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