vmscan: add customisable shrinker batch size
[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/compaction.h>
36 #include <linux/notifier.h>
37 #include <linux/rwsem.h>
38 #include <linux/delay.h>
39 #include <linux/kthread.h>
40 #include <linux/freezer.h>
41 #include <linux/memcontrol.h>
42 #include <linux/delayacct.h>
43 #include <linux/sysctl.h>
44 #include <linux/oom.h>
45 #include <linux/prefetch.h>
46
47 #include <asm/tlbflush.h>
48 #include <asm/div64.h>
49
50 #include <linux/swapops.h>
51
52 #include "internal.h"
53
54 #define CREATE_TRACE_POINTS
55 #include <trace/events/vmscan.h>
56
57 /*
58  * reclaim_mode determines how the inactive list is shrunk
59  * RECLAIM_MODE_SINGLE: Reclaim only order-0 pages
60  * RECLAIM_MODE_ASYNC:  Do not block
61  * RECLAIM_MODE_SYNC:   Allow blocking e.g. call wait_on_page_writeback
62  * RECLAIM_MODE_LUMPYRECLAIM: For high-order allocations, take a reference
63  *                      page from the LRU and reclaim all pages within a
64  *                      naturally aligned range
65  * RECLAIM_MODE_COMPACTION: For high-order allocations, reclaim a number of
66  *                      order-0 pages and then compact the zone
67  */
68 typedef unsigned __bitwise__ reclaim_mode_t;
69 #define RECLAIM_MODE_SINGLE             ((__force reclaim_mode_t)0x01u)
70 #define RECLAIM_MODE_ASYNC              ((__force reclaim_mode_t)0x02u)
71 #define RECLAIM_MODE_SYNC               ((__force reclaim_mode_t)0x04u)
72 #define RECLAIM_MODE_LUMPYRECLAIM       ((__force reclaim_mode_t)0x08u)
73 #define RECLAIM_MODE_COMPACTION         ((__force reclaim_mode_t)0x10u)
74
75 struct scan_control {
76         /* Incremented by the number of inactive pages that were scanned */
77         unsigned long nr_scanned;
78
79         /* Number of pages freed so far during a call to shrink_zones() */
80         unsigned long nr_reclaimed;
81
82         /* How many pages shrink_list() should reclaim */
83         unsigned long nr_to_reclaim;
84
85         unsigned long hibernation_mode;
86
87         /* This context's GFP mask */
88         gfp_t gfp_mask;
89
90         int may_writepage;
91
92         /* Can mapped pages be reclaimed? */
93         int may_unmap;
94
95         /* Can pages be swapped as part of reclaim? */
96         int may_swap;
97
98         int swappiness;
99
100         int order;
101
102         /*
103          * Intend to reclaim enough continuous memory rather than reclaim
104          * enough amount of memory. i.e, mode for high order allocation.
105          */
106         reclaim_mode_t reclaim_mode;
107
108         /* Which cgroup do we reclaim from */
109         struct mem_cgroup *mem_cgroup;
110
111         /*
112          * Nodemask of nodes allowed by the caller. If NULL, all nodes
113          * are scanned.
114          */
115         nodemask_t      *nodemask;
116 };
117
118 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
119
120 #ifdef ARCH_HAS_PREFETCH
121 #define prefetch_prev_lru_page(_page, _base, _field)                    \
122         do {                                                            \
123                 if ((_page)->lru.prev != _base) {                       \
124                         struct page *prev;                              \
125                                                                         \
126                         prev = lru_to_page(&(_page->lru));              \
127                         prefetch(&prev->_field);                        \
128                 }                                                       \
129         } while (0)
130 #else
131 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
132 #endif
133
134 #ifdef ARCH_HAS_PREFETCHW
135 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
136         do {                                                            \
137                 if ((_page)->lru.prev != _base) {                       \
138                         struct page *prev;                              \
139                                                                         \
140                         prev = lru_to_page(&(_page->lru));              \
141                         prefetchw(&prev->_field);                       \
142                 }                                                       \
143         } while (0)
144 #else
145 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
146 #endif
147
148 /*
149  * From 0 .. 100.  Higher means more swappy.
150  */
151 int vm_swappiness = 60;
152 long vm_total_pages;    /* The total number of pages which the VM controls */
153
154 static LIST_HEAD(shrinker_list);
155 static DECLARE_RWSEM(shrinker_rwsem);
156
157 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
158 #define scanning_global_lru(sc) (!(sc)->mem_cgroup)
159 #else
160 #define scanning_global_lru(sc) (1)
161 #endif
162
163 static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
164                                                   struct scan_control *sc)
165 {
166         if (!scanning_global_lru(sc))
167                 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
168
169         return &zone->reclaim_stat;
170 }
171
172 static unsigned long zone_nr_lru_pages(struct zone *zone,
173                                 struct scan_control *sc, enum lru_list lru)
174 {
175         if (!scanning_global_lru(sc))
176                 return mem_cgroup_zone_nr_lru_pages(sc->mem_cgroup, zone, lru);
177
178         return zone_page_state(zone, NR_LRU_BASE + lru);
179 }
180
181
182 /*
183  * Add a shrinker callback to be called from the vm
184  */
185 void register_shrinker(struct shrinker *shrinker)
186 {
187         shrinker->nr = 0;
188         down_write(&shrinker_rwsem);
189         list_add_tail(&shrinker->list, &shrinker_list);
190         up_write(&shrinker_rwsem);
191 }
192 EXPORT_SYMBOL(register_shrinker);
193
194 /*
195  * Remove one
196  */
197 void unregister_shrinker(struct shrinker *shrinker)
198 {
199         down_write(&shrinker_rwsem);
200         list_del(&shrinker->list);
201         up_write(&shrinker_rwsem);
202 }
203 EXPORT_SYMBOL(unregister_shrinker);
204
205 static inline int do_shrinker_shrink(struct shrinker *shrinker,
206                                      struct shrink_control *sc,
207                                      unsigned long nr_to_scan)
208 {
209         sc->nr_to_scan = nr_to_scan;
210         return (*shrinker->shrink)(shrinker, sc);
211 }
212
213 #define SHRINK_BATCH 128
214 /*
215  * Call the shrink functions to age shrinkable caches
216  *
217  * Here we assume it costs one seek to replace a lru page and that it also
218  * takes a seek to recreate a cache object.  With this in mind we age equal
219  * percentages of the lru and ageable caches.  This should balance the seeks
220  * generated by these structures.
221  *
222  * If the vm encountered mapped pages on the LRU it increase the pressure on
223  * slab to avoid swapping.
224  *
225  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
226  *
227  * `lru_pages' represents the number of on-LRU pages in all the zones which
228  * are eligible for the caller's allocation attempt.  It is used for balancing
229  * slab reclaim versus page reclaim.
230  *
231  * Returns the number of slab objects which we shrunk.
232  */
233 unsigned long shrink_slab(struct shrink_control *shrink,
234                           unsigned long nr_pages_scanned,
235                           unsigned long lru_pages)
236 {
237         struct shrinker *shrinker;
238         unsigned long ret = 0;
239
240         if (nr_pages_scanned == 0)
241                 nr_pages_scanned = SWAP_CLUSTER_MAX;
242
243         if (!down_read_trylock(&shrinker_rwsem)) {
244                 /* Assume we'll be able to shrink next time */
245                 ret = 1;
246                 goto out;
247         }
248
249         list_for_each_entry(shrinker, &shrinker_list, list) {
250                 unsigned long long delta;
251                 unsigned long total_scan;
252                 unsigned long max_pass;
253                 int shrink_ret = 0;
254                 long nr;
255                 long new_nr;
256                 long batch_size = shrinker->batch ? shrinker->batch
257                                                   : SHRINK_BATCH;
258
259                 /*
260                  * copy the current shrinker scan count into a local variable
261                  * and zero it so that other concurrent shrinker invocations
262                  * don't also do this scanning work.
263                  */
264                 do {
265                         nr = shrinker->nr;
266                 } while (cmpxchg(&shrinker->nr, nr, 0) != nr);
267
268                 total_scan = nr;
269                 max_pass = do_shrinker_shrink(shrinker, shrink, 0);
270                 delta = (4 * nr_pages_scanned) / shrinker->seeks;
271                 delta *= max_pass;
272                 do_div(delta, lru_pages + 1);
273                 total_scan += delta;
274                 if (total_scan < 0) {
275                         printk(KERN_ERR "shrink_slab: %pF negative objects to "
276                                "delete nr=%ld\n",
277                                shrinker->shrink, total_scan);
278                         total_scan = max_pass;
279                 }
280
281                 /*
282                  * We need to avoid excessive windup on filesystem shrinkers
283                  * due to large numbers of GFP_NOFS allocations causing the
284                  * shrinkers to return -1 all the time. This results in a large
285                  * nr being built up so when a shrink that can do some work
286                  * comes along it empties the entire cache due to nr >>>
287                  * max_pass.  This is bad for sustaining a working set in
288                  * memory.
289                  *
290                  * Hence only allow the shrinker to scan the entire cache when
291                  * a large delta change is calculated directly.
292                  */
293                 if (delta < max_pass / 4)
294                         total_scan = min(total_scan, max_pass / 2);
295
296                 /*
297                  * Avoid risking looping forever due to too large nr value:
298                  * never try to free more than twice the estimate number of
299                  * freeable entries.
300                  */
301                 if (total_scan > max_pass * 2)
302                         total_scan = max_pass * 2;
303
304                 trace_mm_shrink_slab_start(shrinker, shrink, nr,
305                                         nr_pages_scanned, lru_pages,
306                                         max_pass, delta, total_scan);
307
308                 while (total_scan >= batch_size) {
309                         int nr_before;
310
311                         nr_before = do_shrinker_shrink(shrinker, shrink, 0);
312                         shrink_ret = do_shrinker_shrink(shrinker, shrink,
313                                                         batch_size);
314                         if (shrink_ret == -1)
315                                 break;
316                         if (shrink_ret < nr_before)
317                                 ret += nr_before - shrink_ret;
318                         count_vm_events(SLABS_SCANNED, batch_size);
319                         total_scan -= batch_size;
320
321                         cond_resched();
322                 }
323
324                 /*
325                  * move the unused scan count back into the shrinker in a
326                  * manner that handles concurrent updates. If we exhausted the
327                  * scan, there is no need to do an update.
328                  */
329                 do {
330                         nr = shrinker->nr;
331                         new_nr = total_scan + nr;
332                         if (total_scan <= 0)
333                                 break;
334                 } while (cmpxchg(&shrinker->nr, nr, new_nr) != nr);
335
336                 trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr);
337         }
338         up_read(&shrinker_rwsem);
339 out:
340         cond_resched();
341         return ret;
342 }
343
344 static void set_reclaim_mode(int priority, struct scan_control *sc,
345                                    bool sync)
346 {
347         reclaim_mode_t syncmode = sync ? RECLAIM_MODE_SYNC : RECLAIM_MODE_ASYNC;
348
349         /*
350          * Initially assume we are entering either lumpy reclaim or
351          * reclaim/compaction.Depending on the order, we will either set the
352          * sync mode or just reclaim order-0 pages later.
353          */
354         if (COMPACTION_BUILD)
355                 sc->reclaim_mode = RECLAIM_MODE_COMPACTION;
356         else
357                 sc->reclaim_mode = RECLAIM_MODE_LUMPYRECLAIM;
358
359         /*
360          * Avoid using lumpy reclaim or reclaim/compaction if possible by
361          * restricting when its set to either costly allocations or when
362          * under memory pressure
363          */
364         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
365                 sc->reclaim_mode |= syncmode;
366         else if (sc->order && priority < DEF_PRIORITY - 2)
367                 sc->reclaim_mode |= syncmode;
368         else
369                 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
370 }
371
372 static void reset_reclaim_mode(struct scan_control *sc)
373 {
374         sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
375 }
376
377 static inline int is_page_cache_freeable(struct page *page)
378 {
379         /*
380          * A freeable page cache page is referenced only by the caller
381          * that isolated the page, the page cache radix tree and
382          * optional buffer heads at page->private.
383          */
384         return page_count(page) - page_has_private(page) == 2;
385 }
386
387 static int may_write_to_queue(struct backing_dev_info *bdi,
388                               struct scan_control *sc)
389 {
390         if (current->flags & PF_SWAPWRITE)
391                 return 1;
392         if (!bdi_write_congested(bdi))
393                 return 1;
394         if (bdi == current->backing_dev_info)
395                 return 1;
396
397         /* lumpy reclaim for hugepage often need a lot of write */
398         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
399                 return 1;
400         return 0;
401 }
402
403 /*
404  * We detected a synchronous write error writing a page out.  Probably
405  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
406  * fsync(), msync() or close().
407  *
408  * The tricky part is that after writepage we cannot touch the mapping: nothing
409  * prevents it from being freed up.  But we have a ref on the page and once
410  * that page is locked, the mapping is pinned.
411  *
412  * We're allowed to run sleeping lock_page() here because we know the caller has
413  * __GFP_FS.
414  */
415 static void handle_write_error(struct address_space *mapping,
416                                 struct page *page, int error)
417 {
418         lock_page(page);
419         if (page_mapping(page) == mapping)
420                 mapping_set_error(mapping, error);
421         unlock_page(page);
422 }
423
424 /* possible outcome of pageout() */
425 typedef enum {
426         /* failed to write page out, page is locked */
427         PAGE_KEEP,
428         /* move page to the active list, page is locked */
429         PAGE_ACTIVATE,
430         /* page has been sent to the disk successfully, page is unlocked */
431         PAGE_SUCCESS,
432         /* page is clean and locked */
433         PAGE_CLEAN,
434 } pageout_t;
435
436 /*
437  * pageout is called by shrink_page_list() for each dirty page.
438  * Calls ->writepage().
439  */
440 static pageout_t pageout(struct page *page, struct address_space *mapping,
441                          struct scan_control *sc)
442 {
443         /*
444          * If the page is dirty, only perform writeback if that write
445          * will be non-blocking.  To prevent this allocation from being
446          * stalled by pagecache activity.  But note that there may be
447          * stalls if we need to run get_block().  We could test
448          * PagePrivate for that.
449          *
450          * If this process is currently in __generic_file_aio_write() against
451          * this page's queue, we can perform writeback even if that
452          * will block.
453          *
454          * If the page is swapcache, write it back even if that would
455          * block, for some throttling. This happens by accident, because
456          * swap_backing_dev_info is bust: it doesn't reflect the
457          * congestion state of the swapdevs.  Easy to fix, if needed.
458          */
459         if (!is_page_cache_freeable(page))
460                 return PAGE_KEEP;
461         if (!mapping) {
462                 /*
463                  * Some data journaling orphaned pages can have
464                  * page->mapping == NULL while being dirty with clean buffers.
465                  */
466                 if (page_has_private(page)) {
467                         if (try_to_free_buffers(page)) {
468                                 ClearPageDirty(page);
469                                 printk("%s: orphaned page\n", __func__);
470                                 return PAGE_CLEAN;
471                         }
472                 }
473                 return PAGE_KEEP;
474         }
475         if (mapping->a_ops->writepage == NULL)
476                 return PAGE_ACTIVATE;
477         if (!may_write_to_queue(mapping->backing_dev_info, sc))
478                 return PAGE_KEEP;
479
480         if (clear_page_dirty_for_io(page)) {
481                 int res;
482                 struct writeback_control wbc = {
483                         .sync_mode = WB_SYNC_NONE,
484                         .nr_to_write = SWAP_CLUSTER_MAX,
485                         .range_start = 0,
486                         .range_end = LLONG_MAX,
487                         .for_reclaim = 1,
488                 };
489
490                 SetPageReclaim(page);
491                 res = mapping->a_ops->writepage(page, &wbc);
492                 if (res < 0)
493                         handle_write_error(mapping, page, res);
494                 if (res == AOP_WRITEPAGE_ACTIVATE) {
495                         ClearPageReclaim(page);
496                         return PAGE_ACTIVATE;
497                 }
498
499                 /*
500                  * Wait on writeback if requested to. This happens when
501                  * direct reclaiming a large contiguous area and the
502                  * first attempt to free a range of pages fails.
503                  */
504                 if (PageWriteback(page) &&
505                     (sc->reclaim_mode & RECLAIM_MODE_SYNC))
506                         wait_on_page_writeback(page);
507
508                 if (!PageWriteback(page)) {
509                         /* synchronous write or broken a_ops? */
510                         ClearPageReclaim(page);
511                 }
512                 trace_mm_vmscan_writepage(page,
513                         trace_reclaim_flags(page, sc->reclaim_mode));
514                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
515                 return PAGE_SUCCESS;
516         }
517
518         return PAGE_CLEAN;
519 }
520
521 /*
522  * Same as remove_mapping, but if the page is removed from the mapping, it
523  * gets returned with a refcount of 0.
524  */
525 static int __remove_mapping(struct address_space *mapping, struct page *page)
526 {
527         BUG_ON(!PageLocked(page));
528         BUG_ON(mapping != page_mapping(page));
529
530         spin_lock_irq(&mapping->tree_lock);
531         /*
532          * The non racy check for a busy page.
533          *
534          * Must be careful with the order of the tests. When someone has
535          * a ref to the page, it may be possible that they dirty it then
536          * drop the reference. So if PageDirty is tested before page_count
537          * here, then the following race may occur:
538          *
539          * get_user_pages(&page);
540          * [user mapping goes away]
541          * write_to(page);
542          *                              !PageDirty(page)    [good]
543          * SetPageDirty(page);
544          * put_page(page);
545          *                              !page_count(page)   [good, discard it]
546          *
547          * [oops, our write_to data is lost]
548          *
549          * Reversing the order of the tests ensures such a situation cannot
550          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
551          * load is not satisfied before that of page->_count.
552          *
553          * Note that if SetPageDirty is always performed via set_page_dirty,
554          * and thus under tree_lock, then this ordering is not required.
555          */
556         if (!page_freeze_refs(page, 2))
557                 goto cannot_free;
558         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
559         if (unlikely(PageDirty(page))) {
560                 page_unfreeze_refs(page, 2);
561                 goto cannot_free;
562         }
563
564         if (PageSwapCache(page)) {
565                 swp_entry_t swap = { .val = page_private(page) };
566                 __delete_from_swap_cache(page);
567                 spin_unlock_irq(&mapping->tree_lock);
568                 swapcache_free(swap, page);
569         } else {
570                 void (*freepage)(struct page *);
571
572                 freepage = mapping->a_ops->freepage;
573
574                 __delete_from_page_cache(page);
575                 spin_unlock_irq(&mapping->tree_lock);
576                 mem_cgroup_uncharge_cache_page(page);
577
578                 if (freepage != NULL)
579                         freepage(page);
580         }
581
582         return 1;
583
584 cannot_free:
585         spin_unlock_irq(&mapping->tree_lock);
586         return 0;
587 }
588
589 /*
590  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
591  * someone else has a ref on the page, abort and return 0.  If it was
592  * successfully detached, return 1.  Assumes the caller has a single ref on
593  * this page.
594  */
595 int remove_mapping(struct address_space *mapping, struct page *page)
596 {
597         if (__remove_mapping(mapping, page)) {
598                 /*
599                  * Unfreezing the refcount with 1 rather than 2 effectively
600                  * drops the pagecache ref for us without requiring another
601                  * atomic operation.
602                  */
603                 page_unfreeze_refs(page, 1);
604                 return 1;
605         }
606         return 0;
607 }
608
609 /**
610  * putback_lru_page - put previously isolated page onto appropriate LRU list
611  * @page: page to be put back to appropriate lru list
612  *
613  * Add previously isolated @page to appropriate LRU list.
614  * Page may still be unevictable for other reasons.
615  *
616  * lru_lock must not be held, interrupts must be enabled.
617  */
618 void putback_lru_page(struct page *page)
619 {
620         int lru;
621         int active = !!TestClearPageActive(page);
622         int was_unevictable = PageUnevictable(page);
623
624         VM_BUG_ON(PageLRU(page));
625
626 redo:
627         ClearPageUnevictable(page);
628
629         if (page_evictable(page, NULL)) {
630                 /*
631                  * For evictable pages, we can use the cache.
632                  * In event of a race, worst case is we end up with an
633                  * unevictable page on [in]active list.
634                  * We know how to handle that.
635                  */
636                 lru = active + page_lru_base_type(page);
637                 lru_cache_add_lru(page, lru);
638         } else {
639                 /*
640                  * Put unevictable pages directly on zone's unevictable
641                  * list.
642                  */
643                 lru = LRU_UNEVICTABLE;
644                 add_page_to_unevictable_list(page);
645                 /*
646                  * When racing with an mlock clearing (page is
647                  * unlocked), make sure that if the other thread does
648                  * not observe our setting of PG_lru and fails
649                  * isolation, we see PG_mlocked cleared below and move
650                  * the page back to the evictable list.
651                  *
652                  * The other side is TestClearPageMlocked().
653                  */
654                 smp_mb();
655         }
656
657         /*
658          * page's status can change while we move it among lru. If an evictable
659          * page is on unevictable list, it never be freed. To avoid that,
660          * check after we added it to the list, again.
661          */
662         if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
663                 if (!isolate_lru_page(page)) {
664                         put_page(page);
665                         goto redo;
666                 }
667                 /* This means someone else dropped this page from LRU
668                  * So, it will be freed or putback to LRU again. There is
669                  * nothing to do here.
670                  */
671         }
672
673         if (was_unevictable && lru != LRU_UNEVICTABLE)
674                 count_vm_event(UNEVICTABLE_PGRESCUED);
675         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
676                 count_vm_event(UNEVICTABLE_PGCULLED);
677
678         put_page(page);         /* drop ref from isolate */
679 }
680
681 enum page_references {
682         PAGEREF_RECLAIM,
683         PAGEREF_RECLAIM_CLEAN,
684         PAGEREF_KEEP,
685         PAGEREF_ACTIVATE,
686 };
687
688 static enum page_references page_check_references(struct page *page,
689                                                   struct scan_control *sc)
690 {
691         int referenced_ptes, referenced_page;
692         unsigned long vm_flags;
693
694         referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
695         referenced_page = TestClearPageReferenced(page);
696
697         /* Lumpy reclaim - ignore references */
698         if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
699                 return PAGEREF_RECLAIM;
700
701         /*
702          * Mlock lost the isolation race with us.  Let try_to_unmap()
703          * move the page to the unevictable list.
704          */
705         if (vm_flags & VM_LOCKED)
706                 return PAGEREF_RECLAIM;
707
708         if (referenced_ptes) {
709                 if (PageAnon(page))
710                         return PAGEREF_ACTIVATE;
711                 /*
712                  * All mapped pages start out with page table
713                  * references from the instantiating fault, so we need
714                  * to look twice if a mapped file page is used more
715                  * than once.
716                  *
717                  * Mark it and spare it for another trip around the
718                  * inactive list.  Another page table reference will
719                  * lead to its activation.
720                  *
721                  * Note: the mark is set for activated pages as well
722                  * so that recently deactivated but used pages are
723                  * quickly recovered.
724                  */
725                 SetPageReferenced(page);
726
727                 if (referenced_page)
728                         return PAGEREF_ACTIVATE;
729
730                 return PAGEREF_KEEP;
731         }
732
733         /* Reclaim if clean, defer dirty pages to writeback */
734         if (referenced_page && !PageSwapBacked(page))
735                 return PAGEREF_RECLAIM_CLEAN;
736
737         return PAGEREF_RECLAIM;
738 }
739
740 static noinline_for_stack void free_page_list(struct list_head *free_pages)
741 {
742         struct pagevec freed_pvec;
743         struct page *page, *tmp;
744
745         pagevec_init(&freed_pvec, 1);
746
747         list_for_each_entry_safe(page, tmp, free_pages, lru) {
748                 list_del(&page->lru);
749                 if (!pagevec_add(&freed_pvec, page)) {
750                         __pagevec_free(&freed_pvec);
751                         pagevec_reinit(&freed_pvec);
752                 }
753         }
754
755         pagevec_free(&freed_pvec);
756 }
757
758 /*
759  * shrink_page_list() returns the number of reclaimed pages
760  */
761 static unsigned long shrink_page_list(struct list_head *page_list,
762                                       struct zone *zone,
763                                       struct scan_control *sc)
764 {
765         LIST_HEAD(ret_pages);
766         LIST_HEAD(free_pages);
767         int pgactivate = 0;
768         unsigned long nr_dirty = 0;
769         unsigned long nr_congested = 0;
770         unsigned long nr_reclaimed = 0;
771
772         cond_resched();
773
774         while (!list_empty(page_list)) {
775                 enum page_references references;
776                 struct address_space *mapping;
777                 struct page *page;
778                 int may_enter_fs;
779
780                 cond_resched();
781
782                 page = lru_to_page(page_list);
783                 list_del(&page->lru);
784
785                 if (!trylock_page(page))
786                         goto keep;
787
788                 VM_BUG_ON(PageActive(page));
789                 VM_BUG_ON(page_zone(page) != zone);
790
791                 sc->nr_scanned++;
792
793                 if (unlikely(!page_evictable(page, NULL)))
794                         goto cull_mlocked;
795
796                 if (!sc->may_unmap && page_mapped(page))
797                         goto keep_locked;
798
799                 /* Double the slab pressure for mapped and swapcache pages */
800                 if (page_mapped(page) || PageSwapCache(page))
801                         sc->nr_scanned++;
802
803                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
804                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
805
806                 if (PageWriteback(page)) {
807                         /*
808                          * Synchronous reclaim is performed in two passes,
809                          * first an asynchronous pass over the list to
810                          * start parallel writeback, and a second synchronous
811                          * pass to wait for the IO to complete.  Wait here
812                          * for any page for which writeback has already
813                          * started.
814                          */
815                         if ((sc->reclaim_mode & RECLAIM_MODE_SYNC) &&
816                             may_enter_fs)
817                                 wait_on_page_writeback(page);
818                         else {
819                                 unlock_page(page);
820                                 goto keep_lumpy;
821                         }
822                 }
823
824                 references = page_check_references(page, sc);
825                 switch (references) {
826                 case PAGEREF_ACTIVATE:
827                         goto activate_locked;
828                 case PAGEREF_KEEP:
829                         goto keep_locked;
830                 case PAGEREF_RECLAIM:
831                 case PAGEREF_RECLAIM_CLEAN:
832                         ; /* try to reclaim the page below */
833                 }
834
835                 /*
836                  * Anonymous process memory has backing store?
837                  * Try to allocate it some swap space here.
838                  */
839                 if (PageAnon(page) && !PageSwapCache(page)) {
840                         if (!(sc->gfp_mask & __GFP_IO))
841                                 goto keep_locked;
842                         if (!add_to_swap(page))
843                                 goto activate_locked;
844                         may_enter_fs = 1;
845                 }
846
847                 mapping = page_mapping(page);
848
849                 /*
850                  * The page is mapped into the page tables of one or more
851                  * processes. Try to unmap it here.
852                  */
853                 if (page_mapped(page) && mapping) {
854                         switch (try_to_unmap(page, TTU_UNMAP)) {
855                         case SWAP_FAIL:
856                                 goto activate_locked;
857                         case SWAP_AGAIN:
858                                 goto keep_locked;
859                         case SWAP_MLOCK:
860                                 goto cull_mlocked;
861                         case SWAP_SUCCESS:
862                                 ; /* try to free the page below */
863                         }
864                 }
865
866                 if (PageDirty(page)) {
867                         nr_dirty++;
868
869                         if (references == PAGEREF_RECLAIM_CLEAN)
870                                 goto keep_locked;
871                         if (!may_enter_fs)
872                                 goto keep_locked;
873                         if (!sc->may_writepage)
874                                 goto keep_locked;
875
876                         /* Page is dirty, try to write it out here */
877                         switch (pageout(page, mapping, sc)) {
878                         case PAGE_KEEP:
879                                 nr_congested++;
880                                 goto keep_locked;
881                         case PAGE_ACTIVATE:
882                                 goto activate_locked;
883                         case PAGE_SUCCESS:
884                                 if (PageWriteback(page))
885                                         goto keep_lumpy;
886                                 if (PageDirty(page))
887                                         goto keep;
888
889                                 /*
890                                  * A synchronous write - probably a ramdisk.  Go
891                                  * ahead and try to reclaim the page.
892                                  */
893                                 if (!trylock_page(page))
894                                         goto keep;
895                                 if (PageDirty(page) || PageWriteback(page))
896                                         goto keep_locked;
897                                 mapping = page_mapping(page);
898                         case PAGE_CLEAN:
899                                 ; /* try to free the page below */
900                         }
901                 }
902
903                 /*
904                  * If the page has buffers, try to free the buffer mappings
905                  * associated with this page. If we succeed we try to free
906                  * the page as well.
907                  *
908                  * We do this even if the page is PageDirty().
909                  * try_to_release_page() does not perform I/O, but it is
910                  * possible for a page to have PageDirty set, but it is actually
911                  * clean (all its buffers are clean).  This happens if the
912                  * buffers were written out directly, with submit_bh(). ext3
913                  * will do this, as well as the blockdev mapping.
914                  * try_to_release_page() will discover that cleanness and will
915                  * drop the buffers and mark the page clean - it can be freed.
916                  *
917                  * Rarely, pages can have buffers and no ->mapping.  These are
918                  * the pages which were not successfully invalidated in
919                  * truncate_complete_page().  We try to drop those buffers here
920                  * and if that worked, and the page is no longer mapped into
921                  * process address space (page_count == 1) it can be freed.
922                  * Otherwise, leave the page on the LRU so it is swappable.
923                  */
924                 if (page_has_private(page)) {
925                         if (!try_to_release_page(page, sc->gfp_mask))
926                                 goto activate_locked;
927                         if (!mapping && page_count(page) == 1) {
928                                 unlock_page(page);
929                                 if (put_page_testzero(page))
930                                         goto free_it;
931                                 else {
932                                         /*
933                                          * rare race with speculative reference.
934                                          * the speculative reference will free
935                                          * this page shortly, so we may
936                                          * increment nr_reclaimed here (and
937                                          * leave it off the LRU).
938                                          */
939                                         nr_reclaimed++;
940                                         continue;
941                                 }
942                         }
943                 }
944
945                 if (!mapping || !__remove_mapping(mapping, page))
946                         goto keep_locked;
947
948                 /*
949                  * At this point, we have no other references and there is
950                  * no way to pick any more up (removed from LRU, removed
951                  * from pagecache). Can use non-atomic bitops now (and
952                  * we obviously don't have to worry about waking up a process
953                  * waiting on the page lock, because there are no references.
954                  */
955                 __clear_page_locked(page);
956 free_it:
957                 nr_reclaimed++;
958
959                 /*
960                  * Is there need to periodically free_page_list? It would
961                  * appear not as the counts should be low
962                  */
963                 list_add(&page->lru, &free_pages);
964                 continue;
965
966 cull_mlocked:
967                 if (PageSwapCache(page))
968                         try_to_free_swap(page);
969                 unlock_page(page);
970                 putback_lru_page(page);
971                 reset_reclaim_mode(sc);
972                 continue;
973
974 activate_locked:
975                 /* Not a candidate for swapping, so reclaim swap space. */
976                 if (PageSwapCache(page) && vm_swap_full())
977                         try_to_free_swap(page);
978                 VM_BUG_ON(PageActive(page));
979                 SetPageActive(page);
980                 pgactivate++;
981 keep_locked:
982                 unlock_page(page);
983 keep:
984                 reset_reclaim_mode(sc);
985 keep_lumpy:
986                 list_add(&page->lru, &ret_pages);
987                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
988         }
989
990         /*
991          * Tag a zone as congested if all the dirty pages encountered were
992          * backed by a congested BDI. In this case, reclaimers should just
993          * back off and wait for congestion to clear because further reclaim
994          * will encounter the same problem
995          */
996         if (nr_dirty && nr_dirty == nr_congested && scanning_global_lru(sc))
997                 zone_set_flag(zone, ZONE_CONGESTED);
998
999         free_page_list(&free_pages);
1000
1001         list_splice(&ret_pages, page_list);
1002         count_vm_events(PGACTIVATE, pgactivate);
1003         return nr_reclaimed;
1004 }
1005
1006 /*
1007  * Attempt to remove the specified page from its LRU.  Only take this page
1008  * if it is of the appropriate PageActive status.  Pages which are being
1009  * freed elsewhere are also ignored.
1010  *
1011  * page:        page to consider
1012  * mode:        one of the LRU isolation modes defined above
1013  *
1014  * returns 0 on success, -ve errno on failure.
1015  */
1016 int __isolate_lru_page(struct page *page, int mode, int file)
1017 {
1018         int ret = -EINVAL;
1019
1020         /* Only take pages on the LRU. */
1021         if (!PageLRU(page))
1022                 return ret;
1023
1024         /*
1025          * When checking the active state, we need to be sure we are
1026          * dealing with comparible boolean values.  Take the logical not
1027          * of each.
1028          */
1029         if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
1030                 return ret;
1031
1032         if (mode != ISOLATE_BOTH && page_is_file_cache(page) != file)
1033                 return ret;
1034
1035         /*
1036          * When this function is being called for lumpy reclaim, we
1037          * initially look into all LRU pages, active, inactive and
1038          * unevictable; only give shrink_page_list evictable pages.
1039          */
1040         if (PageUnevictable(page))
1041                 return ret;
1042
1043         ret = -EBUSY;
1044
1045         if (likely(get_page_unless_zero(page))) {
1046                 /*
1047                  * Be careful not to clear PageLRU until after we're
1048                  * sure the page is not being freed elsewhere -- the
1049                  * page release code relies on it.
1050                  */
1051                 ClearPageLRU(page);
1052                 ret = 0;
1053         }
1054
1055         return ret;
1056 }
1057
1058 /*
1059  * zone->lru_lock is heavily contended.  Some of the functions that
1060  * shrink the lists perform better by taking out a batch of pages
1061  * and working on them outside the LRU lock.
1062  *
1063  * For pagecache intensive workloads, this function is the hottest
1064  * spot in the kernel (apart from copy_*_user functions).
1065  *
1066  * Appropriate locks must be held before calling this function.
1067  *
1068  * @nr_to_scan: The number of pages to look through on the list.
1069  * @src:        The LRU list to pull pages off.
1070  * @dst:        The temp list to put pages on to.
1071  * @scanned:    The number of pages that were scanned.
1072  * @order:      The caller's attempted allocation order
1073  * @mode:       One of the LRU isolation modes
1074  * @file:       True [1] if isolating file [!anon] pages
1075  *
1076  * returns how many pages were moved onto *@dst.
1077  */
1078 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1079                 struct list_head *src, struct list_head *dst,
1080                 unsigned long *scanned, int order, int mode, int file)
1081 {
1082         unsigned long nr_taken = 0;
1083         unsigned long nr_lumpy_taken = 0;
1084         unsigned long nr_lumpy_dirty = 0;
1085         unsigned long nr_lumpy_failed = 0;
1086         unsigned long scan;
1087
1088         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1089                 struct page *page;
1090                 unsigned long pfn;
1091                 unsigned long end_pfn;
1092                 unsigned long page_pfn;
1093                 int zone_id;
1094
1095                 page = lru_to_page(src);
1096                 prefetchw_prev_lru_page(page, src, flags);
1097
1098                 VM_BUG_ON(!PageLRU(page));
1099
1100                 switch (__isolate_lru_page(page, mode, file)) {
1101                 case 0:
1102                         list_move(&page->lru, dst);
1103                         mem_cgroup_del_lru(page);
1104                         nr_taken += hpage_nr_pages(page);
1105                         break;
1106
1107                 case -EBUSY:
1108                         /* else it is being freed elsewhere */
1109                         list_move(&page->lru, src);
1110                         mem_cgroup_rotate_lru_list(page, page_lru(page));
1111                         continue;
1112
1113                 default:
1114                         BUG();
1115                 }
1116
1117                 if (!order)
1118                         continue;
1119
1120                 /*
1121                  * Attempt to take all pages in the order aligned region
1122                  * surrounding the tag page.  Only take those pages of
1123                  * the same active state as that tag page.  We may safely
1124                  * round the target page pfn down to the requested order
1125                  * as the mem_map is guaranteed valid out to MAX_ORDER,
1126                  * where that page is in a different zone we will detect
1127                  * it from its zone id and abort this block scan.
1128                  */
1129                 zone_id = page_zone_id(page);
1130                 page_pfn = page_to_pfn(page);
1131                 pfn = page_pfn & ~((1 << order) - 1);
1132                 end_pfn = pfn + (1 << order);
1133                 for (; pfn < end_pfn; pfn++) {
1134                         struct page *cursor_page;
1135
1136                         /* The target page is in the block, ignore it. */
1137                         if (unlikely(pfn == page_pfn))
1138                                 continue;
1139
1140                         /* Avoid holes within the zone. */
1141                         if (unlikely(!pfn_valid_within(pfn)))
1142                                 break;
1143
1144                         cursor_page = pfn_to_page(pfn);
1145
1146                         /* Check that we have not crossed a zone boundary. */
1147                         if (unlikely(page_zone_id(cursor_page) != zone_id))
1148                                 break;
1149
1150                         /*
1151                          * If we don't have enough swap space, reclaiming of
1152                          * anon page which don't already have a swap slot is
1153                          * pointless.
1154                          */
1155                         if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
1156                             !PageSwapCache(cursor_page))
1157                                 break;
1158
1159                         if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1160                                 list_move(&cursor_page->lru, dst);
1161                                 mem_cgroup_del_lru(cursor_page);
1162                                 nr_taken += hpage_nr_pages(page);
1163                                 nr_lumpy_taken++;
1164                                 if (PageDirty(cursor_page))
1165                                         nr_lumpy_dirty++;
1166                                 scan++;
1167                         } else {
1168                                 /*
1169                                  * Check if the page is freed already.
1170                                  *
1171                                  * We can't use page_count() as that
1172                                  * requires compound_head and we don't
1173                                  * have a pin on the page here. If a
1174                                  * page is tail, we may or may not
1175                                  * have isolated the head, so assume
1176                                  * it's not free, it'd be tricky to
1177                                  * track the head status without a
1178                                  * page pin.
1179                                  */
1180                                 if (!PageTail(cursor_page) &&
1181                                     !atomic_read(&cursor_page->_count))
1182                                         continue;
1183                                 break;
1184                         }
1185                 }
1186
1187                 /* If we break out of the loop above, lumpy reclaim failed */
1188                 if (pfn < end_pfn)
1189                         nr_lumpy_failed++;
1190         }
1191
1192         *scanned = scan;
1193
1194         trace_mm_vmscan_lru_isolate(order,
1195                         nr_to_scan, scan,
1196                         nr_taken,
1197                         nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1198                         mode);
1199         return nr_taken;
1200 }
1201
1202 static unsigned long isolate_pages_global(unsigned long nr,
1203                                         struct list_head *dst,
1204                                         unsigned long *scanned, int order,
1205                                         int mode, struct zone *z,
1206                                         int active, int file)
1207 {
1208         int lru = LRU_BASE;
1209         if (active)
1210                 lru += LRU_ACTIVE;
1211         if (file)
1212                 lru += LRU_FILE;
1213         return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1214                                                                 mode, file);
1215 }
1216
1217 /*
1218  * clear_active_flags() is a helper for shrink_active_list(), clearing
1219  * any active bits from the pages in the list.
1220  */
1221 static unsigned long clear_active_flags(struct list_head *page_list,
1222                                         unsigned int *count)
1223 {
1224         int nr_active = 0;
1225         int lru;
1226         struct page *page;
1227
1228         list_for_each_entry(page, page_list, lru) {
1229                 int numpages = hpage_nr_pages(page);
1230                 lru = page_lru_base_type(page);
1231                 if (PageActive(page)) {
1232                         lru += LRU_ACTIVE;
1233                         ClearPageActive(page);
1234                         nr_active += numpages;
1235                 }
1236                 if (count)
1237                         count[lru] += numpages;
1238         }
1239
1240         return nr_active;
1241 }
1242
1243 /**
1244  * isolate_lru_page - tries to isolate a page from its LRU list
1245  * @page: page to isolate from its LRU list
1246  *
1247  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1248  * vmstat statistic corresponding to whatever LRU list the page was on.
1249  *
1250  * Returns 0 if the page was removed from an LRU list.
1251  * Returns -EBUSY if the page was not on an LRU list.
1252  *
1253  * The returned page will have PageLRU() cleared.  If it was found on
1254  * the active list, it will have PageActive set.  If it was found on
1255  * the unevictable list, it will have the PageUnevictable bit set. That flag
1256  * may need to be cleared by the caller before letting the page go.
1257  *
1258  * The vmstat statistic corresponding to the list on which the page was
1259  * found will be decremented.
1260  *
1261  * Restrictions:
1262  * (1) Must be called with an elevated refcount on the page. This is a
1263  *     fundamentnal difference from isolate_lru_pages (which is called
1264  *     without a stable reference).
1265  * (2) the lru_lock must not be held.
1266  * (3) interrupts must be enabled.
1267  */
1268 int isolate_lru_page(struct page *page)
1269 {
1270         int ret = -EBUSY;
1271
1272         VM_BUG_ON(!page_count(page));
1273
1274         if (PageLRU(page)) {
1275                 struct zone *zone = page_zone(page);
1276
1277                 spin_lock_irq(&zone->lru_lock);
1278                 if (PageLRU(page)) {
1279                         int lru = page_lru(page);
1280                         ret = 0;
1281                         get_page(page);
1282                         ClearPageLRU(page);
1283
1284                         del_page_from_lru_list(zone, page, lru);
1285                 }
1286                 spin_unlock_irq(&zone->lru_lock);
1287         }
1288         return ret;
1289 }
1290
1291 /*
1292  * Are there way too many processes in the direct reclaim path already?
1293  */
1294 static int too_many_isolated(struct zone *zone, int file,
1295                 struct scan_control *sc)
1296 {
1297         unsigned long inactive, isolated;
1298
1299         if (current_is_kswapd())
1300                 return 0;
1301
1302         if (!scanning_global_lru(sc))
1303                 return 0;
1304
1305         if (file) {
1306                 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1307                 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1308         } else {
1309                 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1310                 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1311         }
1312
1313         return isolated > inactive;
1314 }
1315
1316 /*
1317  * TODO: Try merging with migrations version of putback_lru_pages
1318  */
1319 static noinline_for_stack void
1320 putback_lru_pages(struct zone *zone, struct scan_control *sc,
1321                                 unsigned long nr_anon, unsigned long nr_file,
1322                                 struct list_head *page_list)
1323 {
1324         struct page *page;
1325         struct pagevec pvec;
1326         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1327
1328         pagevec_init(&pvec, 1);
1329
1330         /*
1331          * Put back any unfreeable pages.
1332          */
1333         spin_lock(&zone->lru_lock);
1334         while (!list_empty(page_list)) {
1335                 int lru;
1336                 page = lru_to_page(page_list);
1337                 VM_BUG_ON(PageLRU(page));
1338                 list_del(&page->lru);
1339                 if (unlikely(!page_evictable(page, NULL))) {
1340                         spin_unlock_irq(&zone->lru_lock);
1341                         putback_lru_page(page);
1342                         spin_lock_irq(&zone->lru_lock);
1343                         continue;
1344                 }
1345                 SetPageLRU(page);
1346                 lru = page_lru(page);
1347                 add_page_to_lru_list(zone, page, lru);
1348                 if (is_active_lru(lru)) {
1349                         int file = is_file_lru(lru);
1350                         int numpages = hpage_nr_pages(page);
1351                         reclaim_stat->recent_rotated[file] += numpages;
1352                 }
1353                 if (!pagevec_add(&pvec, page)) {
1354                         spin_unlock_irq(&zone->lru_lock);
1355                         __pagevec_release(&pvec);
1356                         spin_lock_irq(&zone->lru_lock);
1357                 }
1358         }
1359         __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1360         __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1361
1362         spin_unlock_irq(&zone->lru_lock);
1363         pagevec_release(&pvec);
1364 }
1365
1366 static noinline_for_stack void update_isolated_counts(struct zone *zone,
1367                                         struct scan_control *sc,
1368                                         unsigned long *nr_anon,
1369                                         unsigned long *nr_file,
1370                                         struct list_head *isolated_list)
1371 {
1372         unsigned long nr_active;
1373         unsigned int count[NR_LRU_LISTS] = { 0, };
1374         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1375
1376         nr_active = clear_active_flags(isolated_list, count);
1377         __count_vm_events(PGDEACTIVATE, nr_active);
1378
1379         __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1380                               -count[LRU_ACTIVE_FILE]);
1381         __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1382                               -count[LRU_INACTIVE_FILE]);
1383         __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1384                               -count[LRU_ACTIVE_ANON]);
1385         __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1386                               -count[LRU_INACTIVE_ANON]);
1387
1388         *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1389         *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1390         __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1391         __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1392
1393         reclaim_stat->recent_scanned[0] += *nr_anon;
1394         reclaim_stat->recent_scanned[1] += *nr_file;
1395 }
1396
1397 /*
1398  * Returns true if the caller should wait to clean dirty/writeback pages.
1399  *
1400  * If we are direct reclaiming for contiguous pages and we do not reclaim
1401  * everything in the list, try again and wait for writeback IO to complete.
1402  * This will stall high-order allocations noticeably. Only do that when really
1403  * need to free the pages under high memory pressure.
1404  */
1405 static inline bool should_reclaim_stall(unsigned long nr_taken,
1406                                         unsigned long nr_freed,
1407                                         int priority,
1408                                         struct scan_control *sc)
1409 {
1410         int lumpy_stall_priority;
1411
1412         /* kswapd should not stall on sync IO */
1413         if (current_is_kswapd())
1414                 return false;
1415
1416         /* Only stall on lumpy reclaim */
1417         if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
1418                 return false;
1419
1420         /* If we have relaimed everything on the isolated list, no stall */
1421         if (nr_freed == nr_taken)
1422                 return false;
1423
1424         /*
1425          * For high-order allocations, there are two stall thresholds.
1426          * High-cost allocations stall immediately where as lower
1427          * order allocations such as stacks require the scanning
1428          * priority to be much higher before stalling.
1429          */
1430         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1431                 lumpy_stall_priority = DEF_PRIORITY;
1432         else
1433                 lumpy_stall_priority = DEF_PRIORITY / 3;
1434
1435         return priority <= lumpy_stall_priority;
1436 }
1437
1438 /*
1439  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1440  * of reclaimed pages
1441  */
1442 static noinline_for_stack unsigned long
1443 shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1444                         struct scan_control *sc, int priority, int file)
1445 {
1446         LIST_HEAD(page_list);
1447         unsigned long nr_scanned;
1448         unsigned long nr_reclaimed = 0;
1449         unsigned long nr_taken;
1450         unsigned long nr_anon;
1451         unsigned long nr_file;
1452
1453         while (unlikely(too_many_isolated(zone, file, sc))) {
1454                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1455
1456                 /* We are about to die and free our memory. Return now. */
1457                 if (fatal_signal_pending(current))
1458                         return SWAP_CLUSTER_MAX;
1459         }
1460
1461         set_reclaim_mode(priority, sc, false);
1462         lru_add_drain();
1463         spin_lock_irq(&zone->lru_lock);
1464
1465         if (scanning_global_lru(sc)) {
1466                 nr_taken = isolate_pages_global(nr_to_scan,
1467                         &page_list, &nr_scanned, sc->order,
1468                         sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM ?
1469                                         ISOLATE_BOTH : ISOLATE_INACTIVE,
1470                         zone, 0, file);
1471                 zone->pages_scanned += nr_scanned;
1472                 if (current_is_kswapd())
1473                         __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1474                                                nr_scanned);
1475                 else
1476                         __count_zone_vm_events(PGSCAN_DIRECT, zone,
1477                                                nr_scanned);
1478         } else {
1479                 nr_taken = mem_cgroup_isolate_pages(nr_to_scan,
1480                         &page_list, &nr_scanned, sc->order,
1481                         sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM ?
1482                                         ISOLATE_BOTH : ISOLATE_INACTIVE,
1483                         zone, sc->mem_cgroup,
1484                         0, file);
1485                 /*
1486                  * mem_cgroup_isolate_pages() keeps track of
1487                  * scanned pages on its own.
1488                  */
1489         }
1490
1491         if (nr_taken == 0) {
1492                 spin_unlock_irq(&zone->lru_lock);
1493                 return 0;
1494         }
1495
1496         update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
1497
1498         spin_unlock_irq(&zone->lru_lock);
1499
1500         nr_reclaimed = shrink_page_list(&page_list, zone, sc);
1501
1502         /* Check if we should syncronously wait for writeback */
1503         if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1504                 set_reclaim_mode(priority, sc, true);
1505                 nr_reclaimed += shrink_page_list(&page_list, zone, sc);
1506         }
1507
1508         local_irq_disable();
1509         if (current_is_kswapd())
1510                 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1511         __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1512
1513         putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
1514
1515         trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1516                 zone_idx(zone),
1517                 nr_scanned, nr_reclaimed,
1518                 priority,
1519                 trace_shrink_flags(file, sc->reclaim_mode));
1520         return nr_reclaimed;
1521 }
1522
1523 /*
1524  * This moves pages from the active list to the inactive list.
1525  *
1526  * We move them the other way if the page is referenced by one or more
1527  * processes, from rmap.
1528  *
1529  * If the pages are mostly unmapped, the processing is fast and it is
1530  * appropriate to hold zone->lru_lock across the whole operation.  But if
1531  * the pages are mapped, the processing is slow (page_referenced()) so we
1532  * should drop zone->lru_lock around each page.  It's impossible to balance
1533  * this, so instead we remove the pages from the LRU while processing them.
1534  * It is safe to rely on PG_active against the non-LRU pages in here because
1535  * nobody will play with that bit on a non-LRU page.
1536  *
1537  * The downside is that we have to touch page->_count against each page.
1538  * But we had to alter page->flags anyway.
1539  */
1540
1541 static void move_active_pages_to_lru(struct zone *zone,
1542                                      struct list_head *list,
1543                                      enum lru_list lru)
1544 {
1545         unsigned long pgmoved = 0;
1546         struct pagevec pvec;
1547         struct page *page;
1548
1549         pagevec_init(&pvec, 1);
1550
1551         while (!list_empty(list)) {
1552                 page = lru_to_page(list);
1553
1554                 VM_BUG_ON(PageLRU(page));
1555                 SetPageLRU(page);
1556
1557                 list_move(&page->lru, &zone->lru[lru].list);
1558                 mem_cgroup_add_lru_list(page, lru);
1559                 pgmoved += hpage_nr_pages(page);
1560
1561                 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1562                         spin_unlock_irq(&zone->lru_lock);
1563                         if (buffer_heads_over_limit)
1564                                 pagevec_strip(&pvec);
1565                         __pagevec_release(&pvec);
1566                         spin_lock_irq(&zone->lru_lock);
1567                 }
1568         }
1569         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1570         if (!is_active_lru(lru))
1571                 __count_vm_events(PGDEACTIVATE, pgmoved);
1572 }
1573
1574 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1575                         struct scan_control *sc, int priority, int file)
1576 {
1577         unsigned long nr_taken;
1578         unsigned long pgscanned;
1579         unsigned long vm_flags;
1580         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1581         LIST_HEAD(l_active);
1582         LIST_HEAD(l_inactive);
1583         struct page *page;
1584         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1585         unsigned long nr_rotated = 0;
1586
1587         lru_add_drain();
1588         spin_lock_irq(&zone->lru_lock);
1589         if (scanning_global_lru(sc)) {
1590                 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1591                                                 &pgscanned, sc->order,
1592                                                 ISOLATE_ACTIVE, zone,
1593                                                 1, file);
1594                 zone->pages_scanned += pgscanned;
1595         } else {
1596                 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1597                                                 &pgscanned, sc->order,
1598                                                 ISOLATE_ACTIVE, zone,
1599                                                 sc->mem_cgroup, 1, file);
1600                 /*
1601                  * mem_cgroup_isolate_pages() keeps track of
1602                  * scanned pages on its own.
1603                  */
1604         }
1605
1606         reclaim_stat->recent_scanned[file] += nr_taken;
1607
1608         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1609         if (file)
1610                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1611         else
1612                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1613         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1614         spin_unlock_irq(&zone->lru_lock);
1615
1616         while (!list_empty(&l_hold)) {
1617                 cond_resched();
1618                 page = lru_to_page(&l_hold);
1619                 list_del(&page->lru);
1620
1621                 if (unlikely(!page_evictable(page, NULL))) {
1622                         putback_lru_page(page);
1623                         continue;
1624                 }
1625
1626                 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1627                         nr_rotated += hpage_nr_pages(page);
1628                         /*
1629                          * Identify referenced, file-backed active pages and
1630                          * give them one more trip around the active list. So
1631                          * that executable code get better chances to stay in
1632                          * memory under moderate memory pressure.  Anon pages
1633                          * are not likely to be evicted by use-once streaming
1634                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1635                          * so we ignore them here.
1636                          */
1637                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1638                                 list_add(&page->lru, &l_active);
1639                                 continue;
1640                         }
1641                 }
1642
1643                 ClearPageActive(page);  /* we are de-activating */
1644                 list_add(&page->lru, &l_inactive);
1645         }
1646
1647         /*
1648          * Move pages back to the lru list.
1649          */
1650         spin_lock_irq(&zone->lru_lock);
1651         /*
1652          * Count referenced pages from currently used mappings as rotated,
1653          * even though only some of them are actually re-activated.  This
1654          * helps balance scan pressure between file and anonymous pages in
1655          * get_scan_ratio.
1656          */
1657         reclaim_stat->recent_rotated[file] += nr_rotated;
1658
1659         move_active_pages_to_lru(zone, &l_active,
1660                                                 LRU_ACTIVE + file * LRU_FILE);
1661         move_active_pages_to_lru(zone, &l_inactive,
1662                                                 LRU_BASE   + file * LRU_FILE);
1663         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1664         spin_unlock_irq(&zone->lru_lock);
1665 }
1666
1667 #ifdef CONFIG_SWAP
1668 static int inactive_anon_is_low_global(struct zone *zone)
1669 {
1670         unsigned long active, inactive;
1671
1672         active = zone_page_state(zone, NR_ACTIVE_ANON);
1673         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1674
1675         if (inactive * zone->inactive_ratio < active)
1676                 return 1;
1677
1678         return 0;
1679 }
1680
1681 /**
1682  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1683  * @zone: zone to check
1684  * @sc:   scan control of this context
1685  *
1686  * Returns true if the zone does not have enough inactive anon pages,
1687  * meaning some active anon pages need to be deactivated.
1688  */
1689 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1690 {
1691         int low;
1692
1693         /*
1694          * If we don't have swap space, anonymous page deactivation
1695          * is pointless.
1696          */
1697         if (!total_swap_pages)
1698                 return 0;
1699
1700         if (scanning_global_lru(sc))
1701                 low = inactive_anon_is_low_global(zone);
1702         else
1703                 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1704         return low;
1705 }
1706 #else
1707 static inline int inactive_anon_is_low(struct zone *zone,
1708                                         struct scan_control *sc)
1709 {
1710         return 0;
1711 }
1712 #endif
1713
1714 static int inactive_file_is_low_global(struct zone *zone)
1715 {
1716         unsigned long active, inactive;
1717
1718         active = zone_page_state(zone, NR_ACTIVE_FILE);
1719         inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1720
1721         return (active > inactive);
1722 }
1723
1724 /**
1725  * inactive_file_is_low - check if file pages need to be deactivated
1726  * @zone: zone to check
1727  * @sc:   scan control of this context
1728  *
1729  * When the system is doing streaming IO, memory pressure here
1730  * ensures that active file pages get deactivated, until more
1731  * than half of the file pages are on the inactive list.
1732  *
1733  * Once we get to that situation, protect the system's working
1734  * set from being evicted by disabling active file page aging.
1735  *
1736  * This uses a different ratio than the anonymous pages, because
1737  * the page cache uses a use-once replacement algorithm.
1738  */
1739 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1740 {
1741         int low;
1742
1743         if (scanning_global_lru(sc))
1744                 low = inactive_file_is_low_global(zone);
1745         else
1746                 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1747         return low;
1748 }
1749
1750 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1751                                 int file)
1752 {
1753         if (file)
1754                 return inactive_file_is_low(zone, sc);
1755         else
1756                 return inactive_anon_is_low(zone, sc);
1757 }
1758
1759 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1760         struct zone *zone, struct scan_control *sc, int priority)
1761 {
1762         int file = is_file_lru(lru);
1763
1764         if (is_active_lru(lru)) {
1765                 if (inactive_list_is_low(zone, sc, file))
1766                     shrink_active_list(nr_to_scan, zone, sc, priority, file);
1767                 return 0;
1768         }
1769
1770         return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1771 }
1772
1773 /*
1774  * Determine how aggressively the anon and file LRU lists should be
1775  * scanned.  The relative value of each set of LRU lists is determined
1776  * by looking at the fraction of the pages scanned we did rotate back
1777  * onto the active list instead of evict.
1778  *
1779  * nr[0] = anon pages to scan; nr[1] = file pages to scan
1780  */
1781 static void get_scan_count(struct zone *zone, struct scan_control *sc,
1782                                         unsigned long *nr, int priority)
1783 {
1784         unsigned long anon, file, free;
1785         unsigned long anon_prio, file_prio;
1786         unsigned long ap, fp;
1787         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1788         u64 fraction[2], denominator;
1789         enum lru_list l;
1790         int noswap = 0;
1791         int force_scan = 0;
1792
1793
1794         anon  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1795                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1796         file  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1797                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1798
1799         if (((anon + file) >> priority) < SWAP_CLUSTER_MAX) {
1800                 /* kswapd does zone balancing and need to scan this zone */
1801                 if (scanning_global_lru(sc) && current_is_kswapd())
1802                         force_scan = 1;
1803                 /* memcg may have small limit and need to avoid priority drop */
1804                 if (!scanning_global_lru(sc))
1805                         force_scan = 1;
1806         }
1807
1808         /* If we have no swap space, do not bother scanning anon pages. */
1809         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1810                 noswap = 1;
1811                 fraction[0] = 0;
1812                 fraction[1] = 1;
1813                 denominator = 1;
1814                 goto out;
1815         }
1816
1817         if (scanning_global_lru(sc)) {
1818                 free  = zone_page_state(zone, NR_FREE_PAGES);
1819                 /* If we have very few page cache pages,
1820                    force-scan anon pages. */
1821                 if (unlikely(file + free <= high_wmark_pages(zone))) {
1822                         fraction[0] = 1;
1823                         fraction[1] = 0;
1824                         denominator = 1;
1825                         goto out;
1826                 }
1827         }
1828
1829         /*
1830          * With swappiness at 100, anonymous and file have the same priority.
1831          * This scanning priority is essentially the inverse of IO cost.
1832          */
1833         anon_prio = sc->swappiness;
1834         file_prio = 200 - sc->swappiness;
1835
1836         /*
1837          * OK, so we have swap space and a fair amount of page cache
1838          * pages.  We use the recently rotated / recently scanned
1839          * ratios to determine how valuable each cache is.
1840          *
1841          * Because workloads change over time (and to avoid overflow)
1842          * we keep these statistics as a floating average, which ends
1843          * up weighing recent references more than old ones.
1844          *
1845          * anon in [0], file in [1]
1846          */
1847         spin_lock_irq(&zone->lru_lock);
1848         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1849                 reclaim_stat->recent_scanned[0] /= 2;
1850                 reclaim_stat->recent_rotated[0] /= 2;
1851         }
1852
1853         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1854                 reclaim_stat->recent_scanned[1] /= 2;
1855                 reclaim_stat->recent_rotated[1] /= 2;
1856         }
1857
1858         /*
1859          * The amount of pressure on anon vs file pages is inversely
1860          * proportional to the fraction of recently scanned pages on
1861          * each list that were recently referenced and in active use.
1862          */
1863         ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1864         ap /= reclaim_stat->recent_rotated[0] + 1;
1865
1866         fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1867         fp /= reclaim_stat->recent_rotated[1] + 1;
1868         spin_unlock_irq(&zone->lru_lock);
1869
1870         fraction[0] = ap;
1871         fraction[1] = fp;
1872         denominator = ap + fp + 1;
1873 out:
1874         for_each_evictable_lru(l) {
1875                 int file = is_file_lru(l);
1876                 unsigned long scan;
1877
1878                 scan = zone_nr_lru_pages(zone, sc, l);
1879                 if (priority || noswap) {
1880                         scan >>= priority;
1881                         scan = div64_u64(scan * fraction[file], denominator);
1882                 }
1883
1884                 /*
1885                  * If zone is small or memcg is small, nr[l] can be 0.
1886                  * This results no-scan on this priority and priority drop down.
1887                  * For global direct reclaim, it can visit next zone and tend
1888                  * not to have problems. For global kswapd, it's for zone
1889                  * balancing and it need to scan a small amounts. When using
1890                  * memcg, priority drop can cause big latency. So, it's better
1891                  * to scan small amount. See may_noscan above.
1892                  */
1893                 if (!scan && force_scan) {
1894                         if (file)
1895                                 scan = SWAP_CLUSTER_MAX;
1896                         else if (!noswap)
1897                                 scan = SWAP_CLUSTER_MAX;
1898                 }
1899                 nr[l] = scan;
1900         }
1901 }
1902
1903 /*
1904  * Reclaim/compaction depends on a number of pages being freed. To avoid
1905  * disruption to the system, a small number of order-0 pages continue to be
1906  * rotated and reclaimed in the normal fashion. However, by the time we get
1907  * back to the allocator and call try_to_compact_zone(), we ensure that
1908  * there are enough free pages for it to be likely successful
1909  */
1910 static inline bool should_continue_reclaim(struct zone *zone,
1911                                         unsigned long nr_reclaimed,
1912                                         unsigned long nr_scanned,
1913                                         struct scan_control *sc)
1914 {
1915         unsigned long pages_for_compaction;
1916         unsigned long inactive_lru_pages;
1917
1918         /* If not in reclaim/compaction mode, stop */
1919         if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
1920                 return false;
1921
1922         /* Consider stopping depending on scan and reclaim activity */
1923         if (sc->gfp_mask & __GFP_REPEAT) {
1924                 /*
1925                  * For __GFP_REPEAT allocations, stop reclaiming if the
1926                  * full LRU list has been scanned and we are still failing
1927                  * to reclaim pages. This full LRU scan is potentially
1928                  * expensive but a __GFP_REPEAT caller really wants to succeed
1929                  */
1930                 if (!nr_reclaimed && !nr_scanned)
1931                         return false;
1932         } else {
1933                 /*
1934                  * For non-__GFP_REPEAT allocations which can presumably
1935                  * fail without consequence, stop if we failed to reclaim
1936                  * any pages from the last SWAP_CLUSTER_MAX number of
1937                  * pages that were scanned. This will return to the
1938                  * caller faster at the risk reclaim/compaction and
1939                  * the resulting allocation attempt fails
1940                  */
1941                 if (!nr_reclaimed)
1942                         return false;
1943         }
1944
1945         /*
1946          * If we have not reclaimed enough pages for compaction and the
1947          * inactive lists are large enough, continue reclaiming
1948          */
1949         pages_for_compaction = (2UL << sc->order);
1950         inactive_lru_pages = zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON) +
1951                                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1952         if (sc->nr_reclaimed < pages_for_compaction &&
1953                         inactive_lru_pages > pages_for_compaction)
1954                 return true;
1955
1956         /* If compaction would go ahead or the allocation would succeed, stop */
1957         switch (compaction_suitable(zone, sc->order)) {
1958         case COMPACT_PARTIAL:
1959         case COMPACT_CONTINUE:
1960                 return false;
1961         default:
1962                 return true;
1963         }
1964 }
1965
1966 /*
1967  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1968  */
1969 static void shrink_zone(int priority, struct zone *zone,
1970                                 struct scan_control *sc)
1971 {
1972         unsigned long nr[NR_LRU_LISTS];
1973         unsigned long nr_to_scan;
1974         enum lru_list l;
1975         unsigned long nr_reclaimed, nr_scanned;
1976         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1977
1978 restart:
1979         nr_reclaimed = 0;
1980         nr_scanned = sc->nr_scanned;
1981         get_scan_count(zone, sc, nr, priority);
1982
1983         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1984                                         nr[LRU_INACTIVE_FILE]) {
1985                 for_each_evictable_lru(l) {
1986                         if (nr[l]) {
1987                                 nr_to_scan = min_t(unsigned long,
1988                                                    nr[l], SWAP_CLUSTER_MAX);
1989                                 nr[l] -= nr_to_scan;
1990
1991                                 nr_reclaimed += shrink_list(l, nr_to_scan,
1992                                                             zone, sc, priority);
1993                         }
1994                 }
1995                 /*
1996                  * On large memory systems, scan >> priority can become
1997                  * really large. This is fine for the starting priority;
1998                  * we want to put equal scanning pressure on each zone.
1999                  * However, if the VM has a harder time of freeing pages,
2000                  * with multiple processes reclaiming pages, the total
2001                  * freeing target can get unreasonably large.
2002                  */
2003                 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
2004                         break;
2005         }
2006         sc->nr_reclaimed += nr_reclaimed;
2007
2008         /*
2009          * Even if we did not try to evict anon pages at all, we want to
2010          * rebalance the anon lru active/inactive ratio.
2011          */
2012         if (inactive_anon_is_low(zone, sc))
2013                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
2014
2015         /* reclaim/compaction might need reclaim to continue */
2016         if (should_continue_reclaim(zone, nr_reclaimed,
2017                                         sc->nr_scanned - nr_scanned, sc))
2018                 goto restart;
2019
2020         throttle_vm_writeout(sc->gfp_mask);
2021 }
2022
2023 /*
2024  * This is the direct reclaim path, for page-allocating processes.  We only
2025  * try to reclaim pages from zones which will satisfy the caller's allocation
2026  * request.
2027  *
2028  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2029  * Because:
2030  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2031  *    allocation or
2032  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2033  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2034  *    zone defense algorithm.
2035  *
2036  * If a zone is deemed to be full of pinned pages then just give it a light
2037  * scan then give up on it.
2038  */
2039 static void shrink_zones(int priority, struct zonelist *zonelist,
2040                                         struct scan_control *sc)
2041 {
2042         struct zoneref *z;
2043         struct zone *zone;
2044         unsigned long nr_soft_reclaimed;
2045         unsigned long nr_soft_scanned;
2046
2047         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2048                                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2049                 if (!populated_zone(zone))
2050                         continue;
2051                 /*
2052                  * Take care memory controller reclaiming has small influence
2053                  * to global LRU.
2054                  */
2055                 if (scanning_global_lru(sc)) {
2056                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2057                                 continue;
2058                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2059                                 continue;       /* Let kswapd poll it */
2060                         /*
2061                          * This steals pages from memory cgroups over softlimit
2062                          * and returns the number of reclaimed pages and
2063                          * scanned pages. This works for global memory pressure
2064                          * and balancing, not for a memcg's limit.
2065                          */
2066                         nr_soft_scanned = 0;
2067                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2068                                                 sc->order, sc->gfp_mask,
2069                                                 &nr_soft_scanned);
2070                         sc->nr_reclaimed += nr_soft_reclaimed;
2071                         sc->nr_scanned += nr_soft_scanned;
2072                         /* need some check for avoid more shrink_zone() */
2073                 }
2074
2075                 shrink_zone(priority, zone, sc);
2076         }
2077 }
2078
2079 static bool zone_reclaimable(struct zone *zone)
2080 {
2081         return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2082 }
2083
2084 /* All zones in zonelist are unreclaimable? */
2085 static bool all_unreclaimable(struct zonelist *zonelist,
2086                 struct scan_control *sc)
2087 {
2088         struct zoneref *z;
2089         struct zone *zone;
2090
2091         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2092                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2093                 if (!populated_zone(zone))
2094                         continue;
2095                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2096                         continue;
2097                 if (!zone->all_unreclaimable)
2098                         return false;
2099         }
2100
2101         return true;
2102 }
2103
2104 /*
2105  * This is the main entry point to direct page reclaim.
2106  *
2107  * If a full scan of the inactive list fails to free enough memory then we
2108  * are "out of memory" and something needs to be killed.
2109  *
2110  * If the caller is !__GFP_FS then the probability of a failure is reasonably
2111  * high - the zone may be full of dirty or under-writeback pages, which this
2112  * caller can't do much about.  We kick the writeback threads and take explicit
2113  * naps in the hope that some of these pages can be written.  But if the
2114  * allocating task holds filesystem locks which prevent writeout this might not
2115  * work, and the allocation attempt will fail.
2116  *
2117  * returns:     0, if no pages reclaimed
2118  *              else, the number of pages reclaimed
2119  */
2120 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2121                                         struct scan_control *sc,
2122                                         struct shrink_control *shrink)
2123 {
2124         int priority;
2125         unsigned long total_scanned = 0;
2126         struct reclaim_state *reclaim_state = current->reclaim_state;
2127         struct zoneref *z;
2128         struct zone *zone;
2129         unsigned long writeback_threshold;
2130
2131         get_mems_allowed();
2132         delayacct_freepages_start();
2133
2134         if (scanning_global_lru(sc))
2135                 count_vm_event(ALLOCSTALL);
2136
2137         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2138                 sc->nr_scanned = 0;
2139                 if (!priority)
2140                         disable_swap_token(sc->mem_cgroup);
2141                 shrink_zones(priority, zonelist, sc);
2142                 /*
2143                  * Don't shrink slabs when reclaiming memory from
2144                  * over limit cgroups
2145                  */
2146                 if (scanning_global_lru(sc)) {
2147                         unsigned long lru_pages = 0;
2148                         for_each_zone_zonelist(zone, z, zonelist,
2149                                         gfp_zone(sc->gfp_mask)) {
2150                                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2151                                         continue;
2152
2153                                 lru_pages += zone_reclaimable_pages(zone);
2154                         }
2155
2156                         shrink_slab(shrink, sc->nr_scanned, lru_pages);
2157                         if (reclaim_state) {
2158                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2159                                 reclaim_state->reclaimed_slab = 0;
2160                         }
2161                 }
2162                 total_scanned += sc->nr_scanned;
2163                 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2164                         goto out;
2165
2166                 /*
2167                  * Try to write back as many pages as we just scanned.  This
2168                  * tends to cause slow streaming writers to write data to the
2169                  * disk smoothly, at the dirtying rate, which is nice.   But
2170                  * that's undesirable in laptop mode, where we *want* lumpy
2171                  * writeout.  So in laptop mode, write out the whole world.
2172                  */
2173                 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2174                 if (total_scanned > writeback_threshold) {
2175                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
2176                         sc->may_writepage = 1;
2177                 }
2178
2179                 /* Take a nap, wait for some writeback to complete */
2180                 if (!sc->hibernation_mode && sc->nr_scanned &&
2181                     priority < DEF_PRIORITY - 2) {
2182                         struct zone *preferred_zone;
2183
2184                         first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2185                                                 &cpuset_current_mems_allowed,
2186                                                 &preferred_zone);
2187                         wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2188                 }
2189         }
2190
2191 out:
2192         delayacct_freepages_end();
2193         put_mems_allowed();
2194
2195         if (sc->nr_reclaimed)
2196                 return sc->nr_reclaimed;
2197
2198         /*
2199          * As hibernation is going on, kswapd is freezed so that it can't mark
2200          * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2201          * check.
2202          */
2203         if (oom_killer_disabled)
2204                 return 0;
2205
2206         /* top priority shrink_zones still had more to do? don't OOM, then */
2207         if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
2208                 return 1;
2209
2210         return 0;
2211 }
2212
2213 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2214                                 gfp_t gfp_mask, nodemask_t *nodemask)
2215 {
2216         unsigned long nr_reclaimed;
2217         struct scan_control sc = {
2218                 .gfp_mask = gfp_mask,
2219                 .may_writepage = !laptop_mode,
2220                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2221                 .may_unmap = 1,
2222                 .may_swap = 1,
2223                 .swappiness = vm_swappiness,
2224                 .order = order,
2225                 .mem_cgroup = NULL,
2226                 .nodemask = nodemask,
2227         };
2228         struct shrink_control shrink = {
2229                 .gfp_mask = sc.gfp_mask,
2230         };
2231
2232         trace_mm_vmscan_direct_reclaim_begin(order,
2233                                 sc.may_writepage,
2234                                 gfp_mask);
2235
2236         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2237
2238         trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2239
2240         return nr_reclaimed;
2241 }
2242
2243 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2244
2245 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2246                                                 gfp_t gfp_mask, bool noswap,
2247                                                 unsigned int swappiness,
2248                                                 struct zone *zone,
2249                                                 unsigned long *nr_scanned)
2250 {
2251         struct scan_control sc = {
2252                 .nr_scanned = 0,
2253                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2254                 .may_writepage = !laptop_mode,
2255                 .may_unmap = 1,
2256                 .may_swap = !noswap,
2257                 .swappiness = swappiness,
2258                 .order = 0,
2259                 .mem_cgroup = mem,
2260         };
2261
2262         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2263                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2264
2265         trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2266                                                       sc.may_writepage,
2267                                                       sc.gfp_mask);
2268
2269         /*
2270          * NOTE: Although we can get the priority field, using it
2271          * here is not a good idea, since it limits the pages we can scan.
2272          * if we don't reclaim here, the shrink_zone from balance_pgdat
2273          * will pick up pages from other mem cgroup's as well. We hack
2274          * the priority and make it zero.
2275          */
2276         shrink_zone(0, zone, &sc);
2277
2278         trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2279
2280         *nr_scanned = sc.nr_scanned;
2281         return sc.nr_reclaimed;
2282 }
2283
2284 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
2285                                            gfp_t gfp_mask,
2286                                            bool noswap,
2287                                            unsigned int swappiness)
2288 {
2289         struct zonelist *zonelist;
2290         unsigned long nr_reclaimed;
2291         int nid;
2292         struct scan_control sc = {
2293                 .may_writepage = !laptop_mode,
2294                 .may_unmap = 1,
2295                 .may_swap = !noswap,
2296                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2297                 .swappiness = swappiness,
2298                 .order = 0,
2299                 .mem_cgroup = mem_cont,
2300                 .nodemask = NULL, /* we don't care the placement */
2301                 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2302                                 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2303         };
2304         struct shrink_control shrink = {
2305                 .gfp_mask = sc.gfp_mask,
2306         };
2307
2308         /*
2309          * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2310          * take care of from where we get pages. So the node where we start the
2311          * scan does not need to be the current node.
2312          */
2313         nid = mem_cgroup_select_victim_node(mem_cont);
2314
2315         zonelist = NODE_DATA(nid)->node_zonelists;
2316
2317         trace_mm_vmscan_memcg_reclaim_begin(0,
2318                                             sc.may_writepage,
2319                                             sc.gfp_mask);
2320
2321         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2322
2323         trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2324
2325         return nr_reclaimed;
2326 }
2327 #endif
2328
2329 /*
2330  * pgdat_balanced is used when checking if a node is balanced for high-order
2331  * allocations. Only zones that meet watermarks and are in a zone allowed
2332  * by the callers classzone_idx are added to balanced_pages. The total of
2333  * balanced pages must be at least 25% of the zones allowed by classzone_idx
2334  * for the node to be considered balanced. Forcing all zones to be balanced
2335  * for high orders can cause excessive reclaim when there are imbalanced zones.
2336  * The choice of 25% is due to
2337  *   o a 16M DMA zone that is balanced will not balance a zone on any
2338  *     reasonable sized machine
2339  *   o On all other machines, the top zone must be at least a reasonable
2340  *     percentage of the middle zones. For example, on 32-bit x86, highmem
2341  *     would need to be at least 256M for it to be balance a whole node.
2342  *     Similarly, on x86-64 the Normal zone would need to be at least 1G
2343  *     to balance a node on its own. These seemed like reasonable ratios.
2344  */
2345 static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2346                                                 int classzone_idx)
2347 {
2348         unsigned long present_pages = 0;
2349         int i;
2350
2351         for (i = 0; i <= classzone_idx; i++)
2352                 present_pages += pgdat->node_zones[i].present_pages;
2353
2354         /* A special case here: if zone has no page, we think it's balanced */
2355         return balanced_pages >= (present_pages >> 2);
2356 }
2357
2358 /* is kswapd sleeping prematurely? */
2359 static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2360                                         int classzone_idx)
2361 {
2362         int i;
2363         unsigned long balanced = 0;
2364         bool all_zones_ok = true;
2365
2366         /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2367         if (remaining)
2368                 return true;
2369
2370         /* Check the watermark levels */
2371         for (i = 0; i <= classzone_idx; i++) {
2372                 struct zone *zone = pgdat->node_zones + i;
2373
2374                 if (!populated_zone(zone))
2375                         continue;
2376
2377                 /*
2378                  * balance_pgdat() skips over all_unreclaimable after
2379                  * DEF_PRIORITY. Effectively, it considers them balanced so
2380                  * they must be considered balanced here as well if kswapd
2381                  * is to sleep
2382                  */
2383                 if (zone->all_unreclaimable) {
2384                         balanced += zone->present_pages;
2385                         continue;
2386                 }
2387
2388                 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2389                                                         i, 0))
2390                         all_zones_ok = false;
2391                 else
2392                         balanced += zone->present_pages;
2393         }
2394
2395         /*
2396          * For high-order requests, the balanced zones must contain at least
2397          * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2398          * must be balanced
2399          */
2400         if (order)
2401                 return !pgdat_balanced(pgdat, balanced, classzone_idx);
2402         else
2403                 return !all_zones_ok;
2404 }
2405
2406 /*
2407  * For kswapd, balance_pgdat() will work across all this node's zones until
2408  * they are all at high_wmark_pages(zone).
2409  *
2410  * Returns the final order kswapd was reclaiming at
2411  *
2412  * There is special handling here for zones which are full of pinned pages.
2413  * This can happen if the pages are all mlocked, or if they are all used by
2414  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
2415  * What we do is to detect the case where all pages in the zone have been
2416  * scanned twice and there has been zero successful reclaim.  Mark the zone as
2417  * dead and from now on, only perform a short scan.  Basically we're polling
2418  * the zone for when the problem goes away.
2419  *
2420  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
2421  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2422  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2423  * lower zones regardless of the number of free pages in the lower zones. This
2424  * interoperates with the page allocator fallback scheme to ensure that aging
2425  * of pages is balanced across the zones.
2426  */
2427 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2428                                                         int *classzone_idx)
2429 {
2430         int all_zones_ok;
2431         unsigned long balanced;
2432         int priority;
2433         int i;
2434         int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
2435         unsigned long total_scanned;
2436         struct reclaim_state *reclaim_state = current->reclaim_state;
2437         unsigned long nr_soft_reclaimed;
2438         unsigned long nr_soft_scanned;
2439         struct scan_control sc = {
2440                 .gfp_mask = GFP_KERNEL,
2441                 .may_unmap = 1,
2442                 .may_swap = 1,
2443                 /*
2444                  * kswapd doesn't want to be bailed out while reclaim. because
2445                  * we want to put equal scanning pressure on each zone.
2446                  */
2447                 .nr_to_reclaim = ULONG_MAX,
2448                 .swappiness = vm_swappiness,
2449                 .order = order,
2450                 .mem_cgroup = NULL,
2451         };
2452         struct shrink_control shrink = {
2453                 .gfp_mask = sc.gfp_mask,
2454         };
2455 loop_again:
2456         total_scanned = 0;
2457         sc.nr_reclaimed = 0;
2458         sc.may_writepage = !laptop_mode;
2459         count_vm_event(PAGEOUTRUN);
2460
2461         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2462                 unsigned long lru_pages = 0;
2463                 int has_under_min_watermark_zone = 0;
2464
2465                 /* The swap token gets in the way of swapout... */
2466                 if (!priority)
2467                         disable_swap_token(NULL);
2468
2469                 all_zones_ok = 1;
2470                 balanced = 0;
2471
2472                 /*
2473                  * Scan in the highmem->dma direction for the highest
2474                  * zone which needs scanning
2475                  */
2476                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2477                         struct zone *zone = pgdat->node_zones + i;
2478
2479                         if (!populated_zone(zone))
2480                                 continue;
2481
2482                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2483                                 continue;
2484
2485                         /*
2486                          * Do some background aging of the anon list, to give
2487                          * pages a chance to be referenced before reclaiming.
2488                          */
2489                         if (inactive_anon_is_low(zone, &sc))
2490                                 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2491                                                         &sc, priority, 0);
2492
2493                         if (!zone_watermark_ok_safe(zone, order,
2494                                         high_wmark_pages(zone), 0, 0)) {
2495                                 end_zone = i;
2496                                 break;
2497                         }
2498                 }
2499                 if (i < 0)
2500                         goto out;
2501
2502                 for (i = 0; i <= end_zone; i++) {
2503                         struct zone *zone = pgdat->node_zones + i;
2504
2505                         lru_pages += zone_reclaimable_pages(zone);
2506                 }
2507
2508                 /*
2509                  * Now scan the zone in the dma->highmem direction, stopping
2510                  * at the last zone which needs scanning.
2511                  *
2512                  * We do this because the page allocator works in the opposite
2513                  * direction.  This prevents the page allocator from allocating
2514                  * pages behind kswapd's direction of progress, which would
2515                  * cause too much scanning of the lower zones.
2516                  */
2517                 for (i = 0; i <= end_zone; i++) {
2518                         struct zone *zone = pgdat->node_zones + i;
2519                         int nr_slab;
2520                         unsigned long balance_gap;
2521
2522                         if (!populated_zone(zone))
2523                                 continue;
2524
2525                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2526                                 continue;
2527
2528                         sc.nr_scanned = 0;
2529
2530                         nr_soft_scanned = 0;
2531                         /*
2532                          * Call soft limit reclaim before calling shrink_zone.
2533                          */
2534                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2535                                                         order, sc.gfp_mask,
2536                                                         &nr_soft_scanned);
2537                         sc.nr_reclaimed += nr_soft_reclaimed;
2538                         total_scanned += nr_soft_scanned;
2539
2540                         /*
2541                          * We put equal pressure on every zone, unless
2542                          * one zone has way too many pages free
2543                          * already. The "too many pages" is defined
2544                          * as the high wmark plus a "gap" where the
2545                          * gap is either the low watermark or 1%
2546                          * of the zone, whichever is smaller.
2547                          */
2548                         balance_gap = min(low_wmark_pages(zone),
2549                                 (zone->present_pages +
2550                                         KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2551                                 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2552                         if (!zone_watermark_ok_safe(zone, order,
2553                                         high_wmark_pages(zone) + balance_gap,
2554                                         end_zone, 0)) {
2555                                 shrink_zone(priority, zone, &sc);
2556
2557                                 reclaim_state->reclaimed_slab = 0;
2558                                 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2559                                 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2560                                 total_scanned += sc.nr_scanned;
2561
2562                                 if (nr_slab == 0 && !zone_reclaimable(zone))
2563                                         zone->all_unreclaimable = 1;
2564                         }
2565
2566                         /*
2567                          * If we've done a decent amount of scanning and
2568                          * the reclaim ratio is low, start doing writepage
2569                          * even in laptop mode
2570                          */
2571                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2572                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2573                                 sc.may_writepage = 1;
2574
2575                         if (zone->all_unreclaimable) {
2576                                 if (end_zone && end_zone == i)
2577                                         end_zone--;
2578                                 continue;
2579                         }
2580
2581                         if (!zone_watermark_ok_safe(zone, order,
2582                                         high_wmark_pages(zone), end_zone, 0)) {
2583                                 all_zones_ok = 0;
2584                                 /*
2585                                  * We are still under min water mark.  This
2586                                  * means that we have a GFP_ATOMIC allocation
2587                                  * failure risk. Hurry up!
2588                                  */
2589                                 if (!zone_watermark_ok_safe(zone, order,
2590                                             min_wmark_pages(zone), end_zone, 0))
2591                                         has_under_min_watermark_zone = 1;
2592                         } else {
2593                                 /*
2594                                  * If a zone reaches its high watermark,
2595                                  * consider it to be no longer congested. It's
2596                                  * possible there are dirty pages backed by
2597                                  * congested BDIs but as pressure is relieved,
2598                                  * spectulatively avoid congestion waits
2599                                  */
2600                                 zone_clear_flag(zone, ZONE_CONGESTED);
2601                                 if (i <= *classzone_idx)
2602                                         balanced += zone->present_pages;
2603                         }
2604
2605                 }
2606                 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2607                         break;          /* kswapd: all done */
2608                 /*
2609                  * OK, kswapd is getting into trouble.  Take a nap, then take
2610                  * another pass across the zones.
2611                  */
2612                 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2613                         if (has_under_min_watermark_zone)
2614                                 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2615                         else
2616                                 congestion_wait(BLK_RW_ASYNC, HZ/10);
2617                 }
2618
2619                 /*
2620                  * We do this so kswapd doesn't build up large priorities for
2621                  * example when it is freeing in parallel with allocators. It
2622                  * matches the direct reclaim path behaviour in terms of impact
2623                  * on zone->*_priority.
2624                  */
2625                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2626                         break;
2627         }
2628 out:
2629
2630         /*
2631          * order-0: All zones must meet high watermark for a balanced node
2632          * high-order: Balanced zones must make up at least 25% of the node
2633          *             for the node to be balanced
2634          */
2635         if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2636                 cond_resched();
2637
2638                 try_to_freeze();
2639
2640                 /*
2641                  * Fragmentation may mean that the system cannot be
2642                  * rebalanced for high-order allocations in all zones.
2643                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2644                  * it means the zones have been fully scanned and are still
2645                  * not balanced. For high-order allocations, there is
2646                  * little point trying all over again as kswapd may
2647                  * infinite loop.
2648                  *
2649                  * Instead, recheck all watermarks at order-0 as they
2650                  * are the most important. If watermarks are ok, kswapd will go
2651                  * back to sleep. High-order users can still perform direct
2652                  * reclaim if they wish.
2653                  */
2654                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2655                         order = sc.order = 0;
2656
2657                 goto loop_again;
2658         }
2659
2660         /*
2661          * If kswapd was reclaiming at a higher order, it has the option of
2662          * sleeping without all zones being balanced. Before it does, it must
2663          * ensure that the watermarks for order-0 on *all* zones are met and
2664          * that the congestion flags are cleared. The congestion flag must
2665          * be cleared as kswapd is the only mechanism that clears the flag
2666          * and it is potentially going to sleep here.
2667          */
2668         if (order) {
2669                 for (i = 0; i <= end_zone; i++) {
2670                         struct zone *zone = pgdat->node_zones + i;
2671
2672                         if (!populated_zone(zone))
2673                                 continue;
2674
2675                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2676                                 continue;
2677
2678                         /* Confirm the zone is balanced for order-0 */
2679                         if (!zone_watermark_ok(zone, 0,
2680                                         high_wmark_pages(zone), 0, 0)) {
2681                                 order = sc.order = 0;
2682                                 goto loop_again;
2683                         }
2684
2685                         /* If balanced, clear the congested flag */
2686                         zone_clear_flag(zone, ZONE_CONGESTED);
2687                 }
2688         }
2689
2690         /*
2691          * Return the order we were reclaiming at so sleeping_prematurely()
2692          * makes a decision on the order we were last reclaiming at. However,
2693          * if another caller entered the allocator slow path while kswapd
2694          * was awake, order will remain at the higher level
2695          */
2696         *classzone_idx = end_zone;
2697         return order;
2698 }
2699
2700 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2701 {
2702         long remaining = 0;
2703         DEFINE_WAIT(wait);
2704
2705         if (freezing(current) || kthread_should_stop())
2706                 return;
2707
2708         prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2709
2710         /* Try to sleep for a short interval */
2711         if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2712                 remaining = schedule_timeout(HZ/10);
2713                 finish_wait(&pgdat->kswapd_wait, &wait);
2714                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2715         }
2716
2717         /*
2718          * After a short sleep, check if it was a premature sleep. If not, then
2719          * go fully to sleep until explicitly woken up.
2720          */
2721         if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2722                 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2723
2724                 /*
2725                  * vmstat counters are not perfectly accurate and the estimated
2726                  * value for counters such as NR_FREE_PAGES can deviate from the
2727                  * true value by nr_online_cpus * threshold. To avoid the zone
2728                  * watermarks being breached while under pressure, we reduce the
2729                  * per-cpu vmstat threshold while kswapd is awake and restore
2730                  * them before going back to sleep.
2731                  */
2732                 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2733                 schedule();
2734                 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2735         } else {
2736                 if (remaining)
2737                         count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2738                 else
2739                         count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2740         }
2741         finish_wait(&pgdat->kswapd_wait, &wait);
2742 }
2743
2744 /*
2745  * The background pageout daemon, started as a kernel thread
2746  * from the init process.
2747  *
2748  * This basically trickles out pages so that we have _some_
2749  * free memory available even if there is no other activity
2750  * that frees anything up. This is needed for things like routing
2751  * etc, where we otherwise might have all activity going on in
2752  * asynchronous contexts that cannot page things out.
2753  *
2754  * If there are applications that are active memory-allocators
2755  * (most normal use), this basically shouldn't matter.
2756  */
2757 static int kswapd(void *p)
2758 {
2759         unsigned long order, new_order;
2760         int classzone_idx, new_classzone_idx;
2761         pg_data_t *pgdat = (pg_data_t*)p;
2762         struct task_struct *tsk = current;
2763
2764         struct reclaim_state reclaim_state = {
2765                 .reclaimed_slab = 0,
2766         };
2767         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2768
2769         lockdep_set_current_reclaim_state(GFP_KERNEL);
2770
2771         if (!cpumask_empty(cpumask))
2772                 set_cpus_allowed_ptr(tsk, cpumask);
2773         current->reclaim_state = &reclaim_state;
2774
2775         /*
2776          * Tell the memory management that we're a "memory allocator",
2777          * and that if we need more memory we should get access to it
2778          * regardless (see "__alloc_pages()"). "kswapd" should
2779          * never get caught in the normal page freeing logic.
2780          *
2781          * (Kswapd normally doesn't need memory anyway, but sometimes
2782          * you need a small amount of memory in order to be able to
2783          * page out something else, and this flag essentially protects
2784          * us from recursively trying to free more memory as we're
2785          * trying to free the first piece of memory in the first place).
2786          */
2787         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2788         set_freezable();
2789
2790         order = new_order = 0;
2791         classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
2792         for ( ; ; ) {
2793                 int ret;
2794
2795                 /*
2796                  * If the last balance_pgdat was unsuccessful it's unlikely a
2797                  * new request of a similar or harder type will succeed soon
2798                  * so consider going to sleep on the basis we reclaimed at
2799                  */
2800                 if (classzone_idx >= new_classzone_idx && order == new_order) {
2801                         new_order = pgdat->kswapd_max_order;
2802                         new_classzone_idx = pgdat->classzone_idx;
2803                         pgdat->kswapd_max_order =  0;
2804                         pgdat->classzone_idx = pgdat->nr_zones - 1;
2805                 }
2806
2807                 if (order < new_order || classzone_idx > new_classzone_idx) {
2808                         /*
2809                          * Don't sleep if someone wants a larger 'order'
2810                          * allocation or has tigher zone constraints
2811                          */
2812                         order = new_order;
2813                         classzone_idx = new_classzone_idx;
2814                 } else {
2815                         kswapd_try_to_sleep(pgdat, order, classzone_idx);
2816                         order = pgdat->kswapd_max_order;
2817                         classzone_idx = pgdat->classzone_idx;
2818                         pgdat->kswapd_max_order = 0;
2819                         pgdat->classzone_idx = pgdat->nr_zones - 1;
2820                 }
2821
2822                 ret = try_to_freeze();
2823                 if (kthread_should_stop())
2824                         break;
2825
2826                 /*
2827                  * We can speed up thawing tasks if we don't call balance_pgdat
2828                  * after returning from the refrigerator
2829                  */
2830                 if (!ret) {
2831                         trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2832                         order = balance_pgdat(pgdat, order, &classzone_idx);
2833                 }
2834         }
2835         return 0;
2836 }
2837
2838 /*
2839  * A zone is low on free memory, so wake its kswapd task to service it.
2840  */
2841 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
2842 {
2843         pg_data_t *pgdat;
2844
2845         if (!populated_zone(zone))
2846                 return;
2847
2848         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2849                 return;
2850         pgdat = zone->zone_pgdat;
2851         if (pgdat->kswapd_max_order < order) {
2852                 pgdat->kswapd_max_order = order;
2853                 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
2854         }
2855         if (!waitqueue_active(&pgdat->kswapd_wait))
2856                 return;
2857         if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
2858                 return;
2859
2860         trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
2861         wake_up_interruptible(&pgdat->kswapd_wait);
2862 }
2863
2864 /*
2865  * The reclaimable count would be mostly accurate.
2866  * The less reclaimable pages may be
2867  * - mlocked pages, which will be moved to unevictable list when encountered
2868  * - mapped pages, which may require several travels to be reclaimed
2869  * - dirty pages, which is not "instantly" reclaimable
2870  */
2871 unsigned long global_reclaimable_pages(void)
2872 {
2873         int nr;
2874
2875         nr = global_page_state(NR_ACTIVE_FILE) +
2876              global_page_state(NR_INACTIVE_FILE);
2877
2878         if (nr_swap_pages > 0)
2879                 nr += global_page_state(NR_ACTIVE_ANON) +
2880                       global_page_state(NR_INACTIVE_ANON);
2881
2882         return nr;
2883 }
2884
2885 unsigned long zone_reclaimable_pages(struct zone *zone)
2886 {
2887         int nr;
2888
2889         nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2890              zone_page_state(zone, NR_INACTIVE_FILE);
2891
2892         if (nr_swap_pages > 0)
2893                 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2894                       zone_page_state(zone, NR_INACTIVE_ANON);
2895
2896         return nr;
2897 }
2898
2899 #ifdef CONFIG_HIBERNATION
2900 /*
2901  * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2902  * freed pages.
2903  *
2904  * Rather than trying to age LRUs the aim is to preserve the overall
2905  * LRU order by reclaiming preferentially
2906  * inactive > active > active referenced > active mapped
2907  */
2908 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
2909 {
2910         struct reclaim_state reclaim_state;
2911         struct scan_control sc = {
2912                 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2913                 .may_swap = 1,
2914                 .may_unmap = 1,
2915                 .may_writepage = 1,
2916                 .nr_to_reclaim = nr_to_reclaim,
2917                 .hibernation_mode = 1,
2918                 .swappiness = vm_swappiness,
2919                 .order = 0,
2920         };
2921         struct shrink_control shrink = {
2922                 .gfp_mask = sc.gfp_mask,
2923         };
2924         struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2925         struct task_struct *p = current;
2926         unsigned long nr_reclaimed;
2927
2928         p->flags |= PF_MEMALLOC;
2929         lockdep_set_current_reclaim_state(sc.gfp_mask);
2930         reclaim_state.reclaimed_slab = 0;
2931         p->reclaim_state = &reclaim_state;
2932
2933         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2934
2935         p->reclaim_state = NULL;
2936         lockdep_clear_current_reclaim_state();
2937         p->flags &= ~PF_MEMALLOC;
2938
2939         return nr_reclaimed;
2940 }
2941 #endif /* CONFIG_HIBERNATION */
2942
2943 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2944    not required for correctness.  So if the last cpu in a node goes
2945    away, we get changed to run anywhere: as the first one comes back,
2946    restore their cpu bindings. */
2947 static int __devinit cpu_callback(struct notifier_block *nfb,
2948                                   unsigned long action, void *hcpu)
2949 {
2950         int nid;
2951
2952         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2953                 for_each_node_state(nid, N_HIGH_MEMORY) {
2954                         pg_data_t *pgdat = NODE_DATA(nid);
2955                         const struct cpumask *mask;
2956
2957                         mask = cpumask_of_node(pgdat->node_id);
2958
2959                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2960                                 /* One of our CPUs online: restore mask */
2961                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2962                 }
2963         }
2964         return NOTIFY_OK;
2965 }
2966
2967 /*
2968  * This kswapd start function will be called by init and node-hot-add.
2969  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2970  */
2971 int kswapd_run(int nid)
2972 {
2973         pg_data_t *pgdat = NODE_DATA(nid);
2974         int ret = 0;
2975
2976         if (pgdat->kswapd)
2977                 return 0;
2978
2979         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2980         if (IS_ERR(pgdat->kswapd)) {
2981                 /* failure at boot is fatal */
2982                 BUG_ON(system_state == SYSTEM_BOOTING);
2983                 printk("Failed to start kswapd on node %d\n",nid);
2984                 ret = -1;
2985         }
2986         return ret;
2987 }
2988
2989 /*
2990  * Called by memory hotplug when all memory in a node is offlined.
2991  */
2992 void kswapd_stop(int nid)
2993 {
2994         struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
2995
2996         if (kswapd)
2997                 kthread_stop(kswapd);
2998 }
2999
3000 static int __init kswapd_init(void)
3001 {
3002         int nid;
3003
3004         swap_setup();
3005         for_each_node_state(nid, N_HIGH_MEMORY)
3006                 kswapd_run(nid);
3007         hotcpu_notifier(cpu_callback, 0);
3008         return 0;
3009 }
3010
3011 module_init(kswapd_init)
3012
3013 #ifdef CONFIG_NUMA
3014 /*
3015  * Zone reclaim mode
3016  *
3017  * If non-zero call zone_reclaim when the number of free pages falls below
3018  * the watermarks.
3019  */
3020 int zone_reclaim_mode __read_mostly;
3021
3022 #define RECLAIM_OFF 0
3023 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
3024 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
3025 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
3026
3027 /*
3028  * Priority for ZONE_RECLAIM. This determines the fraction of pages
3029  * of a node considered for each zone_reclaim. 4 scans 1/16th of
3030  * a zone.
3031  */
3032 #define ZONE_RECLAIM_PRIORITY 4
3033
3034 /*
3035  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3036  * occur.
3037  */
3038 int sysctl_min_unmapped_ratio = 1;
3039
3040 /*
3041  * If the number of slab pages in a zone grows beyond this percentage then
3042  * slab reclaim needs to occur.
3043  */
3044 int sysctl_min_slab_ratio = 5;
3045
3046 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3047 {
3048         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3049         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3050                 zone_page_state(zone, NR_ACTIVE_FILE);
3051
3052         /*
3053          * It's possible for there to be more file mapped pages than
3054          * accounted for by the pages on the file LRU lists because
3055          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3056          */
3057         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3058 }
3059
3060 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3061 static long zone_pagecache_reclaimable(struct zone *zone)
3062 {
3063         long nr_pagecache_reclaimable;
3064         long delta = 0;
3065
3066         /*
3067          * If RECLAIM_SWAP is set, then all file pages are considered
3068          * potentially reclaimable. Otherwise, we have to worry about
3069          * pages like swapcache and zone_unmapped_file_pages() provides
3070          * a better estimate
3071          */
3072         if (zone_reclaim_mode & RECLAIM_SWAP)
3073                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3074         else
3075                 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3076
3077         /* If we can't clean pages, remove dirty pages from consideration */
3078         if (!(zone_reclaim_mode & RECLAIM_WRITE))
3079                 delta += zone_page_state(zone, NR_FILE_DIRTY);
3080
3081         /* Watch for any possible underflows due to delta */
3082         if (unlikely(delta > nr_pagecache_reclaimable))
3083                 delta = nr_pagecache_reclaimable;
3084
3085         return nr_pagecache_reclaimable - delta;
3086 }
3087
3088 /*
3089  * Try to free up some pages from this zone through reclaim.
3090  */
3091 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3092 {
3093         /* Minimum pages needed in order to stay on node */
3094         const unsigned long nr_pages = 1 << order;
3095         struct task_struct *p = current;
3096         struct reclaim_state reclaim_state;
3097         int priority;
3098         struct scan_control sc = {
3099                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3100                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3101                 .may_swap = 1,
3102                 .nr_to_reclaim = max_t(unsigned long, nr_pages,
3103                                        SWAP_CLUSTER_MAX),
3104                 .gfp_mask = gfp_mask,
3105                 .swappiness = vm_swappiness,
3106                 .order = order,
3107         };
3108         struct shrink_control shrink = {
3109                 .gfp_mask = sc.gfp_mask,
3110         };
3111         unsigned long nr_slab_pages0, nr_slab_pages1;
3112
3113         cond_resched();
3114         /*
3115          * We need to be able to allocate from the reserves for RECLAIM_SWAP
3116          * and we also need to be able to write out pages for RECLAIM_WRITE
3117          * and RECLAIM_SWAP.
3118          */
3119         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3120         lockdep_set_current_reclaim_state(gfp_mask);
3121         reclaim_state.reclaimed_slab = 0;
3122         p->reclaim_state = &reclaim_state;
3123
3124         if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3125                 /*
3126                  * Free memory by calling shrink zone with increasing
3127                  * priorities until we have enough memory freed.
3128                  */
3129                 priority = ZONE_RECLAIM_PRIORITY;
3130                 do {
3131                         shrink_zone(priority, zone, &sc);
3132                         priority--;
3133                 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
3134         }
3135
3136         nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3137         if (nr_slab_pages0 > zone->min_slab_pages) {
3138                 /*
3139                  * shrink_slab() does not currently allow us to determine how
3140                  * many pages were freed in this zone. So we take the current
3141                  * number of slab pages and shake the slab until it is reduced
3142                  * by the same nr_pages that we used for reclaiming unmapped
3143                  * pages.
3144                  *
3145                  * Note that shrink_slab will free memory on all zones and may
3146                  * take a long time.
3147                  */
3148                 for (;;) {
3149                         unsigned long lru_pages = zone_reclaimable_pages(zone);
3150
3151                         /* No reclaimable slab or very low memory pressure */
3152                         if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3153                                 break;
3154
3155                         /* Freed enough memory */
3156                         nr_slab_pages1 = zone_page_state(zone,
3157                                                         NR_SLAB_RECLAIMABLE);
3158                         if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3159                                 break;
3160                 }
3161
3162                 /*
3163                  * Update nr_reclaimed by the number of slab pages we
3164                  * reclaimed from this zone.
3165                  */
3166                 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3167                 if (nr_slab_pages1 < nr_slab_pages0)
3168                         sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3169         }
3170
3171         p->reclaim_state = NULL;
3172         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3173         lockdep_clear_current_reclaim_state();
3174         return sc.nr_reclaimed >= nr_pages;
3175 }
3176
3177 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3178 {
3179         int node_id;
3180         int ret;
3181
3182         /*
3183          * Zone reclaim reclaims unmapped file backed pages and
3184          * slab pages if we are over the defined limits.
3185          *
3186          * A small portion of unmapped file backed pages is needed for
3187          * file I/O otherwise pages read by file I/O will be immediately
3188          * thrown out if the zone is overallocated. So we do not reclaim
3189          * if less than a specified percentage of the zone is used by
3190          * unmapped file backed pages.
3191          */
3192         if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3193             zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3194                 return ZONE_RECLAIM_FULL;
3195
3196         if (zone->all_unreclaimable)
3197                 return ZONE_RECLAIM_FULL;
3198
3199         /*
3200          * Do not scan if the allocation should not be delayed.
3201          */
3202         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3203                 return ZONE_RECLAIM_NOSCAN;
3204
3205         /*
3206          * Only run zone reclaim on the local zone or on zones that do not
3207          * have associated processors. This will favor the local processor
3208          * over remote processors and spread off node memory allocations
3209          * as wide as possible.
3210          */
3211         node_id = zone_to_nid(zone);
3212         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3213                 return ZONE_RECLAIM_NOSCAN;
3214
3215         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3216                 return ZONE_RECLAIM_NOSCAN;
3217
3218         ret = __zone_reclaim(zone, gfp_mask, order);
3219         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3220
3221         if (!ret)
3222                 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3223
3224         return ret;
3225 }
3226 #endif
3227
3228 /*
3229  * page_evictable - test whether a page is evictable
3230  * @page: the page to test
3231  * @vma: the VMA in which the page is or will be mapped, may be NULL
3232  *
3233  * Test whether page is evictable--i.e., should be placed on active/inactive
3234  * lists vs unevictable list.  The vma argument is !NULL when called from the
3235  * fault path to determine how to instantate a new page.
3236  *
3237  * Reasons page might not be evictable:
3238  * (1) page's mapping marked unevictable
3239  * (2) page is part of an mlocked VMA
3240  *
3241  */
3242 int page_evictable(struct page *page, struct vm_area_struct *vma)
3243 {
3244
3245         if (mapping_unevictable(page_mapping(page)))
3246                 return 0;
3247
3248         if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
3249                 return 0;
3250
3251         return 1;
3252 }
3253
3254 /**
3255  * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
3256  * @page: page to check evictability and move to appropriate lru list
3257  * @zone: zone page is in
3258  *
3259  * Checks a page for evictability and moves the page to the appropriate
3260  * zone lru list.
3261  *
3262  * Restrictions: zone->lru_lock must be held, page must be on LRU and must
3263  * have PageUnevictable set.
3264  */
3265 static void check_move_unevictable_page(struct page *page, struct zone *zone)
3266 {
3267         VM_BUG_ON(PageActive(page));
3268
3269 retry:
3270         ClearPageUnevictable(page);
3271         if (page_evictable(page, NULL)) {
3272                 enum lru_list l = page_lru_base_type(page);
3273
3274                 __dec_zone_state(zone, NR_UNEVICTABLE);
3275                 list_move(&page->lru, &zone->lru[l].list);
3276                 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
3277                 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
3278                 __count_vm_event(UNEVICTABLE_PGRESCUED);
3279         } else {
3280                 /*
3281                  * rotate unevictable list
3282                  */
3283                 SetPageUnevictable(page);
3284                 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
3285                 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
3286                 if (page_evictable(page, NULL))
3287                         goto retry;
3288         }
3289 }
3290
3291 /**
3292  * scan_mapping_unevictable_pages - scan an address space for evictable pages
3293  * @mapping: struct address_space to scan for evictable pages
3294  *
3295  * Scan all pages in mapping.  Check unevictable pages for
3296  * evictability and move them to the appropriate zone lru list.
3297  */
3298 void scan_mapping_unevictable_pages(struct address_space *mapping)
3299 {
3300         pgoff_t next = 0;
3301         pgoff_t end   = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
3302                          PAGE_CACHE_SHIFT;
3303         struct zone *zone;
3304         struct pagevec pvec;
3305
3306         if (mapping->nrpages == 0)
3307                 return;
3308
3309         pagevec_init(&pvec, 0);
3310         while (next < end &&
3311                 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
3312                 int i;
3313                 int pg_scanned = 0;
3314
3315                 zone = NULL;
3316
3317                 for (i = 0; i < pagevec_count(&pvec); i++) {
3318                         struct page *page = pvec.pages[i];
3319                         pgoff_t page_index = page->index;
3320                         struct zone *pagezone = page_zone(page);
3321
3322                         pg_scanned++;
3323                         if (page_index > next)
3324                                 next = page_index;
3325                         next++;
3326
3327                         if (pagezone != zone) {
3328                                 if (zone)
3329                                         spin_unlock_irq(&zone->lru_lock);
3330                                 zone = pagezone;
3331                                 spin_lock_irq(&zone->lru_lock);
3332                         }
3333
3334                         if (PageLRU(page) && PageUnevictable(page))
3335                                 check_move_unevictable_page(page, zone);
3336                 }
3337                 if (zone)
3338                         spin_unlock_irq(&zone->lru_lock);
3339                 pagevec_release(&pvec);
3340
3341                 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
3342         }
3343
3344 }
3345
3346 /**
3347  * scan_zone_unevictable_pages - check unevictable list for evictable pages
3348  * @zone - zone of which to scan the unevictable list
3349  *
3350  * Scan @zone's unevictable LRU lists to check for pages that have become
3351  * evictable.  Move those that have to @zone's inactive list where they
3352  * become candidates for reclaim, unless shrink_inactive_zone() decides
3353  * to reactivate them.  Pages that are still unevictable are rotated
3354  * back onto @zone's unevictable list.
3355  */
3356 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
3357 static void scan_zone_unevictable_pages(struct zone *zone)
3358 {
3359         struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
3360         unsigned long scan;
3361         unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
3362
3363         while (nr_to_scan > 0) {
3364                 unsigned long batch_size = min(nr_to_scan,
3365                                                 SCAN_UNEVICTABLE_BATCH_SIZE);
3366
3367                 spin_lock_irq(&zone->lru_lock);
3368                 for (scan = 0;  scan < batch_size; scan++) {
3369                         struct page *page = lru_to_page(l_unevictable);
3370
3371                         if (!trylock_page(page))
3372                                 continue;
3373
3374                         prefetchw_prev_lru_page(page, l_unevictable, flags);
3375
3376                         if (likely(PageLRU(page) && PageUnevictable(page)))
3377                                 check_move_unevictable_page(page, zone);
3378
3379                         unlock_page(page);
3380                 }
3381                 spin_unlock_irq(&zone->lru_lock);
3382
3383                 nr_to_scan -= batch_size;
3384         }
3385 }
3386
3387
3388 /**
3389  * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
3390  *
3391  * A really big hammer:  scan all zones' unevictable LRU lists to check for
3392  * pages that have become evictable.  Move those back to the zones'
3393  * inactive list where they become candidates for reclaim.
3394  * This occurs when, e.g., we have unswappable pages on the unevictable lists,
3395  * and we add swap to the system.  As such, it runs in the context of a task
3396  * that has possibly/probably made some previously unevictable pages
3397  * evictable.
3398  */
3399 static void scan_all_zones_unevictable_pages(void)
3400 {
3401         struct zone *zone;
3402
3403         for_each_zone(zone) {
3404                 scan_zone_unevictable_pages(zone);
3405         }
3406 }
3407
3408 /*
3409  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
3410  * all nodes' unevictable lists for evictable pages
3411  */
3412 unsigned long scan_unevictable_pages;
3413
3414 int scan_unevictable_handler(struct ctl_table *table, int write,
3415                            void __user *buffer,
3416                            size_t *length, loff_t *ppos)
3417 {
3418         proc_doulongvec_minmax(table, write, buffer, length, ppos);
3419
3420         if (write && *(unsigned long *)table->data)
3421                 scan_all_zones_unevictable_pages();
3422
3423         scan_unevictable_pages = 0;
3424         return 0;
3425 }
3426
3427 #ifdef CONFIG_NUMA
3428 /*
3429  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
3430  * a specified node's per zone unevictable lists for evictable pages.
3431  */
3432
3433 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
3434                                           struct sysdev_attribute *attr,
3435                                           char *buf)
3436 {
3437         return sprintf(buf, "0\n");     /* always zero; should fit... */
3438 }
3439
3440 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
3441                                            struct sysdev_attribute *attr,
3442                                         const char *buf, size_t count)
3443 {
3444         struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
3445         struct zone *zone;
3446         unsigned long res;
3447         unsigned long req = strict_strtoul(buf, 10, &res);
3448
3449         if (!req)
3450                 return 1;       /* zero is no-op */
3451
3452         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
3453                 if (!populated_zone(zone))
3454                         continue;
3455                 scan_zone_unevictable_pages(zone);
3456         }
3457         return 1;
3458 }
3459
3460
3461 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3462                         read_scan_unevictable_node,
3463                         write_scan_unevictable_node);
3464
3465 int scan_unevictable_register_node(struct node *node)
3466 {
3467         return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
3468 }
3469
3470 void scan_unevictable_unregister_node(struct node *node)
3471 {
3472         sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
3473 }
3474 #endif