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