Checks a page for evictability and moves the page to the appropriate
2768  * zone lru list.
2769  *
2770  * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2771  * have PageUnevictable set.
2772  */
2773 static void check_move_unevictable_page(struct page *page, struct zone *zone)
2774 {
2775         VM_BUG_ON(PageActive(page));
2776
2777 retry:
2778         ClearPageUnevictable(page);
2779         if (page_evictable(page, NULL)) {
2780                 enum lru_list l = page_lru_base_type(page);
2781
2782                 __dec_zone_state(zone, NR_UNEVICTABLE);
2783                 list_move(&page->lru, &zone->lru[l].list);
2784                 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
2785                 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
2786                 __count_vm_event(UNEVICTABLE_PGRESCUED);
2787         } else {
2788                 /*
2789                  * rotate unevictable list
2790                  */
2791                 SetPageUnevictable(page);
2792                 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
2793                 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
2794                 if (page_evictable(page, NULL))
2795                         goto retry;
2796         }
2797 }
2798
2799 /**
2800  * scan_mapping_unevictable_pages - scan an address space for evictable pages
2801  * @mapping: struct address_space to scan for evictable pages
2802  *
2803  * Scan all pages in mapping.  Check unevictable pages for
2804  * evictability and move them to the appropriate zone lru list.
2805  */
2806 void scan_mapping_unevictable_pages(struct address_space *mapping)
2807 {
2808         pgoff_t next = 0;
2809         pgoff_t end   = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
2810                          PAGE_CACHE_SHIFT;
2811         struct zone *zone;
2812         struct pagevec pvec;
2813
2814         if (mapping->nrpages == 0)
2815                 return;
2816
2817         pagevec_init(&pvec, 0);
2818         while (next < end &&
2819                 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
2820                 int i;
2821                 int pg_scanned = 0;
2822
2823                 zone = NULL;
2824
2825                 for (i = 0; i < pagevec_count(&pvec); i++) {
2826                         struct page *page = pvec.pages[i];
2827                         pgoff_t page_index = page->index;
2828                         struct zone *pagezone = page_zone(page);
2829
2830                         pg_scanned++;
2831                         if (page_index > next)
2832                                 next = page_index;
2833                         next++;
2834
2835                         if (pagezone != zone) {
2836                                 if (zone)
2837                                         spin_unlock_irq(&zone->lru_lock);
2838                                 zone = pagezone;
2839                                 spin_lock_irq(&zone->lru_lock);
2840                         }
2841
2842                         if (PageLRU(page) && PageUnevictable(page))
2843                                 check_move_unevictable_page(page, zone);
2844                 }
2845                 if (zone)
2846                         spin_unlock_irq(&zone->lru_lock);
2847                 pagevec_release(&pvec);
2848
2849                 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
2850         }
2851
2852 }
2853
2854 /**
2855  * scan_zone_unevictable_pages - check unevictable list for evictable pages
2856  * @zone - zone of which to scan the unevictable list
2857  *
2858  * Scan @zone's unevictable LRU lists to check for pages that have become
2859  * evictable.  Move those that have to @zone's inactive list where they
2860  * become candidates for reclaim, unless shrink_inactive_zone() decides
2861  * to reactivate them.  Pages that are still unevictable are rotated
2862  * back onto @zone's unevictable list.
2863  */
2864 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
2865 static void scan_zone_unevictable_pages(struct zone *zone)
2866 {
2867         struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
2868         unsigned long scan;
2869         unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
2870
2871         while (nr_to_scan > 0) {
2872                 unsigned long batch_size = min(nr_to_scan,
2873                                                 SCAN_UNEVICTABLE_BATCH_SIZE);
2874
2875                 spin_lock_irq(&zone->lru_lock);
2876                 for (scan = 0;  scan < batch_size; scan++) {
2877                         struct page *page = lru_to_page(l_unevictable);
2878
2879                         if (!trylock_page(page))
2880                                 continue;
2881
2882                         prefetchw_prev_lru_page(page, l_unevictable, flags);
2883
2884                         if (likely(PageLRU(page) && PageUnevictable(page)))
2885                                 check_move_unevictable_page(page, zone);
2886
2887                         unlock_page(page);
2888                 }
2889                 spin_unlock_irq(&zone->lru_lock);
2890
2891                 nr_to_scan -= batch_size;
2892         }
2893 }
2894
2895
2896 /**
2897  * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
2898  *
2899  * A really big hammer:  scan all zones' unevictable LRU lists to check for
2900  * pages that have become evictable.  Move those back to the zones'
2901  * inactive list where they become candidates for reclaim.
2902  * This occurs when, e.g., we have unswappable pages on the unevictable lists,
2903  * and we add swap to the system.  As such, it runs in the context of a task
2904  * that has possibly/probably made some previously unevictable pages
2905  * evictable.
2906  */
2907 static void scan_all_zones_unevictable_pages(void)
2908 {
2909         struct zone *zone;
2910
2911         for_each_zone(zone) {
2912                 scan_zone_unevictable_pages(zone);
2913         }
2914 }
2915
2916 /*
2917  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
2918  * all nodes' unevictable lists for evictable pages
2919  */
2920 unsigned long scan_unevictable_pages;
2921
2922 int scan_unevictable_handler(struct ctl_table *table, int write,
2923                            void __user *buffer,
2924                            size_t *length, loff_t *ppos)
2925 {
2926         proc_doulongvec_minmax(table, write, buffer, length, ppos);
2927
2928         if (write && *(unsigned long *)table->data)
2929                 scan_all_zones_unevictable_pages();
2930
2931         scan_unevictable_pages = 0;
2932         return 0;
2933 }
2934
2935 /*
2936  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
2937  * a specified node's per zone unevictable lists for evictable pages.
2938  */
2939
2940 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
2941                                           struct sysdev_attribute *attr,
2942                                           char *buf)
2943 {
2944         return sprintf(buf, "0\n");     /* always zero; should fit... */
2945 }
2946
2947 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
2948                                            struct sysdev_attribute *attr,
2949                                         const char *buf, size_t count)
2950 {
2951         struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
2952         struct zone *zone;
2953         unsigned long res;
2954         unsigned long req = strict_strtoul(buf, 10, &res);
2955
2956         if (!req)
2957                 return 1;       /* zero is no-op */
2958
2959         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2960                 if (!populated_zone(zone))
2961                         continue;
2962                 scan_zone_unevictable_pages(zone);
2963         }
2964         return 1;
2965 }
2966
2967
2968 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
2969                         read_scan_unevictable_node,
2970                         write_scan_unevictable_node);
2971
2972 int scan_unevictable_register_node(struct node *node)
2973 {
2974         return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
2975 }
2976
2977 void scan_unevictable_unregister_node(struct node *node)
2978 {
2979         sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
2980 }
2981