tmpfs: copy truncate_inode_pages_range
[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 order;
99
100         /*
101          * Intend to reclaim enough continuous memory rather than reclaim
102          * enough amount of memory. i.e, mode for high order allocation.
103          */
104         reclaim_mode_t reclaim_mode;
105
106         /* Which cgroup do we reclaim from */
107         struct mem_cgroup *mem_cgroup;
108         struct memcg_scanrecord *memcg_record;
109
110         /*
111          * Nodemask of nodes allowed by the caller. If NULL, all nodes
112          * are scanned.
113          */
114         nodemask_t      *nodemask;
115 };
116
117 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
118
119 #ifdef ARCH_HAS_PREFETCH
120 #define prefetch_prev_lru_page(_page, _base, _field)                    \
121         do {                                                            \
122                 if ((_page)->lru.prev != _base) {                       \
123                         struct page *prev;                              \
124                                                                         \
125                         prev = lru_to_page(&(_page->lru));              \
126                         prefetch(&prev->_field);                        \
127                 }                                                       \
128         } while (0)
129 #else
130 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
131 #endif
132
133 #ifdef ARCH_HAS_PREFETCHW
134 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
135         do {                                                            \
136                 if ((_page)->lru.prev != _base) {                       \
137                         struct page *prev;                              \
138                                                                         \
139                         prev = lru_to_page(&(_page->lru));              \
140                         prefetchw(&prev->_field);                       \
141                 }                                                       \
142         } while (0)
143 #else
144 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
145 #endif
146
147 /*
148  * From 0 .. 100.  Higher means more swappy.
149  */
150 int vm_swappiness = 60;
151 long vm_total_pages;    /* The total number of pages which the VM controls */
152
153 static LIST_HEAD(shrinker_list);
154 static DECLARE_RWSEM(shrinker_rwsem);
155
156 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
157 #define scanning_global_lru(sc) (!(sc)->mem_cgroup)
158 #else
159 #define scanning_global_lru(sc) (1)
160 #endif
161
162 static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
163                                                   struct scan_control *sc)
164 {
165         if (!scanning_global_lru(sc))
166                 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
167
168         return &zone->reclaim_stat;
169 }
170
171 static unsigned long zone_nr_lru_pages(struct zone *zone,
172                                 struct scan_control *sc, enum lru_list lru)
173 {
174         if (!scanning_global_lru(sc))
175                 return mem_cgroup_zone_nr_lru_pages(sc->mem_cgroup,
176                                 zone_to_nid(zone), zone_idx(zone), BIT(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                         if (!scanning_global_lru(sc))
1353                                 sc->memcg_record->nr_rotated[file] += numpages;
1354                 }
1355                 if (!pagevec_add(&pvec, page)) {
1356                         spin_unlock_irq(&zone->lru_lock);
1357                         __pagevec_release(&pvec);
1358                         spin_lock_irq(&zone->lru_lock);
1359                 }
1360         }
1361         __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1362         __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1363
1364         spin_unlock_irq(&zone->lru_lock);
1365         pagevec_release(&pvec);
1366 }
1367
1368 static noinline_for_stack void update_isolated_counts(struct zone *zone,
1369                                         struct scan_control *sc,
1370                                         unsigned long *nr_anon,
1371                                         unsigned long *nr_file,
1372                                         struct list_head *isolated_list)
1373 {
1374         unsigned long nr_active;
1375         unsigned int count[NR_LRU_LISTS] = { 0, };
1376         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1377
1378         nr_active = clear_active_flags(isolated_list, count);
1379         __count_vm_events(PGDEACTIVATE, nr_active);
1380
1381         __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1382                               -count[LRU_ACTIVE_FILE]);
1383         __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1384                               -count[LRU_INACTIVE_FILE]);
1385         __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1386                               -count[LRU_ACTIVE_ANON]);
1387         __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1388                               -count[LRU_INACTIVE_ANON]);
1389
1390         *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1391         *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1392         __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1393         __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1394
1395         reclaim_stat->recent_scanned[0] += *nr_anon;
1396         reclaim_stat->recent_scanned[1] += *nr_file;
1397         if (!scanning_global_lru(sc)) {
1398                 sc->memcg_record->nr_scanned[0] += *nr_anon;
1399                 sc->memcg_record->nr_scanned[1] += *nr_file;
1400         }
1401 }
1402
1403 /*
1404  * Returns true if the caller should wait to clean dirty/writeback pages.
1405  *
1406  * If we are direct reclaiming for contiguous pages and we do not reclaim
1407  * everything in the list, try again and wait for writeback IO to complete.
1408  * This will stall high-order allocations noticeably. Only do that when really
1409  * need to free the pages under high memory pressure.
1410  */
1411 static inline bool should_reclaim_stall(unsigned long nr_taken,
1412                                         unsigned long nr_freed,
1413                                         int priority,
1414                                         struct scan_control *sc)
1415 {
1416         int lumpy_stall_priority;
1417
1418         /* kswapd should not stall on sync IO */
1419         if (current_is_kswapd())
1420                 return false;
1421
1422         /* Only stall on lumpy reclaim */
1423         if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
1424                 return false;
1425
1426         /* If we have relaimed everything on the isolated list, no stall */
1427         if (nr_freed == nr_taken)
1428                 return false;
1429
1430         /*
1431          * For high-order allocations, there are two stall thresholds.
1432          * High-cost allocations stall immediately where as lower
1433          * order allocations such as stacks require the scanning
1434          * priority to be much higher before stalling.
1435          */
1436         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1437                 lumpy_stall_priority = DEF_PRIORITY;
1438         else
1439                 lumpy_stall_priority = DEF_PRIORITY / 3;
1440
1441         return priority <= lumpy_stall_priority;
1442 }
1443
1444 /*
1445  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1446  * of reclaimed pages
1447  */
1448 static noinline_for_stack unsigned long
1449 shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1450                         struct scan_control *sc, int priority, int file)
1451 {
1452         LIST_HEAD(page_list);
1453         unsigned long nr_scanned;
1454         unsigned long nr_reclaimed = 0;
1455         unsigned long nr_taken;
1456         unsigned long nr_anon;
1457         unsigned long nr_file;
1458
1459         while (unlikely(too_many_isolated(zone, file, sc))) {
1460                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1461
1462                 /* We are about to die and free our memory. Return now. */
1463                 if (fatal_signal_pending(current))
1464                         return SWAP_CLUSTER_MAX;
1465         }
1466
1467         set_reclaim_mode(priority, sc, false);
1468         lru_add_drain();
1469         spin_lock_irq(&zone->lru_lock);
1470
1471         if (scanning_global_lru(sc)) {
1472                 nr_taken = isolate_pages_global(nr_to_scan,
1473                         &page_list, &nr_scanned, sc->order,
1474                         sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM ?
1475                                         ISOLATE_BOTH : ISOLATE_INACTIVE,
1476                         zone, 0, file);
1477                 zone->pages_scanned += nr_scanned;
1478                 if (current_is_kswapd())
1479                         __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1480                                                nr_scanned);
1481                 else
1482                         __count_zone_vm_events(PGSCAN_DIRECT, zone,
1483                                                nr_scanned);
1484         } else {
1485                 nr_taken = mem_cgroup_isolate_pages(nr_to_scan,
1486                         &page_list, &nr_scanned, sc->order,
1487                         sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM ?
1488                                         ISOLATE_BOTH : ISOLATE_INACTIVE,
1489                         zone, sc->mem_cgroup,
1490                         0, file);
1491                 /*
1492                  * mem_cgroup_isolate_pages() keeps track of
1493                  * scanned pages on its own.
1494                  */
1495         }
1496
1497         if (nr_taken == 0) {
1498                 spin_unlock_irq(&zone->lru_lock);
1499                 return 0;
1500         }
1501
1502         update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
1503
1504         spin_unlock_irq(&zone->lru_lock);
1505
1506         nr_reclaimed = shrink_page_list(&page_list, zone, sc);
1507
1508         /* Check if we should syncronously wait for writeback */
1509         if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1510                 set_reclaim_mode(priority, sc, true);
1511                 nr_reclaimed += shrink_page_list(&page_list, zone, sc);
1512         }
1513
1514         if (!scanning_global_lru(sc))
1515                 sc->memcg_record->nr_freed[file] += nr_reclaimed;
1516
1517         local_irq_disable();
1518         if (current_is_kswapd())
1519                 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1520         __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1521
1522         putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
1523
1524         trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1525                 zone_idx(zone),
1526                 nr_scanned, nr_reclaimed,
1527                 priority,
1528                 trace_shrink_flags(file, sc->reclaim_mode));
1529         return nr_reclaimed;
1530 }
1531
1532 /*
1533  * This moves pages from the active list to the inactive list.
1534  *
1535  * We move them the other way if the page is referenced by one or more
1536  * processes, from rmap.
1537  *
1538  * If the pages are mostly unmapped, the processing is fast and it is
1539  * appropriate to hold zone->lru_lock across the whole operation.  But if
1540  * the pages are mapped, the processing is slow (page_referenced()) so we
1541  * should drop zone->lru_lock around each page.  It's impossible to balance
1542  * this, so instead we remove the pages from the LRU while processing them.
1543  * It is safe to rely on PG_active against the non-LRU pages in here because
1544  * nobody will play with that bit on a non-LRU page.
1545  *
1546  * The downside is that we have to touch page->_count against each page.
1547  * But we had to alter page->flags anyway.
1548  */
1549
1550 static void move_active_pages_to_lru(struct zone *zone,
1551                                      struct list_head *list,
1552                                      enum lru_list lru)
1553 {
1554         unsigned long pgmoved = 0;
1555         struct pagevec pvec;
1556         struct page *page;
1557
1558         pagevec_init(&pvec, 1);
1559
1560         while (!list_empty(list)) {
1561                 page = lru_to_page(list);
1562
1563                 VM_BUG_ON(PageLRU(page));
1564                 SetPageLRU(page);
1565
1566                 list_move(&page->lru, &zone->lru[lru].list);
1567                 mem_cgroup_add_lru_list(page, lru);
1568                 pgmoved += hpage_nr_pages(page);
1569
1570                 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1571                         spin_unlock_irq(&zone->lru_lock);
1572                         if (buffer_heads_over_limit)
1573                                 pagevec_strip(&pvec);
1574                         __pagevec_release(&pvec);
1575                         spin_lock_irq(&zone->lru_lock);
1576                 }
1577         }
1578         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1579         if (!is_active_lru(lru))
1580                 __count_vm_events(PGDEACTIVATE, pgmoved);
1581 }
1582
1583 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1584                         struct scan_control *sc, int priority, int file)
1585 {
1586         unsigned long nr_taken;
1587         unsigned long pgscanned;
1588         unsigned long vm_flags;
1589         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1590         LIST_HEAD(l_active);
1591         LIST_HEAD(l_inactive);
1592         struct page *page;
1593         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1594         unsigned long nr_rotated = 0;
1595
1596         lru_add_drain();
1597         spin_lock_irq(&zone->lru_lock);
1598         if (scanning_global_lru(sc)) {
1599                 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1600                                                 &pgscanned, sc->order,
1601                                                 ISOLATE_ACTIVE, zone,
1602                                                 1, file);
1603                 zone->pages_scanned += pgscanned;
1604         } else {
1605                 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1606                                                 &pgscanned, sc->order,
1607                                                 ISOLATE_ACTIVE, zone,
1608                                                 sc->mem_cgroup, 1, file);
1609                 /*
1610                  * mem_cgroup_isolate_pages() keeps track of
1611                  * scanned pages on its own.
1612                  */
1613         }
1614
1615         reclaim_stat->recent_scanned[file] += nr_taken;
1616         if (!scanning_global_lru(sc))
1617                 sc->memcg_record->nr_scanned[file] += nr_taken;
1618
1619         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1620         if (file)
1621                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1622         else
1623                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1624         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1625         spin_unlock_irq(&zone->lru_lock);
1626
1627         while (!list_empty(&l_hold)) {
1628                 cond_resched();
1629                 page = lru_to_page(&l_hold);
1630                 list_del(&page->lru);
1631
1632                 if (unlikely(!page_evictable(page, NULL))) {
1633                         putback_lru_page(page);
1634                         continue;
1635                 }
1636
1637                 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1638                         nr_rotated += hpage_nr_pages(page);
1639                         /*
1640                          * Identify referenced, file-backed active pages and
1641                          * give them one more trip around the active list. So
1642                          * that executable code get better chances to stay in
1643                          * memory under moderate memory pressure.  Anon pages
1644                          * are not likely to be evicted by use-once streaming
1645                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1646                          * so we ignore them here.
1647                          */
1648                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1649                                 list_add(&page->lru, &l_active);
1650                                 continue;
1651                         }
1652                 }
1653
1654                 ClearPageActive(page);  /* we are de-activating */
1655                 list_add(&page->lru, &l_inactive);
1656         }
1657
1658         /*
1659          * Move pages back to the lru list.
1660          */
1661         spin_lock_irq(&zone->lru_lock);
1662         /*
1663          * Count referenced pages from currently used mappings as rotated,
1664          * even though only some of them are actually re-activated.  This
1665          * helps balance scan pressure between file and anonymous pages in
1666          * get_scan_ratio.
1667          */
1668         reclaim_stat->recent_rotated[file] += nr_rotated;
1669         if (!scanning_global_lru(sc))
1670                 sc->memcg_record->nr_rotated[file] += nr_rotated;
1671
1672         move_active_pages_to_lru(zone, &l_active,
1673                                                 LRU_ACTIVE + file * LRU_FILE);
1674         move_active_pages_to_lru(zone, &l_inactive,
1675                                                 LRU_BASE   + file * LRU_FILE);
1676         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1677         spin_unlock_irq(&zone->lru_lock);
1678 }
1679
1680 #ifdef CONFIG_SWAP
1681 static int inactive_anon_is_low_global(struct zone *zone)
1682 {
1683         unsigned long active, inactive;
1684
1685         active = zone_page_state(zone, NR_ACTIVE_ANON);
1686         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1687
1688         if (inactive * zone->inactive_ratio < active)
1689                 return 1;
1690
1691         return 0;
1692 }
1693
1694 /**
1695  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1696  * @zone: zone to check
1697  * @sc:   scan control of this context
1698  *
1699  * Returns true if the zone does not have enough inactive anon pages,
1700  * meaning some active anon pages need to be deactivated.
1701  */
1702 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1703 {
1704         int low;
1705
1706         /*
1707          * If we don't have swap space, anonymous page deactivation
1708          * is pointless.
1709          */
1710         if (!total_swap_pages)
1711                 return 0;
1712
1713         if (scanning_global_lru(sc))
1714                 low = inactive_anon_is_low_global(zone);
1715         else
1716                 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1717         return low;
1718 }
1719 #else
1720 static inline int inactive_anon_is_low(struct zone *zone,
1721                                         struct scan_control *sc)
1722 {
1723         return 0;
1724 }
1725 #endif
1726
1727 static int inactive_file_is_low_global(struct zone *zone)
1728 {
1729         unsigned long active, inactive;
1730
1731         active = zone_page_state(zone, NR_ACTIVE_FILE);
1732         inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1733
1734         return (active > inactive);
1735 }
1736
1737 /**
1738  * inactive_file_is_low - check if file pages need to be deactivated
1739  * @zone: zone to check
1740  * @sc:   scan control of this context
1741  *
1742  * When the system is doing streaming IO, memory pressure here
1743  * ensures that active file pages get deactivated, until more
1744  * than half of the file pages are on the inactive list.
1745  *
1746  * Once we get to that situation, protect the system's working
1747  * set from being evicted by disabling active file page aging.
1748  *
1749  * This uses a different ratio than the anonymous pages, because
1750  * the page cache uses a use-once replacement algorithm.
1751  */
1752 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1753 {
1754         int low;
1755
1756         if (scanning_global_lru(sc))
1757                 low = inactive_file_is_low_global(zone);
1758         else
1759                 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1760         return low;
1761 }
1762
1763 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1764                                 int file)
1765 {
1766         if (file)
1767                 return inactive_file_is_low(zone, sc);
1768         else
1769                 return inactive_anon_is_low(zone, sc);
1770 }
1771
1772 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1773         struct zone *zone, struct scan_control *sc, int priority)
1774 {
1775         int file = is_file_lru(lru);
1776
1777         if (is_active_lru(lru)) {
1778                 if (inactive_list_is_low(zone, sc, file))
1779                     shrink_active_list(nr_to_scan, zone, sc, priority, file);
1780                 return 0;
1781         }
1782
1783         return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1784 }
1785
1786 static int vmscan_swappiness(struct scan_control *sc)
1787 {
1788         if (scanning_global_lru(sc))
1789                 return vm_swappiness;
1790         return mem_cgroup_swappiness(sc->mem_cgroup);
1791 }
1792
1793 /*
1794  * Determine how aggressively the anon and file LRU lists should be
1795  * scanned.  The relative value of each set of LRU lists is determined
1796  * by looking at the fraction of the pages scanned we did rotate back
1797  * onto the active list instead of evict.
1798  *
1799  * nr[0] = anon pages to scan; nr[1] = file pages to scan
1800  */
1801 static void get_scan_count(struct zone *zone, struct scan_control *sc,
1802                                         unsigned long *nr, int priority)
1803 {
1804         unsigned long anon, file, free;
1805         unsigned long anon_prio, file_prio;
1806         unsigned long ap, fp;
1807         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1808         u64 fraction[2], denominator;
1809         enum lru_list l;
1810         int noswap = 0;
1811         int force_scan = 0;
1812         unsigned long nr_force_scan[2];
1813
1814
1815         anon  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1816                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1817         file  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1818                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1819
1820         if (((anon + file) >> priority) < SWAP_CLUSTER_MAX) {
1821                 /* kswapd does zone balancing and need to scan this zone */
1822                 if (scanning_global_lru(sc) && current_is_kswapd())
1823                         force_scan = 1;
1824                 /* memcg may have small limit and need to avoid priority drop */
1825                 if (!scanning_global_lru(sc))
1826                         force_scan = 1;
1827         }
1828
1829         /* If we have no swap space, do not bother scanning anon pages. */
1830         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1831                 noswap = 1;
1832                 fraction[0] = 0;
1833                 fraction[1] = 1;
1834                 denominator = 1;
1835                 nr_force_scan[0] = 0;
1836                 nr_force_scan[1] = SWAP_CLUSTER_MAX;
1837                 goto out;
1838         }
1839
1840         if (scanning_global_lru(sc)) {
1841                 free  = zone_page_state(zone, NR_FREE_PAGES);
1842                 /* If we have very few page cache pages,
1843                    force-scan anon pages. */
1844                 if (unlikely(file + free <= high_wmark_pages(zone))) {
1845                         fraction[0] = 1;
1846                         fraction[1] = 0;
1847                         denominator = 1;
1848                         nr_force_scan[0] = SWAP_CLUSTER_MAX;
1849                         nr_force_scan[1] = 0;
1850                         goto out;
1851                 }
1852         }
1853
1854         /*
1855          * With swappiness at 100, anonymous and file have the same priority.
1856          * This scanning priority is essentially the inverse of IO cost.
1857          */
1858         anon_prio = vmscan_swappiness(sc);
1859         file_prio = 200 - vmscan_swappiness(sc);
1860
1861         /*
1862          * OK, so we have swap space and a fair amount of page cache
1863          * pages.  We use the recently rotated / recently scanned
1864          * ratios to determine how valuable each cache is.
1865          *
1866          * Because workloads change over time (and to avoid overflow)
1867          * we keep these statistics as a floating average, which ends
1868          * up weighing recent references more than old ones.
1869          *
1870          * anon in [0], file in [1]
1871          */
1872         spin_lock_irq(&zone->lru_lock);
1873         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1874                 reclaim_stat->recent_scanned[0] /= 2;
1875                 reclaim_stat->recent_rotated[0] /= 2;
1876         }
1877
1878         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1879                 reclaim_stat->recent_scanned[1] /= 2;
1880                 reclaim_stat->recent_rotated[1] /= 2;
1881         }
1882
1883         /*
1884          * The amount of pressure on anon vs file pages is inversely
1885          * proportional to the fraction of recently scanned pages on
1886          * each list that were recently referenced and in active use.
1887          */
1888         ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1889         ap /= reclaim_stat->recent_rotated[0] + 1;
1890
1891         fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1892         fp /= reclaim_stat->recent_rotated[1] + 1;
1893         spin_unlock_irq(&zone->lru_lock);
1894
1895         fraction[0] = ap;
1896         fraction[1] = fp;
1897         denominator = ap + fp + 1;
1898         if (force_scan) {
1899                 unsigned long scan = SWAP_CLUSTER_MAX;
1900                 nr_force_scan[0] = div64_u64(scan * ap, denominator);
1901                 nr_force_scan[1] = div64_u64(scan * fp, denominator);
1902         }
1903 out:
1904         for_each_evictable_lru(l) {
1905                 int file = is_file_lru(l);
1906                 unsigned long scan;
1907
1908                 scan = zone_nr_lru_pages(zone, sc, l);
1909                 if (priority || noswap) {
1910                         scan >>= priority;
1911                         scan = div64_u64(scan * fraction[file], denominator);
1912                 }
1913
1914                 /*
1915                  * If zone is small or memcg is small, nr[l] can be 0.
1916                  * This results no-scan on this priority and priority drop down.
1917                  * For global direct reclaim, it can visit next zone and tend
1918                  * not to have problems. For global kswapd, it's for zone
1919                  * balancing and it need to scan a small amounts. When using
1920                  * memcg, priority drop can cause big latency. So, it's better
1921                  * to scan small amount. See may_noscan above.
1922                  */
1923                 if (!scan && force_scan)
1924                         scan = nr_force_scan[file];
1925                 nr[l] = scan;
1926         }
1927 }
1928
1929 /*
1930  * Reclaim/compaction depends on a number of pages being freed. To avoid
1931  * disruption to the system, a small number of order-0 pages continue to be
1932  * rotated and reclaimed in the normal fashion. However, by the time we get
1933  * back to the allocator and call try_to_compact_zone(), we ensure that
1934  * there are enough free pages for it to be likely successful
1935  */
1936 static inline bool should_continue_reclaim(struct zone *zone,
1937                                         unsigned long nr_reclaimed,
1938                                         unsigned long nr_scanned,
1939                                         struct scan_control *sc)
1940 {
1941         unsigned long pages_for_compaction;
1942         unsigned long inactive_lru_pages;
1943
1944         /* If not in reclaim/compaction mode, stop */
1945         if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
1946                 return false;
1947
1948         /* Consider stopping depending on scan and reclaim activity */
1949         if (sc->gfp_mask & __GFP_REPEAT) {
1950                 /*
1951                  * For __GFP_REPEAT allocations, stop reclaiming if the
1952                  * full LRU list has been scanned and we are still failing
1953                  * to reclaim pages. This full LRU scan is potentially
1954                  * expensive but a __GFP_REPEAT caller really wants to succeed
1955                  */
1956                 if (!nr_reclaimed && !nr_scanned)
1957                         return false;
1958         } else {
1959                 /*
1960                  * For non-__GFP_REPEAT allocations which can presumably
1961                  * fail without consequence, stop if we failed to reclaim
1962                  * any pages from the last SWAP_CLUSTER_MAX number of
1963                  * pages that were scanned. This will return to the
1964                  * caller faster at the risk reclaim/compaction and
1965                  * the resulting allocation attempt fails
1966                  */
1967                 if (!nr_reclaimed)
1968                         return false;
1969         }
1970
1971         /*
1972          * If we have not reclaimed enough pages for compaction and the
1973          * inactive lists are large enough, continue reclaiming
1974          */
1975         pages_for_compaction = (2UL << sc->order);
1976         inactive_lru_pages = zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON) +
1977                                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1978         if (sc->nr_reclaimed < pages_for_compaction &&
1979                         inactive_lru_pages > pages_for_compaction)
1980                 return true;
1981
1982         /* If compaction would go ahead or the allocation would succeed, stop */
1983         switch (compaction_suitable(zone, sc->order)) {
1984         case COMPACT_PARTIAL:
1985         case COMPACT_CONTINUE:
1986                 return false;
1987         default:
1988                 return true;
1989         }
1990 }
1991
1992 /*
1993  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1994  */
1995 static void shrink_zone(int priority, struct zone *zone,
1996                                 struct scan_control *sc)
1997 {
1998         unsigned long nr[NR_LRU_LISTS];
1999         unsigned long nr_to_scan;
2000         enum lru_list l;
2001         unsigned long nr_reclaimed, nr_scanned;
2002         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2003
2004 restart:
2005         nr_reclaimed = 0;
2006         nr_scanned = sc->nr_scanned;
2007         get_scan_count(zone, sc, nr, priority);
2008
2009         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2010                                         nr[LRU_INACTIVE_FILE]) {
2011                 for_each_evictable_lru(l) {
2012                         if (nr[l]) {
2013                                 nr_to_scan = min_t(unsigned long,
2014                                                    nr[l], SWAP_CLUSTER_MAX);
2015                                 nr[l] -= nr_to_scan;
2016
2017                                 nr_reclaimed += shrink_list(l, nr_to_scan,
2018                                                             zone, sc, priority);
2019                         }
2020                 }
2021                 /*
2022                  * On large memory systems, scan >> priority can become
2023                  * really large. This is fine for the starting priority;
2024                  * we want to put equal scanning pressure on each zone.
2025                  * However, if the VM has a harder time of freeing pages,
2026                  * with multiple processes reclaiming pages, the total
2027                  * freeing target can get unreasonably large.
2028                  */
2029                 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
2030                         break;
2031         }
2032         sc->nr_reclaimed += nr_reclaimed;
2033
2034         /*
2035          * Even if we did not try to evict anon pages at all, we want to
2036          * rebalance the anon lru active/inactive ratio.
2037          */
2038         if (inactive_anon_is_low(zone, sc))
2039                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
2040
2041         /* reclaim/compaction might need reclaim to continue */
2042         if (should_continue_reclaim(zone, nr_reclaimed,
2043                                         sc->nr_scanned - nr_scanned, sc))
2044                 goto restart;
2045
2046         throttle_vm_writeout(sc->gfp_mask);
2047 }
2048
2049 /*
2050  * This is the direct reclaim path, for page-allocating processes.  We only
2051  * try to reclaim pages from zones which will satisfy the caller's allocation
2052  * request.
2053  *
2054  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2055  * Because:
2056  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2057  *    allocation or
2058  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2059  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2060  *    zone defense algorithm.
2061  *
2062  * If a zone is deemed to be full of pinned pages then just give it a light
2063  * scan then give up on it.
2064  */
2065 static void shrink_zones(int priority, struct zonelist *zonelist,
2066                                         struct scan_control *sc)
2067 {
2068         struct zoneref *z;
2069         struct zone *zone;
2070         unsigned long nr_soft_reclaimed;
2071         unsigned long nr_soft_scanned;
2072
2073         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2074                                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2075                 if (!populated_zone(zone))
2076                         continue;
2077                 /*
2078                  * Take care memory controller reclaiming has small influence
2079                  * to global LRU.
2080                  */
2081                 if (scanning_global_lru(sc)) {
2082                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2083                                 continue;
2084                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2085                                 continue;       /* Let kswapd poll it */
2086                         /*
2087                          * This steals pages from memory cgroups over softlimit
2088                          * and returns the number of reclaimed pages and
2089                          * scanned pages. This works for global memory pressure
2090                          * and balancing, not for a memcg's limit.
2091                          */
2092                         nr_soft_scanned = 0;
2093                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2094                                                 sc->order, sc->gfp_mask,
2095                                                 &nr_soft_scanned);
2096                         sc->nr_reclaimed += nr_soft_reclaimed;
2097                         sc->nr_scanned += nr_soft_scanned;
2098                         /* need some check for avoid more shrink_zone() */
2099                 }
2100
2101                 shrink_zone(priority, zone, sc);
2102         }
2103 }
2104
2105 static bool zone_reclaimable(struct zone *zone)
2106 {
2107         return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2108 }
2109
2110 /* All zones in zonelist are unreclaimable? */
2111 static bool all_unreclaimable(struct zonelist *zonelist,
2112                 struct scan_control *sc)
2113 {
2114         struct zoneref *z;
2115         struct zone *zone;
2116
2117         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2118                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2119                 if (!populated_zone(zone))
2120                         continue;
2121                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2122                         continue;
2123                 if (!zone->all_unreclaimable)
2124                         return false;
2125         }
2126
2127         return true;
2128 }
2129
2130 /*
2131  * This is the main entry point to direct page reclaim.
2132  *
2133  * If a full scan of the inactive list fails to free enough memory then we
2134  * are "out of memory" and something needs to be killed.
2135  *
2136  * If the caller is !__GFP_FS then the probability of a failure is reasonably
2137  * high - the zone may be full of dirty or under-writeback pages, which this
2138  * caller can't do much about.  We kick the writeback threads and take explicit
2139  * naps in the hope that some of these pages can be written.  But if the
2140  * allocating task holds filesystem locks which prevent writeout this might not
2141  * work, and the allocation attempt will fail.
2142  *
2143  * returns:     0, if no pages reclaimed
2144  *              else, the number of pages reclaimed
2145  */
2146 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2147                                         struct scan_control *sc,
2148                                         struct shrink_control *shrink)
2149 {
2150         int priority;
2151         unsigned long total_scanned = 0;
2152         struct reclaim_state *reclaim_state = current->reclaim_state;
2153         struct zoneref *z;
2154         struct zone *zone;
2155         unsigned long writeback_threshold;
2156
2157         get_mems_allowed();
2158         delayacct_freepages_start();
2159
2160         if (scanning_global_lru(sc))
2161                 count_vm_event(ALLOCSTALL);
2162
2163         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2164                 sc->nr_scanned = 0;
2165                 if (!priority)
2166                         disable_swap_token(sc->mem_cgroup);
2167                 shrink_zones(priority, zonelist, sc);
2168                 /*
2169                  * Don't shrink slabs when reclaiming memory from
2170                  * over limit cgroups
2171                  */
2172                 if (scanning_global_lru(sc)) {
2173                         unsigned long lru_pages = 0;
2174                         for_each_zone_zonelist(zone, z, zonelist,
2175                                         gfp_zone(sc->gfp_mask)) {
2176                                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2177                                         continue;
2178
2179                                 lru_pages += zone_reclaimable_pages(zone);
2180                         }
2181
2182                         shrink_slab(shrink, sc->nr_scanned, lru_pages);
2183                         if (reclaim_state) {
2184                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2185                                 reclaim_state->reclaimed_slab = 0;
2186                         }
2187                 }
2188                 total_scanned += sc->nr_scanned;
2189                 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2190                         goto out;
2191
2192                 /*
2193                  * Try to write back as many pages as we just scanned.  This
2194                  * tends to cause slow streaming writers to write data to the
2195                  * disk smoothly, at the dirtying rate, which is nice.   But
2196                  * that's undesirable in laptop mode, where we *want* lumpy
2197                  * writeout.  So in laptop mode, write out the whole world.
2198                  */
2199                 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2200                 if (total_scanned > writeback_threshold) {
2201                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
2202                         sc->may_writepage = 1;
2203                 }
2204
2205                 /* Take a nap, wait for some writeback to complete */
2206                 if (!sc->hibernation_mode && sc->nr_scanned &&
2207                     priority < DEF_PRIORITY - 2) {
2208                         struct zone *preferred_zone;
2209
2210                         first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2211                                                 &cpuset_current_mems_allowed,
2212                                                 &preferred_zone);
2213                         wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2214                 }
2215         }
2216
2217 out:
2218         delayacct_freepages_end();
2219         put_mems_allowed();
2220
2221         if (sc->nr_reclaimed)
2222                 return sc->nr_reclaimed;
2223
2224         /*
2225          * As hibernation is going on, kswapd is freezed so that it can't mark
2226          * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2227          * check.
2228          */
2229         if (oom_killer_disabled)
2230                 return 0;
2231
2232         /* top priority shrink_zones still had more to do? don't OOM, then */
2233         if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
2234                 return 1;
2235
2236         return 0;
2237 }
2238
2239 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2240                                 gfp_t gfp_mask, nodemask_t *nodemask)
2241 {
2242         unsigned long nr_reclaimed;
2243         struct scan_control sc = {
2244                 .gfp_mask = gfp_mask,
2245                 .may_writepage = !laptop_mode,
2246                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2247                 .may_unmap = 1,
2248                 .may_swap = 1,
2249                 .order = order,
2250                 .mem_cgroup = NULL,
2251                 .nodemask = nodemask,
2252         };
2253         struct shrink_control shrink = {
2254                 .gfp_mask = sc.gfp_mask,
2255         };
2256
2257         trace_mm_vmscan_direct_reclaim_begin(order,
2258                                 sc.may_writepage,
2259                                 gfp_mask);
2260
2261         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2262
2263         trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2264
2265         return nr_reclaimed;
2266 }
2267
2268 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2269
2270 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2271                                         gfp_t gfp_mask, bool noswap,
2272                                         struct zone *zone,
2273                                         struct memcg_scanrecord *rec,
2274                                         unsigned long *scanned)
2275 {
2276         struct scan_control sc = {
2277                 .nr_scanned = 0,
2278                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2279                 .may_writepage = !laptop_mode,
2280                 .may_unmap = 1,
2281                 .may_swap = !noswap,
2282                 .order = 0,
2283                 .mem_cgroup = mem,
2284                 .memcg_record = rec,
2285         };
2286         unsigned long start, end;
2287
2288         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2289                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2290
2291         trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2292                                                       sc.may_writepage,
2293                                                       sc.gfp_mask);
2294
2295         start = sched_clock();
2296         /*
2297          * NOTE: Although we can get the priority field, using it
2298          * here is not a good idea, since it limits the pages we can scan.
2299          * if we don't reclaim here, the shrink_zone from balance_pgdat
2300          * will pick up pages from other mem cgroup's as well. We hack
2301          * the priority and make it zero.
2302          */
2303         shrink_zone(0, zone, &sc);
2304         end = sched_clock();
2305
2306         if (rec)
2307                 rec->elapsed += end - start;
2308         *scanned = sc.nr_scanned;
2309
2310         trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2311
2312         return sc.nr_reclaimed;
2313 }
2314
2315 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
2316                                            gfp_t gfp_mask,
2317                                            bool noswap,
2318                                            struct memcg_scanrecord *rec)
2319 {
2320         struct zonelist *zonelist;
2321         unsigned long nr_reclaimed;
2322         unsigned long start, end;
2323         int nid;
2324         struct scan_control sc = {
2325                 .may_writepage = !laptop_mode,
2326                 .may_unmap = 1,
2327                 .may_swap = !noswap,
2328                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2329                 .order = 0,
2330                 .mem_cgroup = mem_cont,
2331                 .memcg_record = rec,
2332                 .nodemask = NULL, /* we don't care the placement */
2333                 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2334                                 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2335         };
2336         struct shrink_control shrink = {
2337                 .gfp_mask = sc.gfp_mask,
2338         };
2339
2340         start = sched_clock();
2341         /*
2342          * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2343          * take care of from where we get pages. So the node where we start the
2344          * scan does not need to be the current node.
2345          */
2346         nid = mem_cgroup_select_victim_node(mem_cont);
2347
2348         zonelist = NODE_DATA(nid)->node_zonelists;
2349
2350         trace_mm_vmscan_memcg_reclaim_begin(0,
2351                                             sc.may_writepage,
2352                                             sc.gfp_mask);
2353
2354         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2355         end = sched_clock();
2356         if (rec)
2357                 rec->elapsed += end - start;
2358
2359         trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2360
2361         return nr_reclaimed;
2362 }
2363 #endif
2364
2365 /*
2366  * pgdat_balanced is used when checking if a node is balanced for high-order
2367  * allocations. Only zones that meet watermarks and are in a zone allowed
2368  * by the callers classzone_idx are added to balanced_pages. The total of
2369  * balanced pages must be at least 25% of the zones allowed by classzone_idx
2370  * for the node to be considered balanced. Forcing all zones to be balanced
2371  * for high orders can cause excessive reclaim when there are imbalanced zones.
2372  * The choice of 25% is due to
2373  *   o a 16M DMA zone that is balanced will not balance a zone on any
2374  *     reasonable sized machine
2375  *   o On all other machines, the top zone must be at least a reasonable
2376  *     percentage of the middle zones. For example, on 32-bit x86, highmem
2377  *     would need to be at least 256M for it to be balance a whole node.
2378  *     Similarly, on x86-64 the Normal zone would need to be at least 1G
2379  *     to balance a node on its own. These seemed like reasonable ratios.
2380  */
2381 static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2382                                                 int classzone_idx)
2383 {
2384         unsigned long present_pages = 0;
2385         int i;
2386
2387         for (i = 0; i <= classzone_idx; i++)
2388                 present_pages += pgdat->node_zones[i].present_pages;
2389
2390         /* A special case here: if zone has no page, we think it's balanced */
2391         return balanced_pages >= (present_pages >> 2);
2392 }
2393
2394 /* is kswapd sleeping prematurely? */
2395 static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2396                                         int classzone_idx)
2397 {
2398         int i;
2399         unsigned long balanced = 0;
2400         bool all_zones_ok = true;
2401
2402         /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2403         if (remaining)
2404                 return true;
2405
2406         /* Check the watermark levels */
2407         for (i = 0; i <= classzone_idx; i++) {
2408                 struct zone *zone = pgdat->node_zones + i;
2409
2410                 if (!populated_zone(zone))
2411                         continue;
2412
2413                 /*
2414                  * balance_pgdat() skips over all_unreclaimable after
2415                  * DEF_PRIORITY. Effectively, it considers them balanced so
2416                  * they must be considered balanced here as well if kswapd
2417                  * is to sleep
2418                  */
2419                 if (zone->all_unreclaimable) {
2420                         balanced += zone->present_pages;
2421                         continue;
2422                 }
2423
2424                 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2425                                                         i, 0))
2426                         all_zones_ok = false;
2427                 else
2428                         balanced += zone->present_pages;
2429         }
2430
2431         /*
2432          * For high-order requests, the balanced zones must contain at least
2433          * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2434          * must be balanced
2435          */
2436         if (order)
2437                 return !pgdat_balanced(pgdat, balanced, classzone_idx);
2438         else
2439                 return !all_zones_ok;
2440 }
2441
2442 /*
2443  * For kswapd, balance_pgdat() will work across all this node's zones until
2444  * they are all at high_wmark_pages(zone).
2445  *
2446  * Returns the final order kswapd was reclaiming at
2447  *
2448  * There is special handling here for zones which are full of pinned pages.
2449  * This can happen if the pages are all mlocked, or if they are all used by
2450  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
2451  * What we do is to detect the case where all pages in the zone have been
2452  * scanned twice and there has been zero successful reclaim.  Mark the zone as
2453  * dead and from now on, only perform a short scan.  Basically we're polling
2454  * the zone for when the problem goes away.
2455  *
2456  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
2457  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2458  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2459  * lower zones regardless of the number of free pages in the lower zones. This
2460  * interoperates with the page allocator fallback scheme to ensure that aging
2461  * of pages is balanced across the zones.
2462  */
2463 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2464                                                         int *classzone_idx)
2465 {
2466         int all_zones_ok;
2467         unsigned long balanced;
2468         int priority;
2469         int i;
2470         int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
2471         unsigned long total_scanned;
2472         struct reclaim_state *reclaim_state = current->reclaim_state;
2473         unsigned long nr_soft_reclaimed;
2474         unsigned long nr_soft_scanned;
2475         struct scan_control sc = {
2476                 .gfp_mask = GFP_KERNEL,
2477                 .may_unmap = 1,
2478                 .may_swap = 1,
2479                 /*
2480                  * kswapd doesn't want to be bailed out while reclaim. because
2481                  * we want to put equal scanning pressure on each zone.
2482                  */
2483                 .nr_to_reclaim = ULONG_MAX,
2484                 .order = order,
2485                 .mem_cgroup = NULL,
2486         };
2487         struct shrink_control shrink = {
2488                 .gfp_mask = sc.gfp_mask,
2489         };
2490 loop_again:
2491         total_scanned = 0;
2492         sc.nr_reclaimed = 0;
2493         sc.may_writepage = !laptop_mode;
2494         count_vm_event(PAGEOUTRUN);
2495
2496         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2497                 unsigned long lru_pages = 0;
2498                 int has_under_min_watermark_zone = 0;
2499
2500                 /* The swap token gets in the way of swapout... */
2501                 if (!priority)
2502                         disable_swap_token(NULL);
2503
2504                 all_zones_ok = 1;
2505                 balanced = 0;
2506
2507                 /*
2508                  * Scan in the highmem->dma direction for the highest
2509                  * zone which needs scanning
2510                  */
2511                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2512                         struct zone *zone = pgdat->node_zones + i;
2513
2514                         if (!populated_zone(zone))
2515                                 continue;
2516
2517                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2518                                 continue;
2519
2520                         /*
2521                          * Do some background aging of the anon list, to give
2522                          * pages a chance to be referenced before reclaiming.
2523                          */
2524                         if (inactive_anon_is_low(zone, &sc))
2525                                 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2526                                                         &sc, priority, 0);
2527
2528                         if (!zone_watermark_ok_safe(zone, order,
2529                                         high_wmark_pages(zone), 0, 0)) {
2530                                 end_zone = i;
2531                                 break;
2532                         }
2533                 }
2534                 if (i < 0)
2535                         goto out;
2536
2537                 for (i = 0; i <= end_zone; i++) {
2538                         struct zone *zone = pgdat->node_zones + i;
2539
2540                         lru_pages += zone_reclaimable_pages(zone);
2541                 }
2542
2543                 /*
2544                  * Now scan the zone in the dma->highmem direction, stopping
2545                  * at the last zone which needs scanning.
2546                  *
2547                  * We do this because the page allocator works in the opposite
2548                  * direction.  This prevents the page allocator from allocating
2549                  * pages behind kswapd's direction of progress, which would
2550                  * cause too much scanning of the lower zones.
2551                  */
2552                 for (i = 0; i <= end_zone; i++) {
2553                         struct zone *zone = pgdat->node_zones + i;
2554                         int nr_slab;
2555                         unsigned long balance_gap;
2556
2557                         if (!populated_zone(zone))
2558                                 continue;
2559
2560                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2561                                 continue;
2562
2563                         sc.nr_scanned = 0;
2564
2565                         nr_soft_scanned = 0;
2566                         /*
2567                          * Call soft limit reclaim before calling shrink_zone.
2568                          */
2569                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2570                                                         order, sc.gfp_mask,
2571                                                         &nr_soft_scanned);
2572                         sc.nr_reclaimed += nr_soft_reclaimed;
2573                         total_scanned += nr_soft_scanned;
2574
2575                         /*
2576                          * We put equal pressure on every zone, unless
2577                          * one zone has way too many pages free
2578                          * already. The "too many pages" is defined
2579                          * as the high wmark plus a "gap" where the
2580                          * gap is either the low watermark or 1%
2581                          * of the zone, whichever is smaller.
2582                          */
2583                         balance_gap = min(low_wmark_pages(zone),
2584                                 (zone->present_pages +
2585                                         KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2586                                 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2587                         if (!zone_watermark_ok_safe(zone, order,
2588                                         high_wmark_pages(zone) + balance_gap,
2589                                         end_zone, 0)) {
2590                                 shrink_zone(priority, zone, &sc);
2591
2592                                 reclaim_state->reclaimed_slab = 0;
2593                                 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2594                                 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2595                                 total_scanned += sc.nr_scanned;
2596
2597                                 if (nr_slab == 0 && !zone_reclaimable(zone))
2598                                         zone->all_unreclaimable = 1;
2599                         }
2600
2601                         /*
2602                          * If we've done a decent amount of scanning and
2603                          * the reclaim ratio is low, start doing writepage
2604                          * even in laptop mode
2605                          */
2606                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2607                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2608                                 sc.may_writepage = 1;
2609
2610                         if (zone->all_unreclaimable) {
2611                                 if (end_zone && end_zone == i)
2612                                         end_zone--;
2613                                 continue;
2614                         }
2615
2616                         if (!zone_watermark_ok_safe(zone, order,
2617                                         high_wmark_pages(zone), end_zone, 0)) {
2618                                 all_zones_ok = 0;
2619                                 /*
2620                                  * We are still under min water mark.  This
2621                                  * means that we have a GFP_ATOMIC allocation
2622                                  * failure risk. Hurry up!
2623                                  */
2624                                 if (!zone_watermark_ok_safe(zone, order,
2625                                             min_wmark_pages(zone), end_zone, 0))
2626                                         has_under_min_watermark_zone = 1;
2627                         } else {
2628                                 /*
2629                                  * If a zone reaches its high watermark,
2630                                  * consider it to be no longer congested. It's
2631                                  * possible there are dirty pages backed by
2632                                  * congested BDIs but as pressure is relieved,
2633                                  * spectulatively avoid congestion waits
2634                                  */
2635                                 zone_clear_flag(zone, ZONE_CONGESTED);
2636                                 if (i <= *classzone_idx)
2637                                         balanced += zone->present_pages;
2638                         }
2639
2640                 }
2641                 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2642                         break;          /* kswapd: all done */
2643                 /*
2644                  * OK, kswapd is getting into trouble.  Take a nap, then take
2645                  * another pass across the zones.
2646                  */
2647                 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2648                         if (has_under_min_watermark_zone)
2649                                 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2650                         else
2651                                 congestion_wait(BLK_RW_ASYNC, HZ/10);
2652                 }
2653
2654                 /*
2655                  * We do this so kswapd doesn't build up large priorities for
2656                  * example when it is freeing in parallel with allocators. It
2657                  * matches the direct reclaim path behaviour in terms of impact
2658                  * on zone->*_priority.
2659                  */
2660                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2661                         break;
2662         }
2663 out:
2664
2665         /*
2666          * order-0: All zones must meet high watermark for a balanced node
2667          * high-order: Balanced zones must make up at least 25% of the node
2668          *             for the node to be balanced
2669          */
2670         if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2671                 cond_resched();
2672
2673                 try_to_freeze();
2674
2675                 /*
2676                  * Fragmentation may mean that the system cannot be
2677                  * rebalanced for high-order allocations in all zones.
2678                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2679                  * it means the zones have been fully scanned and are still
2680                  * not balanced. For high-order allocations, there is
2681                  * little point trying all over again as kswapd may
2682                  * infinite loop.
2683                  *
2684                  * Instead, recheck all watermarks at order-0 as they
2685                  * are the most important. If watermarks are ok, kswapd will go
2686                  * back to sleep. High-order users can still perform direct
2687                  * reclaim if they wish.
2688                  */
2689                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2690                         order = sc.order = 0;
2691
2692                 goto loop_again;
2693         }
2694
2695         /*
2696          * If kswapd was reclaiming at a higher order, it has the option of
2697          * sleeping without all zones being balanced. Before it does, it must
2698          * ensure that the watermarks for order-0 on *all* zones are met and
2699          * that the congestion flags are cleared. The congestion flag must
2700          * be cleared as kswapd is the only mechanism that clears the flag
2701          * and it is potentially going to sleep here.
2702          */
2703         if (order) {
2704                 for (i = 0; i <= end_zone; i++) {
2705                         struct zone *zone = pgdat->node_zones + i;
2706
2707                         if (!populated_zone(zone))
2708                                 continue;
2709
2710                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2711                                 continue;
2712
2713                         /* Confirm the zone is balanced for order-0 */
2714                         if (!zone_watermark_ok(zone, 0,
2715                                         high_wmark_pages(zone), 0, 0)) {
2716                                 order = sc.order = 0;
2717                                 goto loop_again;
2718                         }
2719
2720                         /* If balanced, clear the congested flag */
2721                         zone_clear_flag(zone, ZONE_CONGESTED);
2722                 }
2723         }
2724
2725         /*
2726          * Return the order we were reclaiming at so sleeping_prematurely()
2727          * makes a decision on the order we were last reclaiming at. However,
2728          * if another caller entered the allocator slow path while kswapd
2729          * was awake, order will remain at the higher level
2730          */
2731         *classzone_idx = end_zone;
2732         return order;
2733 }
2734
2735 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2736 {
2737         long remaining = 0;
2738         DEFINE_WAIT(wait);
2739
2740         if (freezing(current) || kthread_should_stop())
2741                 return;
2742
2743         prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2744
2745         /* Try to sleep for a short interval */
2746         if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2747                 remaining = schedule_timeout(HZ/10);
2748                 finish_wait(&pgdat->kswapd_wait, &wait);
2749                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2750         }
2751
2752         /*
2753          * After a short sleep, check if it was a premature sleep. If not, then
2754          * go fully to sleep until explicitly woken up.
2755          */
2756         if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2757                 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2758
2759                 /*
2760                  * vmstat counters are not perfectly accurate and the estimated
2761                  * value for counters such as NR_FREE_PAGES can deviate from the
2762                  * true value by nr_online_cpus * threshold. To avoid the zone
2763                  * watermarks being breached while under pressure, we reduce the
2764                  * per-cpu vmstat threshold while kswapd is awake and restore
2765                  * them before going back to sleep.
2766                  */
2767                 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2768                 schedule();
2769                 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2770         } else {
2771                 if (remaining)
2772                         count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2773                 else
2774                         count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2775         }
2776         finish_wait(&pgdat->kswapd_wait, &wait);
2777 }
2778
2779 /*
2780  * The background pageout daemon, started as a kernel thread
2781  * from the init process.
2782  *
2783  * This basically trickles out pages so that we have _some_
2784  * free memory available even if there is no other activity
2785  * that frees anything up. This is needed for things like routing
2786  * etc, where we otherwise might have all activity going on in
2787  * asynchronous contexts that cannot page things out.
2788  *
2789  * If there are applications that are active memory-allocators
2790  * (most normal use), this basically shouldn't matter.
2791  */
2792 static int kswapd(void *p)
2793 {
2794         unsigned long order, new_order;
2795         int classzone_idx, new_classzone_idx;
2796         pg_data_t *pgdat = (pg_data_t*)p;
2797         struct task_struct *tsk = current;
2798
2799         struct reclaim_state reclaim_state = {
2800                 .reclaimed_slab = 0,
2801         };
2802         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2803
2804         lockdep_set_current_reclaim_state(GFP_KERNEL);
2805
2806         if (!cpumask_empty(cpumask))
2807                 set_cpus_allowed_ptr(tsk, cpumask);
2808         current->reclaim_state = &reclaim_state;
2809
2810         /*
2811          * Tell the memory management that we're a "memory allocator",
2812          * and that if we need more memory we should get access to it
2813          * regardless (see "__alloc_pages()"). "kswapd" should
2814          * never get caught in the normal page freeing logic.
2815          *
2816          * (Kswapd normally doesn't need memory anyway, but sometimes
2817          * you need a small amount of memory in order to be able to
2818          * page out something else, and this flag essentially protects
2819          * us from recursively trying to free more memory as we're
2820          * trying to free the first piece of memory in the first place).
2821          */
2822         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2823         set_freezable();
2824
2825         order = new_order = 0;
2826         classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
2827         for ( ; ; ) {
2828                 int ret;
2829
2830                 /*
2831                  * If the last balance_pgdat was unsuccessful it's unlikely a
2832                  * new request of a similar or harder type will succeed soon
2833                  * so consider going to sleep on the basis we reclaimed at
2834                  */
2835                 if (classzone_idx >= new_classzone_idx && order == new_order) {
2836                         new_order = pgdat->kswapd_max_order;
2837                         new_classzone_idx = pgdat->classzone_idx;
2838                         pgdat->kswapd_max_order =  0;
2839                         pgdat->classzone_idx = pgdat->nr_zones - 1;
2840                 }
2841
2842                 if (order < new_order || classzone_idx > new_classzone_idx) {
2843                         /*
2844                          * Don't sleep if someone wants a larger 'order'
2845                          * allocation or has tigher zone constraints
2846                          */
2847                         order = new_order;
2848                         classzone_idx = new_classzone_idx;
2849                 } else {
2850                         kswapd_try_to_sleep(pgdat, order, classzone_idx);
2851                         order = pgdat->kswapd_max_order;
2852                         classzone_idx = pgdat->classzone_idx;
2853                         pgdat->kswapd_max_order = 0;
2854                         pgdat->classzone_idx = pgdat->nr_zones - 1;
2855                 }
2856
2857                 ret = try_to_freeze();
2858                 if (kthread_should_stop())
2859                         break;
2860
2861                 /*
2862                  * We can speed up thawing tasks if we don't call balance_pgdat
2863                  * after returning from the refrigerator
2864                  */
2865                 if (!ret) {
2866                         trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2867                         order = balance_pgdat(pgdat, order, &classzone_idx);
2868                 }
2869         }
2870         return 0;
2871 }
2872
2873 /*
2874  * A zone is low on free memory, so wake its kswapd task to service it.
2875  */
2876 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
2877 {
2878         pg_data_t *pgdat;
2879
2880         if (!populated_zone(zone))
2881                 return;
2882
2883         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2884                 return;
2885         pgdat = zone->zone_pgdat;
2886         if (pgdat->kswapd_max_order < order) {
2887                 pgdat->kswapd_max_order = order;
2888                 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
2889         }
2890         if (!waitqueue_active(&pgdat->kswapd_wait))
2891                 return;
2892         if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
2893                 return;
2894
2895         trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
2896         wake_up_interruptible(&pgdat->kswapd_wait);
2897 }
2898
2899 /*
2900  * The reclaimable count would be mostly accurate.
2901  * The less reclaimable pages may be
2902  * - mlocked pages, which will be moved to unevictable list when encountered
2903  * - mapped pages, which may require several travels to be reclaimed
2904  * - dirty pages, which is not "instantly" reclaimable
2905  */
2906 unsigned long global_reclaimable_pages(void)
2907 {
2908         int nr;
2909
2910         nr = global_page_state(NR_ACTIVE_FILE) +
2911              global_page_state(NR_INACTIVE_FILE);
2912
2913         if (nr_swap_pages > 0)
2914                 nr += global_page_state(NR_ACTIVE_ANON) +
2915                       global_page_state(NR_INACTIVE_ANON);
2916
2917         return nr;
2918 }
2919
2920 unsigned long zone_reclaimable_pages(struct zone *zone)
2921 {
2922         int nr;
2923
2924         nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2925              zone_page_state(zone, NR_INACTIVE_FILE);
2926
2927         if (nr_swap_pages > 0)
2928                 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2929                       zone_page_state(zone, NR_INACTIVE_ANON);
2930
2931         return nr;
2932 }
2933
2934 #ifdef CONFIG_HIBERNATION
2935 /*
2936  * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2937  * freed pages.
2938  *
2939  * Rather than trying to age LRUs the aim is to preserve the overall
2940  * LRU order by reclaiming preferentially
2941  * inactive > active > active referenced > active mapped
2942  */
2943 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
2944 {
2945         struct reclaim_state reclaim_state;
2946         struct scan_control sc = {
2947                 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2948                 .may_swap = 1,
2949                 .may_unmap = 1,
2950                 .may_writepage = 1,
2951                 .nr_to_reclaim = nr_to_reclaim,
2952                 .hibernation_mode = 1,
2953                 .order = 0,
2954         };
2955         struct shrink_control shrink = {
2956                 .gfp_mask = sc.gfp_mask,
2957         };
2958         struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2959         struct task_struct *p = current;
2960         unsigned long nr_reclaimed;
2961
2962         p->flags |= PF_MEMALLOC;
2963         lockdep_set_current_reclaim_state(sc.gfp_mask);
2964         reclaim_state.reclaimed_slab = 0;
2965         p->reclaim_state = &reclaim_state;
2966
2967         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2968
2969         p->reclaim_state = NULL;
2970         lockdep_clear_current_reclaim_state();
2971         p->flags &= ~PF_MEMALLOC;
2972
2973         return nr_reclaimed;
2974 }
2975 #endif /* CONFIG_HIBERNATION */
2976
2977 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2978    not required for correctness.  So if the last cpu in a node goes
2979    away, we get changed to run anywhere: as the first one comes back,
2980    restore their cpu bindings. */
2981 static int __devinit cpu_callback(struct notifier_block *nfb,
2982                                   unsigned long action, void *hcpu)
2983 {
2984         int nid;
2985
2986         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2987                 for_each_node_state(nid, N_HIGH_MEMORY) {
2988                         pg_data_t *pgdat = NODE_DATA(nid);
2989                         const struct cpumask *mask;
2990
2991                         mask = cpumask_of_node(pgdat->node_id);
2992
2993                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2994                                 /* One of our CPUs online: restore mask */
2995                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2996                 }
2997         }
2998         return NOTIFY_OK;
2999 }
3000
3001 /*
3002  * This kswapd start function will be called by init and node-hot-add.
3003  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3004  */
3005 int kswapd_run(int nid)
3006 {
3007         pg_data_t *pgdat = NODE_DATA(nid);
3008         int ret = 0;
3009
3010         if (pgdat->kswapd)
3011                 return 0;
3012
3013         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3014         if (IS_ERR(pgdat->kswapd)) {
3015                 /* failure at boot is fatal */
3016                 BUG_ON(system_state == SYSTEM_BOOTING);
3017                 printk("Failed to start kswapd on node %d\n",nid);
3018                 ret = -1;
3019         }
3020         return ret;
3021 }
3022
3023 /*
3024  * Called by memory hotplug when all memory in a node is offlined.
3025  */
3026 void kswapd_stop(int nid)
3027 {
3028         struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3029
3030         if (kswapd)
3031                 kthread_stop(kswapd);
3032 }
3033
3034 static int __init kswapd_init(void)
3035 {
3036         int nid;
3037
3038         swap_setup();
3039         for_each_node_state(nid, N_HIGH_MEMORY)
3040                 kswapd_run(nid);
3041         hotcpu_notifier(cpu_callback, 0);
3042         return 0;
3043 }
3044
3045 module_init(kswapd_init)
3046
3047 #ifdef CONFIG_NUMA
3048 /*
3049  * Zone reclaim mode
3050  *
3051  * If non-zero call zone_reclaim when the number of free pages falls below
3052  * the watermarks.
3053  */
3054 int zone_reclaim_mode __read_mostly;
3055
3056 #define RECLAIM_OFF 0
3057 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
3058 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
3059 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
3060
3061 /*
3062  * Priority for ZONE_RECLAIM. This determines the fraction of pages
3063  * of a node considered for each zone_reclaim. 4 scans 1/16th of
3064  * a zone.
3065  */
3066 #define ZONE_RECLAIM_PRIORITY 4
3067
3068 /*
3069  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3070  * occur.
3071  */
3072 int sysctl_min_unmapped_ratio = 1;
3073
3074 /*
3075  * If the number of slab pages in a zone grows beyond this percentage then
3076  * slab reclaim needs to occur.
3077  */
3078 int sysctl_min_slab_ratio = 5;
3079
3080 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3081 {
3082         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3083         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3084                 zone_page_state(zone, NR_ACTIVE_FILE);
3085
3086         /*
3087          * It's possible for there to be more file mapped pages than
3088          * accounted for by the pages on the file LRU lists because
3089          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3090          */
3091         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3092 }
3093
3094 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3095 static long zone_pagecache_reclaimable(struct zone *zone)
3096 {
3097         long nr_pagecache_reclaimable;
3098         long delta = 0;
3099
3100         /*
3101          * If RECLAIM_SWAP is set, then all file pages are considered
3102          * potentially reclaimable. Otherwise, we have to worry about
3103          * pages like swapcache and zone_unmapped_file_pages() provides
3104          * a better estimate
3105          */
3106         if (zone_reclaim_mode & RECLAIM_SWAP)
3107                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3108         else
3109                 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3110
3111         /* If we can't clean pages, remove dirty pages from consideration */
3112         if (!(zone_reclaim_mode & RECLAIM_WRITE))
3113                 delta += zone_page_state(zone, NR_FILE_DIRTY);
3114
3115         /* Watch for any possible underflows due to delta */
3116         if (unlikely(delta > nr_pagecache_reclaimable))
3117                 delta = nr_pagecache_reclaimable;
3118
3119         return nr_pagecache_reclaimable - delta;
3120 }
3121
3122 /*
3123  * Try to free up some pages from this zone through reclaim.
3124  */
3125 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3126 {
3127         /* Minimum pages needed in order to stay on node */
3128         const unsigned long nr_pages = 1 << order;
3129         struct task_struct *p = current;
3130         struct reclaim_state reclaim_state;
3131         int priority;
3132         struct scan_control sc = {
3133                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3134                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3135                 .may_swap = 1,
3136                 .nr_to_reclaim = max_t(unsigned long, nr_pages,
3137                                        SWAP_CLUSTER_MAX),
3138                 .gfp_mask = gfp_mask,
3139                 .order = order,
3140         };
3141         struct shrink_control shrink = {
3142                 .gfp_mask = sc.gfp_mask,
3143         };
3144         unsigned long nr_slab_pages0, nr_slab_pages1;
3145
3146         cond_resched();
3147         /*
3148          * We need to be able to allocate from the reserves for RECLAIM_SWAP
3149          * and we also need to be able to write out pages for RECLAIM_WRITE
3150          * and RECLAIM_SWAP.
3151          */
3152         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3153         lockdep_set_current_reclaim_state(gfp_mask);
3154         reclaim_state.reclaimed_slab = 0;
3155         p->reclaim_state = &reclaim_state;
3156
3157         if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3158                 /*
3159                  * Free memory by calling shrink zone with increasing
3160                  * priorities until we have enough memory freed.
3161                  */
3162                 priority = ZONE_RECLAIM_PRIORITY;
3163                 do {
3164                         shrink_zone(priority, zone, &sc);
3165                         priority--;
3166                 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
3167         }
3168
3169         nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3170         if (nr_slab_pages0 > zone->min_slab_pages) {
3171                 /*
3172                  * shrink_slab() does not currently allow us to determine how
3173                  * many pages were freed in this zone. So we take the current
3174                  * number of slab pages and shake the slab until it is reduced
3175                  * by the same nr_pages that we used for reclaiming unmapped
3176                  * pages.
3177                  *
3178                  * Note that shrink_slab will free memory on all zones and may
3179                  * take a long time.
3180                  */
3181                 for (;;) {
3182                         unsigned long lru_pages = zone_reclaimable_pages(zone);
3183
3184                         /* No reclaimable slab or very low memory pressure */
3185                         if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3186                                 break;
3187
3188                         /* Freed enough memory */
3189                         nr_slab_pages1 = zone_page_state(zone,
3190                                                         NR_SLAB_RECLAIMABLE);
3191                         if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3192                                 break;
3193                 }
3194
3195                 /*
3196                  * Update nr_reclaimed by the number of slab pages we
3197                  * reclaimed from this zone.
3198                  */
3199                 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3200                 if (nr_slab_pages1 < nr_slab_pages0)
3201                         sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3202         }
3203
3204         p->reclaim_state = NULL;
3205         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3206         lockdep_clear_current_reclaim_state();
3207         return sc.nr_reclaimed >= nr_pages;
3208 }
3209
3210 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3211 {
3212         int node_id;
3213         int ret;
3214
3215         /*
3216          * Zone reclaim reclaims unmapped file backed pages and
3217          * slab pages if we are over the defined limits.
3218          *
3219          * A small portion of unmapped file backed pages is needed for
3220          * file I/O otherwise pages read by file I/O will be immediately
3221          * thrown out if the zone is overallocated. So we do not reclaim
3222          * if less than a specified percentage of the zone is used by
3223          * unmapped file backed pages.
3224          */
3225         if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3226             zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3227                 return ZONE_RECLAIM_FULL;
3228
3229         if (zone->all_unreclaimable)
3230                 return ZONE_RECLAIM_FULL;
3231
3232         /*
3233          * Do not scan if the allocation should not be delayed.
3234          */
3235         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3236                 return ZONE_RECLAIM_NOSCAN;
3237
3238         /*
3239          * Only run zone reclaim on the local zone or on zones that do not
3240          * have associated processors. This will favor the local processor
3241          * over remote processors and spread off node memory allocations
3242          * as wide as possible.
3243          */
3244         node_id = zone_to_nid(zone);
3245         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3246                 return ZONE_RECLAIM_NOSCAN;
3247
3248         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3249                 return ZONE_RECLAIM_NOSCAN;
3250
3251         ret = __zone_reclaim(zone, gfp_mask, order);
3252         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3253
3254         if (!ret)
3255                 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3256
3257         return ret;
3258 }
3259 #endif
3260
3261 /*
3262  * page_evictable - test whether a page is evictable
3263  * @page: the page to test
3264  * @vma: the VMA in which the page is or will be mapped, may be NULL
3265  *
3266  * Test whether page is evictable--i.e., should be placed on active/inactive
3267  * lists vs unevictable list.  The vma argument is !NULL when called from the
3268  * fault path to determine how to instantate a new page.
3269  *
3270  * Reasons page might not be evictable:
3271  * (1) page's mapping marked unevictable
3272  * (2) page is part of an mlocked VMA
3273  *
3274  */
3275 int page_evictable(struct page *page, struct vm_area_struct *vma)
3276 {
3277
3278         if (mapping_unevictable(page_mapping(page)))
3279                 return 0;
3280
3281         if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
3282                 return 0;
3283
3284         return 1;
3285 }
3286
3287 /**
3288  * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
3289  * @page: page to check evictability and move to appropriate lru list
3290  * @zone: zone page is in
3291  *
3292  * Checks a page for evictability and moves the page to the appropriate
3293  * zone lru list.
3294  *
3295  * Restrictions: zone->lru_lock must be held, page must be on LRU and must
3296  * have PageUnevictable set.
3297  */
3298 static void check_move_unevictable_page(struct page *page, struct zone *zone)
3299 {
3300         VM_BUG_ON(PageActive(page));
3301
3302 retry:
3303         ClearPageUnevictable(page);
3304         if (page_evictable(page, NULL)) {
3305                 enum lru_list l = page_lru_base_type(page);
3306
3307                 __dec_zone_state(zone, NR_UNEVICTABLE);
3308                 list_move(&page->lru, &zone->lru[l].list);
3309                 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
3310                 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
3311                 __count_vm_event(UNEVICTABLE_PGRESCUED);
3312         } else {
3313                 /*
3314                  * rotate unevictable list
3315                  */
3316                 SetPageUnevictable(page);
3317                 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
3318                 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
3319                 if (page_evictable(page, NULL))
3320                         goto retry;
3321         }
3322 }
3323
3324 /**
3325  * scan_mapping_unevictable_pages - scan an address space for evictable pages
3326  * @mapping: struct address_space to scan for evictable pages
3327  *
3328  * Scan all pages in mapping.  Check unevictable pages for
3329  * evictability and move them to the appropriate zone lru list.
3330  */
3331 void scan_mapping_unevictable_pages(struct address_space *mapping)
3332 {
3333         pgoff_t next = 0;
3334         pgoff_t end   = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
3335                          PAGE_CACHE_SHIFT;
3336         struct zone *zone;
3337         struct pagevec pvec;
3338
3339         if (mapping->nrpages == 0)
3340                 return;
3341
3342         pagevec_init(&pvec, 0);
3343         while (next < end &&
3344                 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
3345                 int i;
3346                 int pg_scanned = 0;
3347
3348                 zone = NULL;
3349
3350                 for (i = 0; i < pagevec_count(&pvec); i++) {
3351                         struct page *page = pvec.pages[i];
3352                         pgoff_t page_index = page->index;
3353                         struct zone *pagezone = page_zone(page);
3354
3355                         pg_scanned++;
3356                         if (page_index > next)
3357                                 next = page_index;
3358                         next++;
3359
3360                         if (pagezone != zone) {
3361                                 if (zone)
3362                                         spin_unlock_irq(&zone->lru_lock);
3363                                 zone = pagezone;
3364                                 spin_lock_irq(&zone->lru_lock);
3365                         }
3366
3367                         if (PageLRU(page) && PageUnevictable(page))
3368                                 check_move_unevictable_page(page, zone);
3369                 }
3370                 if (zone)
3371                         spin_unlock_irq(&zone->lru_lock);
3372                 pagevec_release(&pvec);
3373
3374                 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
3375         }
3376
3377 }
3378
3379 /**
3380  * scan_zone_unevictable_pages - check unevictable list for evictable pages
3381  * @zone - zone of which to scan the unevictable list
3382  *
3383  * Scan @zone's unevictable LRU lists to check for pages that have become
3384  * evictable.  Move those that have to @zone's inactive list where they
3385  * become candidates for reclaim, unless shrink_inactive_zone() decides
3386  * to reactivate them.  Pages that are still unevictable are rotated
3387  * back onto @zone's unevictable list.
3388  */
3389 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
3390 static void scan_zone_unevictable_pages(struct zone *zone)
3391 {
3392         struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
3393         unsigned long scan;
3394         unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
3395
3396         while (nr_to_scan > 0) {
3397                 unsigned long batch_size = min(nr_to_scan,
3398                                                 SCAN_UNEVICTABLE_BATCH_SIZE);
3399
3400                 spin_lock_irq(&zone->lru_lock);
3401                 for (scan = 0;  scan < batch_size; scan++) {
3402                         struct page *page = lru_to_page(l_unevictable);
3403
3404                         if (!trylock_page(page))
3405                                 continue;
3406
3407                         prefetchw_prev_lru_page(page, l_unevictable, flags);
3408
3409                         if (likely(PageLRU(page) && PageUnevictable(page)))
3410                                 check_move_unevictable_page(page, zone);
3411
3412                         unlock_page(page);
3413                 }
3414                 spin_unlock_irq(&zone->lru_lock);
3415
3416                 nr_to_scan -= batch_size;
3417         }
3418 }
3419
3420
3421 /**
3422  * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
3423  *
3424  * A really big hammer:  scan all zones' unevictable LRU lists to check for
3425  * pages that have become evictable.  Move those back to the zones'
3426  * inactive list where they become candidates for reclaim.
3427  * This occurs when, e.g., we have unswappable pages on the unevictable lists,
3428  * and we add swap to the system.  As such, it runs in the context of a task
3429  * that has possibly/probably made some previously unevictable pages
3430  * evictable.
3431  */
3432 static void scan_all_zones_unevictable_pages(void)
3433 {
3434         struct zone *zone;
3435
3436         for_each_zone(zone) {
3437                 scan_zone_unevictable_pages(zone);
3438         }
3439 }
3440
3441 /*
3442  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
3443  * all nodes' unevictable lists for evictable pages
3444  */
3445 unsigned long scan_unevictable_pages;
3446
3447 int scan_unevictable_handler(struct ctl_table *table, int write,
3448                            void __user *buffer,
3449                            size_t *length, loff_t *ppos)
3450 {
3451         proc_doulongvec_minmax(table, write, buffer, length, ppos);
3452
3453         if (write && *(unsigned long *)table->data)
3454                 scan_all_zones_unevictable_pages();
3455
3456         scan_unevictable_pages = 0;
3457         return 0;
3458 }
3459
3460 #ifdef CONFIG_NUMA
3461 /*
3462  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
3463  * a specified node's per zone unevictable lists for evictable pages.
3464  */
3465
3466 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
3467                                           struct sysdev_attribute *attr,
3468                                           char *buf)
3469 {
3470         return sprintf(buf, "0\n");     /* always zero; should fit... */
3471 }
3472
3473 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
3474                                            struct sysdev_attribute *attr,
3475                                         const char *buf, size_t count)
3476 {
3477         struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
3478         struct zone *zone;
3479         unsigned long res;
3480         unsigned long req = strict_strtoul(buf, 10, &res);
3481
3482         if (!req)
3483                 return 1;       /* zero is no-op */
3484
3485         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
3486                 if (!populated_zone(zone))
3487                         continue;
3488                 scan_zone_unevictable_pages(zone);
3489         }
3490         return 1;
3491 }
3492
3493
3494 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3495                         read_scan_unevictable_node,
3496                         write_scan_unevictable_node);
3497
3498 int scan_unevictable_register_node(struct node *node)
3499 {
3500         return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
3501 }
3502
3503 void scan_unevictable_unregister_node(struct node *node)
3504 {
3505         sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
3506 }
3507 #endif