4a6dccb57586f73596dd8c00311acb229274eee0
[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 continuous memory rather than reclaim
83          * enough amount of memory. i.e, 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                         .for_reclaim = 1,
380                 };
381
382                 SetPageReclaim(page);
383                 res = mapping->a_ops->writepage(page, &wbc);
384                 if (res < 0)
385                         handle_write_error(mapping, page, res);
386                 if (res == AOP_WRITEPAGE_ACTIVATE) {
387                         ClearPageReclaim(page);
388                         return PAGE_ACTIVATE;
389                 }
390
391                 /*
392                  * Wait on writeback if requested to. This happens when
393                  * direct reclaiming a large contiguous area and the
394                  * first attempt to free a range of pages fails.
395                  */
396                 if (PageWriteback(page) && sync_writeback == PAGEOUT_IO_SYNC)
397                         wait_on_page_writeback(page);
398
399                 if (!PageWriteback(page)) {
400                         /* synchronous write or broken a_ops? */
401                         ClearPageReclaim(page);
402                 }
403                 trace_mm_vmscan_writepage(page,
404                         trace_reclaim_flags(page, sync_writeback));
405                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
406                 return PAGE_SUCCESS;
407         }
408
409         return PAGE_CLEAN;
410 }
411
412 /*
413  * Same as remove_mapping, but if the page is removed from the mapping, it
414  * gets returned with a refcount of 0.
415  */
416 static int __remove_mapping(struct address_space *mapping, struct page *page)
417 {
418         BUG_ON(!PageLocked(page));
419         BUG_ON(mapping != page_mapping(page));
420
421         spin_lock_irq(&mapping->tree_lock);
422         /*
423          * The non racy check for a busy page.
424          *
425          * Must be careful with the order of the tests. When someone has
426          * a ref to the page, it may be possible that they dirty it then
427          * drop the reference. So if PageDirty is tested before page_count
428          * here, then the following race may occur:
429          *
430          * get_user_pages(&page);
431          * [user mapping goes away]
432          * write_to(page);
433          *                              !PageDirty(page)    [good]
434          * SetPageDirty(page);
435          * put_page(page);
436          *                              !page_count(page)   [good, discard it]
437          *
438          * [oops, our write_to data is lost]
439          *
440          * Reversing the order of the tests ensures such a situation cannot
441          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
442          * load is not satisfied before that of page->_count.
443          *
444          * Note that if SetPageDirty is always performed via set_page_dirty,
445          * and thus under tree_lock, then this ordering is not required.
446          */
447         if (!page_freeze_refs(page, 2))
448                 goto cannot_free;
449         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
450         if (unlikely(PageDirty(page))) {
451                 page_unfreeze_refs(page, 2);
452                 goto cannot_free;
453         }
454
455         if (PageSwapCache(page)) {
456                 swp_entry_t swap = { .val = page_private(page) };
457                 __delete_from_swap_cache(page);
458                 spin_unlock_irq(&mapping->tree_lock);
459                 swapcache_free(swap, page);
460         } else {
461                 __remove_from_page_cache(page);
462                 spin_unlock_irq(&mapping->tree_lock);
463                 mem_cgroup_uncharge_cache_page(page);
464         }
465
466         return 1;
467
468 cannot_free:
469         spin_unlock_irq(&mapping->tree_lock);
470         return 0;
471 }
472
473 /*
474  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
475  * someone else has a ref on the page, abort and return 0.  If it was
476  * successfully detached, return 1.  Assumes the caller has a single ref on
477  * this page.
478  */
479 int remove_mapping(struct address_space *mapping, struct page *page)
480 {
481         if (__remove_mapping(mapping, page)) {
482                 /*
483                  * Unfreezing the refcount with 1 rather than 2 effectively
484                  * drops the pagecache ref for us without requiring another
485                  * atomic operation.
486                  */
487                 page_unfreeze_refs(page, 1);
488                 return 1;
489         }
490         return 0;
491 }
492
493 /**
494  * putback_lru_page - put previously isolated page onto appropriate LRU list
495  * @page: page to be put back to appropriate lru list
496  *
497  * Add previously isolated @page to appropriate LRU list.
498  * Page may still be unevictable for other reasons.
499  *
500  * lru_lock must not be held, interrupts must be enabled.
501  */
502 void putback_lru_page(struct page *page)
503 {
504         int lru;
505         int active = !!TestClearPageActive(page);
506         int was_unevictable = PageUnevictable(page);
507
508         VM_BUG_ON(PageLRU(page));
509
510 redo:
511         ClearPageUnevictable(page);
512
513         if (page_evictable(page, NULL)) {
514                 /*
515                  * For evictable pages, we can use the cache.
516                  * In event of a race, worst case is we end up with an
517                  * unevictable page on [in]active list.
518                  * We know how to handle that.
519                  */
520                 lru = active + page_lru_base_type(page);
521                 lru_cache_add_lru(page, lru);
522         } else {
523                 /*
524                  * Put unevictable pages directly on zone's unevictable
525                  * list.
526                  */
527                 lru = LRU_UNEVICTABLE;
528                 add_page_to_unevictable_list(page);
529                 /*
530                  * When racing with an mlock clearing (page is
531                  * unlocked), make sure that if the other thread does
532                  * not observe our setting of PG_lru and fails
533                  * isolation, we see PG_mlocked cleared below and move
534                  * the page back to the evictable list.
535                  *
536                  * The other side is TestClearPageMlocked().
537                  */
538                 smp_mb();
539         }
540
541         /*
542          * page's status can change while we move it among lru. If an evictable
543          * page is on unevictable list, it never be freed. To avoid that,
544          * check after we added it to the list, again.
545          */
546         if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
547                 if (!isolate_lru_page(page)) {
548                         put_page(page);
549                         goto redo;
550                 }
551                 /* This means someone else dropped this page from LRU
552                  * So, it will be freed or putback to LRU again. There is
553                  * nothing to do here.
554                  */
555         }
556
557         if (was_unevictable && lru != LRU_UNEVICTABLE)
558                 count_vm_event(UNEVICTABLE_PGRESCUED);
559         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
560                 count_vm_event(UNEVICTABLE_PGCULLED);
561
562         put_page(page);         /* drop ref from isolate */
563 }
564
565 enum page_references {
566         PAGEREF_RECLAIM,
567         PAGEREF_RECLAIM_CLEAN,
568         PAGEREF_KEEP,
569         PAGEREF_ACTIVATE,
570 };
571
572 static enum page_references page_check_references(struct page *page,
573                                                   struct scan_control *sc)
574 {
575         int referenced_ptes, referenced_page;
576         unsigned long vm_flags;
577
578         referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
579         referenced_page = TestClearPageReferenced(page);
580
581         /* Lumpy reclaim - ignore references */
582         if (sc->lumpy_reclaim_mode)
583                 return PAGEREF_RECLAIM;
584
585         /*
586          * Mlock lost the isolation race with us.  Let try_to_unmap()
587          * move the page to the unevictable list.
588          */
589         if (vm_flags & VM_LOCKED)
590                 return PAGEREF_RECLAIM;
591
592         if (referenced_ptes) {
593                 if (PageAnon(page))
594                         return PAGEREF_ACTIVATE;
595                 /*
596                  * All mapped pages start out with page table
597                  * references from the instantiating fault, so we need
598                  * to look twice if a mapped file page is used more
599                  * than once.
600                  *
601                  * Mark it and spare it for another trip around the
602                  * inactive list.  Another page table reference will
603                  * lead to its activation.
604                  *
605                  * Note: the mark is set for activated pages as well
606                  * so that recently deactivated but used pages are
607                  * quickly recovered.
608                  */
609                 SetPageReferenced(page);
610
611                 if (referenced_page)
612                         return PAGEREF_ACTIVATE;
613
614                 return PAGEREF_KEEP;
615         }
616
617         /* Reclaim if clean, defer dirty pages to writeback */
618         if (referenced_page)
619                 return PAGEREF_RECLAIM_CLEAN;
620
621         return PAGEREF_RECLAIM;
622 }
623
624 static noinline_for_stack void free_page_list(struct list_head *free_pages)
625 {
626         struct pagevec freed_pvec;
627         struct page *page, *tmp;
628
629         pagevec_init(&freed_pvec, 1);
630
631         list_for_each_entry_safe(page, tmp, free_pages, lru) {
632                 list_del(&page->lru);
633                 if (!pagevec_add(&freed_pvec, page)) {
634                         __pagevec_free(&freed_pvec);
635                         pagevec_reinit(&freed_pvec);
636                 }
637         }
638
639         pagevec_free(&freed_pvec);
640 }
641
642 /*
643  * shrink_page_list() returns the number of reclaimed pages
644  */
645 static unsigned long shrink_page_list(struct list_head *page_list,
646                                         struct scan_control *sc,
647                                         enum pageout_io sync_writeback)
648 {
649         LIST_HEAD(ret_pages);
650         LIST_HEAD(free_pages);
651         int pgactivate = 0;
652         unsigned long nr_reclaimed = 0;
653
654         cond_resched();
655
656         while (!list_empty(page_list)) {
657                 enum page_references references;
658                 struct address_space *mapping;
659                 struct page *page;
660                 int may_enter_fs;
661
662                 cond_resched();
663
664                 page = lru_to_page(page_list);
665                 list_del(&page->lru);
666
667                 if (!trylock_page(page))
668                         goto keep;
669
670                 VM_BUG_ON(PageActive(page));
671
672                 sc->nr_scanned++;
673
674                 if (unlikely(!page_evictable(page, NULL)))
675                         goto cull_mlocked;
676
677                 if (!sc->may_unmap && page_mapped(page))
678                         goto keep_locked;
679
680                 /* Double the slab pressure for mapped and swapcache pages */
681                 if (page_mapped(page) || PageSwapCache(page))
682                         sc->nr_scanned++;
683
684                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
685                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
686
687                 if (PageWriteback(page)) {
688                         /*
689                          * Synchronous reclaim is performed in two passes,
690                          * first an asynchronous pass over the list to
691                          * start parallel writeback, and a second synchronous
692                          * pass to wait for the IO to complete.  Wait here
693                          * for any page for which writeback has already
694                          * started.
695                          */
696                         if (sync_writeback == PAGEOUT_IO_SYNC && may_enter_fs)
697                                 wait_on_page_writeback(page);
698                         else
699                                 goto keep_locked;
700                 }
701
702                 references = page_check_references(page, sc);
703                 switch (references) {
704                 case PAGEREF_ACTIVATE:
705                         goto activate_locked;
706                 case PAGEREF_KEEP:
707                         goto keep_locked;
708                 case PAGEREF_RECLAIM:
709                 case PAGEREF_RECLAIM_CLEAN:
710                         ; /* try to reclaim the page below */
711                 }
712
713                 /*
714                  * Anonymous process memory has backing store?
715                  * Try to allocate it some swap space here.
716                  */
717                 if (PageAnon(page) && !PageSwapCache(page)) {
718                         if (!(sc->gfp_mask & __GFP_IO))
719                                 goto keep_locked;
720                         if (!add_to_swap(page))
721                                 goto activate_locked;
722                         may_enter_fs = 1;
723                 }
724
725                 mapping = page_mapping(page);
726
727                 /*
728                  * The page is mapped into the page tables of one or more
729                  * processes. Try to unmap it here.
730                  */
731                 if (page_mapped(page) && mapping) {
732                         switch (try_to_unmap(page, TTU_UNMAP)) {
733                         case SWAP_FAIL:
734                                 goto activate_locked;
735                         case SWAP_AGAIN:
736                                 goto keep_locked;
737                         case SWAP_MLOCK:
738                                 goto cull_mlocked;
739                         case SWAP_SUCCESS:
740                                 ; /* try to free the page below */
741                         }
742                 }
743
744                 if (PageDirty(page)) {
745                         if (references == PAGEREF_RECLAIM_CLEAN)
746                                 goto keep_locked;
747                         if (!may_enter_fs)
748                                 goto keep_locked;
749                         if (!sc->may_writepage)
750                                 goto keep_locked;
751
752                         /* Page is dirty, try to write it out here */
753                         switch (pageout(page, mapping, sync_writeback)) {
754                         case PAGE_KEEP:
755                                 goto keep_locked;
756                         case PAGE_ACTIVATE:
757                                 goto activate_locked;
758                         case PAGE_SUCCESS:
759                                 if (PageWriteback(page) || PageDirty(page))
760                                         goto keep;
761                                 /*
762                                  * A synchronous write - probably a ramdisk.  Go
763                                  * ahead and try to reclaim the page.
764                                  */
765                                 if (!trylock_page(page))
766                                         goto keep;
767                                 if (PageDirty(page) || PageWriteback(page))
768                                         goto keep_locked;
769                                 mapping = page_mapping(page);
770                         case PAGE_CLEAN:
771                                 ; /* try to free the page below */
772                         }
773                 }
774
775                 /*
776                  * If the page has buffers, try to free the buffer mappings
777                  * associated with this page. If we succeed we try to free
778                  * the page as well.
779                  *
780                  * We do this even if the page is PageDirty().
781                  * try_to_release_page() does not perform I/O, but it is
782                  * possible for a page to have PageDirty set, but it is actually
783                  * clean (all its buffers are clean).  This happens if the
784                  * buffers were written out directly, with submit_bh(). ext3
785                  * will do this, as well as the blockdev mapping.
786                  * try_to_release_page() will discover that cleanness and will
787                  * drop the buffers and mark the page clean - it can be freed.
788                  *
789                  * Rarely, pages can have buffers and no ->mapping.  These are
790                  * the pages which were not successfully invalidated in
791                  * truncate_complete_page().  We try to drop those buffers here
792                  * and if that worked, and the page is no longer mapped into
793                  * process address space (page_count == 1) it can be freed.
794                  * Otherwise, leave the page on the LRU so it is swappable.
795                  */
796                 if (page_has_private(page)) {
797                         if (!try_to_release_page(page, sc->gfp_mask))
798                                 goto activate_locked;
799                         if (!mapping && page_count(page) == 1) {
800                                 unlock_page(page);
801                                 if (put_page_testzero(page))
802                                         goto free_it;
803                                 else {
804                                         /*
805                                          * rare race with speculative reference.
806                                          * the speculative reference will free
807                                          * this page shortly, so we may
808                                          * increment nr_reclaimed here (and
809                                          * leave it off the LRU).
810                                          */
811                                         nr_reclaimed++;
812                                         continue;
813                                 }
814                         }
815                 }
816
817                 if (!mapping || !__remove_mapping(mapping, page))
818                         goto keep_locked;
819
820                 /*
821                  * At this point, we have no other references and there is
822                  * no way to pick any more up (removed from LRU, removed
823                  * from pagecache). Can use non-atomic bitops now (and
824                  * we obviously don't have to worry about waking up a process
825                  * waiting on the page lock, because there are no references.
826                  */
827                 __clear_page_locked(page);
828 free_it:
829                 nr_reclaimed++;
830
831                 /*
832                  * Is there need to periodically free_page_list? It would
833                  * appear not as the counts should be low
834                  */
835                 list_add(&page->lru, &free_pages);
836                 continue;
837
838 cull_mlocked:
839                 if (PageSwapCache(page))
840                         try_to_free_swap(page);
841                 unlock_page(page);
842                 putback_lru_page(page);
843                 continue;
844
845 activate_locked:
846                 /* Not a candidate for swapping, so reclaim swap space. */
847                 if (PageSwapCache(page) && vm_swap_full())
848                         try_to_free_swap(page);
849                 VM_BUG_ON(PageActive(page));
850                 SetPageActive(page);
851                 pgactivate++;
852 keep_locked:
853                 unlock_page(page);
854 keep:
855                 list_add(&page->lru, &ret_pages);
856                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
857         }
858
859         free_page_list(&free_pages);
860
861         list_splice(&ret_pages, page_list);
862         count_vm_events(PGACTIVATE, pgactivate);
863         return nr_reclaimed;
864 }
865
866 /*
867  * Attempt to remove the specified page from its LRU.  Only take this page
868  * if it is of the appropriate PageActive status.  Pages which are being
869  * freed elsewhere are also ignored.
870  *
871  * page:        page to consider
872  * mode:        one of the LRU isolation modes defined above
873  *
874  * returns 0 on success, -ve errno on failure.
875  */
876 int __isolate_lru_page(struct page *page, int mode, int file)
877 {
878         int ret = -EINVAL;
879
880         /* Only take pages on the LRU. */
881         if (!PageLRU(page))
882                 return ret;
883
884         /*
885          * When checking the active state, we need to be sure we are
886          * dealing with comparible boolean values.  Take the logical not
887          * of each.
888          */
889         if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
890                 return ret;
891
892         if (mode != ISOLATE_BOTH && page_is_file_cache(page) != file)
893                 return ret;
894
895         /*
896          * When this function is being called for lumpy reclaim, we
897          * initially look into all LRU pages, active, inactive and
898          * unevictable; only give shrink_page_list evictable pages.
899          */
900         if (PageUnevictable(page))
901                 return ret;
902
903         ret = -EBUSY;
904
905         if (likely(get_page_unless_zero(page))) {
906                 /*
907                  * Be careful not to clear PageLRU until after we're
908                  * sure the page is not being freed elsewhere -- the
909                  * page release code relies on it.
910                  */
911                 ClearPageLRU(page);
912                 ret = 0;
913         }
914
915         return ret;
916 }
917
918 /*
919  * zone->lru_lock is heavily contended.  Some of the functions that
920  * shrink the lists perform better by taking out a batch of pages
921  * and working on them outside the LRU lock.
922  *
923  * For pagecache intensive workloads, this function is the hottest
924  * spot in the kernel (apart from copy_*_user functions).
925  *
926  * Appropriate locks must be held before calling this function.
927  *
928  * @nr_to_scan: The number of pages to look through on the list.
929  * @src:        The LRU list to pull pages off.
930  * @dst:        The temp list to put pages on to.
931  * @scanned:    The number of pages that were scanned.
932  * @order:      The caller's attempted allocation order
933  * @mode:       One of the LRU isolation modes
934  * @file:       True [1] if isolating file [!anon] pages
935  *
936  * returns how many pages were moved onto *@dst.
937  */
938 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
939                 struct list_head *src, struct list_head *dst,
940                 unsigned long *scanned, int order, int mode, int file)
941 {
942         unsigned long nr_taken = 0;
943         unsigned long nr_lumpy_taken = 0;
944         unsigned long nr_lumpy_dirty = 0;
945         unsigned long nr_lumpy_failed = 0;
946         unsigned long scan;
947
948         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
949                 struct page *page;
950                 unsigned long pfn;
951                 unsigned long end_pfn;
952                 unsigned long page_pfn;
953                 int zone_id;
954
955                 page = lru_to_page(src);
956                 prefetchw_prev_lru_page(page, src, flags);
957
958                 VM_BUG_ON(!PageLRU(page));
959
960                 switch (__isolate_lru_page(page, mode, file)) {
961                 case 0:
962                         list_move(&page->lru, dst);
963                         mem_cgroup_del_lru(page);
964                         nr_taken++;
965                         break;
966
967                 case -EBUSY:
968                         /* else it is being freed elsewhere */
969                         list_move(&page->lru, src);
970                         mem_cgroup_rotate_lru_list(page, page_lru(page));
971                         continue;
972
973                 default:
974                         BUG();
975                 }
976
977                 if (!order)
978                         continue;
979
980                 /*
981                  * Attempt to take all pages in the order aligned region
982                  * surrounding the tag page.  Only take those pages of
983                  * the same active state as that tag page.  We may safely
984                  * round the target page pfn down to the requested order
985                  * as the mem_map is guarenteed valid out to MAX_ORDER,
986                  * where that page is in a different zone we will detect
987                  * it from its zone id and abort this block scan.
988                  */
989                 zone_id = page_zone_id(page);
990                 page_pfn = page_to_pfn(page);
991                 pfn = page_pfn & ~((1 << order) - 1);
992                 end_pfn = pfn + (1 << order);
993                 for (; pfn < end_pfn; pfn++) {
994                         struct page *cursor_page;
995
996                         /* The target page is in the block, ignore it. */
997                         if (unlikely(pfn == page_pfn))
998                                 continue;
999
1000                         /* Avoid holes within the zone. */
1001                         if (unlikely(!pfn_valid_within(pfn)))
1002                                 break;
1003
1004                         cursor_page = pfn_to_page(pfn);
1005
1006                         /* Check that we have not crossed a zone boundary. */
1007                         if (unlikely(page_zone_id(cursor_page) != zone_id))
1008                                 continue;
1009
1010                         /*
1011                          * If we don't have enough swap space, reclaiming of
1012                          * anon page which don't already have a swap slot is
1013                          * pointless.
1014                          */
1015                         if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
1016                                         !PageSwapCache(cursor_page))
1017                                 continue;
1018
1019                         if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1020                                 list_move(&cursor_page->lru, dst);
1021                                 mem_cgroup_del_lru(cursor_page);
1022                                 nr_taken++;
1023                                 nr_lumpy_taken++;
1024                                 if (PageDirty(cursor_page))
1025                                         nr_lumpy_dirty++;
1026                                 scan++;
1027                         } else {
1028                                 if (mode == ISOLATE_BOTH &&
1029                                                 page_count(cursor_page))
1030                                         nr_lumpy_failed++;
1031                         }
1032                 }
1033         }
1034
1035         *scanned = scan;
1036
1037         trace_mm_vmscan_lru_isolate(order,
1038                         nr_to_scan, scan,
1039                         nr_taken,
1040                         nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1041                         mode);
1042         return nr_taken;
1043 }
1044
1045 static unsigned long isolate_pages_global(unsigned long nr,
1046                                         struct list_head *dst,
1047                                         unsigned long *scanned, int order,
1048                                         int mode, struct zone *z,
1049                                         int active, int file)
1050 {
1051         int lru = LRU_BASE;
1052         if (active)
1053                 lru += LRU_ACTIVE;
1054         if (file)
1055                 lru += LRU_FILE;
1056         return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1057                                                                 mode, file);
1058 }
1059
1060 /*
1061  * clear_active_flags() is a helper for shrink_active_list(), clearing
1062  * any active bits from the pages in the list.
1063  */
1064 static unsigned long clear_active_flags(struct list_head *page_list,
1065                                         unsigned int *count)
1066 {
1067         int nr_active = 0;
1068         int lru;
1069         struct page *page;
1070
1071         list_for_each_entry(page, page_list, lru) {
1072                 lru = page_lru_base_type(page);
1073                 if (PageActive(page)) {
1074                         lru += LRU_ACTIVE;
1075                         ClearPageActive(page);
1076                         nr_active++;
1077                 }
1078                 if (count)
1079                         count[lru]++;
1080         }
1081
1082         return nr_active;
1083 }
1084
1085 /**
1086  * isolate_lru_page - tries to isolate a page from its LRU list
1087  * @page: page to isolate from its LRU list
1088  *
1089  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1090  * vmstat statistic corresponding to whatever LRU list the page was on.
1091  *
1092  * Returns 0 if the page was removed from an LRU list.
1093  * Returns -EBUSY if the page was not on an LRU list.
1094  *
1095  * The returned page will have PageLRU() cleared.  If it was found on
1096  * the active list, it will have PageActive set.  If it was found on
1097  * the unevictable list, it will have the PageUnevictable bit set. That flag
1098  * may need to be cleared by the caller before letting the page go.
1099  *
1100  * The vmstat statistic corresponding to the list on which the page was
1101  * found will be decremented.
1102  *
1103  * Restrictions:
1104  * (1) Must be called with an elevated refcount on the page. This is a
1105  *     fundamentnal difference from isolate_lru_pages (which is called
1106  *     without a stable reference).
1107  * (2) the lru_lock must not be held.
1108  * (3) interrupts must be enabled.
1109  */
1110 int isolate_lru_page(struct page *page)
1111 {
1112         int ret = -EBUSY;
1113
1114         if (PageLRU(page)) {
1115                 struct zone *zone = page_zone(page);
1116
1117                 spin_lock_irq(&zone->lru_lock);
1118                 if (PageLRU(page) && get_page_unless_zero(page)) {
1119                         int lru = page_lru(page);
1120                         ret = 0;
1121                         ClearPageLRU(page);
1122
1123                         del_page_from_lru_list(zone, page, lru);
1124                 }
1125                 spin_unlock_irq(&zone->lru_lock);
1126         }
1127         return ret;
1128 }
1129
1130 /*
1131  * Are there way too many processes in the direct reclaim path already?
1132  */
1133 static int too_many_isolated(struct zone *zone, int file,
1134                 struct scan_control *sc)
1135 {
1136         unsigned long inactive, isolated;
1137
1138         if (current_is_kswapd())
1139                 return 0;
1140
1141         if (!scanning_global_lru(sc))
1142                 return 0;
1143
1144         if (file) {
1145                 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1146                 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1147         } else {
1148                 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1149                 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1150         }
1151
1152         return isolated > inactive;
1153 }
1154
1155 /*
1156  * TODO: Try merging with migrations version of putback_lru_pages
1157  */
1158 static noinline_for_stack void
1159 putback_lru_pages(struct zone *zone, struct scan_control *sc,
1160                                 unsigned long nr_anon, unsigned long nr_file,
1161                                 struct list_head *page_list)
1162 {
1163         struct page *page;
1164         struct pagevec pvec;
1165         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1166
1167         pagevec_init(&pvec, 1);
1168
1169         /*
1170          * Put back any unfreeable pages.
1171          */
1172         spin_lock(&zone->lru_lock);
1173         while (!list_empty(page_list)) {
1174                 int lru;
1175                 page = lru_to_page(page_list);
1176                 VM_BUG_ON(PageLRU(page));
1177                 list_del(&page->lru);
1178                 if (unlikely(!page_evictable(page, NULL))) {
1179                         spin_unlock_irq(&zone->lru_lock);
1180                         putback_lru_page(page);
1181                         spin_lock_irq(&zone->lru_lock);
1182                         continue;
1183                 }
1184                 SetPageLRU(page);
1185                 lru = page_lru(page);
1186                 add_page_to_lru_list(zone, page, lru);
1187                 if (is_active_lru(lru)) {
1188                         int file = is_file_lru(lru);
1189                         reclaim_stat->recent_rotated[file]++;
1190                 }
1191                 if (!pagevec_add(&pvec, page)) {
1192                         spin_unlock_irq(&zone->lru_lock);
1193                         __pagevec_release(&pvec);
1194                         spin_lock_irq(&zone->lru_lock);
1195                 }
1196         }
1197         __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1198         __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1199
1200         spin_unlock_irq(&zone->lru_lock);
1201         pagevec_release(&pvec);
1202 }
1203
1204 static noinline_for_stack void update_isolated_counts(struct zone *zone,
1205                                         struct scan_control *sc,
1206                                         unsigned long *nr_anon,
1207                                         unsigned long *nr_file,
1208                                         struct list_head *isolated_list)
1209 {
1210         unsigned long nr_active;
1211         unsigned int count[NR_LRU_LISTS] = { 0, };
1212         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1213
1214         nr_active = clear_active_flags(isolated_list, count);
1215         __count_vm_events(PGDEACTIVATE, nr_active);
1216
1217         __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1218                               -count[LRU_ACTIVE_FILE]);
1219         __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1220                               -count[LRU_INACTIVE_FILE]);
1221         __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1222                               -count[LRU_ACTIVE_ANON]);
1223         __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1224                               -count[LRU_INACTIVE_ANON]);
1225
1226         *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1227         *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1228         __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1229         __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1230
1231         reclaim_stat->recent_scanned[0] += *nr_anon;
1232         reclaim_stat->recent_scanned[1] += *nr_file;
1233 }
1234
1235 /*
1236  * Returns true if the caller should wait to clean dirty/writeback pages.
1237  *
1238  * If we are direct reclaiming for contiguous pages and we do not reclaim
1239  * everything in the list, try again and wait for writeback IO to complete.
1240  * This will stall high-order allocations noticeably. Only do that when really
1241  * need to free the pages under high memory pressure.
1242  */
1243 static inline bool should_reclaim_stall(unsigned long nr_taken,
1244                                         unsigned long nr_freed,
1245                                         int priority,
1246                                         struct scan_control *sc)
1247 {
1248         int lumpy_stall_priority;
1249
1250         /* kswapd should not stall on sync IO */
1251         if (current_is_kswapd())
1252                 return false;
1253
1254         /* Only stall on lumpy reclaim */
1255         if (!sc->lumpy_reclaim_mode)
1256                 return false;
1257
1258         /* If we have relaimed everything on the isolated list, no stall */
1259         if (nr_freed == nr_taken)
1260                 return false;
1261
1262         /*
1263          * For high-order allocations, there are two stall thresholds.
1264          * High-cost allocations stall immediately where as lower
1265          * order allocations such as stacks require the scanning
1266          * priority to be much higher before stalling.
1267          */
1268         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1269                 lumpy_stall_priority = DEF_PRIORITY;
1270         else
1271                 lumpy_stall_priority = DEF_PRIORITY / 3;
1272
1273         return priority <= lumpy_stall_priority;
1274 }
1275
1276 /*
1277  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1278  * of reclaimed pages
1279  */
1280 static noinline_for_stack unsigned long
1281 shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1282                         struct scan_control *sc, int priority, int file)
1283 {
1284         LIST_HEAD(page_list);
1285         unsigned long nr_scanned;
1286         unsigned long nr_reclaimed = 0;
1287         unsigned long nr_taken;
1288         unsigned long nr_active;
1289         unsigned long nr_anon;
1290         unsigned long nr_file;
1291
1292         while (unlikely(too_many_isolated(zone, file, sc))) {
1293                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1294
1295                 /* We are about to die and free our memory. Return now. */
1296                 if (fatal_signal_pending(current))
1297                         return SWAP_CLUSTER_MAX;
1298         }
1299
1300
1301         lru_add_drain();
1302         spin_lock_irq(&zone->lru_lock);
1303
1304         if (scanning_global_lru(sc)) {
1305                 nr_taken = isolate_pages_global(nr_to_scan,
1306                         &page_list, &nr_scanned, sc->order,
1307                         sc->lumpy_reclaim_mode ?
1308                                 ISOLATE_BOTH : ISOLATE_INACTIVE,
1309                         zone, 0, file);
1310                 zone->pages_scanned += nr_scanned;
1311                 if (current_is_kswapd())
1312                         __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1313                                                nr_scanned);
1314                 else
1315                         __count_zone_vm_events(PGSCAN_DIRECT, zone,
1316                                                nr_scanned);
1317         } else {
1318                 nr_taken = mem_cgroup_isolate_pages(nr_to_scan,
1319                         &page_list, &nr_scanned, sc->order,
1320                         sc->lumpy_reclaim_mode ?
1321                                 ISOLATE_BOTH : ISOLATE_INACTIVE,
1322                         zone, sc->mem_cgroup,
1323                         0, file);
1324                 /*
1325                  * mem_cgroup_isolate_pages() keeps track of
1326                  * scanned pages on its own.
1327                  */
1328         }
1329
1330         if (nr_taken == 0) {
1331                 spin_unlock_irq(&zone->lru_lock);
1332                 return 0;
1333         }
1334
1335         update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
1336
1337         spin_unlock_irq(&zone->lru_lock);
1338
1339         nr_reclaimed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
1340
1341         /* Check if we should syncronously wait for writeback */
1342         if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1343                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1344
1345                 /*
1346                  * The attempt at page out may have made some
1347                  * of the pages active, mark them inactive again.
1348                  */
1349                 nr_active = clear_active_flags(&page_list, NULL);
1350                 count_vm_events(PGDEACTIVATE, nr_active);
1351
1352                 nr_reclaimed += shrink_page_list(&page_list, sc, PAGEOUT_IO_SYNC);
1353         }
1354
1355         local_irq_disable();
1356         if (current_is_kswapd())
1357                 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1358         __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1359
1360         putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
1361
1362         trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1363                 zone_idx(zone),
1364                 nr_scanned, nr_reclaimed,
1365                 priority,
1366                 trace_shrink_flags(file, sc->lumpy_reclaim_mode));
1367         return nr_reclaimed;
1368 }
1369
1370 /*
1371  * This moves pages from the active list to the inactive list.
1372  *
1373  * We move them the other way if the page is referenced by one or more
1374  * processes, from rmap.
1375  *
1376  * If the pages are mostly unmapped, the processing is fast and it is
1377  * appropriate to hold zone->lru_lock across the whole operation.  But if
1378  * the pages are mapped, the processing is slow (page_referenced()) so we
1379  * should drop zone->lru_lock around each page.  It's impossible to balance
1380  * this, so instead we remove the pages from the LRU while processing them.
1381  * It is safe to rely on PG_active against the non-LRU pages in here because
1382  * nobody will play with that bit on a non-LRU page.
1383  *
1384  * The downside is that we have to touch page->_count against each page.
1385  * But we had to alter page->flags anyway.
1386  */
1387
1388 static void move_active_pages_to_lru(struct zone *zone,
1389                                      struct list_head *list,
1390                                      enum lru_list lru)
1391 {
1392         unsigned long pgmoved = 0;
1393         struct pagevec pvec;
1394         struct page *page;
1395
1396         pagevec_init(&pvec, 1);
1397
1398         while (!list_empty(list)) {
1399                 page = lru_to_page(list);
1400
1401                 VM_BUG_ON(PageLRU(page));
1402                 SetPageLRU(page);
1403
1404                 list_move(&page->lru, &zone->lru[lru].list);
1405                 mem_cgroup_add_lru_list(page, lru);
1406                 pgmoved++;
1407
1408                 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1409                         spin_unlock_irq(&zone->lru_lock);
1410                         if (buffer_heads_over_limit)
1411                                 pagevec_strip(&pvec);
1412                         __pagevec_release(&pvec);
1413                         spin_lock_irq(&zone->lru_lock);
1414                 }
1415         }
1416         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1417         if (!is_active_lru(lru))
1418                 __count_vm_events(PGDEACTIVATE, pgmoved);
1419 }
1420
1421 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1422                         struct scan_control *sc, int priority, int file)
1423 {
1424         unsigned long nr_taken;
1425         unsigned long pgscanned;
1426         unsigned long vm_flags;
1427         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1428         LIST_HEAD(l_active);
1429         LIST_HEAD(l_inactive);
1430         struct page *page;
1431         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1432         unsigned long nr_rotated = 0;
1433
1434         lru_add_drain();
1435         spin_lock_irq(&zone->lru_lock);
1436         if (scanning_global_lru(sc)) {
1437                 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1438                                                 &pgscanned, sc->order,
1439                                                 ISOLATE_ACTIVE, zone,
1440                                                 1, file);
1441                 zone->pages_scanned += pgscanned;
1442         } else {
1443                 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1444                                                 &pgscanned, sc->order,
1445                                                 ISOLATE_ACTIVE, zone,
1446                                                 sc->mem_cgroup, 1, file);
1447                 /*
1448                  * mem_cgroup_isolate_pages() keeps track of
1449                  * scanned pages on its own.
1450                  */
1451         }
1452
1453         reclaim_stat->recent_scanned[file] += nr_taken;
1454
1455         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1456         if (file)
1457                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1458         else
1459                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1460         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1461         spin_unlock_irq(&zone->lru_lock);
1462
1463         while (!list_empty(&l_hold)) {
1464                 cond_resched();
1465                 page = lru_to_page(&l_hold);
1466                 list_del(&page->lru);
1467
1468                 if (unlikely(!page_evictable(page, NULL))) {
1469                         putback_lru_page(page);
1470                         continue;
1471                 }
1472
1473                 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1474                         nr_rotated++;
1475                         /*
1476                          * Identify referenced, file-backed active pages and
1477                          * give them one more trip around the active list. So
1478                          * that executable code get better chances to stay in
1479                          * memory under moderate memory pressure.  Anon pages
1480                          * are not likely to be evicted by use-once streaming
1481                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1482                          * so we ignore them here.
1483                          */
1484                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1485                                 list_add(&page->lru, &l_active);
1486                                 continue;
1487                         }
1488                 }
1489
1490                 ClearPageActive(page);  /* we are de-activating */
1491                 list_add(&page->lru, &l_inactive);
1492         }
1493
1494         /*
1495          * Move pages back to the lru list.
1496          */
1497         spin_lock_irq(&zone->lru_lock);
1498         /*
1499          * Count referenced pages from currently used mappings as rotated,
1500          * even though only some of them are actually re-activated.  This
1501          * helps balance scan pressure between file and anonymous pages in
1502          * get_scan_ratio.
1503          */
1504         reclaim_stat->recent_rotated[file] += nr_rotated;
1505
1506         move_active_pages_to_lru(zone, &l_active,
1507                                                 LRU_ACTIVE + file * LRU_FILE);
1508         move_active_pages_to_lru(zone, &l_inactive,
1509                                                 LRU_BASE   + file * LRU_FILE);
1510         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1511         spin_unlock_irq(&zone->lru_lock);
1512 }
1513
1514 #ifdef CONFIG_SWAP
1515 static int inactive_anon_is_low_global(struct zone *zone)
1516 {
1517         unsigned long active, inactive;
1518
1519         active = zone_page_state(zone, NR_ACTIVE_ANON);
1520         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1521
1522         if (inactive * zone->inactive_ratio < active)
1523                 return 1;
1524
1525         return 0;
1526 }
1527
1528 /**
1529  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1530  * @zone: zone to check
1531  * @sc:   scan control of this context
1532  *
1533  * Returns true if the zone does not have enough inactive anon pages,
1534  * meaning some active anon pages need to be deactivated.
1535  */
1536 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1537 {
1538         int low;
1539
1540         /*
1541          * If we don't have swap space, anonymous page deactivation
1542          * is pointless.
1543          */
1544         if (!total_swap_pages)
1545                 return 0;
1546
1547         if (scanning_global_lru(sc))
1548                 low = inactive_anon_is_low_global(zone);
1549         else
1550                 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1551         return low;
1552 }
1553 #else
1554 static inline int inactive_anon_is_low(struct zone *zone,
1555                                         struct scan_control *sc)
1556 {
1557         return 0;
1558 }
1559 #endif
1560
1561 static int inactive_file_is_low_global(struct zone *zone)
1562 {
1563         unsigned long active, inactive;
1564
1565         active = zone_page_state(zone, NR_ACTIVE_FILE);
1566         inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1567
1568         return (active > inactive);
1569 }
1570
1571 /**
1572  * inactive_file_is_low - check if file pages need to be deactivated
1573  * @zone: zone to check
1574  * @sc:   scan control of this context
1575  *
1576  * When the system is doing streaming IO, memory pressure here
1577  * ensures that active file pages get deactivated, until more
1578  * than half of the file pages are on the inactive list.
1579  *
1580  * Once we get to that situation, protect the system's working
1581  * set from being evicted by disabling active file page aging.
1582  *
1583  * This uses a different ratio than the anonymous pages, because
1584  * the page cache uses a use-once replacement algorithm.
1585  */
1586 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1587 {
1588         int low;
1589
1590         if (scanning_global_lru(sc))
1591                 low = inactive_file_is_low_global(zone);
1592         else
1593                 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1594         return low;
1595 }
1596
1597 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1598                                 int file)
1599 {
1600         if (file)
1601                 return inactive_file_is_low(zone, sc);
1602         else
1603                 return inactive_anon_is_low(zone, sc);
1604 }
1605
1606 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1607         struct zone *zone, struct scan_control *sc, int priority)
1608 {
1609         int file = is_file_lru(lru);
1610
1611         if (is_active_lru(lru)) {
1612                 if (inactive_list_is_low(zone, sc, file))
1613                     shrink_active_list(nr_to_scan, zone, sc, priority, file);
1614                 return 0;
1615         }
1616
1617         return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1618 }
1619
1620 /*
1621  * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1622  * until we collected @swap_cluster_max pages to scan.
1623  */
1624 static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
1625                                        unsigned long *nr_saved_scan)
1626 {
1627         unsigned long nr;
1628
1629         *nr_saved_scan += nr_to_scan;
1630         nr = *nr_saved_scan;
1631
1632         if (nr >= SWAP_CLUSTER_MAX)
1633                 *nr_saved_scan = 0;
1634         else
1635                 nr = 0;
1636
1637         return nr;
1638 }
1639
1640 /*
1641  * Determine how aggressively the anon and file LRU lists should be
1642  * scanned.  The relative value of each set of LRU lists is determined
1643  * by looking at the fraction of the pages scanned we did rotate back
1644  * onto the active list instead of evict.
1645  *
1646  * nr[0] = anon pages to scan; nr[1] = file pages to scan
1647  */
1648 static void get_scan_count(struct zone *zone, struct scan_control *sc,
1649                                         unsigned long *nr, int priority)
1650 {
1651         unsigned long anon, file, free;
1652         unsigned long anon_prio, file_prio;
1653         unsigned long ap, fp;
1654         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1655         u64 fraction[2], denominator;
1656         enum lru_list l;
1657         int noswap = 0;
1658
1659         /* If we have no swap space, do not bother scanning anon pages. */
1660         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1661                 noswap = 1;
1662                 fraction[0] = 0;
1663                 fraction[1] = 1;
1664                 denominator = 1;
1665                 goto out;
1666         }
1667
1668         anon  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1669                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1670         file  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1671                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1672
1673         if (scanning_global_lru(sc)) {
1674                 free  = zone_page_state(zone, NR_FREE_PAGES);
1675                 /* If we have very few page cache pages,
1676                    force-scan anon pages. */
1677                 if (unlikely(file + free <= high_wmark_pages(zone))) {
1678                         fraction[0] = 1;
1679                         fraction[1] = 0;
1680                         denominator = 1;
1681                         goto out;
1682                 }
1683         }
1684
1685         /*
1686          * With swappiness at 100, anonymous and file have the same priority.
1687          * This scanning priority is essentially the inverse of IO cost.
1688          */
1689         anon_prio = sc->swappiness;
1690         file_prio = 200 - sc->swappiness;
1691
1692         /*
1693          * OK, so we have swap space and a fair amount of page cache
1694          * pages.  We use the recently rotated / recently scanned
1695          * ratios to determine how valuable each cache is.
1696          *
1697          * Because workloads change over time (and to avoid overflow)
1698          * we keep these statistics as a floating average, which ends
1699          * up weighing recent references more than old ones.
1700          *
1701          * anon in [0], file in [1]
1702          */
1703         spin_lock_irq(&zone->lru_lock);
1704         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1705                 reclaim_stat->recent_scanned[0] /= 2;
1706                 reclaim_stat->recent_rotated[0] /= 2;
1707         }
1708
1709         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1710                 reclaim_stat->recent_scanned[1] /= 2;
1711                 reclaim_stat->recent_rotated[1] /= 2;
1712         }
1713
1714         /*
1715          * The amount of pressure on anon vs file pages is inversely
1716          * proportional to the fraction of recently scanned pages on
1717          * each list that were recently referenced and in active use.
1718          */
1719         ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1720         ap /= reclaim_stat->recent_rotated[0] + 1;
1721
1722         fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1723         fp /= reclaim_stat->recent_rotated[1] + 1;
1724         spin_unlock_irq(&zone->lru_lock);
1725
1726         fraction[0] = ap;
1727         fraction[1] = fp;
1728         denominator = ap + fp + 1;
1729 out:
1730         for_each_evictable_lru(l) {
1731                 int file = is_file_lru(l);
1732                 unsigned long scan;
1733
1734                 scan = zone_nr_lru_pages(zone, sc, l);
1735                 if (priority || noswap) {
1736                         scan >>= priority;
1737                         scan = div64_u64(scan * fraction[file], denominator);
1738                 }
1739                 nr[l] = nr_scan_try_batch(scan,
1740                                           &reclaim_stat->nr_saved_scan[l]);
1741         }
1742 }
1743
1744 static void set_lumpy_reclaim_mode(int priority, struct scan_control *sc)
1745 {
1746         /*
1747          * If we need a large contiguous chunk of memory, or have
1748          * trouble getting a small set of contiguous pages, we
1749          * will reclaim both active and inactive pages.
1750          */
1751         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1752                 sc->lumpy_reclaim_mode = 1;
1753         else if (sc->order && priority < DEF_PRIORITY - 2)
1754                 sc->lumpy_reclaim_mode = 1;
1755         else
1756                 sc->lumpy_reclaim_mode = 0;
1757 }
1758
1759 /*
1760  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1761  */
1762 static void shrink_zone(int priority, struct zone *zone,
1763                                 struct scan_control *sc)
1764 {
1765         unsigned long nr[NR_LRU_LISTS];
1766         unsigned long nr_to_scan;
1767         enum lru_list l;
1768         unsigned long nr_reclaimed = sc->nr_reclaimed;
1769         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1770
1771         get_scan_count(zone, sc, nr, priority);
1772
1773         set_lumpy_reclaim_mode(priority, sc);
1774
1775         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1776                                         nr[LRU_INACTIVE_FILE]) {
1777                 for_each_evictable_lru(l) {
1778                         if (nr[l]) {
1779                                 nr_to_scan = min_t(unsigned long,
1780                                                    nr[l], SWAP_CLUSTER_MAX);
1781                                 nr[l] -= nr_to_scan;
1782
1783                                 nr_reclaimed += shrink_list(l, nr_to_scan,
1784                                                             zone, sc, priority);
1785                         }
1786                 }
1787                 /*
1788                  * On large memory systems, scan >> priority can become
1789                  * really large. This is fine for the starting priority;
1790                  * we want to put equal scanning pressure on each zone.
1791                  * However, if the VM has a harder time of freeing pages,
1792                  * with multiple processes reclaiming pages, the total
1793                  * freeing target can get unreasonably large.
1794                  */
1795                 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
1796                         break;
1797         }
1798
1799         sc->nr_reclaimed = nr_reclaimed;
1800
1801         /*
1802          * Even if we did not try to evict anon pages at all, we want to
1803          * rebalance the anon lru active/inactive ratio.
1804          */
1805         if (inactive_anon_is_low(zone, sc))
1806                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1807
1808         throttle_vm_writeout(sc->gfp_mask);
1809 }
1810
1811 /*
1812  * This is the direct reclaim path, for page-allocating processes.  We only
1813  * try to reclaim pages from zones which will satisfy the caller's allocation
1814  * request.
1815  *
1816  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1817  * Because:
1818  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1819  *    allocation or
1820  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1821  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1822  *    zone defense algorithm.
1823  *
1824  * If a zone is deemed to be full of pinned pages then just give it a light
1825  * scan then give up on it.
1826  */
1827 static void shrink_zones(int priority, struct zonelist *zonelist,
1828                                         struct scan_control *sc)
1829 {
1830         struct zoneref *z;
1831         struct zone *zone;
1832
1833         for_each_zone_zonelist_nodemask(zone, z, zonelist,
1834                                         gfp_zone(sc->gfp_mask), sc->nodemask) {
1835                 if (!populated_zone(zone))
1836                         continue;
1837                 /*
1838                  * Take care memory controller reclaiming has small influence
1839                  * to global LRU.
1840                  */
1841                 if (scanning_global_lru(sc)) {
1842                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1843                                 continue;
1844                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1845                                 continue;       /* Let kswapd poll it */
1846                 }
1847
1848                 shrink_zone(priority, zone, sc);
1849         }
1850 }
1851
1852 static bool zone_reclaimable(struct zone *zone)
1853 {
1854         return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
1855 }
1856
1857 /*
1858  * As hibernation is going on, kswapd is freezed so that it can't mark
1859  * the zone into all_unreclaimable. It can't handle OOM during hibernation.
1860  * So let's check zone's unreclaimable in direct reclaim as well as kswapd.
1861  */
1862 static bool all_unreclaimable(struct zonelist *zonelist,
1863                 struct scan_control *sc)
1864 {
1865         struct zoneref *z;
1866         struct zone *zone;
1867         bool all_unreclaimable = true;
1868
1869         for_each_zone_zonelist_nodemask(zone, z, zonelist,
1870                         gfp_zone(sc->gfp_mask), sc->nodemask) {
1871                 if (!populated_zone(zone))
1872                         continue;
1873                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1874                         continue;
1875                 if (zone_reclaimable(zone)) {
1876                         all_unreclaimable = false;
1877                         break;
1878                 }
1879         }
1880
1881         return all_unreclaimable;
1882 }
1883
1884 /*
1885  * This is the main entry point to direct page reclaim.
1886  *
1887  * If a full scan of the inactive list fails to free enough memory then we
1888  * are "out of memory" and something needs to be killed.
1889  *
1890  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1891  * high - the zone may be full of dirty or under-writeback pages, which this
1892  * caller can't do much about.  We kick the writeback threads and take explicit
1893  * naps in the hope that some of these pages can be written.  But if the
1894  * allocating task holds filesystem locks which prevent writeout this might not
1895  * work, and the allocation attempt will fail.
1896  *
1897  * returns:     0, if no pages reclaimed
1898  *              else, the number of pages reclaimed
1899  */
1900 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
1901                                         struct scan_control *sc)
1902 {
1903         int priority;
1904         unsigned long total_scanned = 0;
1905         struct reclaim_state *reclaim_state = current->reclaim_state;
1906         struct zoneref *z;
1907         struct zone *zone;
1908         unsigned long writeback_threshold;
1909
1910         get_mems_allowed();
1911         delayacct_freepages_start();
1912
1913         if (scanning_global_lru(sc))
1914                 count_vm_event(ALLOCSTALL);
1915
1916         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1917                 sc->nr_scanned = 0;
1918                 if (!priority)
1919                         disable_swap_token();
1920                 shrink_zones(priority, zonelist, sc);
1921                 /*
1922                  * Don't shrink slabs when reclaiming memory from
1923                  * over limit cgroups
1924                  */
1925                 if (scanning_global_lru(sc)) {
1926                         unsigned long lru_pages = 0;
1927                         for_each_zone_zonelist(zone, z, zonelist,
1928                                         gfp_zone(sc->gfp_mask)) {
1929                                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1930                                         continue;
1931
1932                                 lru_pages += zone_reclaimable_pages(zone);
1933                         }
1934
1935                         shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
1936                         if (reclaim_state) {
1937                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
1938                                 reclaim_state->reclaimed_slab = 0;
1939                         }
1940                 }
1941                 total_scanned += sc->nr_scanned;
1942                 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
1943                         goto out;
1944
1945                 /*
1946                  * Try to write back as many pages as we just scanned.  This
1947                  * tends to cause slow streaming writers to write data to the
1948                  * disk smoothly, at the dirtying rate, which is nice.   But
1949                  * that's undesirable in laptop mode, where we *want* lumpy
1950                  * writeout.  So in laptop mode, write out the whole world.
1951                  */
1952                 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
1953                 if (total_scanned > writeback_threshold) {
1954                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
1955                         sc->may_writepage = 1;
1956                 }
1957
1958                 /* Take a nap, wait for some writeback to complete */
1959                 if (!sc->hibernation_mode && sc->nr_scanned &&
1960                     priority < DEF_PRIORITY - 2)
1961                         congestion_wait(BLK_RW_ASYNC, HZ/10);
1962         }
1963
1964 out:
1965         delayacct_freepages_end();
1966         put_mems_allowed();
1967
1968         if (sc->nr_reclaimed)
1969                 return sc->nr_reclaimed;
1970
1971         /* top priority shrink_zones still had more to do? don't OOM, then */
1972         if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
1973                 return 1;
1974
1975         return 0;
1976 }
1977
1978 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
1979                                 gfp_t gfp_mask, nodemask_t *nodemask)
1980 {
1981         unsigned long nr_reclaimed;
1982         struct scan_control sc = {
1983                 .gfp_mask = gfp_mask,
1984                 .may_writepage = !laptop_mode,
1985                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
1986                 .may_unmap = 1,
1987                 .may_swap = 1,
1988                 .swappiness = vm_swappiness,
1989                 .order = order,
1990                 .mem_cgroup = NULL,
1991                 .nodemask = nodemask,
1992         };
1993
1994         trace_mm_vmscan_direct_reclaim_begin(order,
1995                                 sc.may_writepage,
1996                                 gfp_mask);
1997
1998         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
1999
2000         trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2001
2002         return nr_reclaimed;
2003 }
2004
2005 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2006
2007 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2008                                                 gfp_t gfp_mask, bool noswap,
2009                                                 unsigned int swappiness,
2010                                                 struct zone *zone)
2011 {
2012         struct scan_control sc = {
2013                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2014                 .may_writepage = !laptop_mode,
2015                 .may_unmap = 1,
2016                 .may_swap = !noswap,
2017                 .swappiness = swappiness,
2018                 .order = 0,
2019                 .mem_cgroup = mem,
2020         };
2021         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2022                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2023
2024         trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2025                                                       sc.may_writepage,
2026                                                       sc.gfp_mask);
2027
2028         /*
2029          * NOTE: Although we can get the priority field, using it
2030          * here is not a good idea, since it limits the pages we can scan.
2031          * if we don't reclaim here, the shrink_zone from balance_pgdat
2032          * will pick up pages from other mem cgroup's as well. We hack
2033          * the priority and make it zero.
2034          */
2035         shrink_zone(0, zone, &sc);
2036
2037         trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2038
2039         return sc.nr_reclaimed;
2040 }
2041
2042 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
2043                                            gfp_t gfp_mask,
2044                                            bool noswap,
2045                                            unsigned int swappiness)
2046 {
2047         struct zonelist *zonelist;
2048         unsigned long nr_reclaimed;
2049         struct scan_control sc = {
2050                 .may_writepage = !laptop_mode,
2051                 .may_unmap = 1,
2052                 .may_swap = !noswap,
2053                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2054                 .swappiness = swappiness,
2055                 .order = 0,
2056                 .mem_cgroup = mem_cont,
2057                 .nodemask = NULL, /* we don't care the placement */
2058         };
2059
2060         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2061                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2062         zonelist = NODE_DATA(numa_node_id())->node_zonelists;
2063
2064         trace_mm_vmscan_memcg_reclaim_begin(0,
2065                                             sc.may_writepage,
2066                                             sc.gfp_mask);
2067
2068         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2069
2070         trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2071
2072         return nr_reclaimed;
2073 }
2074 #endif
2075
2076 /* is kswapd sleeping prematurely? */
2077 static int sleeping_prematurely(pg_data_t *pgdat, int order, long remaining)
2078 {
2079         int i;
2080
2081         /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2082         if (remaining)
2083                 return 1;
2084
2085         /* If after HZ/10, a zone is below the high mark, it's premature */
2086         for (i = 0; i < pgdat->nr_zones; i++) {
2087                 struct zone *zone = pgdat->node_zones + i;
2088
2089                 if (!populated_zone(zone))
2090                         continue;
2091
2092                 if (zone->all_unreclaimable)
2093                         continue;
2094
2095                 if (!zone_watermark_ok(zone, order, high_wmark_pages(zone),
2096                                                                 0, 0))
2097                         return 1;
2098         }
2099
2100         return 0;
2101 }
2102
2103 /*
2104  * For kswapd, balance_pgdat() will work across all this node's zones until
2105  * they are all at high_wmark_pages(zone).
2106  *
2107  * Returns the number of pages which were actually freed.
2108  *
2109  * There is special handling here for zones which are full of pinned pages.
2110  * This can happen if the pages are all mlocked, or if they are all used by
2111  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
2112  * What we do is to detect the case where all pages in the zone have been
2113  * scanned twice and there has been zero successful reclaim.  Mark the zone as
2114  * dead and from now on, only perform a short scan.  Basically we're polling
2115  * the zone for when the problem goes away.
2116  *
2117  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
2118  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2119  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2120  * lower zones regardless of the number of free pages in the lower zones. This
2121  * interoperates with the page allocator fallback scheme to ensure that aging
2122  * of pages is balanced across the zones.
2123  */
2124 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
2125 {
2126         int all_zones_ok;
2127         int priority;
2128         int i;
2129         unsigned long total_scanned;
2130         struct reclaim_state *reclaim_state = current->reclaim_state;
2131         struct scan_control sc = {
2132                 .gfp_mask = GFP_KERNEL,
2133                 .may_unmap = 1,
2134                 .may_swap = 1,
2135                 /*
2136                  * kswapd doesn't want to be bailed out while reclaim. because
2137                  * we want to put equal scanning pressure on each zone.
2138                  */
2139                 .nr_to_reclaim = ULONG_MAX,
2140                 .swappiness = vm_swappiness,
2141                 .order = order,
2142                 .mem_cgroup = NULL,
2143         };
2144 loop_again:
2145         total_scanned = 0;
2146         sc.nr_reclaimed = 0;
2147         sc.may_writepage = !laptop_mode;
2148         count_vm_event(PAGEOUTRUN);
2149
2150         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2151                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
2152                 unsigned long lru_pages = 0;
2153                 int has_under_min_watermark_zone = 0;
2154
2155                 /* The swap token gets in the way of swapout... */
2156                 if (!priority)
2157                         disable_swap_token();
2158
2159                 all_zones_ok = 1;
2160
2161                 /*
2162                  * Scan in the highmem->dma direction for the highest
2163                  * zone which needs scanning
2164                  */
2165                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2166                         struct zone *zone = pgdat->node_zones + i;
2167
2168                         if (!populated_zone(zone))
2169                                 continue;
2170
2171                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2172                                 continue;
2173
2174                         /*
2175                          * Do some background aging of the anon list, to give
2176                          * pages a chance to be referenced before reclaiming.
2177                          */
2178                         if (inactive_anon_is_low(zone, &sc))
2179                                 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2180                                                         &sc, priority, 0);
2181
2182                         if (!zone_watermark_ok(zone, order,
2183                                         high_wmark_pages(zone), 0, 0)) {
2184                                 end_zone = i;
2185                                 break;
2186                         }
2187                 }
2188                 if (i < 0)
2189                         goto out;
2190
2191                 for (i = 0; i <= end_zone; i++) {
2192                         struct zone *zone = pgdat->node_zones + i;
2193
2194                         lru_pages += zone_reclaimable_pages(zone);
2195                 }
2196
2197                 /*
2198                  * Now scan the zone in the dma->highmem direction, stopping
2199                  * at the last zone which needs scanning.
2200                  *
2201                  * We do this because the page allocator works in the opposite
2202                  * direction.  This prevents the page allocator from allocating
2203                  * pages behind kswapd's direction of progress, which would
2204                  * cause too much scanning of the lower zones.
2205                  */
2206                 for (i = 0; i <= end_zone; i++) {
2207                         struct zone *zone = pgdat->node_zones + i;
2208                         int nr_slab;
2209
2210                         if (!populated_zone(zone))
2211                                 continue;
2212
2213                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2214                                 continue;
2215
2216                         sc.nr_scanned = 0;
2217
2218                         /*
2219                          * Call soft limit reclaim before calling shrink_zone.
2220                          * For now we ignore the return value
2221                          */
2222                         mem_cgroup_soft_limit_reclaim(zone, order, sc.gfp_mask);
2223
2224                         /*
2225                          * We put equal pressure on every zone, unless one
2226                          * zone has way too many pages free already.
2227                          */
2228                         if (!zone_watermark_ok(zone, order,
2229                                         8*high_wmark_pages(zone), end_zone, 0))
2230                                 shrink_zone(priority, zone, &sc);
2231                         reclaim_state->reclaimed_slab = 0;
2232                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
2233                                                 lru_pages);
2234                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2235                         total_scanned += sc.nr_scanned;
2236                         if (zone->all_unreclaimable)
2237                                 continue;
2238                         if (nr_slab == 0 && !zone_reclaimable(zone))
2239                                 zone->all_unreclaimable = 1;
2240                         /*
2241                          * If we've done a decent amount of scanning and
2242                          * the reclaim ratio is low, start doing writepage
2243                          * even in laptop mode
2244                          */
2245                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2246                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2247                                 sc.may_writepage = 1;
2248
2249                         if (!zone_watermark_ok(zone, order,
2250                                         high_wmark_pages(zone), end_zone, 0)) {
2251                                 all_zones_ok = 0;
2252                                 /*
2253                                  * We are still under min water mark.  This
2254                                  * means that we have a GFP_ATOMIC allocation
2255                                  * failure risk. Hurry up!
2256                                  */
2257                                 if (!zone_watermark_ok(zone, order,
2258                                             min_wmark_pages(zone), end_zone, 0))
2259                                         has_under_min_watermark_zone = 1;
2260                         }
2261
2262                 }
2263                 if (all_zones_ok)
2264                         break;          /* kswapd: all done */
2265                 /*
2266                  * OK, kswapd is getting into trouble.  Take a nap, then take
2267                  * another pass across the zones.
2268                  */
2269                 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2270                         if (has_under_min_watermark_zone)
2271                                 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2272                         else
2273                                 congestion_wait(BLK_RW_ASYNC, HZ/10);
2274                 }
2275
2276                 /*
2277                  * We do this so kswapd doesn't build up large priorities for
2278                  * example when it is freeing in parallel with allocators. It
2279                  * matches the direct reclaim path behaviour in terms of impact
2280                  * on zone->*_priority.
2281                  */
2282                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2283                         break;
2284         }
2285 out:
2286         if (!all_zones_ok) {
2287                 cond_resched();
2288
2289                 try_to_freeze();
2290
2291                 /*
2292                  * Fragmentation may mean that the system cannot be
2293                  * rebalanced for high-order allocations in all zones.
2294                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2295                  * it means the zones have been fully scanned and are still
2296                  * not balanced. For high-order allocations, there is
2297                  * little point trying all over again as kswapd may
2298                  * infinite loop.
2299                  *
2300                  * Instead, recheck all watermarks at order-0 as they
2301                  * are the most important. If watermarks are ok, kswapd will go
2302                  * back to sleep. High-order users can still perform direct
2303                  * reclaim if they wish.
2304                  */
2305                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2306                         order = sc.order = 0;
2307
2308                 goto loop_again;
2309         }
2310
2311         return sc.nr_reclaimed;
2312 }
2313
2314 /*
2315  * The background pageout daemon, started as a kernel thread
2316  * from the init process.
2317  *
2318  * This basically trickles out pages so that we have _some_
2319  * free memory available even if there is no other activity
2320  * that frees anything up. This is needed for things like routing
2321  * etc, where we otherwise might have all activity going on in
2322  * asynchronous contexts that cannot page things out.
2323  *
2324  * If there are applications that are active memory-allocators
2325  * (most normal use), this basically shouldn't matter.
2326  */
2327 static int kswapd(void *p)
2328 {
2329         unsigned long order;
2330         pg_data_t *pgdat = (pg_data_t*)p;
2331         struct task_struct *tsk = current;
2332         DEFINE_WAIT(wait);
2333         struct reclaim_state reclaim_state = {
2334                 .reclaimed_slab = 0,
2335         };
2336         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2337
2338         lockdep_set_current_reclaim_state(GFP_KERNEL);
2339
2340         if (!cpumask_empty(cpumask))
2341                 set_cpus_allowed_ptr(tsk, cpumask);
2342         current->reclaim_state = &reclaim_state;
2343
2344         /*
2345          * Tell the memory management that we're a "memory allocator",
2346          * and that if we need more memory we should get access to it
2347          * regardless (see "__alloc_pages()"). "kswapd" should
2348          * never get caught in the normal page freeing logic.
2349          *
2350          * (Kswapd normally doesn't need memory anyway, but sometimes
2351          * you need a small amount of memory in order to be able to
2352          * page out something else, and this flag essentially protects
2353          * us from recursively trying to free more memory as we're
2354          * trying to free the first piece of memory in the first place).
2355          */
2356         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2357         set_freezable();
2358
2359         order = 0;
2360         for ( ; ; ) {
2361                 unsigned long new_order;
2362                 int ret;
2363
2364                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2365                 new_order = pgdat->kswapd_max_order;
2366                 pgdat->kswapd_max_order = 0;
2367                 if (order < new_order) {
2368                         /*
2369                          * Don't sleep if someone wants a larger 'order'
2370                          * allocation
2371                          */
2372                         order = new_order;
2373                 } else {
2374                         if (!freezing(current) && !kthread_should_stop()) {
2375                                 long remaining = 0;
2376
2377                                 /* Try to sleep for a short interval */
2378                                 if (!sleeping_prematurely(pgdat, order, remaining)) {
2379                                         remaining = schedule_timeout(HZ/10);
2380                                         finish_wait(&pgdat->kswapd_wait, &wait);
2381                                         prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2382                                 }
2383
2384                                 /*
2385                                  * After a short sleep, check if it was a
2386                                  * premature sleep. If not, then go fully
2387                                  * to sleep until explicitly woken up
2388                                  */
2389                                 if (!sleeping_prematurely(pgdat, order, remaining)) {
2390                                         trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2391                                         schedule();
2392                                 } else {
2393                                         if (remaining)
2394                                                 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2395                                         else
2396                                                 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2397                                 }
2398                         }
2399
2400                         order = pgdat->kswapd_max_order;
2401                 }
2402                 finish_wait(&pgdat->kswapd_wait, &wait);
2403
2404                 ret = try_to_freeze();
2405                 if (kthread_should_stop())
2406                         break;
2407
2408                 /*
2409                  * We can speed up thawing tasks if we don't call balance_pgdat
2410                  * after returning from the refrigerator
2411                  */
2412                 if (!ret) {
2413                         trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2414                         balance_pgdat(pgdat, order);
2415                 }
2416         }
2417         return 0;
2418 }
2419
2420 /*
2421  * A zone is low on free memory, so wake its kswapd task to service it.
2422  */
2423 void wakeup_kswapd(struct zone *zone, int order)
2424 {
2425         pg_data_t *pgdat;
2426
2427         if (!populated_zone(zone))
2428                 return;
2429
2430         pgdat = zone->zone_pgdat;
2431         if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
2432                 return;
2433         if (pgdat->kswapd_max_order < order)
2434                 pgdat->kswapd_max_order = order;
2435         trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
2436         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2437                 return;
2438         if (!waitqueue_active(&pgdat->kswapd_wait))
2439                 return;
2440         wake_up_interruptible(&pgdat->kswapd_wait);
2441 }
2442
2443 /*
2444  * The reclaimable count would be mostly accurate.
2445  * The less reclaimable pages may be
2446  * - mlocked pages, which will be moved to unevictable list when encountered
2447  * - mapped pages, which may require several travels to be reclaimed
2448  * - dirty pages, which is not "instantly" reclaimable
2449  */
2450 unsigned long global_reclaimable_pages(void)
2451 {
2452         int nr;
2453
2454         nr = global_page_state(NR_ACTIVE_FILE) +
2455              global_page_state(NR_INACTIVE_FILE);
2456
2457         if (nr_swap_pages > 0)
2458                 nr += global_page_state(NR_ACTIVE_ANON) +
2459                       global_page_state(NR_INACTIVE_ANON);
2460
2461         return nr;
2462 }
2463
2464 unsigned long zone_reclaimable_pages(struct zone *zone)
2465 {
2466         int nr;
2467
2468         nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2469              zone_page_state(zone, NR_INACTIVE_FILE);
2470
2471         if (nr_swap_pages > 0)
2472                 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2473                       zone_page_state(zone, NR_INACTIVE_ANON);
2474
2475         return nr;
2476 }
2477
2478 #ifdef CONFIG_HIBERNATION
2479 /*
2480  * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2481  * freed pages.
2482  *
2483  * Rather than trying to age LRUs the aim is to preserve the overall
2484  * LRU order by reclaiming preferentially
2485  * inactive > active > active referenced > active mapped
2486  */
2487 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
2488 {
2489         struct reclaim_state reclaim_state;
2490         struct scan_control sc = {
2491                 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2492                 .may_swap = 1,
2493                 .may_unmap = 1,
2494                 .may_writepage = 1,
2495                 .nr_to_reclaim = nr_to_reclaim,
2496                 .hibernation_mode = 1,
2497                 .swappiness = vm_swappiness,
2498                 .order = 0,
2499         };
2500         struct zonelist * zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2501         struct task_struct *p = current;
2502         unsigned long nr_reclaimed;
2503
2504         p->flags |= PF_MEMALLOC;
2505         lockdep_set_current_reclaim_state(sc.gfp_mask);
2506         reclaim_state.reclaimed_slab = 0;
2507         p->reclaim_state = &reclaim_state;
2508
2509         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2510
2511         p->reclaim_state = NULL;
2512         lockdep_clear_current_reclaim_state();
2513         p->flags &= ~PF_MEMALLOC;
2514
2515         return nr_reclaimed;
2516 }
2517 #endif /* CONFIG_HIBERNATION */
2518
2519 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2520    not required for correctness.  So if the last cpu in a node goes
2521    away, we get changed to run anywhere: as the first one comes back,
2522    restore their cpu bindings. */
2523 static int __devinit cpu_callback(struct notifier_block *nfb,
2524                                   unsigned long action, void *hcpu)
2525 {
2526         int nid;
2527
2528         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2529                 for_each_node_state(nid, N_HIGH_MEMORY) {
2530                         pg_data_t *pgdat = NODE_DATA(nid);
2531                         const struct cpumask *mask;
2532
2533                         mask = cpumask_of_node(pgdat->node_id);
2534
2535                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2536                                 /* One of our CPUs online: restore mask */
2537                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2538                 }
2539         }
2540         return NOTIFY_OK;
2541 }
2542
2543 /*
2544  * This kswapd start function will be called by init and node-hot-add.
2545  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2546  */
2547 int kswapd_run(int nid)
2548 {
2549         pg_data_t *pgdat = NODE_DATA(nid);
2550         int ret = 0;
2551
2552         if (pgdat->kswapd)
2553                 return 0;
2554
2555         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2556         if (IS_ERR(pgdat->kswapd)) {
2557                 /* failure at boot is fatal */
2558                 BUG_ON(system_state == SYSTEM_BOOTING);
2559                 printk("Failed to start kswapd on node %d\n",nid);
2560                 ret = -1;
2561         }
2562         return ret;
2563 }
2564
2565 /*
2566  * Called by memory hotplug when all memory in a node is offlined.
2567  */
2568 void kswapd_stop(int nid)
2569 {
2570         struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
2571
2572         if (kswapd)
2573                 kthread_stop(kswapd);
2574 }
2575
2576 static int __init kswapd_init(void)
2577 {
2578         int nid;
2579
2580         swap_setup();
2581         for_each_node_state(nid, N_HIGH_MEMORY)
2582                 kswapd_run(nid);
2583         hotcpu_notifier(cpu_callback, 0);
2584         return 0;
2585 }
2586
2587 module_init(kswapd_init)
2588
2589 #ifdef CONFIG_NUMA
2590 /*
2591  * Zone reclaim mode
2592  *
2593  * If non-zero call zone_reclaim when the number of free pages falls below
2594  * the watermarks.
2595  */
2596 int zone_reclaim_mode __read_mostly;
2597
2598 #define RECLAIM_OFF 0
2599 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
2600 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
2601 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
2602
2603 /*
2604  * Priority for ZONE_RECLAIM. This determines the fraction of pages
2605  * of a node considered for each zone_reclaim. 4 scans 1/16th of
2606  * a zone.
2607  */
2608 #define ZONE_RECLAIM_PRIORITY 4
2609
2610 /*
2611  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2612  * occur.
2613  */
2614 int sysctl_min_unmapped_ratio = 1;
2615
2616 /*
2617  * If the number of slab pages in a zone grows beyond this percentage then
2618  * slab reclaim needs to occur.
2619  */
2620 int sysctl_min_slab_ratio = 5;
2621
2622 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
2623 {
2624         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
2625         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
2626                 zone_page_state(zone, NR_ACTIVE_FILE);
2627
2628         /*
2629          * It's possible for there to be more file mapped pages than
2630          * accounted for by the pages on the file LRU lists because
2631          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2632          */
2633         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
2634 }
2635
2636 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
2637 static long zone_pagecache_reclaimable(struct zone *zone)
2638 {
2639         long nr_pagecache_reclaimable;
2640         long delta = 0;
2641
2642         /*
2643          * If RECLAIM_SWAP is set, then all file pages are considered
2644          * potentially reclaimable. Otherwise, we have to worry about
2645          * pages like swapcache and zone_unmapped_file_pages() provides
2646          * a better estimate
2647          */
2648         if (zone_reclaim_mode & RECLAIM_SWAP)
2649                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
2650         else
2651                 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
2652
2653         /* If we can't clean pages, remove dirty pages from consideration */
2654         if (!(zone_reclaim_mode & RECLAIM_WRITE))
2655                 delta += zone_page_state(zone, NR_FILE_DIRTY);
2656
2657         /* Watch for any possible underflows due to delta */
2658         if (unlikely(delta > nr_pagecache_reclaimable))
2659                 delta = nr_pagecache_reclaimable;
2660
2661         return nr_pagecache_reclaimable - delta;
2662 }
2663
2664 /*
2665  * Try to free up some pages from this zone through reclaim.
2666  */
2667 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2668 {
2669         /* Minimum pages needed in order to stay on node */
2670         const unsigned long nr_pages = 1 << order;
2671         struct task_struct *p = current;
2672         struct reclaim_state reclaim_state;
2673         int priority;
2674         struct scan_control sc = {
2675                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2676                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2677                 .may_swap = 1,
2678                 .nr_to_reclaim = max_t(unsigned long, nr_pages,
2679                                        SWAP_CLUSTER_MAX),
2680                 .gfp_mask = gfp_mask,
2681                 .swappiness = vm_swappiness,
2682                 .order = order,
2683         };
2684         unsigned long nr_slab_pages0, nr_slab_pages1;
2685
2686         cond_resched();
2687         /*
2688          * We need to be able to allocate from the reserves for RECLAIM_SWAP
2689          * and we also need to be able to write out pages for RECLAIM_WRITE
2690          * and RECLAIM_SWAP.
2691          */
2692         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2693         lockdep_set_current_reclaim_state(gfp_mask);
2694         reclaim_state.reclaimed_slab = 0;
2695         p->reclaim_state = &reclaim_state;
2696
2697         if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
2698                 /*
2699                  * Free memory by calling shrink zone with increasing
2700                  * priorities until we have enough memory freed.
2701                  */
2702                 priority = ZONE_RECLAIM_PRIORITY;
2703                 do {
2704                         shrink_zone(priority, zone, &sc);
2705                         priority--;
2706                 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
2707         }
2708
2709         nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2710         if (nr_slab_pages0 > zone->min_slab_pages) {
2711                 /*
2712                  * shrink_slab() does not currently allow us to determine how
2713                  * many pages were freed in this zone. So we take the current
2714                  * number of slab pages and shake the slab until it is reduced
2715                  * by the same nr_pages that we used for reclaiming unmapped
2716                  * pages.
2717                  *
2718                  * Note that shrink_slab will free memory on all zones and may
2719                  * take a long time.
2720                  */
2721                 for (;;) {
2722                         unsigned long lru_pages = zone_reclaimable_pages(zone);
2723
2724                         /* No reclaimable slab or very low memory pressure */
2725                         if (!shrink_slab(sc.nr_scanned, gfp_mask, lru_pages))
2726                                 break;
2727
2728                         /* Freed enough memory */
2729                         nr_slab_pages1 = zone_page_state(zone,
2730                                                         NR_SLAB_RECLAIMABLE);
2731                         if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
2732                                 break;
2733                 }
2734
2735                 /*
2736                  * Update nr_reclaimed by the number of slab pages we
2737                  * reclaimed from this zone.
2738                  */
2739                 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2740                 if (nr_slab_pages1 < nr_slab_pages0)
2741                         sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
2742         }
2743
2744         p->reclaim_state = NULL;
2745         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
2746         lockdep_clear_current_reclaim_state();
2747         return sc.nr_reclaimed >= nr_pages;
2748 }
2749
2750 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2751 {
2752         int node_id;
2753         int ret;
2754
2755         /*
2756          * Zone reclaim reclaims unmapped file backed pages and
2757          * slab pages if we are over the defined limits.
2758          *
2759          * A small portion of unmapped file backed pages is needed for
2760          * file I/O otherwise pages read by file I/O will be immediately
2761          * thrown out if the zone is overallocated. So we do not reclaim
2762          * if less than a specified percentage of the zone is used by
2763          * unmapped file backed pages.
2764          */
2765         if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
2766             zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
2767                 return ZONE_RECLAIM_FULL;
2768
2769         if (zone->all_unreclaimable)
2770                 return ZONE_RECLAIM_FULL;
2771
2772         /*
2773          * Do not scan if the allocation should not be delayed.
2774          */
2775         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
2776                 return ZONE_RECLAIM_NOSCAN;
2777
2778         /*
2779          * Only run zone reclaim on the local zone or on zones that do not
2780          * have associated processors. This will favor the local processor
2781          * over remote processors and spread off node memory allocations
2782          * as wide as possible.
2783          */
2784         node_id = zone_to_nid(zone);
2785         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
2786                 return ZONE_RECLAIM_NOSCAN;
2787
2788         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2789                 return ZONE_RECLAIM_NOSCAN;
2790
2791         ret = __zone_reclaim(zone, gfp_mask, order);
2792         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2793
2794         if (!ret)
2795                 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
2796
2797         return ret;
2798 }
2799 #endif
2800
2801 /*
2802  * page_evictable - test whether a page is evictable
2803  * @page: the page to test
2804  * @vma: the VMA in which the page is or will be mapped, may be NULL
2805  *
2806  * Test whether page is evictable--i.e., should be placed on active/inactive
2807  * lists vs unevictable list.  The vma argument is !NULL when called from the
2808  * fault path to determine how to instantate a new page.
2809  *
2810  * Reasons page might not be evictable:
2811  * (1) page's mapping marked unevictable
2812  * (2) page is part of an mlocked VMA
2813  *
2814  */
2815 int page_evictable(struct page *page, struct vm_area_struct *vma)
2816 {
2817
2818         if (mapping_unevictable(page_mapping(page)))
2819                 return 0;
2820
2821         if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2822                 return 0;
2823
2824         return 1;
2825 }
2826
2827 /**
2828  * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2829  * @page: page to check evictability and move to appropriate lru list
2830  * @zone: zone page is in
2831  *
2832  * Checks a page for evictability and moves the page to the appropriate
2833  * zone lru list.
2834  *
2835  * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2836  * have PageUnevictable set.
2837  */
2838 static void check_move_unevictable_page(struct page *page, struct zone *zone)
2839 {
2840         VM_BUG_ON(PageActive(page));
2841
2842 retry:
2843         ClearPageUnevictable(page);
2844         if (page_evictable(page, NULL)) {
2845                 enum lru_list l = page_lru_base_type(page);
2846
2847                 __dec_zone_state(zone, NR_UNEVICTABLE);
2848                 list_move(&page->lru, &zone->lru[l].list);
2849                 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
2850                 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
2851                 __count_vm_event(UNEVICTABLE_PGRESCUED);
2852         } else {
2853                 /*
2854                  * rotate unevictable list
2855                  */
2856                 SetPageUnevictable(page);
2857                 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
2858                 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
2859                 if (page_evictable(page, NULL))
2860                         goto retry;
2861         }
2862 }
2863
2864 /**
2865  * scan_mapping_unevictable_pages - scan an address space for evictable pages
2866  * @mapping: struct address_space to scan for evictable pages
2867  *
2868  * Scan all pages in mapping.  Check unevictable pages for
2869  * evictability and move them to the appropriate zone lru list.
2870  */
2871 void scan_mapping_unevictable_pages(struct address_space *mapping)
2872 {
2873         pgoff_t next = 0;
2874         pgoff_t end   = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
2875                          PAGE_CACHE_SHIFT;
2876         struct zone *zone;
2877         struct pagevec pvec;
2878
2879         if (mapping->nrpages == 0)
2880                 return;
2881
2882         pagevec_init(&pvec, 0);
2883         while (next < end &&
2884                 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
2885                 int i;
2886                 int pg_scanned = 0;
2887
2888                 zone = NULL;
2889
2890                 for (i = 0; i < pagevec_count(&pvec); i++) {
2891                         struct page *page = pvec.pages[i];
2892                         pgoff_t page_index = page->index;
2893                         struct zone *pagezone = page_zone(page);
2894
2895                         pg_scanned++;
2896                         if (page_index > next)
2897                                 next = page_index;
2898                         next++;
2899
2900                         if (pagezone != zone) {
2901                                 if (zone)
2902                                         spin_unlock_irq(&zone->lru_lock);
2903                                 zone = pagezone;
2904                                 spin_lock_irq(&zone->lru_lock);
2905                         }
2906
2907                         if (PageLRU(page) && PageUnevictable(page))
2908                                 check_move_unevictable_page(page, zone);
2909                 }
2910                 if (zone)
2911                         spin_unlock_irq(&zone->lru_lock);
2912                 pagevec_release(&pvec);
2913
2914                 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
2915         }
2916
2917 }
2918
2919 /**
2920  * scan_zone_unevictable_pages - check unevictable list for evictable pages
2921  * @zone - zone of which to scan the unevictable list
2922  *
2923  * Scan @zone's unevictable LRU lists to check for pages that have become
2924  * evictable.  Move those that have to @zone's inactive list where they
2925  * become candidates for reclaim, unless shrink_inactive_zone() decides
2926  * to reactivate them.  Pages that are still unevictable are rotated
2927  * back onto @zone's unevictable list.
2928  */
2929 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
2930 static void scan_zone_unevictable_pages(struct zone *zone)
2931 {
2932         struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
2933         unsigned long scan;
2934         unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
2935
2936         while (nr_to_scan > 0) {
2937                 unsigned long batch_size = min(nr_to_scan,
2938                                                 SCAN_UNEVICTABLE_BATCH_SIZE);
2939
2940                 spin_lock_irq(&zone->lru_lock);
2941                 for (scan = 0;  scan < batch_size; scan++) {
2942                         struct page *page = lru_to_page(l_unevictable);
2943
2944                         if (!trylock_page(page))
2945                                 continue;
2946
2947                         prefetchw_prev_lru_page(page, l_unevictable, flags);
2948
2949                         if (likely(PageLRU(page) && PageUnevictable(page)))
2950                                 check_move_unevictable_page(page, zone);
2951
2952                         unlock_page(page);
2953                 }
2954                 spin_unlock_irq(&zone->lru_lock);
2955
2956                 nr_to_scan -= batch_size;
2957         }
2958 }
2959
2960
2961 /**
2962  * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
2963  *
2964  * A really big hammer:  scan all zones' unevictable LRU lists to check for
2965  * pages that have become evictable.  Move those back to the zones'
2966  * inactive list where they become candidates for reclaim.
2967  * This occurs when, e.g., we have unswappable pages on the unevictable lists,
2968  * and we add swap to the system.  As such, it runs in the context of a task
2969  * that has possibly/probably made some previously unevictable pages
2970  * evictable.
2971  */
2972 static void scan_all_zones_unevictable_pages(void)
2973 {
2974         struct zone *zone;
2975
2976         for_each_zone(zone) {
2977                 scan_zone_unevictable_pages(zone);
2978         }
2979 }
2980
2981 /*
2982  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
2983  * all nodes' unevictable lists for evictable pages
2984  */
2985 unsigned long scan_unevictable_pages;
2986
2987 int scan_unevictable_handler(struct ctl_table *table, int write,
2988                            void __user *buffer,
2989                            size_t *length, loff_t *ppos)
2990 {
2991         proc_doulongvec_minmax(table, write, buffer, length, ppos);
2992
2993         if (write && *(unsigned long *)table->data)
2994                 scan_all_zones_unevictable_pages();
2995
2996         scan_unevictable_pages = 0;
2997         return 0;
2998 }
2999
3000 #ifdef CONFIG_NUMA
3001 /*
3002  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
3003  * a specified node's per zone unevictable lists for evictable pages.
3004  */
3005
3006 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
3007                                           struct sysdev_attribute *attr,
3008                                           char *buf)
3009 {
3010         return sprintf(buf, "0\n");     /* always zero; should fit... */
3011 }
3012
3013 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
3014                                            struct sysdev_attribute *attr,
3015                                         const char *buf, size_t count)
3016 {
3017         struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
3018         struct zone *zone;
3019         unsigned long res;
3020         unsigned long req = strict_strtoul(buf, 10, &res);
3021
3022         if (!req)
3023                 return 1;       /* zero is no-op */
3024
3025         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
3026                 if (!populated_zone(zone))
3027                         continue;
3028                 scan_zone_unevictable_pages(zone);
3029         }
3030         return 1;
3031 }
3032
3033
3034 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3035                         read_scan_unevictable_node,
3036                         write_scan_unevictable_node);
3037
3038 int scan_unevictable_register_node(struct node *node)
3039 {
3040         return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
3041 }
3042
3043 void scan_unevictable_unregister_node(struct node *node)
3044 {
3045         sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
3046 }
3047 #endif