8522327bd04f4ecf28da3ccdfccf398a9d810336
[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 /*
626  * shrink_page_list() returns the number of reclaimed pages
627  */
628 static unsigned long shrink_page_list(struct list_head *page_list,
629                                         struct scan_control *sc,
630                                         enum pageout_io sync_writeback)
631 {
632         LIST_HEAD(ret_pages);
633         struct pagevec freed_pvec;
634         int pgactivate = 0;
635         unsigned long nr_reclaimed = 0;
636
637         cond_resched();
638
639         pagevec_init(&freed_pvec, 1);
640         while (!list_empty(page_list)) {
641                 enum page_references references;
642                 struct address_space *mapping;
643                 struct page *page;
644                 int may_enter_fs;
645
646                 cond_resched();
647
648                 page = lru_to_page(page_list);
649                 list_del(&page->lru);
650
651                 if (!trylock_page(page))
652                         goto keep;
653
654                 VM_BUG_ON(PageActive(page));
655
656                 sc->nr_scanned++;
657
658                 if (unlikely(!page_evictable(page, NULL)))
659                         goto cull_mlocked;
660
661                 if (!sc->may_unmap && page_mapped(page))
662                         goto keep_locked;
663
664                 /* Double the slab pressure for mapped and swapcache pages */
665                 if (page_mapped(page) || PageSwapCache(page))
666                         sc->nr_scanned++;
667
668                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
669                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
670
671                 if (PageWriteback(page)) {
672                         /*
673                          * Synchronous reclaim is performed in two passes,
674                          * first an asynchronous pass over the list to
675                          * start parallel writeback, and a second synchronous
676                          * pass to wait for the IO to complete.  Wait here
677                          * for any page for which writeback has already
678                          * started.
679                          */
680                         if (sync_writeback == PAGEOUT_IO_SYNC && may_enter_fs)
681                                 wait_on_page_writeback(page);
682                         else
683                                 goto keep_locked;
684                 }
685
686                 references = page_check_references(page, sc);
687                 switch (references) {
688                 case PAGEREF_ACTIVATE:
689                         goto activate_locked;
690                 case PAGEREF_KEEP:
691                         goto keep_locked;
692                 case PAGEREF_RECLAIM:
693                 case PAGEREF_RECLAIM_CLEAN:
694                         ; /* try to reclaim the page below */
695                 }
696
697                 /*
698                  * Anonymous process memory has backing store?
699                  * Try to allocate it some swap space here.
700                  */
701                 if (PageAnon(page) && !PageSwapCache(page)) {
702                         if (!(sc->gfp_mask & __GFP_IO))
703                                 goto keep_locked;
704                         if (!add_to_swap(page))
705                                 goto activate_locked;
706                         may_enter_fs = 1;
707                 }
708
709                 mapping = page_mapping(page);
710
711                 /*
712                  * The page is mapped into the page tables of one or more
713                  * processes. Try to unmap it here.
714                  */
715                 if (page_mapped(page) && mapping) {
716                         switch (try_to_unmap(page, TTU_UNMAP)) {
717                         case SWAP_FAIL:
718                                 goto activate_locked;
719                         case SWAP_AGAIN:
720                                 goto keep_locked;
721                         case SWAP_MLOCK:
722                                 goto cull_mlocked;
723                         case SWAP_SUCCESS:
724                                 ; /* try to free the page below */
725                         }
726                 }
727
728                 if (PageDirty(page)) {
729                         if (references == PAGEREF_RECLAIM_CLEAN)
730                                 goto keep_locked;
731                         if (!may_enter_fs)
732                                 goto keep_locked;
733                         if (!sc->may_writepage)
734                                 goto keep_locked;
735
736                         /* Page is dirty, try to write it out here */
737                         switch (pageout(page, mapping, sync_writeback)) {
738                         case PAGE_KEEP:
739                                 goto keep_locked;
740                         case PAGE_ACTIVATE:
741                                 goto activate_locked;
742                         case PAGE_SUCCESS:
743                                 if (PageWriteback(page) || PageDirty(page))
744                                         goto keep;
745                                 /*
746                                  * A synchronous write - probably a ramdisk.  Go
747                                  * ahead and try to reclaim the page.
748                                  */
749                                 if (!trylock_page(page))
750                                         goto keep;
751                                 if (PageDirty(page) || PageWriteback(page))
752                                         goto keep_locked;
753                                 mapping = page_mapping(page);
754                         case PAGE_CLEAN:
755                                 ; /* try to free the page below */
756                         }
757                 }
758
759                 /*
760                  * If the page has buffers, try to free the buffer mappings
761                  * associated with this page. If we succeed we try to free
762                  * the page as well.
763                  *
764                  * We do this even if the page is PageDirty().
765                  * try_to_release_page() does not perform I/O, but it is
766                  * possible for a page to have PageDirty set, but it is actually
767                  * clean (all its buffers are clean).  This happens if the
768                  * buffers were written out directly, with submit_bh(). ext3
769                  * will do this, as well as the blockdev mapping.
770                  * try_to_release_page() will discover that cleanness and will
771                  * drop the buffers and mark the page clean - it can be freed.
772                  *
773                  * Rarely, pages can have buffers and no ->mapping.  These are
774                  * the pages which were not successfully invalidated in
775                  * truncate_complete_page().  We try to drop those buffers here
776                  * and if that worked, and the page is no longer mapped into
777                  * process address space (page_count == 1) it can be freed.
778                  * Otherwise, leave the page on the LRU so it is swappable.
779                  */
780                 if (page_has_private(page)) {
781                         if (!try_to_release_page(page, sc->gfp_mask))
782                                 goto activate_locked;
783                         if (!mapping && page_count(page) == 1) {
784                                 unlock_page(page);
785                                 if (put_page_testzero(page))
786                                         goto free_it;
787                                 else {
788                                         /*
789                                          * rare race with speculative reference.
790                                          * the speculative reference will free
791                                          * this page shortly, so we may
792                                          * increment nr_reclaimed here (and
793                                          * leave it off the LRU).
794                                          */
795                                         nr_reclaimed++;
796                                         continue;
797                                 }
798                         }
799                 }
800
801                 if (!mapping || !__remove_mapping(mapping, page))
802                         goto keep_locked;
803
804                 /*
805                  * At this point, we have no other references and there is
806                  * no way to pick any more up (removed from LRU, removed
807                  * from pagecache). Can use non-atomic bitops now (and
808                  * we obviously don't have to worry about waking up a process
809                  * waiting on the page lock, because there are no references.
810                  */
811                 __clear_page_locked(page);
812 free_it:
813                 nr_reclaimed++;
814                 if (!pagevec_add(&freed_pvec, page)) {
815                         __pagevec_free(&freed_pvec);
816                         pagevec_reinit(&freed_pvec);
817                 }
818                 continue;
819
820 cull_mlocked:
821                 if (PageSwapCache(page))
822                         try_to_free_swap(page);
823                 unlock_page(page);
824                 putback_lru_page(page);
825                 continue;
826
827 activate_locked:
828                 /* Not a candidate for swapping, so reclaim swap space. */
829                 if (PageSwapCache(page) && vm_swap_full())
830                         try_to_free_swap(page);
831                 VM_BUG_ON(PageActive(page));
832                 SetPageActive(page);
833                 pgactivate++;
834 keep_locked:
835                 unlock_page(page);
836 keep:
837                 list_add(&page->lru, &ret_pages);
838                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
839         }
840         list_splice(&ret_pages, page_list);
841         if (pagevec_count(&freed_pvec))
842                 __pagevec_free(&freed_pvec);
843         count_vm_events(PGACTIVATE, pgactivate);
844         return nr_reclaimed;
845 }
846
847 /*
848  * Attempt to remove the specified page from its LRU.  Only take this page
849  * if it is of the appropriate PageActive status.  Pages which are being
850  * freed elsewhere are also ignored.
851  *
852  * page:        page to consider
853  * mode:        one of the LRU isolation modes defined above
854  *
855  * returns 0 on success, -ve errno on failure.
856  */
857 int __isolate_lru_page(struct page *page, int mode, int file)
858 {
859         int ret = -EINVAL;
860
861         /* Only take pages on the LRU. */
862         if (!PageLRU(page))
863                 return ret;
864
865         /*
866          * When checking the active state, we need to be sure we are
867          * dealing with comparible boolean values.  Take the logical not
868          * of each.
869          */
870         if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
871                 return ret;
872
873         if (mode != ISOLATE_BOTH && page_is_file_cache(page) != file)
874                 return ret;
875
876         /*
877          * When this function is being called for lumpy reclaim, we
878          * initially look into all LRU pages, active, inactive and
879          * unevictable; only give shrink_page_list evictable pages.
880          */
881         if (PageUnevictable(page))
882                 return ret;
883
884         ret = -EBUSY;
885
886         if (likely(get_page_unless_zero(page))) {
887                 /*
888                  * Be careful not to clear PageLRU until after we're
889                  * sure the page is not being freed elsewhere -- the
890                  * page release code relies on it.
891                  */
892                 ClearPageLRU(page);
893                 ret = 0;
894         }
895
896         return ret;
897 }
898
899 /*
900  * zone->lru_lock is heavily contended.  Some of the functions that
901  * shrink the lists perform better by taking out a batch of pages
902  * and working on them outside the LRU lock.
903  *
904  * For pagecache intensive workloads, this function is the hottest
905  * spot in the kernel (apart from copy_*_user functions).
906  *
907  * Appropriate locks must be held before calling this function.
908  *
909  * @nr_to_scan: The number of pages to look through on the list.
910  * @src:        The LRU list to pull pages off.
911  * @dst:        The temp list to put pages on to.
912  * @scanned:    The number of pages that were scanned.
913  * @order:      The caller's attempted allocation order
914  * @mode:       One of the LRU isolation modes
915  * @file:       True [1] if isolating file [!anon] pages
916  *
917  * returns how many pages were moved onto *@dst.
918  */
919 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
920                 struct list_head *src, struct list_head *dst,
921                 unsigned long *scanned, int order, int mode, int file)
922 {
923         unsigned long nr_taken = 0;
924         unsigned long nr_lumpy_taken = 0;
925         unsigned long nr_lumpy_dirty = 0;
926         unsigned long nr_lumpy_failed = 0;
927         unsigned long scan;
928
929         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
930                 struct page *page;
931                 unsigned long pfn;
932                 unsigned long end_pfn;
933                 unsigned long page_pfn;
934                 int zone_id;
935
936                 page = lru_to_page(src);
937                 prefetchw_prev_lru_page(page, src, flags);
938
939                 VM_BUG_ON(!PageLRU(page));
940
941                 switch (__isolate_lru_page(page, mode, file)) {
942                 case 0:
943                         list_move(&page->lru, dst);
944                         mem_cgroup_del_lru(page);
945                         nr_taken++;
946                         break;
947
948                 case -EBUSY:
949                         /* else it is being freed elsewhere */
950                         list_move(&page->lru, src);
951                         mem_cgroup_rotate_lru_list(page, page_lru(page));
952                         continue;
953
954                 default:
955                         BUG();
956                 }
957
958                 if (!order)
959                         continue;
960
961                 /*
962                  * Attempt to take all pages in the order aligned region
963                  * surrounding the tag page.  Only take those pages of
964                  * the same active state as that tag page.  We may safely
965                  * round the target page pfn down to the requested order
966                  * as the mem_map is guarenteed valid out to MAX_ORDER,
967                  * where that page is in a different zone we will detect
968                  * it from its zone id and abort this block scan.
969                  */
970                 zone_id = page_zone_id(page);
971                 page_pfn = page_to_pfn(page);
972                 pfn = page_pfn & ~((1 << order) - 1);
973                 end_pfn = pfn + (1 << order);
974                 for (; pfn < end_pfn; pfn++) {
975                         struct page *cursor_page;
976
977                         /* The target page is in the block, ignore it. */
978                         if (unlikely(pfn == page_pfn))
979                                 continue;
980
981                         /* Avoid holes within the zone. */
982                         if (unlikely(!pfn_valid_within(pfn)))
983                                 break;
984
985                         cursor_page = pfn_to_page(pfn);
986
987                         /* Check that we have not crossed a zone boundary. */
988                         if (unlikely(page_zone_id(cursor_page) != zone_id))
989                                 continue;
990
991                         /*
992                          * If we don't have enough swap space, reclaiming of
993                          * anon page which don't already have a swap slot is
994                          * pointless.
995                          */
996                         if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
997                                         !PageSwapCache(cursor_page))
998                                 continue;
999
1000                         if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1001                                 list_move(&cursor_page->lru, dst);
1002                                 mem_cgroup_del_lru(cursor_page);
1003                                 nr_taken++;
1004                                 nr_lumpy_taken++;
1005                                 if (PageDirty(cursor_page))
1006                                         nr_lumpy_dirty++;
1007                                 scan++;
1008                         } else {
1009                                 if (mode == ISOLATE_BOTH &&
1010                                                 page_count(cursor_page))
1011                                         nr_lumpy_failed++;
1012                         }
1013                 }
1014         }
1015
1016         *scanned = scan;
1017
1018         trace_mm_vmscan_lru_isolate(order,
1019                         nr_to_scan, scan,
1020                         nr_taken,
1021                         nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1022                         mode);
1023         return nr_taken;
1024 }
1025
1026 static unsigned long isolate_pages_global(unsigned long nr,
1027                                         struct list_head *dst,
1028                                         unsigned long *scanned, int order,
1029                                         int mode, struct zone *z,
1030                                         int active, int file)
1031 {
1032         int lru = LRU_BASE;
1033         if (active)
1034                 lru += LRU_ACTIVE;
1035         if (file)
1036                 lru += LRU_FILE;
1037         return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1038                                                                 mode, file);
1039 }
1040
1041 /*
1042  * clear_active_flags() is a helper for shrink_active_list(), clearing
1043  * any active bits from the pages in the list.
1044  */
1045 static unsigned long clear_active_flags(struct list_head *page_list,
1046                                         unsigned int *count)
1047 {
1048         int nr_active = 0;
1049         int lru;
1050         struct page *page;
1051
1052         list_for_each_entry(page, page_list, lru) {
1053                 lru = page_lru_base_type(page);
1054                 if (PageActive(page)) {
1055                         lru += LRU_ACTIVE;
1056                         ClearPageActive(page);
1057                         nr_active++;
1058                 }
1059                 count[lru]++;
1060         }
1061
1062         return nr_active;
1063 }
1064
1065 /**
1066  * isolate_lru_page - tries to isolate a page from its LRU list
1067  * @page: page to isolate from its LRU list
1068  *
1069  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1070  * vmstat statistic corresponding to whatever LRU list the page was on.
1071  *
1072  * Returns 0 if the page was removed from an LRU list.
1073  * Returns -EBUSY if the page was not on an LRU list.
1074  *
1075  * The returned page will have PageLRU() cleared.  If it was found on
1076  * the active list, it will have PageActive set.  If it was found on
1077  * the unevictable list, it will have the PageUnevictable bit set. That flag
1078  * may need to be cleared by the caller before letting the page go.
1079  *
1080  * The vmstat statistic corresponding to the list on which the page was
1081  * found will be decremented.
1082  *
1083  * Restrictions:
1084  * (1) Must be called with an elevated refcount on the page. This is a
1085  *     fundamentnal difference from isolate_lru_pages (which is called
1086  *     without a stable reference).
1087  * (2) the lru_lock must not be held.
1088  * (3) interrupts must be enabled.
1089  */
1090 int isolate_lru_page(struct page *page)
1091 {
1092         int ret = -EBUSY;
1093
1094         if (PageLRU(page)) {
1095                 struct zone *zone = page_zone(page);
1096
1097                 spin_lock_irq(&zone->lru_lock);
1098                 if (PageLRU(page) && get_page_unless_zero(page)) {
1099                         int lru = page_lru(page);
1100                         ret = 0;
1101                         ClearPageLRU(page);
1102
1103                         del_page_from_lru_list(zone, page, lru);
1104                 }
1105                 spin_unlock_irq(&zone->lru_lock);
1106         }
1107         return ret;
1108 }
1109
1110 /*
1111  * Are there way too many processes in the direct reclaim path already?
1112  */
1113 static int too_many_isolated(struct zone *zone, int file,
1114                 struct scan_control *sc)
1115 {
1116         unsigned long inactive, isolated;
1117
1118         if (current_is_kswapd())
1119                 return 0;
1120
1121         if (!scanning_global_lru(sc))
1122                 return 0;
1123
1124         if (file) {
1125                 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1126                 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1127         } else {
1128                 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1129                 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1130         }
1131
1132         return isolated > inactive;
1133 }
1134
1135 /*
1136  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1137  * of reclaimed pages
1138  */
1139 static unsigned long shrink_inactive_list(unsigned long nr_to_scan,
1140                         struct zone *zone, struct scan_control *sc,
1141                         int priority, int file)
1142 {
1143         LIST_HEAD(page_list);
1144         struct pagevec pvec;
1145         unsigned long nr_scanned;
1146         unsigned long nr_reclaimed = 0;
1147         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1148         struct page *page;
1149         unsigned long nr_taken;
1150         unsigned long nr_active;
1151         unsigned int count[NR_LRU_LISTS] = { 0, };
1152         unsigned long nr_anon;
1153         unsigned long nr_file;
1154
1155         while (unlikely(too_many_isolated(zone, file, sc))) {
1156                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1157
1158                 /* We are about to die and free our memory. Return now. */
1159                 if (fatal_signal_pending(current))
1160                         return SWAP_CLUSTER_MAX;
1161         }
1162
1163
1164         pagevec_init(&pvec, 1);
1165
1166         lru_add_drain();
1167         spin_lock_irq(&zone->lru_lock);
1168
1169         if (scanning_global_lru(sc)) {
1170                 nr_taken = isolate_pages_global(nr_to_scan,
1171                         &page_list, &nr_scanned, sc->order,
1172                         sc->lumpy_reclaim_mode ?
1173                                 ISOLATE_BOTH : ISOLATE_INACTIVE,
1174                         zone, 0, file);
1175                 zone->pages_scanned += nr_scanned;
1176                 if (current_is_kswapd())
1177                         __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1178                                                nr_scanned);
1179                 else
1180                         __count_zone_vm_events(PGSCAN_DIRECT, zone,
1181                                                nr_scanned);
1182         } else {
1183                 nr_taken = mem_cgroup_isolate_pages(nr_to_scan,
1184                         &page_list, &nr_scanned, sc->order,
1185                         sc->lumpy_reclaim_mode ?
1186                                 ISOLATE_BOTH : ISOLATE_INACTIVE,
1187                         zone, sc->mem_cgroup,
1188                         0, file);
1189                 /*
1190                  * mem_cgroup_isolate_pages() keeps track of
1191                  * scanned pages on its own.
1192                  */
1193         }
1194
1195         if (nr_taken == 0)
1196                 goto done;
1197
1198         nr_active = clear_active_flags(&page_list, count);
1199         __count_vm_events(PGDEACTIVATE, nr_active);
1200
1201         __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1202                                         -count[LRU_ACTIVE_FILE]);
1203         __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1204                                         -count[LRU_INACTIVE_FILE]);
1205         __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1206                                         -count[LRU_ACTIVE_ANON]);
1207         __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1208                                         -count[LRU_INACTIVE_ANON]);
1209
1210         nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1211         nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1212         __mod_zone_page_state(zone, NR_ISOLATED_ANON, nr_anon);
1213         __mod_zone_page_state(zone, NR_ISOLATED_FILE, nr_file);
1214
1215         reclaim_stat->recent_scanned[0] += nr_anon;
1216         reclaim_stat->recent_scanned[1] += nr_file;
1217
1218         spin_unlock_irq(&zone->lru_lock);
1219
1220         nr_reclaimed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
1221
1222         /*
1223          * If we are direct reclaiming for contiguous pages and we do
1224          * not reclaim everything in the list, try again and wait
1225          * for IO to complete. This will stall high-order allocations
1226          * but that should be acceptable to the caller
1227          */
1228         if (nr_reclaimed < nr_taken && !current_is_kswapd() &&
1229                         sc->lumpy_reclaim_mode) {
1230                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1231
1232                 /*
1233                  * The attempt at page out may have made some
1234                  * of the pages active, mark them inactive again.
1235                  */
1236                 nr_active = clear_active_flags(&page_list, count);
1237                 count_vm_events(PGDEACTIVATE, nr_active);
1238
1239                 nr_reclaimed += shrink_page_list(&page_list, sc, PAGEOUT_IO_SYNC);
1240         }
1241
1242         local_irq_disable();
1243         if (current_is_kswapd())
1244                 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1245         __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1246
1247         spin_lock(&zone->lru_lock);
1248         /*
1249          * Put back any unfreeable pages.
1250          */
1251         while (!list_empty(&page_list)) {
1252                 int lru;
1253                 page = lru_to_page(&page_list);
1254                 VM_BUG_ON(PageLRU(page));
1255                 list_del(&page->lru);
1256                 if (unlikely(!page_evictable(page, NULL))) {
1257                         spin_unlock_irq(&zone->lru_lock);
1258                         putback_lru_page(page);
1259                         spin_lock_irq(&zone->lru_lock);
1260                         continue;
1261                 }
1262                 SetPageLRU(page);
1263                 lru = page_lru(page);
1264                 add_page_to_lru_list(zone, page, lru);
1265                 if (is_active_lru(lru)) {
1266                         int file = is_file_lru(lru);
1267                         reclaim_stat->recent_rotated[file]++;
1268                 }
1269                 if (!pagevec_add(&pvec, page)) {
1270                         spin_unlock_irq(&zone->lru_lock);
1271                         __pagevec_release(&pvec);
1272                         spin_lock_irq(&zone->lru_lock);
1273                 }
1274         }
1275         __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1276         __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1277
1278 done:
1279         spin_unlock_irq(&zone->lru_lock);
1280         pagevec_release(&pvec);
1281         return nr_reclaimed;
1282 }
1283
1284 /*
1285  * This moves pages from the active list to the inactive list.
1286  *
1287  * We move them the other way if the page is referenced by one or more
1288  * processes, from rmap.
1289  *
1290  * If the pages are mostly unmapped, the processing is fast and it is
1291  * appropriate to hold zone->lru_lock across the whole operation.  But if
1292  * the pages are mapped, the processing is slow (page_referenced()) so we
1293  * should drop zone->lru_lock around each page.  It's impossible to balance
1294  * this, so instead we remove the pages from the LRU while processing them.
1295  * It is safe to rely on PG_active against the non-LRU pages in here because
1296  * nobody will play with that bit on a non-LRU page.
1297  *
1298  * The downside is that we have to touch page->_count against each page.
1299  * But we had to alter page->flags anyway.
1300  */
1301
1302 static void move_active_pages_to_lru(struct zone *zone,
1303                                      struct list_head *list,
1304                                      enum lru_list lru)
1305 {
1306         unsigned long pgmoved = 0;
1307         struct pagevec pvec;
1308         struct page *page;
1309
1310         pagevec_init(&pvec, 1);
1311
1312         while (!list_empty(list)) {
1313                 page = lru_to_page(list);
1314
1315                 VM_BUG_ON(PageLRU(page));
1316                 SetPageLRU(page);
1317
1318                 list_move(&page->lru, &zone->lru[lru].list);
1319                 mem_cgroup_add_lru_list(page, lru);
1320                 pgmoved++;
1321
1322                 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1323                         spin_unlock_irq(&zone->lru_lock);
1324                         if (buffer_heads_over_limit)
1325                                 pagevec_strip(&pvec);
1326                         __pagevec_release(&pvec);
1327                         spin_lock_irq(&zone->lru_lock);
1328                 }
1329         }
1330         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1331         if (!is_active_lru(lru))
1332                 __count_vm_events(PGDEACTIVATE, pgmoved);
1333 }
1334
1335 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1336                         struct scan_control *sc, int priority, int file)
1337 {
1338         unsigned long nr_taken;
1339         unsigned long pgscanned;
1340         unsigned long vm_flags;
1341         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1342         LIST_HEAD(l_active);
1343         LIST_HEAD(l_inactive);
1344         struct page *page;
1345         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1346         unsigned long nr_rotated = 0;
1347
1348         lru_add_drain();
1349         spin_lock_irq(&zone->lru_lock);
1350         if (scanning_global_lru(sc)) {
1351                 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1352                                                 &pgscanned, sc->order,
1353                                                 ISOLATE_ACTIVE, zone,
1354                                                 1, file);
1355                 zone->pages_scanned += pgscanned;
1356         } else {
1357                 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1358                                                 &pgscanned, sc->order,
1359                                                 ISOLATE_ACTIVE, zone,
1360                                                 sc->mem_cgroup, 1, file);
1361                 /*
1362                  * mem_cgroup_isolate_pages() keeps track of
1363                  * scanned pages on its own.
1364                  */
1365         }
1366
1367         reclaim_stat->recent_scanned[file] += nr_taken;
1368
1369         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1370         if (file)
1371                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1372         else
1373                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1374         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1375         spin_unlock_irq(&zone->lru_lock);
1376
1377         while (!list_empty(&l_hold)) {
1378                 cond_resched();
1379                 page = lru_to_page(&l_hold);
1380                 list_del(&page->lru);
1381
1382                 if (unlikely(!page_evictable(page, NULL))) {
1383                         putback_lru_page(page);
1384                         continue;
1385                 }
1386
1387                 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1388                         nr_rotated++;
1389                         /*
1390                          * Identify referenced, file-backed active pages and
1391                          * give them one more trip around the active list. So
1392                          * that executable code get better chances to stay in
1393                          * memory under moderate memory pressure.  Anon pages
1394                          * are not likely to be evicted by use-once streaming
1395                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1396                          * so we ignore them here.
1397                          */
1398                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1399                                 list_add(&page->lru, &l_active);
1400                                 continue;
1401                         }
1402                 }
1403
1404                 ClearPageActive(page);  /* we are de-activating */
1405                 list_add(&page->lru, &l_inactive);
1406         }
1407
1408         /*
1409          * Move pages back to the lru list.
1410          */
1411         spin_lock_irq(&zone->lru_lock);
1412         /*
1413          * Count referenced pages from currently used mappings as rotated,
1414          * even though only some of them are actually re-activated.  This
1415          * helps balance scan pressure between file and anonymous pages in
1416          * get_scan_ratio.
1417          */
1418         reclaim_stat->recent_rotated[file] += nr_rotated;
1419
1420         move_active_pages_to_lru(zone, &l_active,
1421                                                 LRU_ACTIVE + file * LRU_FILE);
1422         move_active_pages_to_lru(zone, &l_inactive,
1423                                                 LRU_BASE   + file * LRU_FILE);
1424         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1425         spin_unlock_irq(&zone->lru_lock);
1426 }
1427
1428 static int inactive_anon_is_low_global(struct zone *zone)
1429 {
1430         unsigned long active, inactive;
1431
1432         active = zone_page_state(zone, NR_ACTIVE_ANON);
1433         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1434
1435         if (inactive * zone->inactive_ratio < active)
1436                 return 1;
1437
1438         return 0;
1439 }
1440
1441 /**
1442  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1443  * @zone: zone to check
1444  * @sc:   scan control of this context
1445  *
1446  * Returns true if the zone does not have enough inactive anon pages,
1447  * meaning some active anon pages need to be deactivated.
1448  */
1449 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1450 {
1451         int low;
1452
1453         if (scanning_global_lru(sc))
1454                 low = inactive_anon_is_low_global(zone);
1455         else
1456                 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1457         return low;
1458 }
1459
1460 static int inactive_file_is_low_global(struct zone *zone)
1461 {
1462         unsigned long active, inactive;
1463
1464         active = zone_page_state(zone, NR_ACTIVE_FILE);
1465         inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1466
1467         return (active > inactive);
1468 }
1469
1470 /**
1471  * inactive_file_is_low - check if file pages need to be deactivated
1472  * @zone: zone to check
1473  * @sc:   scan control of this context
1474  *
1475  * When the system is doing streaming IO, memory pressure here
1476  * ensures that active file pages get deactivated, until more
1477  * than half of the file pages are on the inactive list.
1478  *
1479  * Once we get to that situation, protect the system's working
1480  * set from being evicted by disabling active file page aging.
1481  *
1482  * This uses a different ratio than the anonymous pages, because
1483  * the page cache uses a use-once replacement algorithm.
1484  */
1485 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1486 {
1487         int low;
1488
1489         if (scanning_global_lru(sc))
1490                 low = inactive_file_is_low_global(zone);
1491         else
1492                 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1493         return low;
1494 }
1495
1496 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1497                                 int file)
1498 {
1499         if (file)
1500                 return inactive_file_is_low(zone, sc);
1501         else
1502                 return inactive_anon_is_low(zone, sc);
1503 }
1504
1505 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1506         struct zone *zone, struct scan_control *sc, int priority)
1507 {
1508         int file = is_file_lru(lru);
1509
1510         if (is_active_lru(lru)) {
1511                 if (inactive_list_is_low(zone, sc, file))
1512                     shrink_active_list(nr_to_scan, zone, sc, priority, file);
1513                 return 0;
1514         }
1515
1516         return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1517 }
1518
1519 /*
1520  * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1521  * until we collected @swap_cluster_max pages to scan.
1522  */
1523 static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
1524                                        unsigned long *nr_saved_scan)
1525 {
1526         unsigned long nr;
1527
1528         *nr_saved_scan += nr_to_scan;
1529         nr = *nr_saved_scan;
1530
1531         if (nr >= SWAP_CLUSTER_MAX)
1532                 *nr_saved_scan = 0;
1533         else
1534                 nr = 0;
1535
1536         return nr;
1537 }
1538
1539 /*
1540  * Determine how aggressively the anon and file LRU lists should be
1541  * scanned.  The relative value of each set of LRU lists is determined
1542  * by looking at the fraction of the pages scanned we did rotate back
1543  * onto the active list instead of evict.
1544  *
1545  * nr[0] = anon pages to scan; nr[1] = file pages to scan
1546  */
1547 static void get_scan_count(struct zone *zone, struct scan_control *sc,
1548                                         unsigned long *nr, int priority)
1549 {
1550         unsigned long anon, file, free;
1551         unsigned long anon_prio, file_prio;
1552         unsigned long ap, fp;
1553         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1554         u64 fraction[2], denominator;
1555         enum lru_list l;
1556         int noswap = 0;
1557
1558         /* If we have no swap space, do not bother scanning anon pages. */
1559         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1560                 noswap = 1;
1561                 fraction[0] = 0;
1562                 fraction[1] = 1;
1563                 denominator = 1;
1564                 goto out;
1565         }
1566
1567         anon  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1568                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1569         file  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1570                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1571
1572         if (scanning_global_lru(sc)) {
1573                 free  = zone_page_state(zone, NR_FREE_PAGES);
1574                 /* If we have very few page cache pages,
1575                    force-scan anon pages. */
1576                 if (unlikely(file + free <= high_wmark_pages(zone))) {
1577                         fraction[0] = 1;
1578                         fraction[1] = 0;
1579                         denominator = 1;
1580                         goto out;
1581                 }
1582         }
1583
1584         /*
1585          * OK, so we have swap space and a fair amount of page cache
1586          * pages.  We use the recently rotated / recently scanned
1587          * ratios to determine how valuable each cache is.
1588          *
1589          * Because workloads change over time (and to avoid overflow)
1590          * we keep these statistics as a floating average, which ends
1591          * up weighing recent references more than old ones.
1592          *
1593          * anon in [0], file in [1]
1594          */
1595         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1596                 spin_lock_irq(&zone->lru_lock);
1597                 reclaim_stat->recent_scanned[0] /= 2;
1598                 reclaim_stat->recent_rotated[0] /= 2;
1599                 spin_unlock_irq(&zone->lru_lock);
1600         }
1601
1602         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1603                 spin_lock_irq(&zone->lru_lock);
1604                 reclaim_stat->recent_scanned[1] /= 2;
1605                 reclaim_stat->recent_rotated[1] /= 2;
1606                 spin_unlock_irq(&zone->lru_lock);
1607         }
1608
1609         /*
1610          * With swappiness at 100, anonymous and file have the same priority.
1611          * This scanning priority is essentially the inverse of IO cost.
1612          */
1613         anon_prio = sc->swappiness;
1614         file_prio = 200 - sc->swappiness;
1615
1616         /*
1617          * The amount of pressure on anon vs file pages is inversely
1618          * proportional to the fraction of recently scanned pages on
1619          * each list that were recently referenced and in active use.
1620          */
1621         ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1622         ap /= reclaim_stat->recent_rotated[0] + 1;
1623
1624         fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1625         fp /= reclaim_stat->recent_rotated[1] + 1;
1626
1627         fraction[0] = ap;
1628         fraction[1] = fp;
1629         denominator = ap + fp + 1;
1630 out:
1631         for_each_evictable_lru(l) {
1632                 int file = is_file_lru(l);
1633                 unsigned long scan;
1634
1635                 scan = zone_nr_lru_pages(zone, sc, l);
1636                 if (priority || noswap) {
1637                         scan >>= priority;
1638                         scan = div64_u64(scan * fraction[file], denominator);
1639                 }
1640                 nr[l] = nr_scan_try_batch(scan,
1641                                           &reclaim_stat->nr_saved_scan[l]);
1642         }
1643 }
1644
1645 static void set_lumpy_reclaim_mode(int priority, struct scan_control *sc)
1646 {
1647         /*
1648          * If we need a large contiguous chunk of memory, or have
1649          * trouble getting a small set of contiguous pages, we
1650          * will reclaim both active and inactive pages.
1651          */
1652         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1653                 sc->lumpy_reclaim_mode = 1;
1654         else if (sc->order && priority < DEF_PRIORITY - 2)
1655                 sc->lumpy_reclaim_mode = 1;
1656         else
1657                 sc->lumpy_reclaim_mode = 0;
1658 }
1659
1660 /*
1661  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1662  */
1663 static void shrink_zone(int priority, struct zone *zone,
1664                                 struct scan_control *sc)
1665 {
1666         unsigned long nr[NR_LRU_LISTS];
1667         unsigned long nr_to_scan;
1668         enum lru_list l;
1669         unsigned long nr_reclaimed = sc->nr_reclaimed;
1670         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1671
1672         get_scan_count(zone, sc, nr, priority);
1673
1674         set_lumpy_reclaim_mode(priority, sc);
1675
1676         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1677                                         nr[LRU_INACTIVE_FILE]) {
1678                 for_each_evictable_lru(l) {
1679                         if (nr[l]) {
1680                                 nr_to_scan = min_t(unsigned long,
1681                                                    nr[l], SWAP_CLUSTER_MAX);
1682                                 nr[l] -= nr_to_scan;
1683
1684                                 nr_reclaimed += shrink_list(l, nr_to_scan,
1685                                                             zone, sc, priority);
1686                         }
1687                 }
1688                 /*
1689                  * On large memory systems, scan >> priority can become
1690                  * really large. This is fine for the starting priority;
1691                  * we want to put equal scanning pressure on each zone.
1692                  * However, if the VM has a harder time of freeing pages,
1693                  * with multiple processes reclaiming pages, the total
1694                  * freeing target can get unreasonably large.
1695                  */
1696                 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
1697                         break;
1698         }
1699
1700         sc->nr_reclaimed = nr_reclaimed;
1701
1702         /*
1703          * Even if we did not try to evict anon pages at all, we want to
1704          * rebalance the anon lru active/inactive ratio.
1705          */
1706         if (inactive_anon_is_low(zone, sc) && nr_swap_pages > 0)
1707                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1708
1709         throttle_vm_writeout(sc->gfp_mask);
1710 }
1711
1712 /*
1713  * This is the direct reclaim path, for page-allocating processes.  We only
1714  * try to reclaim pages from zones which will satisfy the caller's allocation
1715  * request.
1716  *
1717  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1718  * Because:
1719  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1720  *    allocation or
1721  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1722  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1723  *    zone defense algorithm.
1724  *
1725  * If a zone is deemed to be full of pinned pages then just give it a light
1726  * scan then give up on it.
1727  */
1728 static bool shrink_zones(int priority, struct zonelist *zonelist,
1729                                         struct scan_control *sc)
1730 {
1731         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1732         struct zoneref *z;
1733         struct zone *zone;
1734         bool all_unreclaimable = true;
1735
1736         for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1737                                         sc->nodemask) {
1738                 if (!populated_zone(zone))
1739                         continue;
1740                 /*
1741                  * Take care memory controller reclaiming has small influence
1742                  * to global LRU.
1743                  */
1744                 if (scanning_global_lru(sc)) {
1745                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1746                                 continue;
1747                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1748                                 continue;       /* Let kswapd poll it */
1749                 }
1750
1751                 shrink_zone(priority, zone, sc);
1752                 all_unreclaimable = false;
1753         }
1754         return all_unreclaimable;
1755 }
1756
1757 /*
1758  * This is the main entry point to direct page reclaim.
1759  *
1760  * If a full scan of the inactive list fails to free enough memory then we
1761  * are "out of memory" and something needs to be killed.
1762  *
1763  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1764  * high - the zone may be full of dirty or under-writeback pages, which this
1765  * caller can't do much about.  We kick the writeback threads and take explicit
1766  * naps in the hope that some of these pages can be written.  But if the
1767  * allocating task holds filesystem locks which prevent writeout this might not
1768  * work, and the allocation attempt will fail.
1769  *
1770  * returns:     0, if no pages reclaimed
1771  *              else, the number of pages reclaimed
1772  */
1773 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
1774                                         struct scan_control *sc)
1775 {
1776         int priority;
1777         bool all_unreclaimable;
1778         unsigned long total_scanned = 0;
1779         struct reclaim_state *reclaim_state = current->reclaim_state;
1780         struct zoneref *z;
1781         struct zone *zone;
1782         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1783         unsigned long writeback_threshold;
1784
1785         get_mems_allowed();
1786         delayacct_freepages_start();
1787
1788         if (scanning_global_lru(sc))
1789                 count_vm_event(ALLOCSTALL);
1790
1791         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1792                 sc->nr_scanned = 0;
1793                 if (!priority)
1794                         disable_swap_token();
1795                 all_unreclaimable = shrink_zones(priority, zonelist, sc);
1796                 /*
1797                  * Don't shrink slabs when reclaiming memory from
1798                  * over limit cgroups
1799                  */
1800                 if (scanning_global_lru(sc)) {
1801                         unsigned long lru_pages = 0;
1802                         for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1803                                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1804                                         continue;
1805
1806                                 lru_pages += zone_reclaimable_pages(zone);
1807                         }
1808
1809                         shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
1810                         if (reclaim_state) {
1811                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
1812                                 reclaim_state->reclaimed_slab = 0;
1813                         }
1814                 }
1815                 total_scanned += sc->nr_scanned;
1816                 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
1817                         goto out;
1818
1819                 /*
1820                  * Try to write back as many pages as we just scanned.  This
1821                  * tends to cause slow streaming writers to write data to the
1822                  * disk smoothly, at the dirtying rate, which is nice.   But
1823                  * that's undesirable in laptop mode, where we *want* lumpy
1824                  * writeout.  So in laptop mode, write out the whole world.
1825                  */
1826                 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
1827                 if (total_scanned > writeback_threshold) {
1828                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
1829                         sc->may_writepage = 1;
1830                 }
1831
1832                 /* Take a nap, wait for some writeback to complete */
1833                 if (!sc->hibernation_mode && sc->nr_scanned &&
1834                     priority < DEF_PRIORITY - 2)
1835                         congestion_wait(BLK_RW_ASYNC, HZ/10);
1836         }
1837
1838 out:
1839         /*
1840          * Now that we've scanned all the zones at this priority level, note
1841          * that level within the zone so that the next thread which performs
1842          * scanning of this zone will immediately start out at this priority
1843          * level.  This affects only the decision whether or not to bring
1844          * mapped pages onto the inactive list.
1845          */
1846         if (priority < 0)
1847                 priority = 0;
1848
1849         delayacct_freepages_end();
1850         put_mems_allowed();
1851
1852         if (sc->nr_reclaimed)
1853                 return sc->nr_reclaimed;
1854
1855         /* top priority shrink_zones still had more to do? don't OOM, then */
1856         if (scanning_global_lru(sc) && !all_unreclaimable)
1857                 return 1;
1858
1859         return 0;
1860 }
1861
1862 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
1863                                 gfp_t gfp_mask, nodemask_t *nodemask)
1864 {
1865         unsigned long nr_reclaimed;
1866         struct scan_control sc = {
1867                 .gfp_mask = gfp_mask,
1868                 .may_writepage = !laptop_mode,
1869                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
1870                 .may_unmap = 1,
1871                 .may_swap = 1,
1872                 .swappiness = vm_swappiness,
1873                 .order = order,
1874                 .mem_cgroup = NULL,
1875                 .nodemask = nodemask,
1876         };
1877
1878         trace_mm_vmscan_direct_reclaim_begin(order,
1879                                 sc.may_writepage,
1880                                 gfp_mask);
1881
1882         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
1883
1884         trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
1885
1886         return nr_reclaimed;
1887 }
1888
1889 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
1890
1891 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
1892                                                 gfp_t gfp_mask, bool noswap,
1893                                                 unsigned int swappiness,
1894                                                 struct zone *zone, int nid)
1895 {
1896         struct scan_control sc = {
1897                 .may_writepage = !laptop_mode,
1898                 .may_unmap = 1,
1899                 .may_swap = !noswap,
1900                 .swappiness = swappiness,
1901                 .order = 0,
1902                 .mem_cgroup = mem,
1903         };
1904         nodemask_t nm  = nodemask_of_node(nid);
1905
1906         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1907                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1908         sc.nodemask = &nm;
1909         sc.nr_reclaimed = 0;
1910         sc.nr_scanned = 0;
1911         /*
1912          * NOTE: Although we can get the priority field, using it
1913          * here is not a good idea, since it limits the pages we can scan.
1914          * if we don't reclaim here, the shrink_zone from balance_pgdat
1915          * will pick up pages from other mem cgroup's as well. We hack
1916          * the priority and make it zero.
1917          */
1918         shrink_zone(0, zone, &sc);
1919         return sc.nr_reclaimed;
1920 }
1921
1922 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
1923                                            gfp_t gfp_mask,
1924                                            bool noswap,
1925                                            unsigned int swappiness)
1926 {
1927         struct zonelist *zonelist;
1928         struct scan_control sc = {
1929                 .may_writepage = !laptop_mode,
1930                 .may_unmap = 1,
1931                 .may_swap = !noswap,
1932                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
1933                 .swappiness = swappiness,
1934                 .order = 0,
1935                 .mem_cgroup = mem_cont,
1936                 .nodemask = NULL, /* we don't care the placement */
1937         };
1938
1939         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1940                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1941         zonelist = NODE_DATA(numa_node_id())->node_zonelists;
1942         return do_try_to_free_pages(zonelist, &sc);
1943 }
1944 #endif
1945
1946 /* is kswapd sleeping prematurely? */
1947 static int sleeping_prematurely(pg_data_t *pgdat, int order, long remaining)
1948 {
1949         int i;
1950
1951         /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
1952         if (remaining)
1953                 return 1;
1954
1955         /* If after HZ/10, a zone is below the high mark, it's premature */
1956         for (i = 0; i < pgdat->nr_zones; i++) {
1957                 struct zone *zone = pgdat->node_zones + i;
1958
1959                 if (!populated_zone(zone))
1960                         continue;
1961
1962                 if (zone->all_unreclaimable)
1963                         continue;
1964
1965                 if (!zone_watermark_ok(zone, order, high_wmark_pages(zone),
1966                                                                 0, 0))
1967                         return 1;
1968         }
1969
1970         return 0;
1971 }
1972
1973 /*
1974  * For kswapd, balance_pgdat() will work across all this node's zones until
1975  * they are all at high_wmark_pages(zone).
1976  *
1977  * Returns the number of pages which were actually freed.
1978  *
1979  * There is special handling here for zones which are full of pinned pages.
1980  * This can happen if the pages are all mlocked, or if they are all used by
1981  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1982  * What we do is to detect the case where all pages in the zone have been
1983  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1984  * dead and from now on, only perform a short scan.  Basically we're polling
1985  * the zone for when the problem goes away.
1986  *
1987  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
1988  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
1989  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
1990  * lower zones regardless of the number of free pages in the lower zones. This
1991  * interoperates with the page allocator fallback scheme to ensure that aging
1992  * of pages is balanced across the zones.
1993  */
1994 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
1995 {
1996         int all_zones_ok;
1997         int priority;
1998         int i;
1999         unsigned long total_scanned;
2000         struct reclaim_state *reclaim_state = current->reclaim_state;
2001         struct scan_control sc = {
2002                 .gfp_mask = GFP_KERNEL,
2003                 .may_unmap = 1,
2004                 .may_swap = 1,
2005                 /*
2006                  * kswapd doesn't want to be bailed out while reclaim. because
2007                  * we want to put equal scanning pressure on each zone.
2008                  */
2009                 .nr_to_reclaim = ULONG_MAX,
2010                 .swappiness = vm_swappiness,
2011                 .order = order,
2012                 .mem_cgroup = NULL,
2013         };
2014 loop_again:
2015         total_scanned = 0;
2016         sc.nr_reclaimed = 0;
2017         sc.may_writepage = !laptop_mode;
2018         count_vm_event(PAGEOUTRUN);
2019
2020         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2021                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
2022                 unsigned long lru_pages = 0;
2023                 int has_under_min_watermark_zone = 0;
2024
2025                 /* The swap token gets in the way of swapout... */
2026                 if (!priority)
2027                         disable_swap_token();
2028
2029                 all_zones_ok = 1;
2030
2031                 /*
2032                  * Scan in the highmem->dma direction for the highest
2033                  * zone which needs scanning
2034                  */
2035                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2036                         struct zone *zone = pgdat->node_zones + i;
2037
2038                         if (!populated_zone(zone))
2039                                 continue;
2040
2041                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2042                                 continue;
2043
2044                         /*
2045                          * Do some background aging of the anon list, to give
2046                          * pages a chance to be referenced before reclaiming.
2047                          */
2048                         if (inactive_anon_is_low(zone, &sc))
2049                                 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2050                                                         &sc, priority, 0);
2051
2052                         if (!zone_watermark_ok(zone, order,
2053                                         high_wmark_pages(zone), 0, 0)) {
2054                                 end_zone = i;
2055                                 break;
2056                         }
2057                 }
2058                 if (i < 0)
2059                         goto out;
2060
2061                 for (i = 0; i <= end_zone; i++) {
2062                         struct zone *zone = pgdat->node_zones + i;
2063
2064                         lru_pages += zone_reclaimable_pages(zone);
2065                 }
2066
2067                 /*
2068                  * Now scan the zone in the dma->highmem direction, stopping
2069                  * at the last zone which needs scanning.
2070                  *
2071                  * We do this because the page allocator works in the opposite
2072                  * direction.  This prevents the page allocator from allocating
2073                  * pages behind kswapd's direction of progress, which would
2074                  * cause too much scanning of the lower zones.
2075                  */
2076                 for (i = 0; i <= end_zone; i++) {
2077                         struct zone *zone = pgdat->node_zones + i;
2078                         int nr_slab;
2079                         int nid, zid;
2080
2081                         if (!populated_zone(zone))
2082                                 continue;
2083
2084                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2085                                 continue;
2086
2087                         sc.nr_scanned = 0;
2088
2089                         nid = pgdat->node_id;
2090                         zid = zone_idx(zone);
2091                         /*
2092                          * Call soft limit reclaim before calling shrink_zone.
2093                          * For now we ignore the return value
2094                          */
2095                         mem_cgroup_soft_limit_reclaim(zone, order, sc.gfp_mask,
2096                                                         nid, zid);
2097                         /*
2098                          * We put equal pressure on every zone, unless one
2099                          * zone has way too many pages free already.
2100                          */
2101                         if (!zone_watermark_ok(zone, order,
2102                                         8*high_wmark_pages(zone), end_zone, 0))
2103                                 shrink_zone(priority, zone, &sc);
2104                         reclaim_state->reclaimed_slab = 0;
2105                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
2106                                                 lru_pages);
2107                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2108                         total_scanned += sc.nr_scanned;
2109                         if (zone->all_unreclaimable)
2110                                 continue;
2111                         if (nr_slab == 0 &&
2112                             zone->pages_scanned >= (zone_reclaimable_pages(zone) * 6))
2113                                 zone->all_unreclaimable = 1;
2114                         /*
2115                          * If we've done a decent amount of scanning and
2116                          * the reclaim ratio is low, start doing writepage
2117                          * even in laptop mode
2118                          */
2119                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2120                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2121                                 sc.may_writepage = 1;
2122
2123                         if (!zone_watermark_ok(zone, order,
2124                                         high_wmark_pages(zone), end_zone, 0)) {
2125                                 all_zones_ok = 0;
2126                                 /*
2127                                  * We are still under min water mark.  This
2128                                  * means that we have a GFP_ATOMIC allocation
2129                                  * failure risk. Hurry up!
2130                                  */
2131                                 if (!zone_watermark_ok(zone, order,
2132                                             min_wmark_pages(zone), end_zone, 0))
2133                                         has_under_min_watermark_zone = 1;
2134                         }
2135
2136                 }
2137                 if (all_zones_ok)
2138                         break;          /* kswapd: all done */
2139                 /*
2140                  * OK, kswapd is getting into trouble.  Take a nap, then take
2141                  * another pass across the zones.
2142                  */
2143                 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2144                         if (has_under_min_watermark_zone)
2145                                 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2146                         else
2147                                 congestion_wait(BLK_RW_ASYNC, HZ/10);
2148                 }
2149
2150                 /*
2151                  * We do this so kswapd doesn't build up large priorities for
2152                  * example when it is freeing in parallel with allocators. It
2153                  * matches the direct reclaim path behaviour in terms of impact
2154                  * on zone->*_priority.
2155                  */
2156                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2157                         break;
2158         }
2159 out:
2160         if (!all_zones_ok) {
2161                 cond_resched();
2162
2163                 try_to_freeze();
2164
2165                 /*
2166                  * Fragmentation may mean that the system cannot be
2167                  * rebalanced for high-order allocations in all zones.
2168                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2169                  * it means the zones have been fully scanned and are still
2170                  * not balanced. For high-order allocations, there is
2171                  * little point trying all over again as kswapd may
2172                  * infinite loop.
2173                  *
2174                  * Instead, recheck all watermarks at order-0 as they
2175                  * are the most important. If watermarks are ok, kswapd will go
2176                  * back to sleep. High-order users can still perform direct
2177                  * reclaim if they wish.
2178                  */
2179                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2180                         order = sc.order = 0;
2181
2182                 goto loop_again;
2183         }
2184
2185         return sc.nr_reclaimed;
2186 }
2187
2188 /*
2189  * The background pageout daemon, started as a kernel thread
2190  * from the init process.
2191  *
2192  * This basically trickles out pages so that we have _some_
2193  * free memory available even if there is no other activity
2194  * that frees anything up. This is needed for things like routing
2195  * etc, where we otherwise might have all activity going on in
2196  * asynchronous contexts that cannot page things out.
2197  *
2198  * If there are applications that are active memory-allocators
2199  * (most normal use), this basically shouldn't matter.
2200  */
2201 static int kswapd(void *p)
2202 {
2203         unsigned long order;
2204         pg_data_t *pgdat = (pg_data_t*)p;
2205         struct task_struct *tsk = current;
2206         DEFINE_WAIT(wait);
2207         struct reclaim_state reclaim_state = {
2208                 .reclaimed_slab = 0,
2209         };
2210         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2211
2212         lockdep_set_current_reclaim_state(GFP_KERNEL);
2213
2214         if (!cpumask_empty(cpumask))
2215                 set_cpus_allowed_ptr(tsk, cpumask);
2216         current->reclaim_state = &reclaim_state;
2217
2218         /*
2219          * Tell the memory management that we're a "memory allocator",
2220          * and that if we need more memory we should get access to it
2221          * regardless (see "__alloc_pages()"). "kswapd" should
2222          * never get caught in the normal page freeing logic.
2223          *
2224          * (Kswapd normally doesn't need memory anyway, but sometimes
2225          * you need a small amount of memory in order to be able to
2226          * page out something else, and this flag essentially protects
2227          * us from recursively trying to free more memory as we're
2228          * trying to free the first piece of memory in the first place).
2229          */
2230         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2231         set_freezable();
2232
2233         order = 0;
2234         for ( ; ; ) {
2235                 unsigned long new_order;
2236                 int ret;
2237
2238                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2239                 new_order = pgdat->kswapd_max_order;
2240                 pgdat->kswapd_max_order = 0;
2241                 if (order < new_order) {
2242                         /*
2243                          * Don't sleep if someone wants a larger 'order'
2244                          * allocation
2245                          */
2246                         order = new_order;
2247                 } else {
2248                         if (!freezing(current) && !kthread_should_stop()) {
2249                                 long remaining = 0;
2250
2251                                 /* Try to sleep for a short interval */
2252                                 if (!sleeping_prematurely(pgdat, order, remaining)) {
2253                                         remaining = schedule_timeout(HZ/10);
2254                                         finish_wait(&pgdat->kswapd_wait, &wait);
2255                                         prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2256                                 }
2257
2258                                 /*
2259                                  * After a short sleep, check if it was a
2260                                  * premature sleep. If not, then go fully
2261                                  * to sleep until explicitly woken up
2262                                  */
2263                                 if (!sleeping_prematurely(pgdat, order, remaining)) {
2264                                         trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2265                                         schedule();
2266                                 } else {
2267                                         if (remaining)
2268                                                 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2269                                         else
2270                                                 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2271                                 }
2272                         }
2273
2274                         order = pgdat->kswapd_max_order;
2275                 }
2276                 finish_wait(&pgdat->kswapd_wait, &wait);
2277
2278                 ret = try_to_freeze();
2279                 if (kthread_should_stop())
2280                         break;
2281
2282                 /*
2283                  * We can speed up thawing tasks if we don't call balance_pgdat
2284                  * after returning from the refrigerator
2285                  */
2286                 if (!ret) {
2287                         trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2288                         balance_pgdat(pgdat, order);
2289                 }
2290         }
2291         return 0;
2292 }
2293
2294 /*
2295  * A zone is low on free memory, so wake its kswapd task to service it.
2296  */
2297 void wakeup_kswapd(struct zone *zone, int order)
2298 {
2299         pg_data_t *pgdat;
2300
2301         if (!populated_zone(zone))
2302                 return;
2303
2304         pgdat = zone->zone_pgdat;
2305         if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
2306                 return;
2307         if (pgdat->kswapd_max_order < order)
2308                 pgdat->kswapd_max_order = order;
2309         trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
2310         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2311                 return;
2312         if (!waitqueue_active(&pgdat->kswapd_wait))
2313                 return;
2314         wake_up_interruptible(&pgdat->kswapd_wait);
2315 }
2316
2317 /*
2318  * The reclaimable count would be mostly accurate.
2319  * The less reclaimable pages may be
2320  * - mlocked pages, which will be moved to unevictable list when encountered
2321  * - mapped pages, which may require several travels to be reclaimed
2322  * - dirty pages, which is not "instantly" reclaimable
2323  */
2324 unsigned long global_reclaimable_pages(void)
2325 {
2326         int nr;
2327
2328         nr = global_page_state(NR_ACTIVE_FILE) +
2329              global_page_state(NR_INACTIVE_FILE);
2330
2331         if (nr_swap_pages > 0)
2332                 nr += global_page_state(NR_ACTIVE_ANON) +
2333                       global_page_state(NR_INACTIVE_ANON);
2334
2335         return nr;
2336 }
2337
2338 unsigned long zone_reclaimable_pages(struct zone *zone)
2339 {
2340         int nr;
2341
2342         nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2343              zone_page_state(zone, NR_INACTIVE_FILE);
2344
2345         if (nr_swap_pages > 0)
2346                 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2347                       zone_page_state(zone, NR_INACTIVE_ANON);
2348
2349         return nr;
2350 }
2351
2352 #ifdef CONFIG_HIBERNATION
2353 /*
2354  * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2355  * freed pages.
2356  *
2357  * Rather than trying to age LRUs the aim is to preserve the overall
2358  * LRU order by reclaiming preferentially
2359  * inactive > active > active referenced > active mapped
2360  */
2361 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
2362 {
2363         struct reclaim_state reclaim_state;
2364         struct scan_control sc = {
2365                 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2366                 .may_swap = 1,
2367                 .may_unmap = 1,
2368                 .may_writepage = 1,
2369                 .nr_to_reclaim = nr_to_reclaim,
2370                 .hibernation_mode = 1,
2371                 .swappiness = vm_swappiness,
2372                 .order = 0,
2373         };
2374         struct zonelist * zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2375         struct task_struct *p = current;
2376         unsigned long nr_reclaimed;
2377
2378         p->flags |= PF_MEMALLOC;
2379         lockdep_set_current_reclaim_state(sc.gfp_mask);
2380         reclaim_state.reclaimed_slab = 0;
2381         p->reclaim_state = &reclaim_state;
2382
2383         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2384
2385         p->reclaim_state = NULL;
2386         lockdep_clear_current_reclaim_state();
2387         p->flags &= ~PF_MEMALLOC;
2388
2389         return nr_reclaimed;
2390 }
2391 #endif /* CONFIG_HIBERNATION */
2392
2393 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2394    not required for correctness.  So if the last cpu in a node goes
2395    away, we get changed to run anywhere: as the first one comes back,
2396    restore their cpu bindings. */
2397 static int __devinit cpu_callback(struct notifier_block *nfb,
2398                                   unsigned long action, void *hcpu)
2399 {
2400         int nid;
2401
2402         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2403                 for_each_node_state(nid, N_HIGH_MEMORY) {
2404                         pg_data_t *pgdat = NODE_DATA(nid);
2405                         const struct cpumask *mask;
2406
2407                         mask = cpumask_of_node(pgdat->node_id);
2408
2409                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2410                                 /* One of our CPUs online: restore mask */
2411                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2412                 }
2413         }
2414         return NOTIFY_OK;
2415 }
2416
2417 /*
2418  * This kswapd start function will be called by init and node-hot-add.
2419  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2420  */
2421 int kswapd_run(int nid)
2422 {
2423         pg_data_t *pgdat = NODE_DATA(nid);
2424         int ret = 0;
2425
2426         if (pgdat->kswapd)
2427                 return 0;
2428
2429         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2430         if (IS_ERR(pgdat->kswapd)) {
2431                 /* failure at boot is fatal */
2432                 BUG_ON(system_state == SYSTEM_BOOTING);
2433                 printk("Failed to start kswapd on node %d\n",nid);
2434                 ret = -1;
2435         }
2436         return ret;
2437 }
2438
2439 /*
2440  * Called by memory hotplug when all memory in a node is offlined.
2441  */
2442 void kswapd_stop(int nid)
2443 {
2444         struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
2445
2446         if (kswapd)
2447                 kthread_stop(kswapd);
2448 }
2449
2450 static int __init kswapd_init(void)
2451 {
2452         int nid;
2453
2454         swap_setup();
2455         for_each_node_state(nid, N_HIGH_MEMORY)
2456                 kswapd_run(nid);
2457         hotcpu_notifier(cpu_callback, 0);
2458         return 0;
2459 }
2460
2461 module_init(kswapd_init)
2462
2463 #ifdef CONFIG_NUMA
2464 /*
2465  * Zone reclaim mode
2466  *
2467  * If non-zero call zone_reclaim when the number of free pages falls below
2468  * the watermarks.
2469  */
2470 int zone_reclaim_mode __read_mostly;
2471
2472 #define RECLAIM_OFF 0
2473 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
2474 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
2475 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
2476
2477 /*
2478  * Priority for ZONE_RECLAIM. This determines the fraction of pages
2479  * of a node considered for each zone_reclaim. 4 scans 1/16th of
2480  * a zone.
2481  */
2482 #define ZONE_RECLAIM_PRIORITY 4
2483
2484 /*
2485  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2486  * occur.
2487  */
2488 int sysctl_min_unmapped_ratio = 1;
2489
2490 /*
2491  * If the number of slab pages in a zone grows beyond this percentage then
2492  * slab reclaim needs to occur.
2493  */
2494 int sysctl_min_slab_ratio = 5;
2495
2496 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
2497 {
2498         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
2499         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
2500                 zone_page_state(zone, NR_ACTIVE_FILE);
2501
2502         /*
2503          * It's possible for there to be more file mapped pages than
2504          * accounted for by the pages on the file LRU lists because
2505          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2506          */
2507         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
2508 }
2509
2510 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
2511 static long zone_pagecache_reclaimable(struct zone *zone)
2512 {
2513         long nr_pagecache_reclaimable;
2514         long delta = 0;
2515
2516         /*
2517          * If RECLAIM_SWAP is set, then all file pages are considered
2518          * potentially reclaimable. Otherwise, we have to worry about
2519          * pages like swapcache and zone_unmapped_file_pages() provides
2520          * a better estimate
2521          */
2522         if (zone_reclaim_mode & RECLAIM_SWAP)
2523                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
2524         else
2525                 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
2526
2527         /* If we can't clean pages, remove dirty pages from consideration */
2528         if (!(zone_reclaim_mode & RECLAIM_WRITE))
2529                 delta += zone_page_state(zone, NR_FILE_DIRTY);
2530
2531         /* Watch for any possible underflows due to delta */
2532         if (unlikely(delta > nr_pagecache_reclaimable))
2533                 delta = nr_pagecache_reclaimable;
2534
2535         return nr_pagecache_reclaimable - delta;
2536 }
2537
2538 /*
2539  * Try to free up some pages from this zone through reclaim.
2540  */
2541 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2542 {
2543         /* Minimum pages needed in order to stay on node */
2544         const unsigned long nr_pages = 1 << order;
2545         struct task_struct *p = current;
2546         struct reclaim_state reclaim_state;
2547         int priority;
2548         struct scan_control sc = {
2549                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2550                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2551                 .may_swap = 1,
2552                 .nr_to_reclaim = max_t(unsigned long, nr_pages,
2553                                        SWAP_CLUSTER_MAX),
2554                 .gfp_mask = gfp_mask,
2555                 .swappiness = vm_swappiness,
2556                 .order = order,
2557         };
2558         unsigned long slab_reclaimable;
2559
2560         cond_resched();
2561         /*
2562          * We need to be able to allocate from the reserves for RECLAIM_SWAP
2563          * and we also need to be able to write out pages for RECLAIM_WRITE
2564          * and RECLAIM_SWAP.
2565          */
2566         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2567         lockdep_set_current_reclaim_state(gfp_mask);
2568         reclaim_state.reclaimed_slab = 0;
2569         p->reclaim_state = &reclaim_state;
2570
2571         if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
2572                 /*
2573                  * Free memory by calling shrink zone with increasing
2574                  * priorities until we have enough memory freed.
2575                  */
2576                 priority = ZONE_RECLAIM_PRIORITY;
2577                 do {
2578                         shrink_zone(priority, zone, &sc);
2579                         priority--;
2580                 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
2581         }
2582
2583         slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2584         if (slab_reclaimable > zone->min_slab_pages) {
2585                 /*
2586                  * shrink_slab() does not currently allow us to determine how
2587                  * many pages were freed in this zone. So we take the current
2588                  * number of slab pages and shake the slab until it is reduced
2589                  * by the same nr_pages that we used for reclaiming unmapped
2590                  * pages.
2591                  *
2592                  * Note that shrink_slab will free memory on all zones and may
2593                  * take a long time.
2594                  */
2595                 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
2596                         zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
2597                                 slab_reclaimable - nr_pages)
2598                         ;
2599
2600                 /*
2601                  * Update nr_reclaimed by the number of slab pages we
2602                  * reclaimed from this zone.
2603                  */
2604                 sc.nr_reclaimed += slab_reclaimable -
2605                         zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2606         }
2607
2608         p->reclaim_state = NULL;
2609         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
2610         lockdep_clear_current_reclaim_state();
2611         return sc.nr_reclaimed >= nr_pages;
2612 }
2613
2614 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2615 {
2616         int node_id;
2617         int ret;
2618
2619         /*
2620          * Zone reclaim reclaims unmapped file backed pages and
2621          * slab pages if we are over the defined limits.
2622          *
2623          * A small portion of unmapped file backed pages is needed for
2624          * file I/O otherwise pages read by file I/O will be immediately
2625          * thrown out if the zone is overallocated. So we do not reclaim
2626          * if less than a specified percentage of the zone is used by
2627          * unmapped file backed pages.
2628          */
2629         if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
2630             zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
2631                 return ZONE_RECLAIM_FULL;
2632
2633         if (zone->all_unreclaimable)
2634                 return ZONE_RECLAIM_FULL;
2635
2636         /*
2637          * Do not scan if the allocation should not be delayed.
2638          */
2639         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
2640                 return ZONE_RECLAIM_NOSCAN;
2641
2642         /*
2643          * Only run zone reclaim on the local zone or on zones that do not
2644          * have associated processors. This will favor the local processor
2645          * over remote processors and spread off node memory allocations
2646          * as wide as possible.
2647          */
2648         node_id = zone_to_nid(zone);
2649         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
2650                 return ZONE_RECLAIM_NOSCAN;
2651
2652         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2653                 return ZONE_RECLAIM_NOSCAN;
2654
2655         ret = __zone_reclaim(zone, gfp_mask, order);
2656         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2657
2658         if (!ret)
2659                 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
2660
2661         return ret;
2662 }
2663 #endif
2664
2665 /*
2666  * page_evictable - test whether a page is evictable
2667  * @page: the page to test
2668  * @vma: the VMA in which the page is or will be mapped, may be NULL
2669  *
2670  * Test whether page is evictable--i.e., should be placed on active/inactive
2671  * lists vs unevictable list.  The vma argument is !NULL when called from the
2672  * fault path to determine how to instantate a new page.
2673  *
2674  * Reasons page might not be evictable:
2675  * (1) page's mapping marked unevictable
2676  * (2) page is part of an mlocked VMA
2677  *
2678  */
2679 int page_evictable(struct page *page, struct vm_area_struct *vma)
2680 {
2681
2682         if (mapping_unevictable(page_mapping(page)))
2683                 return 0;
2684
2685         if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2686                 return 0;
2687
2688         return 1;
2689 }
2690
2691 /**
2692  * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2693  * @page: page to check evictability and move to appropriate lru list
2694  * @zone: zone page is in
2695  *
2696  * Checks a page for evictability and moves the page to the appropriate
2697  * zone lru list.
2698  *
2699  * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2700  * have PageUnevictable set.
2701  */
2702 static void check_move_unevictable_page(struct page *page, struct zone *zone)
2703 {
2704         VM_BUG_ON(PageActive(page));
2705
2706 retry:
2707         ClearPageUnevictable(page);
2708         if (page_evictable(page, NULL)) {
2709                 enum lru_list l = page_lru_base_type(page);
2710
2711                 __dec_zone_state(zone, NR_UNEVICTABLE);
2712                 list_move(&page->lru, &zone->lru[l].list);
2713                 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
2714                 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
2715                 __count_vm_event(UNEVICTABLE_PGRESCUED);
2716         } else {
2717                 /*
2718                  * rotate unevictable list
2719                  */
2720                 SetPageUnevictable(page);
2721                 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
2722                 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
2723                 if (page_evictable(page, NULL))
2724                         goto retry;
2725         }
2726 }
2727
2728 /**
2729  * scan_mapping_unevictable_pages - scan an address space for evictable pages
2730  * @mapping: struct address_space to scan for evictable pages
2731  *
2732  * Scan all pages in mapping.  Check unevictable pages for
2733  * evictability and move them to the appropriate zone lru list.
2734  */
2735 void scan_mapping_unevictable_pages(struct address_space *mapping)
2736 {
2737         pgoff_t next = 0;
2738         pgoff_t end   = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
2739                          PAGE_CACHE_SHIFT;
2740         struct zone *zone;
2741         struct pagevec pvec;
2742
2743         if (mapping->nrpages == 0)
2744                 return;
2745
2746         pagevec_init(&pvec, 0);
2747         while (next < end &&
2748                 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
2749                 int i;
2750                 int pg_scanned = 0;
2751
2752                 zone = NULL;
2753
2754                 for (i = 0; i < pagevec_count(&pvec); i++) {
2755                         struct page *page = pvec.pages[i];
2756                         pgoff_t page_index = page->index;
2757                         struct zone *pagezone = page_zone(page);
2758
2759                         pg_scanned++;
2760                         if (page_index > next)
2761                                 next = page_index;
2762                         next++;
2763
2764                         if (pagezone != zone) {
2765                                 if (zone)
2766                                         spin_unlock_irq(&zone->lru_lock);
2767                                 zone = pagezone;
2768                                 spin_lock_irq(&zone->lru_lock);
2769                         }
2770
2771                         if (PageLRU(page) && PageUnevictable(page))
2772                                 check_move_unevictable_page(page, zone);
2773                 }
2774                 if (zone)
2775                         spin_unlock_irq(&zone->lru_lock);
2776                 pagevec_release(&pvec);
2777
2778                 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
2779         }
2780
2781 }
2782
2783 /**
2784  * scan_zone_unevictable_pages - check unevictable list for evictable pages
2785  * @zone - zone of which to scan the unevictable list
2786  *
2787  * Scan @zone's unevictable LRU lists to check for pages that have become
2788  * evictable.  Move those that have to @zone's inactive list where they
2789  * become candidates for reclaim, unless shrink_inactive_zone() decides
2790  * to reactivate them.  Pages that are still unevictable are rotated
2791  * back onto @zone's unevictable list.
2792  */
2793 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
2794 static void scan_zone_unevictable_pages(struct zone *zone)
2795 {
2796         struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
2797         unsigned long scan;
2798         unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
2799
2800         while (nr_to_scan > 0) {
2801                 unsigned long batch_size = min(nr_to_scan,
2802                                                 SCAN_UNEVICTABLE_BATCH_SIZE);
2803
2804                 spin_lock_irq(&zone->lru_lock);
2805                 for (scan = 0;  scan < batch_size; scan++) {
2806                         struct page *page = lru_to_page(l_unevictable);
2807
2808                         if (!trylock_page(page))
2809                                 continue;
2810
2811                         prefetchw_prev_lru_page(page, l_unevictable, flags);
2812
2813                         if (likely(PageLRU(page) && PageUnevictable(page)))
2814                                 check_move_unevictable_page(page, zone);
2815
2816                         unlock_page(page);
2817                 }
2818                 spin_unlock_irq(&zone->lru_lock);
2819
2820                 nr_to_scan -= batch_size;
2821         }
2822 }
2823
2824
2825 /**
2826  * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
2827  *
2828  * A really big hammer:  scan all zones' unevictable LRU lists to check for
2829  * pages that have become evictable.  Move those back to the zones'
2830  * inactive list where they become candidates for reclaim.
2831  * This occurs when, e.g., we have unswappable pages on the unevictable lists,
2832  * and we add swap to the system.  As such, it runs in the context of a task
2833  * that has possibly/probably made some previously unevictable pages
2834  * evictable.
2835  */
2836 static void scan_all_zones_unevictable_pages(void)
2837 {
2838         struct zone *zone;
2839
2840         for_each_zone(zone) {
2841                 scan_zone_unevictable_pages(zone);
2842         }
2843 }
2844
2845 /*
2846  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
2847  * all nodes' unevictable lists for evictable pages
2848  */
2849 unsigned long scan_unevictable_pages;
2850
2851 int scan_unevictable_handler(struct ctl_table *table, int write,
2852                            void __user *buffer,
2853                            size_t *length, loff_t *ppos)
2854 {
2855         proc_doulongvec_minmax(table, write, buffer, length, ppos);
2856
2857         if (write && *(unsigned long *)table->data)
2858                 scan_all_zones_unevictable_pages();
2859
2860         scan_unevictable_pages = 0;
2861         return 0;
2862 }
2863
2864 /*
2865  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
2866  * a specified node's per zone unevictable lists for evictable pages.
2867  */
2868
2869 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
2870                                           struct sysdev_attribute *attr,
2871                                           char *buf)
2872 {
2873         return sprintf(buf, "0\n");     /* always zero; should fit... */
2874 }
2875
2876 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
2877                                            struct sysdev_attribute *attr,
2878                                         const char *buf, size_t count)
2879 {
2880         struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
2881         struct zone *zone;
2882         unsigned long res;
2883         unsigned long req = strict_strtoul(buf, 10, &res);
2884
2885         if (!req)
2886                 return 1;       /* zero is no-op */
2887
2888         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2889                 if (!populated_zone(zone))
2890                         continue;
2891                 scan_zone_unevictable_pages(zone);
2892         }
2893         return 1;
2894 }
2895
2896
2897 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
2898                         read_scan_unevictable_node,
2899                         write_scan_unevictable_node);
2900
2901 int scan_unevictable_register_node(struct node *node)
2902 {
2903         return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
2904 }
2905
2906 void scan_unevictable_unregister_node(struct node *node)
2907 {
2908         sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
2909 }
2910