vmscan: reduce wind up shrinker->nr when shrinker can't do work
[linux-2.6.git] / mm / vmscan.c
1 /*
2  *  linux/mm/vmscan.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  *
6  *  Swap reorganised 29.12.95, Stephen Tweedie.
7  *  kswapd added: 7.1.96  sct
8  *  Removed kswapd_ctl limits, and swap out as many pages as needed
9  *  to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10  *  Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11  *  Multiqueue VM started 5.8.00, Rik van Riel.
12  */
13
14 #include <linux/mm.h>
15 #include <linux/module.h>
16 #include <linux/gfp.h>
17 #include <linux/kernel_stat.h>
18 #include <linux/swap.h>
19 #include <linux/pagemap.h>
20 #include <linux/init.h>
21 #include <linux/highmem.h>
22 #include <linux/vmstat.h>
23 #include <linux/file.h>
24 #include <linux/writeback.h>
25 #include <linux/blkdev.h>
26 #include <linux/buffer_head.h>  /* for try_to_release_page(),
27                                         buffer_heads_over_limit */
28 #include <linux/mm_inline.h>
29 #include <linux/pagevec.h>
30 #include <linux/backing-dev.h>
31 #include <linux/rmap.h>
32 #include <linux/topology.h>
33 #include <linux/cpu.h>
34 #include <linux/cpuset.h>
35 #include <linux/compaction.h>
36 #include <linux/notifier.h>
37 #include <linux/rwsem.h>
38 #include <linux/delay.h>
39 #include <linux/kthread.h>
40 #include <linux/freezer.h>
41 #include <linux/memcontrol.h>
42 #include <linux/delayacct.h>
43 #include <linux/sysctl.h>
44 #include <linux/oom.h>
45 #include <linux/prefetch.h>
46
47 #include <asm/tlbflush.h>
48 #include <asm/div64.h>
49
50 #include <linux/swapops.h>
51
52 #include "internal.h"
53
54 #define CREATE_TRACE_POINTS
55 #include <trace/events/vmscan.h>
56
57 /*
58  * reclaim_mode determines how the inactive list is shrunk
59  * RECLAIM_MODE_SINGLE: Reclaim only order-0 pages
60  * RECLAIM_MODE_ASYNC:  Do not block
61  * RECLAIM_MODE_SYNC:   Allow blocking e.g. call wait_on_page_writeback
62  * RECLAIM_MODE_LUMPYRECLAIM: For high-order allocations, take a reference
63  *                      page from the LRU and reclaim all pages within a
64  *                      naturally aligned range
65  * RECLAIM_MODE_COMPACTION: For high-order allocations, reclaim a number of
66  *                      order-0 pages and then compact the zone
67  */
68 typedef unsigned __bitwise__ reclaim_mode_t;
69 #define RECLAIM_MODE_SINGLE             ((__force reclaim_mode_t)0x01u)
70 #define RECLAIM_MODE_ASYNC              ((__force reclaim_mode_t)0x02u)
71 #define RECLAIM_MODE_SYNC               ((__force reclaim_mode_t)0x04u)
72 #define RECLAIM_MODE_LUMPYRECLAIM       ((__force reclaim_mode_t)0x08u)
73 #define RECLAIM_MODE_COMPACTION         ((__force reclaim_mode_t)0x10u)
74
75 struct scan_control {
76         /* Incremented by the number of inactive pages that were scanned */
77         unsigned long nr_scanned;
78
79         /* Number of pages freed so far during a call to shrink_zones() */
80         unsigned long nr_reclaimed;
81
82         /* How many pages shrink_list() should reclaim */
83         unsigned long nr_to_reclaim;
84
85         unsigned long hibernation_mode;
86
87         /* This context's GFP mask */
88         gfp_t gfp_mask;
89
90         int may_writepage;
91
92         /* Can mapped pages be reclaimed? */
93         int may_unmap;
94
95         /* Can pages be swapped as part of reclaim? */
96         int may_swap;
97
98         int swappiness;
99
100         int order;
101
102         /*
103          * Intend to reclaim enough continuous memory rather than reclaim
104          * enough amount of memory. i.e, mode for high order allocation.
105          */
106         reclaim_mode_t reclaim_mode;
107
108         /* Which cgroup do we reclaim from */
109         struct mem_cgroup *mem_cgroup;
110
111         /*
112          * Nodemask of nodes allowed by the caller. If NULL, all nodes
113          * are scanned.
114          */
115         nodemask_t      *nodemask;
116 };
117
118 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
119
120 #ifdef ARCH_HAS_PREFETCH
121 #define prefetch_prev_lru_page(_page, _base, _field)                    \
122         do {                                                            \
123                 if ((_page)->lru.prev != _base) {                       \
124                         struct page *prev;                              \
125                                                                         \
126                         prev = lru_to_page(&(_page->lru));              \
127                         prefetch(&prev->_field);                        \
128                 }                                                       \
129         } while (0)
130 #else
131 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
132 #endif
133
134 #ifdef ARCH_HAS_PREFETCHW
135 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
136         do {                                                            \
137                 if ((_page)->lru.prev != _base) {                       \
138                         struct page *prev;                              \
139                                                                         \
140                         prev = lru_to_page(&(_page->lru));              \
141                         prefetchw(&prev->_field);                       \
142                 }                                                       \
143         } while (0)
144 #else
145 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
146 #endif
147
148 /*
149  * From 0 .. 100.  Higher means more swappy.
150  */
151 int vm_swappiness = 60;
152 long vm_total_pages;    /* The total number of pages which the VM controls */
153
154 static LIST_HEAD(shrinker_list);
155 static DECLARE_RWSEM(shrinker_rwsem);
156
157 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
158 #define scanning_global_lru(sc) (!(sc)->mem_cgroup)
159 #else
160 #define scanning_global_lru(sc) (1)
161 #endif
162
163 static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
164                                                   struct scan_control *sc)
165 {
166         if (!scanning_global_lru(sc))
167                 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
168
169         return &zone->reclaim_stat;
170 }
171
172 static unsigned long zone_nr_lru_pages(struct zone *zone,
173                                 struct scan_control *sc, enum lru_list lru)
174 {
175         if (!scanning_global_lru(sc))
176                 return mem_cgroup_zone_nr_lru_pages(sc->mem_cgroup, zone, lru);
177
178         return zone_page_state(zone, NR_LRU_BASE + lru);
179 }
180
181
182 /*
183  * Add a shrinker callback to be called from the vm
184  */
185 void register_shrinker(struct shrinker *shrinker)
186 {
187         shrinker->nr = 0;
188         down_write(&shrinker_rwsem);
189         list_add_tail(&shrinker->list, &shrinker_list);
190         up_write(&shrinker_rwsem);
191 }
192 EXPORT_SYMBOL(register_shrinker);
193
194 /*
195  * Remove one
196  */
197 void unregister_shrinker(struct shrinker *shrinker)
198 {
199         down_write(&shrinker_rwsem);
200         list_del(&shrinker->list);
201         up_write(&shrinker_rwsem);
202 }
203 EXPORT_SYMBOL(unregister_shrinker);
204
205 static inline int do_shrinker_shrink(struct shrinker *shrinker,
206                                      struct shrink_control *sc,
207                                      unsigned long nr_to_scan)
208 {
209         sc->nr_to_scan = nr_to_scan;
210         return (*shrinker->shrink)(shrinker, sc);
211 }
212
213 #define SHRINK_BATCH 128
214 /*
215  * Call the shrink functions to age shrinkable caches
216  *
217  * Here we assume it costs one seek to replace a lru page and that it also
218  * takes a seek to recreate a cache object.  With this in mind we age equal
219  * percentages of the lru and ageable caches.  This should balance the seeks
220  * generated by these structures.
221  *
222  * If the vm encountered mapped pages on the LRU it increase the pressure on
223  * slab to avoid swapping.
224  *
225  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
226  *
227  * `lru_pages' represents the number of on-LRU pages in all the zones which
228  * are eligible for the caller's allocation attempt.  It is used for balancing
229  * slab reclaim versus page reclaim.
230  *
231  * Returns the number of slab objects which we shrunk.
232  */
233 unsigned long shrink_slab(struct shrink_control *shrink,
234                           unsigned long nr_pages_scanned,
235                           unsigned long lru_pages)
236 {
237         struct shrinker *shrinker;
238         unsigned long ret = 0;
239
240         if (nr_pages_scanned == 0)
241                 nr_pages_scanned = SWAP_CLUSTER_MAX;
242
243         if (!down_read_trylock(&shrinker_rwsem)) {
244                 /* Assume we'll be able to shrink next time */
245                 ret = 1;
246                 goto out;
247         }
248
249         list_for_each_entry(shrinker, &shrinker_list, list) {
250                 unsigned long long delta;
251                 unsigned long total_scan;
252                 unsigned long max_pass;
253                 int shrink_ret = 0;
254                 long nr;
255                 long new_nr;
256
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 >= SHRINK_BATCH) {
307                         long this_scan = SHRINK_BATCH;
308                         int nr_before;
309
310                         nr_before = do_shrinker_shrink(shrinker, shrink, 0);
311                         shrink_ret = do_shrinker_shrink(shrinker, shrink,
312                                                         this_scan);
313                         if (shrink_ret == -1)
314                                 break;
315                         if (shrink_ret < nr_before)
316                                 ret += nr_before - shrink_ret;
317                         count_vm_events(SLABS_SCANNED, this_scan);
318                         total_scan -= this_scan;
319
320                         cond_resched();
321                 }
322
323                 /*
324                  * move the unused scan count back into the shrinker in a
325                  * manner that handles concurrent updates. If we exhausted the
326                  * scan, there is no need to do an update.
327                  */
328                 do {
329                         nr = shrinker->nr;
330                         new_nr = total_scan + nr;
331                         if (total_scan <= 0)
332                                 break;
333                 } while (cmpxchg(&shrinker->nr, nr, new_nr) != nr);
334
335                 trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr);
336         }
337         up_read(&shrinker_rwsem);
338 out:
339         cond_resched();
340         return ret;
341 }
342
343 static void set_reclaim_mode(int priority, struct scan_control *sc,
344                                    bool sync)
345 {
346         reclaim_mode_t syncmode = sync ? RECLAIM_MODE_SYNC : RECLAIM_MODE_ASYNC;
347
348         /*
349          * Initially assume we are entering either lumpy reclaim or
350          * reclaim/compaction.Depending on the order, we will either set the
351          * sync mode or just reclaim order-0 pages later.
352          */
353         if (COMPACTION_BUILD)
354                 sc->reclaim_mode = RECLAIM_MODE_COMPACTION;
355         else
356                 sc->reclaim_mode = RECLAIM_MODE_LUMPYRECLAIM;
357
358         /*
359          * Avoid using lumpy reclaim or reclaim/compaction if possible by
360          * restricting when its set to either costly allocations or when
361          * under memory pressure
362          */
363         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
364                 sc->reclaim_mode |= syncmode;
365         else if (sc->order && priority < DEF_PRIORITY - 2)
366                 sc->reclaim_mode |= syncmode;
367         else
368                 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
369 }
370
371 static void reset_reclaim_mode(struct scan_control *sc)
372 {
373         sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
374 }
375
376 static inline int is_page_cache_freeable(struct page *page)
377 {
378         /*
379          * A freeable page cache page is referenced only by the caller
380          * that isolated the page, the page cache radix tree and
381          * optional buffer heads at page->private.
382          */
383         return page_count(page) - page_has_private(page) == 2;
384 }
385
386 static int may_write_to_queue(struct backing_dev_info *bdi,
387                               struct scan_control *sc)
388 {
389         if (current->flags & PF_SWAPWRITE)
390                 return 1;
391         if (!bdi_write_congested(bdi))
392                 return 1;
393         if (bdi == current->backing_dev_info)
394                 return 1;
395
396         /* lumpy reclaim for hugepage often need a lot of write */
397         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
398                 return 1;
399         return 0;
400 }
401
402 /*
403  * We detected a synchronous write error writing a page out.  Probably
404  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
405  * fsync(), msync() or close().
406  *
407  * The tricky part is that after writepage we cannot touch the mapping: nothing
408  * prevents it from being freed up.  But we have a ref on the page and once
409  * that page is locked, the mapping is pinned.
410  *
411  * We're allowed to run sleeping lock_page() here because we know the caller has
412  * __GFP_FS.
413  */
414 static void handle_write_error(struct address_space *mapping,
415                                 struct page *page, int error)
416 {
417         lock_page(page);
418         if (page_mapping(page) == mapping)
419                 mapping_set_error(mapping, error);
420         unlock_page(page);
421 }
422
423 /* possible outcome of pageout() */
424 typedef enum {
425         /* failed to write page out, page is locked */
426         PAGE_KEEP,
427         /* move page to the active list, page is locked */
428         PAGE_ACTIVATE,
429         /* page has been sent to the disk successfully, page is unlocked */
430         PAGE_SUCCESS,
431         /* page is clean and locked */
432         PAGE_CLEAN,
433 } pageout_t;
434
435 /*
436  * pageout is called by shrink_page_list() for each dirty page.
437  * Calls ->writepage().
438  */
439 static pageout_t pageout(struct page *page, struct address_space *mapping,
440                          struct scan_control *sc)
441 {
442         /*
443          * If the page is dirty, only perform writeback if that write
444          * will be non-blocking.  To prevent this allocation from being
445          * stalled by pagecache activity.  But note that there may be
446          * stalls if we need to run get_block().  We could test
447          * PagePrivate for that.
448          *
449          * If this process is currently in __generic_file_aio_write() against
450          * this page's queue, we can perform writeback even if that
451          * will block.
452          *
453          * If the page is swapcache, write it back even if that would
454          * block, for some throttling. This happens by accident, because
455          * swap_backing_dev_info is bust: it doesn't reflect the
456          * congestion state of the swapdevs.  Easy to fix, if needed.
457          */
458         if (!is_page_cache_freeable(page))
459                 return PAGE_KEEP;
460         if (!mapping) {
461                 /*
462                  * Some data journaling orphaned pages can have
463                  * page->mapping == NULL while being dirty with clean buffers.
464                  */
465                 if (page_has_private(page)) {
466                         if (try_to_free_buffers(page)) {
467                                 ClearPageDirty(page);
468                                 printk("%s: orphaned page\n", __func__);
469                                 return PAGE_CLEAN;
470                         }
471                 }
472                 return PAGE_KEEP;
473         }
474         if (mapping->a_ops->writepage == NULL)
475                 return PAGE_ACTIVATE;
476         if (!may_write_to_queue(mapping->backing_dev_info, sc))
477                 return PAGE_KEEP;
478
479         if (clear_page_dirty_for_io(page)) {
480                 int res;
481                 struct writeback_control wbc = {
482                         .sync_mode = WB_SYNC_NONE,
483                         .nr_to_write = SWAP_CLUSTER_MAX,
484                         .range_start = 0,
485                         .range_end = LLONG_MAX,
486                         .for_reclaim = 1,
487                 };
488
489                 SetPageReclaim(page);
490                 res = mapping->a_ops->writepage(page, &wbc);
491                 if (res < 0)
492                         handle_write_error(mapping, page, res);
493                 if (res == AOP_WRITEPAGE_ACTIVATE) {
494                         ClearPageReclaim(page);
495                         return PAGE_ACTIVATE;
496                 }
497
498                 /*
499                  * Wait on writeback if requested to. This happens when
500                  * direct reclaiming a large contiguous area and the
501                  * first attempt to free a range of pages fails.
502                  */
503                 if (PageWriteback(page) &&
504                     (sc->reclaim_mode & RECLAIM_MODE_SYNC))
505                         wait_on_page_writeback(page);
506
507                 if (!PageWriteback(page)) {
508                         /* synchronous write or broken a_ops? */
509                         ClearPageReclaim(page);
510                 }
511                 trace_mm_vmscan_writepage(page,
512                         trace_reclaim_flags(page, sc->reclaim_mode));
513                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
514                 return PAGE_SUCCESS;
515         }
516
517         return PAGE_CLEAN;
518 }
519
520 /*
521  * Same as remove_mapping, but if the page is removed from the mapping, it
522  * gets returned with a refcount of 0.
523  */
524 static int __remove_mapping(struct address_space *mapping, struct page *page)
525 {
526         BUG_ON(!PageLocked(page));
527         BUG_ON(mapping != page_mapping(page));
528
529         spin_lock_irq(&mapping->tree_lock);
530         /*
531          * The non racy check for a busy page.
532          *
533          * Must be careful with the order of the tests. When someone has
534          * a ref to the page, it may be possible that they dirty it then
535          * drop the reference. So if PageDirty is tested before page_count
536          * here, then the following race may occur:
537          *
538          * get_user_pages(&page);
539          * [user mapping goes away]
540          * write_to(page);
541          *                              !PageDirty(page)    [good]
542          * SetPageDirty(page);
543          * put_page(page);
544          *                              !page_count(page)   [good, discard it]
545          *
546          * [oops, our write_to data is lost]
547          *
548          * Reversing the order of the tests ensures such a situation cannot
549          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
550          * load is not satisfied before that of page->_count.
551          *
552          * Note that if SetPageDirty is always performed via set_page_dirty,
553          * and thus under tree_lock, then this ordering is not required.
554          */
555         if (!page_freeze_refs(page, 2))
556                 goto cannot_free;
557         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
558         if (unlikely(PageDirty(page))) {
559                 page_unfreeze_refs(page, 2);
560                 goto cannot_free;
561         }
562
563         if (PageSwapCache(page)) {
564                 swp_entry_t swap = { .val = page_private(page) };
565                 __delete_from_swap_cache(page);
566                 spin_unlock_irq(&mapping->tree_lock);
567                 swapcache_free(swap, page);
568         } else {
569                 void (*freepage)(struct page *);
570
571                 freepage = mapping->a_ops->freepage;
572
573                 __delete_from_page_cache(page);
574                 spin_unlock_irq(&mapping->tree_lock);
575                 mem_cgroup_uncharge_cache_page(page);
576
577                 if (freepage != NULL)
578                         freepage(page);
579         }
580
581         return 1;
582
583 cannot_free:
584         spin_unlock_irq(&mapping->tree_lock);
585         return 0;
586 }
587
588 /*
589  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
590  * someone else has a ref on the page, abort and return 0.  If it was
591  * successfully detached, return 1.  Assumes the caller has a single ref on
592  * this page.
593  */
594 int remove_mapping(struct address_space *mapping, struct page *page)
595 {
596         if (__remove_mapping(mapping, page)) {
597                 /*
598                  * Unfreezing the refcount with 1 rather than 2 effectively
599                  * drops the pagecache ref for us without requiring another
600                  * atomic operation.
601                  */
602                 page_unfreeze_refs(page, 1);
603                 return 1;
604         }
605         return 0;
606 }
607
608 /**
609  * putback_lru_page - put previously isolated page onto appropriate LRU list
610  * @page: page to be put back to appropriate lru list
611  *
612  * Add previously isolated @page to appropriate LRU list.
613  * Page may still be unevictable for other reasons.
614  *
615  * lru_lock must not be held, interrupts must be enabled.
616  */
617 void putback_lru_page(struct page *page)
618 {
619         int lru;
620         int active = !!TestClearPageActive(page);
621         int was_unevictable = PageUnevictable(page);
622
623         VM_BUG_ON(PageLRU(page));
624
625 redo:
626         ClearPageUnevictable(page);
627
628         if (page_evictable(page, NULL)) {
629                 /*
630                  * For evictable pages, we can use the cache.
631                  * In event of a race, worst case is we end up with an
632                  * unevictable page on [in]active list.
633                  * We know how to handle that.
634                  */
635                 lru = active + page_lru_base_type(page);
636                 lru_cache_add_lru(page, lru);
637         } else {
638                 /*
639                  * Put unevictable pages directly on zone's unevictable
640                  * list.
641                  */
642                 lru = LRU_UNEVICTABLE;
643                 add_page_to_unevictable_list(page);
644                 /*
645                  * When racing with an mlock clearing (page is
646                  * unlocked), make sure that if the other thread does
647                  * not observe our setting of PG_lru and fails
648                  * isolation, we see PG_mlocked cleared below and move
649                  * the page back to the evictable list.
650                  *
651                  * The other side is TestClearPageMlocked().
652                  */
653                 smp_mb();
654         }
655
656         /*
657          * page's status can change while we move it among lru. If an evictable
658          * page is on unevictable list, it never be freed. To avoid that,
659          * check after we added it to the list, again.
660          */
661         if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
662                 if (!isolate_lru_page(page)) {
663                         put_page(page);
664                         goto redo;
665                 }
666                 /* This means someone else dropped this page from LRU
667                  * So, it will be freed or putback to LRU again. There is
668                  * nothing to do here.
669                  */
670         }
671
672         if (was_unevictable && lru != LRU_UNEVICTABLE)
673                 count_vm_event(UNEVICTABLE_PGRESCUED);
674         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
675                 count_vm_event(UNEVICTABLE_PGCULLED);
676
677         put_page(page);         /* drop ref from isolate */
678 }
679
680 enum page_references {
681         PAGEREF_RECLAIM,
682         PAGEREF_RECLAIM_CLEAN,
683         PAGEREF_KEEP,
684         PAGEREF_ACTIVATE,
685 };
686
687 static enum page_references page_check_references(struct page *page,
688                                                   struct scan_control *sc)
689 {
690         int referenced_ptes, referenced_page;
691         unsigned long vm_flags;
692
693         referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
694         referenced_page = TestClearPageReferenced(page);
695
696         /* Lumpy reclaim - ignore references */
697         if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
698                 return PAGEREF_RECLAIM;
699
700         /*
701          * Mlock lost the isolation race with us.  Let try_to_unmap()
702          * move the page to the unevictable list.
703          */
704         if (vm_flags & VM_LOCKED)
705                 return PAGEREF_RECLAIM;
706
707         if (referenced_ptes) {
708                 if (PageAnon(page))
709                         return PAGEREF_ACTIVATE;
710                 /*
711                  * All mapped pages start out with page table
712                  * references from the instantiating fault, so we need
713                  * to look twice if a mapped file page is used more
714                  * than once.
715                  *
716                  * Mark it and spare it for another trip around the
717                  * inactive list.  Another page table reference will
718                  * lead to its activation.
719                  *
720                  * Note: the mark is set for activated pages as well
721                  * so that recently deactivated but used pages are
722                  * quickly recovered.
723                  */
724                 SetPageReferenced(page);
725
726                 if (referenced_page)
727                         return PAGEREF_ACTIVATE;
728
729                 return PAGEREF_KEEP;
730         }
731
732         /* Reclaim if clean, defer dirty pages to writeback */
733         if (referenced_page && !PageSwapBacked(page))
734                 return PAGEREF_RECLAIM_CLEAN;
735
736         return PAGEREF_RECLAIM;
737 }
738
739 static noinline_for_stack void free_page_list(struct list_head *free_pages)
740 {
741         struct pagevec freed_pvec;
742         struct page *page, *tmp;
743
744         pagevec_init(&freed_pvec, 1);
745
746         list_for_each_entry_safe(page, tmp, free_pages, lru) {
747                 list_del(&page->lru);
748                 if (!pagevec_add(&freed_pvec, page)) {
749                         __pagevec_free(&freed_pvec);
750                         pagevec_reinit(&freed_pvec);
751                 }
752         }
753
754         pagevec_free(&freed_pvec);
755 }
756
757 /*
758  * shrink_page_list() returns the number of reclaimed pages
759  */
760 static unsigned long shrink_page_list(struct list_head *page_list,
761                                       struct zone *zone,
762                                       struct scan_control *sc)
763 {
764         LIST_HEAD(ret_pages);
765         LIST_HEAD(free_pages);
766         int pgactivate = 0;
767         unsigned long nr_dirty = 0;
768         unsigned long nr_congested = 0;
769         unsigned long nr_reclaimed = 0;
770
771         cond_resched();
772
773         while (!list_empty(page_list)) {
774                 enum page_references references;
775                 struct address_space *mapping;
776                 struct page *page;
777                 int may_enter_fs;
778
779                 cond_resched();
780
781                 page = lru_to_page(page_list);
782                 list_del(&page->lru);
783
784                 if (!trylock_page(page))
785                         goto keep;
786
787                 VM_BUG_ON(PageActive(page));
788                 VM_BUG_ON(page_zone(page) != zone);
789
790                 sc->nr_scanned++;
791
792                 if (unlikely(!page_evictable(page, NULL)))
793                         goto cull_mlocked;
794
795                 if (!sc->may_unmap && page_mapped(page))
796                         goto keep_locked;
797
798                 /* Double the slab pressure for mapped and swapcache pages */
799                 if (page_mapped(page) || PageSwapCache(page))
800                         sc->nr_scanned++;
801
802                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
803                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
804
805                 if (PageWriteback(page)) {
806                         /*
807                          * Synchronous reclaim is performed in two passes,
808                          * first an asynchronous pass over the list to
809                          * start parallel writeback, and a second synchronous
810                          * pass to wait for the IO to complete.  Wait here
811                          * for any page for which writeback has already
812                          * started.
813                          */
814                         if ((sc->reclaim_mode & RECLAIM_MODE_SYNC) &&
815                             may_enter_fs)
816                                 wait_on_page_writeback(page);
817                         else {
818                                 unlock_page(page);
819                                 goto keep_lumpy;
820                         }
821                 }
822
823                 references = page_check_references(page, sc);
824                 switch (references) {
825                 case PAGEREF_ACTIVATE:
826                         goto activate_locked;
827                 case PAGEREF_KEEP:
828                         goto keep_locked;
829                 case PAGEREF_RECLAIM:
830                 case PAGEREF_RECLAIM_CLEAN:
831                         ; /* try to reclaim the page below */
832                 }
833
834                 /*
835                  * Anonymous process memory has backing store?
836                  * Try to allocate it some swap space here.
837                  */
838                 if (PageAnon(page) && !PageSwapCache(page)) {
839                         if (!(sc->gfp_mask & __GFP_IO))
840                                 goto keep_locked;
841                         if (!add_to_swap(page))
842                                 goto activate_locked;
843                         may_enter_fs = 1;
844                 }
845
846                 mapping = page_mapping(page);
847
848                 /*
849                  * The page is mapped into the page tables of one or more
850                  * processes. Try to unmap it here.
851                  */
852                 if (page_mapped(page) && mapping) {
853                         switch (try_to_unmap(page, TTU_UNMAP)) {
854                         case SWAP_FAIL:
855                                 goto activate_locked;
856                         case SWAP_AGAIN:
857                                 goto keep_locked;
858                         case SWAP_MLOCK:
859                                 goto cull_mlocked;
860                         case SWAP_SUCCESS:
861                                 ; /* try to free the page below */
862                         }
863                 }
864
865                 if (PageDirty(page)) {
866                         nr_dirty++;
867
868                         if (references == PAGEREF_RECLAIM_CLEAN)
869                                 goto keep_locked;
870                         if (!may_enter_fs)
871                                 goto keep_locked;
872                         if (!sc->may_writepage)
873                                 goto keep_locked;
874
875                         /* Page is dirty, try to write it out here */
876                         switch (pageout(page, mapping, sc)) {
877                         case PAGE_KEEP:
878                                 nr_congested++;
879                                 goto keep_locked;
880                         case PAGE_ACTIVATE:
881                                 goto activate_locked;
882                         case PAGE_SUCCESS:
883                                 if (PageWriteback(page))
884                                         goto keep_lumpy;
885                                 if (PageDirty(page))
886                                         goto keep;
887
888                                 /*
889                                  * A synchronous write - probably a ramdisk.  Go
890                                  * ahead and try to reclaim the page.
891                                  */
892                                 if (!trylock_page(page))
893                                         goto keep;
894                                 if (PageDirty(page) || PageWriteback(page))
895                                         goto keep_locked;
896                                 mapping = page_mapping(page);
897                         case PAGE_CLEAN:
898                                 ; /* try to free the page below */
899                         }
900                 }
901
902                 /*
903                  * If the page has buffers, try to free the buffer mappings
904                  * associated with this page. If we succeed we try to free
905                  * the page as well.
906                  *
907                  * We do this even if the page is PageDirty().
908                  * try_to_release_page() does not perform I/O, but it is
909                  * possible for a page to have PageDirty set, but it is actually
910                  * clean (all its buffers are clean).  This happens if the
911                  * buffers were written out directly, with submit_bh(). ext3
912                  * will do this, as well as the blockdev mapping.
913                  * try_to_release_page() will discover that cleanness and will
914                  * drop the buffers and mark the page clean - it can be freed.
915                  *
916                  * Rarely, pages can have buffers and no ->mapping.  These are
917                  * the pages which were not successfully invalidated in
918                  * truncate_complete_page().  We try to drop those buffers here
919                  * and if that worked, and the page is no longer mapped into
920                  * process address space (page_count == 1) it can be freed.
921                  * Otherwise, leave the page on the LRU so it is swappable.
922                  */
923                 if (page_has_private(page)) {
924                         if (!try_to_release_page(page, sc->gfp_mask))
925                                 goto activate_locked;
926                         if (!mapping && page_count(page) == 1) {
927                                 unlock_page(page);
928                                 if (put_page_testzero(page))
929                                         goto free_it;
930                                 else {
931                                         /*
932                                          * rare race with speculative reference.
933                                          * the speculative reference will free
934                                          * this page shortly, so we may
935                                          * increment nr_reclaimed here (and
936                                          * leave it off the LRU).
937                                          */
938                                         nr_reclaimed++;
939                                         continue;
940                                 }
941                         }
942                 }
943
944                 if (!mapping || !__remove_mapping(mapping, page))
945                         goto keep_locked;
946
947                 /*
948                  * At this point, we have no other references and there is
949                  * no way to pick any more up (removed from LRU, removed
950                  * from pagecache). Can use non-atomic bitops now (and
951                  * we obviously don't have to worry about waking up a process
952                  * waiting on the page lock, because there are no references.
953                  */
954                 __clear_page_locked(page);
955 free_it:
956                 nr_reclaimed++;
957
958                 /*
959                  * Is there need to periodically free_page_list? It would
960                  * appear not as the counts should be low
961                  */
962                 list_add(&page->lru, &free_pages);
963                 continue;
964
965 cull_mlocked:
966                 if (PageSwapCache(page))
967                         try_to_free_swap(page);
968                 unlock_page(page);
969                 putback_lru_page(page);
970                 reset_reclaim_mode(sc);
971                 continue;
972
973 activate_locked:
974                 /* Not a candidate for swapping, so reclaim swap space. */
975                 if (PageSwapCache(page) && vm_swap_full())
976                         try_to_free_swap(page);
977                 VM_BUG_ON(PageActive(page));
978                 SetPageActive(page);
979                 pgactivate++;
980 keep_locked:
981                 unlock_page(page);
982 keep:
983                 reset_reclaim_mode(sc);
984 keep_lumpy:
985                 list_add(&page->lru, &ret_pages);
986                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
987         }
988
989         /*
990          * Tag a zone as congested if all the dirty pages encountered were
991          * backed by a congested BDI. In this case, reclaimers should just
992          * back off and wait for congestion to clear because further reclaim
993          * will encounter the same problem
994          */
995         if (nr_dirty && nr_dirty == nr_congested && scanning_global_lru(sc))
996                 zone_set_flag(zone, ZONE_CONGESTED);
997
998         free_page_list(&free_pages);
999
1000         list_splice(&ret_pages, page_list);
1001         count_vm_events(PGACTIVATE, pgactivate);
1002         return nr_reclaimed;
1003 }
1004
1005 /*
1006  * Attempt to remove the specified page from its LRU.  Only take this page
1007  * if it is of the appropriate PageActive status.  Pages which are being
1008  * freed elsewhere are also ignored.
1009  *
1010  * page:        page to consider
1011  * mode:        one of the LRU isolation modes defined above
1012  *
1013  * returns 0 on success, -ve errno on failure.
1014  */
1015 int __isolate_lru_page(struct page *page, int mode, int file)
1016 {
1017         int ret = -EINVAL;
1018
1019         /* Only take pages on the LRU. */
1020         if (!PageLRU(page))
1021                 return ret;
1022
1023         /*
1024          * When checking the active state, we need to be sure we are
1025          * dealing with comparible boolean values.  Take the logical not
1026          * of each.
1027          */
1028         if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
1029                 return ret;
1030
1031         if (mode != ISOLATE_BOTH && page_is_file_cache(page) != file)
1032                 return ret;
1033
1034         /*
1035          * When this function is being called for lumpy reclaim, we
1036          * initially look into all LRU pages, active, inactive and
1037          * unevictable; only give shrink_page_list evictable pages.
1038          */
1039         if (PageUnevictable(page))
1040                 return ret;
1041
1042         ret = -EBUSY;
1043
1044         if (likely(get_page_unless_zero(page))) {
1045                 /*
1046                  * Be careful not to clear PageLRU until after we're
1047                  * sure the page is not being freed elsewhere -- the
1048                  * page release code relies on it.
1049                  */
1050                 ClearPageLRU(page);
1051                 ret = 0;
1052         }
1053
1054         return ret;
1055 }
1056
1057 /*
1058  * zone->lru_lock is heavily contended.  Some of the functions that
1059  * shrink the lists perform better by taking out a batch of pages
1060  * and working on them outside the LRU lock.
1061  *
1062  * For pagecache intensive workloads, this function is the hottest
1063  * spot in the kernel (apart from copy_*_user functions).
1064  *
1065  * Appropriate locks must be held before calling this function.
1066  *
1067  * @nr_to_scan: The number of pages to look through on the list.
1068  * @src:        The LRU list to pull pages off.
1069  * @dst:        The temp list to put pages on to.
1070  * @scanned:    The number of pages that were scanned.
1071  * @order:      The caller's attempted allocation order
1072  * @mode:       One of the LRU isolation modes
1073  * @file:       True [1] if isolating file [!anon] pages
1074  *
1075  * returns how many pages were moved onto *@dst.
1076  */
1077 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1078                 struct list_head *src, struct list_head *dst,
1079                 unsigned long *scanned, int order, int mode, int file)
1080 {
1081         unsigned long nr_taken = 0;
1082         unsigned long nr_lumpy_taken = 0;
1083         unsigned long nr_lumpy_dirty = 0;
1084         unsigned long nr_lumpy_failed = 0;
1085         unsigned long scan;
1086
1087         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1088                 struct page *page;
1089                 unsigned long pfn;
1090                 unsigned long end_pfn;
1091                 unsigned long page_pfn;
1092                 int zone_id;
1093
1094                 page = lru_to_page(src);
1095                 prefetchw_prev_lru_page(page, src, flags);
1096
1097                 VM_BUG_ON(!PageLRU(page));
1098
1099                 switch (__isolate_lru_page(page, mode, file)) {
1100                 case 0:
1101                         list_move(&page->lru, dst);
1102                         mem_cgroup_del_lru(page);
1103                         nr_taken += hpage_nr_pages(page);
1104                         break;
1105
1106                 case -EBUSY:
1107                         /* else it is being freed elsewhere */
1108                         list_move(&page->lru, src);
1109                         mem_cgroup_rotate_lru_list(page, page_lru(page));
1110                         continue;
1111
1112                 default:
1113                         BUG();
1114                 }
1115
1116                 if (!order)
1117                         continue;
1118
1119                 /*
1120                  * Attempt to take all pages in the order aligned region
1121                  * surrounding the tag page.  Only take those pages of
1122                  * the same active state as that tag page.  We may safely
1123                  * round the target page pfn down to the requested order
1124                  * as the mem_map is guaranteed valid out to MAX_ORDER,
1125                  * where that page is in a different zone we will detect
1126                  * it from its zone id and abort this block scan.
1127                  */
1128                 zone_id = page_zone_id(page);
1129                 page_pfn = page_to_pfn(page);
1130                 pfn = page_pfn & ~((1 << order) - 1);
1131                 end_pfn = pfn + (1 << order);
1132                 for (; pfn < end_pfn; pfn++) {
1133                         struct page *cursor_page;
1134
1135                         /* The target page is in the block, ignore it. */
1136                         if (unlikely(pfn == page_pfn))
1137                                 continue;
1138
1139                         /* Avoid holes within the zone. */
1140                         if (unlikely(!pfn_valid_within(pfn)))
1141                                 break;
1142
1143                         cursor_page = pfn_to_page(pfn);
1144
1145                         /* Check that we have not crossed a zone boundary. */
1146                         if (unlikely(page_zone_id(cursor_page) != zone_id))
1147                                 break;
1148
1149                         /*
1150                          * If we don't have enough swap space, reclaiming of
1151                          * anon page which don't already have a swap slot is
1152                          * pointless.
1153                          */
1154                         if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
1155                             !PageSwapCache(cursor_page))
1156                                 break;
1157
1158                         if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1159                                 list_move(&cursor_page->lru, dst);
1160                                 mem_cgroup_del_lru(cursor_page);
1161                                 nr_taken += hpage_nr_pages(page);
1162                                 nr_lumpy_taken++;
1163                                 if (PageDirty(cursor_page))
1164                                         nr_lumpy_dirty++;
1165                                 scan++;
1166                         } else {
1167                                 /*
1168                                  * Check if the page is freed already.
1169                                  *
1170                                  * We can't use page_count() as that
1171                                  * requires compound_head and we don't
1172                                  * have a pin on the page here. If a
1173                                  * page is tail, we may or may not
1174                                  * have isolated the head, so assume
1175                                  * it's not free, it'd be tricky to
1176                                  * track the head status without a
1177                                  * page pin.
1178                                  */
1179                                 if (!PageTail(cursor_page) &&
1180                                     !atomic_read(&cursor_page->_count))
1181                                         continue;
1182                                 break;
1183                         }
1184                 }
1185
1186                 /* If we break out of the loop above, lumpy reclaim failed */
1187                 if (pfn < end_pfn)
1188                         nr_lumpy_failed++;
1189         }
1190
1191         *scanned = scan;
1192
1193         trace_mm_vmscan_lru_isolate(order,
1194                         nr_to_scan, scan,
1195                         nr_taken,
1196                         nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1197                         mode);
1198         return nr_taken;
1199 }
1200
1201 static unsigned long isolate_pages_global(unsigned long nr,
1202                                         struct list_head *dst,
1203                                         unsigned long *scanned, int order,
1204                                         int mode, struct zone *z,
1205                                         int active, int file)
1206 {
1207         int lru = LRU_BASE;
1208         if (active)
1209                 lru += LRU_ACTIVE;
1210         if (file)
1211                 lru += LRU_FILE;
1212         return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1213                                                                 mode, file);
1214 }
1215
1216 /*
1217  * clear_active_flags() is a helper for shrink_active_list(), clearing
1218  * any active bits from the pages in the list.
1219  */
1220 static unsigned long clear_active_flags(struct list_head *page_list,
1221                                         unsigned int *count)
1222 {
1223         int nr_active = 0;
1224         int lru;
1225         struct page *page;
1226
1227         list_for_each_entry(page, page_list, lru) {
1228                 int numpages = hpage_nr_pages(page);
1229                 lru = page_lru_base_type(page);
1230                 if (PageActive(page)) {
1231                         lru += LRU_ACTIVE;
1232                         ClearPageActive(page);
1233                         nr_active += numpages;
1234                 }
1235                 if (count)
1236                         count[lru] += numpages;
1237         }
1238
1239         return nr_active;
1240 }
1241
1242 /**
1243  * isolate_lru_page - tries to isolate a page from its LRU list
1244  * @page: page to isolate from its LRU list
1245  *
1246  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1247  * vmstat statistic corresponding to whatever LRU list the page was on.
1248  *
1249  * Returns 0 if the page was removed from an LRU list.
1250  * Returns -EBUSY if the page was not on an LRU list.
1251  *
1252  * The returned page will have PageLRU() cleared.  If it was found on
1253  * the active list, it will have PageActive set.  If it was found on
1254  * the unevictable list, it will have the PageUnevictable bit set. That flag
1255  * may need to be cleared by the caller before letting the page go.
1256  *
1257  * The vmstat statistic corresponding to the list on which the page was
1258  * found will be decremented.
1259  *
1260  * Restrictions:
1261  * (1) Must be called with an elevated refcount on the page. This is a
1262  *     fundamentnal difference from isolate_lru_pages (which is called
1263  *     without a stable reference).
1264  * (2) the lru_lock must not be held.
1265  * (3) interrupts must be enabled.
1266  */
1267 int isolate_lru_page(struct page *page)
1268 {
1269         int ret = -EBUSY;
1270
1271         VM_BUG_ON(!page_count(page));
1272
1273         if (PageLRU(page)) {
1274                 struct zone *zone = page_zone(page);
1275
1276                 spin_lock_irq(&zone->lru_lock);
1277                 if (PageLRU(page)) {
1278                         int lru = page_lru(page);
1279                         ret = 0;
1280                         get_page(page);
1281                         ClearPageLRU(page);
1282
1283                         del_page_from_lru_list(zone, page, lru);
1284                 }
1285                 spin_unlock_irq(&zone->lru_lock);
1286         }
1287         return ret;
1288 }
1289
1290 /*
1291  * Are there way too many processes in the direct reclaim path already?
1292  */
1293 static int too_many_isolated(struct zone *zone, int file,
1294                 struct scan_control *sc)
1295 {
1296         unsigned long inactive, isolated;
1297
1298         if (current_is_kswapd())
1299                 return 0;
1300
1301         if (!scanning_global_lru(sc))
1302                 return 0;
1303
1304         if (file) {
1305                 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1306                 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1307         } else {
1308                 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1309                 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1310         }
1311
1312         return isolated > inactive;
1313 }
1314
1315 /*
1316  * TODO: Try merging with migrations version of putback_lru_pages
1317  */
1318 static noinline_for_stack void
1319 putback_lru_pages(struct zone *zone, struct scan_control *sc,
1320                                 unsigned long nr_anon, unsigned long nr_file,
1321                                 struct list_head *page_list)
1322 {
1323         struct page *page;
1324         struct pagevec pvec;
1325         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1326
1327         pagevec_init(&pvec, 1);
1328
1329         /*
1330          * Put back any unfreeable pages.
1331          */
1332         spin_lock(&zone->lru_lock);
1333         while (!list_empty(page_list)) {
1334                 int lru;
1335                 page = lru_to_page(page_list);
1336                 VM_BUG_ON(PageLRU(page));
1337                 list_del(&page->lru);
1338                 if (unlikely(!page_evictable(page, NULL))) {
1339                         spin_unlock_irq(&zone->lru_lock);
1340                         putback_lru_page(page);
1341                         spin_lock_irq(&zone->lru_lock);
1342                         continue;
1343                 }
1344                 SetPageLRU(page);
1345                 lru = page_lru(page);
1346                 add_page_to_lru_list(zone, page, lru);
1347                 if (is_active_lru(lru)) {
1348                         int file = is_file_lru(lru);
1349                         int numpages = hpage_nr_pages(page);
1350                         reclaim_stat->recent_rotated[file] += numpages;
1351                 }
1352                 if (!pagevec_add(&pvec, page)) {
1353                         spin_unlock_irq(&zone->lru_lock);
1354                         __pagevec_release(&pvec);
1355                         spin_lock_irq(&zone->lru_lock);
1356                 }
1357         }
1358         __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1359         __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1360
1361         spin_unlock_irq(&zone->lru_lock);
1362         pagevec_release(&pvec);
1363 }
1364
1365 static noinline_for_stack void update_isolated_counts(struct zone *zone,
1366                                         struct scan_control *sc,
1367                                         unsigned long *nr_anon,
1368                                         unsigned long *nr_file,
1369                                         struct list_head *isolated_list)
1370 {
1371         unsigned long nr_active;
1372         unsigned int count[NR_LRU_LISTS] = { 0, };
1373         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1374
1375         nr_active = clear_active_flags(isolated_list, count);
1376         __count_vm_events(PGDEACTIVATE, nr_active);
1377
1378         __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1379                               -count[LRU_ACTIVE_FILE]);
1380         __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1381                               -count[LRU_INACTIVE_FILE]);
1382         __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1383                               -count[LRU_ACTIVE_ANON]);
1384         __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1385                               -count[LRU_INACTIVE_ANON]);
1386
1387         *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1388         *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1389         __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1390         __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1391
1392         reclaim_stat->recent_scanned[0] += *nr_anon;
1393         reclaim_stat->recent_scanned[1] += *nr_file;
1394 }
1395
1396 /*
1397  * Returns true if the caller should wait to clean dirty/writeback pages.
1398  *
1399  * If we are direct reclaiming for contiguous pages and we do not reclaim
1400  * everything in the list, try again and wait for writeback IO to complete.
1401  * This will stall high-order allocations noticeably. Only do that when really
1402  * need to free the pages under high memory pressure.
1403  */
1404 static inline bool should_reclaim_stall(unsigned long nr_taken,
1405                                         unsigned long nr_freed,
1406                                         int priority,
1407                                         struct scan_control *sc)
1408 {
1409         int lumpy_stall_priority;
1410
1411         /* kswapd should not stall on sync IO */
1412         if (current_is_kswapd())
1413                 return false;
1414
1415         /* Only stall on lumpy reclaim */
1416         if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
1417                 return false;
1418
1419         /* If we have relaimed everything on the isolated list, no stall */
1420         if (nr_freed == nr_taken)
1421                 return false;
1422
1423         /*
1424          * For high-order allocations, there are two stall thresholds.
1425          * High-cost allocations stall immediately where as lower
1426          * order allocations such as stacks require the scanning
1427          * priority to be much higher before stalling.
1428          */
1429         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1430                 lumpy_stall_priority = DEF_PRIORITY;
1431         else
1432                 lumpy_stall_priority = DEF_PRIORITY / 3;
1433
1434         return priority <= lumpy_stall_priority;
1435 }
1436
1437 /*
1438  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1439  * of reclaimed pages
1440  */
1441 static noinline_for_stack unsigned long
1442 shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1443                         struct scan_control *sc, int priority, int file)
1444 {
1445         LIST_HEAD(page_list);
1446         unsigned long nr_scanned;
1447         unsigned long nr_reclaimed = 0;
1448         unsigned long nr_taken;
1449         unsigned long nr_anon;
1450         unsigned long nr_file;
1451
1452         while (unlikely(too_many_isolated(zone, file, sc))) {
1453                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1454
1455                 /* We are about to die and free our memory. Return now. */
1456                 if (fatal_signal_pending(current))
1457                         return SWAP_CLUSTER_MAX;
1458         }
1459
1460         set_reclaim_mode(priority, sc, false);
1461         lru_add_drain();
1462         spin_lock_irq(&zone->lru_lock);
1463
1464         if (scanning_global_lru(sc)) {
1465                 nr_taken = isolate_pages_global(nr_to_scan,
1466                         &page_list, &nr_scanned, sc->order,
1467                         sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM ?
1468                                         ISOLATE_BOTH : ISOLATE_INACTIVE,
1469                         zone, 0, file);
1470                 zone->pages_scanned += nr_scanned;
1471                 if (current_is_kswapd())
1472                         __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1473                                                nr_scanned);
1474                 else
1475                         __count_zone_vm_events(PGSCAN_DIRECT, zone,
1476                                                nr_scanned);
1477         } else {
1478                 nr_taken = mem_cgroup_isolate_pages(nr_to_scan,
1479                         &page_list, &nr_scanned, sc->order,
1480                         sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM ?
1481                                         ISOLATE_BOTH : ISOLATE_INACTIVE,
1482                         zone, sc->mem_cgroup,
1483                         0, file);
1484                 /*
1485                  * mem_cgroup_isolate_pages() keeps track of
1486                  * scanned pages on its own.
1487                  */
1488         }
1489
1490         if (nr_taken == 0) {
1491                 spin_unlock_irq(&zone->lru_lock);
1492                 return 0;
1493         }
1494
1495         update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
1496
1497         spin_unlock_irq(&zone->lru_lock);
1498
1499         nr_reclaimed = shrink_page_list(&page_list, zone, sc);
1500
1501         /* Check if we should syncronously wait for writeback */
1502         if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1503                 set_reclaim_mode(priority, sc, true);
1504                 nr_reclaimed += shrink_page_list(&page_list, zone, sc);
1505         }
1506
1507         local_irq_disable();
1508         if (current_is_kswapd())
1509                 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1510         __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1511
1512         putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
1513
1514         trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1515                 zone_idx(zone),
1516                 nr_scanned, nr_reclaimed,
1517                 priority,
1518                 trace_shrink_flags(file, sc->reclaim_mode));
1519         return nr_reclaimed;
1520 }
1521
1522 /*
1523  * This moves pages from the active list to the inactive list.
1524  *
1525  * We move them the other way if the page is referenced by one or more
1526  * processes, from rmap.
1527  *
1528  * If the pages are mostly unmapped, the processing is fast and it is
1529  * appropriate to hold zone->lru_lock across the whole operation.  But if
1530  * the pages are mapped, the processing is slow (page_referenced()) so we
1531  * should drop zone->lru_lock around each page.  It's impossible to balance
1532  * this, so instead we remove the pages from the LRU while processing them.
1533  * It is safe to rely on PG_active against the non-LRU pages in here because
1534  * nobody will play with that bit on a non-LRU page.
1535  *
1536  * The downside is that we have to touch page->_count against each page.
1537  * But we had to alter page->flags anyway.
1538  */
1539
1540 static void move_active_pages_to_lru(struct zone *zone,
1541                                      struct list_head *list,
1542                                      enum lru_list lru)
1543 {
1544         unsigned long pgmoved = 0;
1545         struct pagevec pvec;
1546         struct page *page;
1547
1548         pagevec_init(&pvec, 1);
1549
1550         while (!list_empty(list)) {
1551                 page = lru_to_page(list);
1552
1553                 VM_BUG_ON(PageLRU(page));
1554                 SetPageLRU(page);
1555
1556                 list_move(&page->lru, &zone->lru[lru].list);
1557                 mem_cgroup_add_lru_list(page, lru);
1558                 pgmoved += hpage_nr_pages(page);
1559
1560                 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1561                         spin_unlock_irq(&zone->lru_lock);
1562                         if (buffer_heads_over_limit)
1563                                 pagevec_strip(&pvec);
1564                         __pagevec_release(&pvec);
1565                         spin_lock_irq(&zone->lru_lock);
1566                 }
1567         }
1568         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1569         if (!is_active_lru(lru))
1570                 __count_vm_events(PGDEACTIVATE, pgmoved);
1571 }
1572
1573 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1574                         struct scan_control *sc, int priority, int file)
1575 {
1576         unsigned long nr_taken;
1577         unsigned long pgscanned;
1578         unsigned long vm_flags;
1579         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1580         LIST_HEAD(l_active);
1581         LIST_HEAD(l_inactive);
1582         struct page *page;
1583         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1584         unsigned long nr_rotated = 0;
1585
1586         lru_add_drain();
1587         spin_lock_irq(&zone->lru_lock);
1588         if (scanning_global_lru(sc)) {
1589                 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1590                                                 &pgscanned, sc->order,
1591                                                 ISOLATE_ACTIVE, zone,
1592                                                 1, file);
1593                 zone->pages_scanned += pgscanned;
1594         } else {
1595                 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1596                                                 &pgscanned, sc->order,
1597                                                 ISOLATE_ACTIVE, zone,
1598                                                 sc->mem_cgroup, 1, file);
1599                 /*
1600                  * mem_cgroup_isolate_pages() keeps track of
1601                  * scanned pages on its own.
1602                  */
1603         }
1604
1605         reclaim_stat->recent_scanned[file] += nr_taken;
1606
1607         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1608         if (file)
1609                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1610         else
1611                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1612         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1613         spin_unlock_irq(&zone->lru_lock);
1614
1615         while (!list_empty(&l_hold)) {
1616                 cond_resched();
1617                 page = lru_to_page(&l_hold);
1618                 list_del(&page->lru);
1619
1620                 if (unlikely(!page_evictable(page, NULL))) {
1621                         putback_lru_page(page);
1622                         continue;
1623                 }
1624
1625                 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1626                         nr_rotated += hpage_nr_pages(page);
1627                         /*
1628                          * Identify referenced, file-backed active pages and
1629                          * give them one more trip around the active list. So
1630                          * that executable code get better chances to stay in
1631                          * memory under moderate memory pressure.  Anon pages
1632                          * are not likely to be evicted by use-once streaming
1633                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1634                          * so we ignore them here.
1635                          */
1636                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1637                                 list_add(&page->lru, &l_active);
1638                                 continue;
1639                         }
1640                 }
1641
1642                 ClearPageActive(page);  /* we are de-activating */
1643                 list_add(&page->lru, &l_inactive);
1644         }
1645
1646         /*
1647          * Move pages back to the lru list.
1648          */
1649         spin_lock_irq(&zone->lru_lock);
1650         /*
1651          * Count referenced pages from currently used mappings as rotated,
1652          * even though only some of them are actually re-activated.  This
1653          * helps balance scan pressure between file and anonymous pages in
1654          * get_scan_ratio.
1655          */
1656         reclaim_stat->recent_rotated[file] += nr_rotated;
1657
1658         move_active_pages_to_lru(zone, &l_active,
1659                                                 LRU_ACTIVE + file * LRU_FILE);
1660         move_active_pages_to_lru(zone, &l_inactive,
1661                                                 LRU_BASE   + file * LRU_FILE);
1662         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1663         spin_unlock_irq(&zone->lru_lock);
1664 }
1665
1666 #ifdef CONFIG_SWAP
1667 static int inactive_anon_is_low_global(struct zone *zone)
1668 {
1669         unsigned long active, inactive;
1670
1671         active = zone_page_state(zone, NR_ACTIVE_ANON);
1672         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1673
1674         if (inactive * zone->inactive_ratio < active)
1675                 return 1;
1676
1677         return 0;
1678 }
1679
1680 /**
1681  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1682  * @zone: zone to check
1683  * @sc:   scan control of this context
1684  *
1685  * Returns true if the zone does not have enough inactive anon pages,
1686  * meaning some active anon pages need to be deactivated.
1687  */
1688 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1689 {
1690         int low;
1691
1692         /*
1693          * If we don't have swap space, anonymous page deactivation
1694          * is pointless.
1695          */
1696         if (!total_swap_pages)
1697                 return 0;
1698
1699         if (scanning_global_lru(sc))
1700                 low = inactive_anon_is_low_global(zone);
1701         else
1702                 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1703         return low;
1704 }
1705 #else
1706 static inline int inactive_anon_is_low(struct zone *zone,
1707                                         struct scan_control *sc)
1708 {
1709         return 0;
1710 }
1711 #endif
1712
1713 static int inactive_file_is_low_global(struct zone *zone)
1714 {
1715         unsigned long active, inactive;
1716
1717         active = zone_page_state(zone, NR_ACTIVE_FILE);
1718         inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1719
1720         return (active > inactive);
1721 }
1722
1723 /**
1724  * inactive_file_is_low - check if file pages need to be deactivated
1725  * @zone: zone to check
1726  * @sc:   scan control of this context
1727  *
1728  * When the system is doing streaming IO, memory pressure here
1729  * ensures that active file pages get deactivated, until more
1730  * than half of the file pages are on the inactive list.
1731  *
1732  * Once we get to that situation, protect the system's working
1733  * set from being evicted by disabling active file page aging.
1734  *
1735  * This uses a different ratio than the anonymous pages, because
1736  * the page cache uses a use-once replacement algorithm.
1737  */
1738 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1739 {
1740         int low;
1741
1742         if (scanning_global_lru(sc))
1743                 low = inactive_file_is_low_global(zone);
1744         else
1745                 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1746         return low;
1747 }
1748
1749 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1750                                 int file)
1751 {
1752         if (file)
1753                 return inactive_file_is_low(zone, sc);
1754         else
1755                 return inactive_anon_is_low(zone, sc);
1756 }
1757
1758 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1759         struct zone *zone, struct scan_control *sc, int priority)
1760 {
1761         int file = is_file_lru(lru);
1762
1763         if (is_active_lru(lru)) {
1764                 if (inactive_list_is_low(zone, sc, file))
1765                     shrink_active_list(nr_to_scan, zone, sc, priority, file);
1766                 return 0;
1767         }
1768
1769         return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1770 }
1771
1772 /*
1773  * Determine how aggressively the anon and file LRU lists should be
1774  * scanned.  The relative value of each set of LRU lists is determined
1775  * by looking at the fraction of the pages scanned we did rotate back
1776  * onto the active list instead of evict.
1777  *
1778  * nr[0] = anon pages to scan; nr[1] = file pages to scan
1779  */
1780 static void get_scan_count(struct zone *zone, struct scan_control *sc,
1781                                         unsigned long *nr, int priority)
1782 {
1783         unsigned long anon, file, free;
1784         unsigned long anon_prio, file_prio;
1785         unsigned long ap, fp;
1786         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1787         u64 fraction[2], denominator;
1788         enum lru_list l;
1789         int noswap = 0;
1790         int force_scan = 0;
1791
1792
1793         anon  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1794                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1795         file  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1796                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1797
1798         if (((anon + file) >> priority) < SWAP_CLUSTER_MAX) {
1799                 /* kswapd does zone balancing and need to scan this zone */
1800                 if (scanning_global_lru(sc) && current_is_kswapd())
1801                         force_scan = 1;
1802                 /* memcg may have small limit and need to avoid priority drop */
1803                 if (!scanning_global_lru(sc))
1804                         force_scan = 1;
1805         }
1806
1807         /* If we have no swap space, do not bother scanning anon pages. */
1808         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1809                 noswap = 1;
1810                 fraction[0] = 0;
1811                 fraction[1] = 1;
1812                 denominator = 1;
1813                 goto out;
1814         }
1815
1816         if (scanning_global_lru(sc)) {
1817                 free  = zone_page_state(zone, NR_FREE_PAGES);
1818                 /* If we have very few page cache pages,
1819                    force-scan anon pages. */
1820                 if (unlikely(file + free <= high_wmark_pages(zone))) {
1821                         fraction[0] = 1;
1822                         fraction[1] = 0;
1823                         denominator = 1;
1824                         goto out;
1825                 }
1826         }
1827
1828         /*
1829          * With swappiness at 100, anonymous and file have the same priority.
1830          * This scanning priority is essentially the inverse of IO cost.
1831          */
1832         anon_prio = sc->swappiness;
1833         file_prio = 200 - sc->swappiness;
1834
1835         /*
1836          * OK, so we have swap space and a fair amount of page cache
1837          * pages.  We use the recently rotated / recently scanned
1838          * ratios to determine how valuable each cache is.
1839          *
1840          * Because workloads change over time (and to avoid overflow)
1841          * we keep these statistics as a floating average, which ends
1842          * up weighing recent references more than old ones.
1843          *
1844          * anon in [0], file in [1]
1845          */
1846         spin_lock_irq(&zone->lru_lock);
1847         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1848                 reclaim_stat->recent_scanned[0] /= 2;
1849                 reclaim_stat->recent_rotated[0] /= 2;
1850         }
1851
1852         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1853                 reclaim_stat->recent_scanned[1] /= 2;
1854                 reclaim_stat->recent_rotated[1] /= 2;
1855         }
1856
1857         /*
1858          * The amount of pressure on anon vs file pages is inversely
1859          * proportional to the fraction of recently scanned pages on
1860          * each list that were recently referenced and in active use.
1861          */
1862         ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1863         ap /= reclaim_stat->recent_rotated[0] + 1;
1864
1865         fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1866         fp /= reclaim_stat->recent_rotated[1] + 1;
1867         spin_unlock_irq(&zone->lru_lock);
1868
1869         fraction[0] = ap;
1870         fraction[1] = fp;
1871         denominator = ap + fp + 1;
1872 out:
1873         for_each_evictable_lru(l) {
1874                 int file = is_file_lru(l);
1875                 unsigned long scan;
1876
1877                 scan = zone_nr_lru_pages(zone, sc, l);
1878                 if (priority || noswap) {
1879                         scan >>= priority;
1880                         scan = div64_u64(scan * fraction[file], denominator);
1881                 }
1882
1883                 /*
1884                  * If zone is small or memcg is small, nr[l] can be 0.
1885                  * This results no-scan on this priority and priority drop down.
1886                  * For global direct reclaim, it can visit next zone and tend
1887                  * not to have problems. For global kswapd, it's for zone
1888                  * balancing and it need to scan a small amounts. When using
1889                  * memcg, priority drop can cause big latency. So, it's better
1890                  * to scan small amount. See may_noscan above.
1891                  */
1892                 if (!scan && force_scan) {
1893                         if (file)
1894                                 scan = SWAP_CLUSTER_MAX;
1895                         else if (!noswap)
1896                                 scan = SWAP_CLUSTER_MAX;
1897                 }
1898                 nr[l] = scan;
1899         }
1900 }
1901
1902 /*
1903  * Reclaim/compaction depends on a number of pages being freed. To avoid
1904  * disruption to the system, a small number of order-0 pages continue to be
1905  * rotated and reclaimed in the normal fashion. However, by the time we get
1906  * back to the allocator and call try_to_compact_zone(), we ensure that
1907  * there are enough free pages for it to be likely successful
1908  */
1909 static inline bool should_continue_reclaim(struct zone *zone,
1910                                         unsigned long nr_reclaimed,
1911                                         unsigned long nr_scanned,
1912                                         struct scan_control *sc)
1913 {
1914         unsigned long pages_for_compaction;
1915         unsigned long inactive_lru_pages;
1916
1917         /* If not in reclaim/compaction mode, stop */
1918         if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
1919                 return false;
1920
1921         /* Consider stopping depending on scan and reclaim activity */
1922         if (sc->gfp_mask & __GFP_REPEAT) {
1923                 /*
1924                  * For __GFP_REPEAT allocations, stop reclaiming if the
1925                  * full LRU list has been scanned and we are still failing
1926                  * to reclaim pages. This full LRU scan is potentially
1927                  * expensive but a __GFP_REPEAT caller really wants to succeed
1928                  */
1929                 if (!nr_reclaimed && !nr_scanned)
1930                         return false;
1931         } else {
1932                 /*
1933                  * For non-__GFP_REPEAT allocations which can presumably
1934                  * fail without consequence, stop if we failed to reclaim
1935                  * any pages from the last SWAP_CLUSTER_MAX number of
1936                  * pages that were scanned. This will return to the
1937                  * caller faster at the risk reclaim/compaction and
1938                  * the resulting allocation attempt fails
1939                  */
1940                 if (!nr_reclaimed)
1941                         return false;
1942         }
1943
1944         /*
1945          * If we have not reclaimed enough pages for compaction and the
1946          * inactive lists are large enough, continue reclaiming
1947          */
1948         pages_for_compaction = (2UL << sc->order);
1949         inactive_lru_pages = zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON) +
1950                                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1951         if (sc->nr_reclaimed < pages_for_compaction &&
1952                         inactive_lru_pages > pages_for_compaction)
1953                 return true;
1954
1955         /* If compaction would go ahead or the allocation would succeed, stop */
1956         switch (compaction_suitable(zone, sc->order)) {
1957         case COMPACT_PARTIAL:
1958         case COMPACT_CONTINUE:
1959                 return false;
1960         default:
1961                 return true;
1962         }
1963 }
1964
1965 /*
1966  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1967  */
1968 static void shrink_zone(int priority, struct zone *zone,
1969                                 struct scan_control *sc)
1970 {
1971         unsigned long nr[NR_LRU_LISTS];
1972         unsigned long nr_to_scan;
1973         enum lru_list l;
1974         unsigned long nr_reclaimed, nr_scanned;
1975         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1976
1977 restart:
1978         nr_reclaimed = 0;
1979         nr_scanned = sc->nr_scanned;
1980         get_scan_count(zone, sc, nr, priority);
1981
1982         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1983                                         nr[LRU_INACTIVE_FILE]) {
1984                 for_each_evictable_lru(l) {
1985                         if (nr[l]) {
1986                                 nr_to_scan = min_t(unsigned long,
1987                                                    nr[l], SWAP_CLUSTER_MAX);
1988                                 nr[l] -= nr_to_scan;
1989
1990                                 nr_reclaimed += shrink_list(l, nr_to_scan,
1991                                                             zone, sc, priority);
1992                         }
1993                 }
1994                 /*
1995                  * On large memory systems, scan >> priority can become
1996                  * really large. This is fine for the starting priority;
1997                  * we want to put equal scanning pressure on each zone.
1998                  * However, if the VM has a harder time of freeing pages,
1999                  * with multiple processes reclaiming pages, the total
2000                  * freeing target can get unreasonably large.
2001                  */
2002                 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
2003                         break;
2004         }
2005         sc->nr_reclaimed += nr_reclaimed;
2006
2007         /*
2008          * Even if we did not try to evict anon pages at all, we want to
2009          * rebalance the anon lru active/inactive ratio.
2010          */
2011         if (inactive_anon_is_low(zone, sc))
2012                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
2013
2014         /* reclaim/compaction might need reclaim to continue */
2015         if (should_continue_reclaim(zone, nr_reclaimed,
2016                                         sc->nr_scanned - nr_scanned, sc))
2017                 goto restart;
2018
2019         throttle_vm_writeout(sc->gfp_mask);
2020 }
2021
2022 /*
2023  * This is the direct reclaim path, for page-allocating processes.  We only
2024  * try to reclaim pages from zones which will satisfy the caller's allocation
2025  * request.
2026  *
2027  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2028  * Because:
2029  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2030  *    allocation or
2031  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2032  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2033  *    zone defense algorithm.
2034  *
2035  * If a zone is deemed to be full of pinned pages then just give it a light
2036  * scan then give up on it.
2037  */
2038 static void shrink_zones(int priority, struct zonelist *zonelist,
2039                                         struct scan_control *sc)
2040 {
2041         struct zoneref *z;
2042         struct zone *zone;
2043         unsigned long nr_soft_reclaimed;
2044         unsigned long nr_soft_scanned;
2045
2046         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2047                                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2048                 if (!populated_zone(zone))
2049                         continue;
2050                 /*
2051                  * Take care memory controller reclaiming has small influence
2052                  * to global LRU.
2053                  */
2054                 if (scanning_global_lru(sc)) {
2055                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2056                                 continue;
2057                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2058                                 continue;       /* Let kswapd poll it */
2059                         /*
2060                          * This steals pages from memory cgroups over softlimit
2061                          * and returns the number of reclaimed pages and
2062                          * scanned pages. This works for global memory pressure
2063                          * and balancing, not for a memcg's limit.
2064                          */
2065                         nr_soft_scanned = 0;
2066                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2067                                                 sc->order, sc->gfp_mask,
2068                                                 &nr_soft_scanned);
2069                         sc->nr_reclaimed += nr_soft_reclaimed;
2070                         sc->nr_scanned += nr_soft_scanned;
2071                         /* need some check for avoid more shrink_zone() */
2072                 }
2073
2074                 shrink_zone(priority, zone, sc);
2075         }
2076 }
2077
2078 static bool zone_reclaimable(struct zone *zone)
2079 {
2080         return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2081 }
2082
2083 /* All zones in zonelist are unreclaimable? */
2084 static bool all_unreclaimable(struct zonelist *zonelist,
2085                 struct scan_control *sc)
2086 {
2087         struct zoneref *z;
2088         struct zone *zone;
2089
2090         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2091                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2092                 if (!populated_zone(zone))
2093                         continue;
2094                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2095                         continue;
2096                 if (!zone->all_unreclaimable)
2097                         return false;
2098         }
2099
2100         return true;
2101 }
2102
2103 /*
2104  * This is the main entry point to direct page reclaim.
2105  *
2106  * If a full scan of the inactive list fails to free enough memory then we
2107  * are "out of memory" and something needs to be killed.
2108  *
2109  * If the caller is !__GFP_FS then the probability of a failure is reasonably
2110  * high - the zone may be full of dirty or under-writeback pages, which this
2111  * caller can't do much about.  We kick the writeback threads and take explicit
2112  * naps in the hope that some of these pages can be written.  But if the
2113  * allocating task holds filesystem locks which prevent writeout this might not
2114  * work, and the allocation attempt will fail.
2115  *
2116  * returns:     0, if no pages reclaimed
2117  *              else, the number of pages reclaimed
2118  */
2119 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2120                                         struct scan_control *sc,
2121                                         struct shrink_control *shrink)
2122 {
2123         int priority;
2124         unsigned long total_scanned = 0;
2125         struct reclaim_state *reclaim_state = current->reclaim_state;
2126         struct zoneref *z;
2127         struct zone *zone;
2128         unsigned long writeback_threshold;
2129
2130         get_mems_allowed();
2131         delayacct_freepages_start();
2132
2133         if (scanning_global_lru(sc))
2134                 count_vm_event(ALLOCSTALL);
2135
2136         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2137                 sc->nr_scanned = 0;
2138                 if (!priority)
2139                         disable_swap_token(sc->mem_cgroup);
2140                 shrink_zones(priority, zonelist, sc);
2141                 /*
2142                  * Don't shrink slabs when reclaiming memory from
2143                  * over limit cgroups
2144                  */
2145                 if (scanning_global_lru(sc)) {
2146                         unsigned long lru_pages = 0;
2147                         for_each_zone_zonelist(zone, z, zonelist,
2148                                         gfp_zone(sc->gfp_mask)) {
2149                                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2150                                         continue;
2151
2152                                 lru_pages += zone_reclaimable_pages(zone);
2153                         }
2154
2155                         shrink_slab(shrink, sc->nr_scanned, lru_pages);
2156                         if (reclaim_state) {
2157                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2158                                 reclaim_state->reclaimed_slab = 0;
2159                         }
2160                 }
2161                 total_scanned += sc->nr_scanned;
2162                 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2163                         goto out;
2164
2165                 /*
2166                  * Try to write back as many pages as we just scanned.  This
2167                  * tends to cause slow streaming writers to write data to the
2168                  * disk smoothly, at the dirtying rate, which is nice.   But
2169                  * that's undesirable in laptop mode, where we *want* lumpy
2170                  * writeout.  So in laptop mode, write out the whole world.
2171                  */
2172                 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2173                 if (total_scanned > writeback_threshold) {
2174                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
2175                         sc->may_writepage = 1;
2176                 }
2177
2178                 /* Take a nap, wait for some writeback to complete */
2179                 if (!sc->hibernation_mode && sc->nr_scanned &&
2180                     priority < DEF_PRIORITY - 2) {
2181                         struct zone *preferred_zone;
2182
2183                         first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2184                                                 &cpuset_current_mems_allowed,
2185                                                 &preferred_zone);
2186                         wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2187                 }
2188         }
2189
2190 out:
2191         delayacct_freepages_end();
2192         put_mems_allowed();
2193
2194         if (sc->nr_reclaimed)
2195                 return sc->nr_reclaimed;
2196
2197         /*
2198          * As hibernation is going on, kswapd is freezed so that it can't mark
2199          * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2200          * check.
2201          */
2202         if (oom_killer_disabled)
2203                 return 0;
2204
2205         /* top priority shrink_zones still had more to do? don't OOM, then */
2206         if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
2207                 return 1;
2208
2209         return 0;
2210 }
2211
2212 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2213                                 gfp_t gfp_mask, nodemask_t *nodemask)
2214 {
2215         unsigned long nr_reclaimed;
2216         struct scan_control sc = {
2217                 .gfp_mask = gfp_mask,
2218                 .may_writepage = !laptop_mode,
2219                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2220                 .may_unmap = 1,
2221                 .may_swap = 1,
2222                 .swappiness = vm_swappiness,
2223                 .order = order,
2224                 .mem_cgroup = NULL,
2225                 .nodemask = nodemask,
2226         };
2227         struct shrink_control shrink = {
2228                 .gfp_mask = sc.gfp_mask,
2229         };
2230
2231         trace_mm_vmscan_direct_reclaim_begin(order,
2232                                 sc.may_writepage,
2233                                 gfp_mask);
2234
2235         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2236
2237         trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2238
2239         return nr_reclaimed;
2240 }
2241
2242 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2243
2244 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2245                                                 gfp_t gfp_mask, bool noswap,
2246                                                 unsigned int swappiness,
2247                                                 struct zone *zone,
2248                                                 unsigned long *nr_scanned)
2249 {
2250         struct scan_control sc = {
2251                 .nr_scanned = 0,
2252                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2253                 .may_writepage = !laptop_mode,
2254                 .may_unmap = 1,
2255                 .may_swap = !noswap,
2256                 .swappiness = swappiness,
2257                 .order = 0,
2258                 .mem_cgroup = mem,
2259         };
2260
2261         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2262                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2263
2264         trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2265                                                       sc.may_writepage,
2266                                                       sc.gfp_mask);
2267
2268         /*
2269          * NOTE: Although we can get the priority field, using it
2270          * here is not a good idea, since it limits the pages we can scan.
2271          * if we don't reclaim here, the shrink_zone from balance_pgdat
2272          * will pick up pages from other mem cgroup's as well. We hack
2273          * the priority and make it zero.
2274          */
2275         shrink_zone(0, zone, &sc);
2276
2277         trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2278
2279         *nr_scanned = sc.nr_scanned;
2280         return sc.nr_reclaimed;
2281 }
2282
2283 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
2284                                            gfp_t gfp_mask,
2285                                            bool noswap,
2286                                            unsigned int swappiness)
2287 {
2288         struct zonelist *zonelist;
2289         unsigned long nr_reclaimed;
2290         int nid;
2291         struct scan_control sc = {
2292                 .may_writepage = !laptop_mode,
2293                 .may_unmap = 1,
2294                 .may_swap = !noswap,
2295                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2296                 .swappiness = swappiness,
2297                 .order = 0,
2298                 .mem_cgroup = mem_cont,
2299                 .nodemask = NULL, /* we don't care the placement */
2300                 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2301                                 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2302         };
2303         struct shrink_control shrink = {
2304                 .gfp_mask = sc.gfp_mask,
2305         };
2306
2307         /*
2308          * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2309          * take care of from where we get pages. So the node where we start the
2310          * scan does not need to be the current node.
2311          */
2312         nid = mem_cgroup_select_victim_node(mem_cont);
2313
2314         zonelist = NODE_DATA(nid)->node_zonelists;
2315
2316         trace_mm_vmscan_memcg_reclaim_begin(0,
2317                                             sc.may_writepage,
2318                                             sc.gfp_mask);
2319
2320         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2321
2322         trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2323
2324         return nr_reclaimed;
2325 }
2326 #endif
2327
2328 /*
2329  * pgdat_balanced is used when checking if a node is balanced for high-order
2330  * allocations. Only zones that meet watermarks and are in a zone allowed
2331  * by the callers classzone_idx are added to balanced_pages. The total of
2332  * balanced pages must be at least 25% of the zones allowed by classzone_idx
2333  * for the node to be considered balanced. Forcing all zones to be balanced
2334  * for high orders can cause excessive reclaim when there are imbalanced zones.
2335  * The choice of 25% is due to
2336  *   o a 16M DMA zone that is balanced will not balance a zone on any
2337  *     reasonable sized machine
2338  *   o On all other machines, the top zone must be at least a reasonable
2339  *     percentage of the middle zones. For example, on 32-bit x86, highmem
2340  *     would need to be at least 256M for it to be balance a whole node.
2341  *     Similarly, on x86-64 the Normal zone would need to be at least 1G
2342  *     to balance a node on its own. These seemed like reasonable ratios.
2343  */
2344 static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2345                                                 int classzone_idx)
2346 {
2347         unsigned long present_pages = 0;
2348         int i;
2349
2350         for (i = 0; i <= classzone_idx; i++)
2351                 present_pages += pgdat->node_zones[i].present_pages;
2352
2353         /* A special case here: if zone has no page, we think it's balanced */
2354         return balanced_pages >= (present_pages >> 2);
2355 }
2356
2357 /* is kswapd sleeping prematurely? */
2358 static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2359                                         int classzone_idx)
2360 {
2361         int i;
2362         unsigned long balanced = 0;
2363         bool all_zones_ok = true;
2364
2365         /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2366         if (remaining)
2367                 return true;
2368
2369         /* Check the watermark levels */
2370         for (i = 0; i <= classzone_idx; i++) {
2371                 struct zone *zone = pgdat->node_zones + i;
2372
2373                 if (!populated_zone(zone))
2374                         continue;
2375
2376                 /*
2377                  * balance_pgdat() skips over all_unreclaimable after
2378                  * DEF_PRIORITY. Effectively, it considers them balanced so
2379                  * they must be considered balanced here as well if kswapd
2380                  * is to sleep
2381                  */
2382                 if (zone->all_unreclaimable) {
2383                         balanced += zone->present_pages;
2384                         continue;
2385                 }
2386
2387                 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2388                                                         i, 0))
2389                         all_zones_ok = false;
2390                 else
2391                         balanced += zone->present_pages;
2392         }
2393
2394         /*
2395          * For high-order requests, the balanced zones must contain at least
2396          * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2397          * must be balanced
2398          */
2399         if (order)
2400                 return !pgdat_balanced(pgdat, balanced, classzone_idx);
2401         else
2402                 return !all_zones_ok;
2403 }
2404
2405 /*
2406  * For kswapd, balance_pgdat() will work across all this node's zones until
2407  * they are all at high_wmark_pages(zone).
2408  *
2409  * Returns the final order kswapd was reclaiming at
2410  *
2411  * There is special handling here for zones which are full of pinned pages.
2412  * This can happen if the pages are all mlocked, or if they are all used by
2413  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
2414  * What we do is to detect the case where all pages in the zone have been
2415  * scanned twice and there has been zero successful reclaim.  Mark the zone as
2416  * dead and from now on, only perform a short scan.  Basically we're polling
2417  * the zone for when the problem goes away.
2418  *
2419  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
2420  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2421  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2422  * lower zones regardless of the number of free pages in the lower zones. This
2423  * interoperates with the page allocator fallback scheme to ensure that aging
2424  * of pages is balanced across the zones.
2425  */
2426 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2427                                                         int *classzone_idx)
2428 {
2429         int all_zones_ok;
2430         unsigned long balanced;
2431         int priority;
2432         int i;
2433         int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
2434         unsigned long total_scanned;
2435         struct reclaim_state *reclaim_state = current->reclaim_state;
2436         unsigned long nr_soft_reclaimed;
2437         unsigned long nr_soft_scanned;
2438         struct scan_control sc = {
2439                 .gfp_mask = GFP_KERNEL,
2440                 .may_unmap = 1,
2441                 .may_swap = 1,
2442                 /*
2443                  * kswapd doesn't want to be bailed out while reclaim. because
2444                  * we want to put equal scanning pressure on each zone.
2445                  */
2446                 .nr_to_reclaim = ULONG_MAX,
2447                 .swappiness = vm_swappiness,
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                 .swappiness = vm_swappiness,
2918                 .order = 0,
2919         };
2920         struct shrink_control shrink = {
2921                 .gfp_mask = sc.gfp_mask,
2922         };
2923         struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2924         struct task_struct *p = current;
2925         unsigned long nr_reclaimed;
2926
2927         p->flags |= PF_MEMALLOC;
2928         lockdep_set_current_reclaim_state(sc.gfp_mask);
2929         reclaim_state.reclaimed_slab = 0;
2930         p->reclaim_state = &reclaim_state;
2931
2932         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2933
2934         p->reclaim_state = NULL;
2935         lockdep_clear_current_reclaim_state();
2936         p->flags &= ~PF_MEMALLOC;
2937
2938         return nr_reclaimed;
2939 }
2940 #endif /* CONFIG_HIBERNATION */
2941
2942 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2943    not required for correctness.  So if the last cpu in a node goes
2944    away, we get changed to run anywhere: as the first one comes back,
2945    restore their cpu bindings. */
2946 static int __devinit cpu_callback(struct notifier_block *nfb,
2947                                   unsigned long action, void *hcpu)
2948 {
2949         int nid;
2950
2951         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2952                 for_each_node_state(nid, N_HIGH_MEMORY) {
2953                         pg_data_t *pgdat = NODE_DATA(nid);
2954                         const struct cpumask *mask;
2955
2956                         mask = cpumask_of_node(pgdat->node_id);
2957
2958                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2959                                 /* One of our CPUs online: restore mask */
2960                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2961                 }
2962         }
2963         return NOTIFY_OK;
2964 }
2965
2966 /*
2967  * This kswapd start function will be called by init and node-hot-add.
2968  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2969  */
2970 int kswapd_run(int nid)
2971 {
2972         pg_data_t *pgdat = NODE_DATA(nid);
2973         int ret = 0;
2974
2975         if (pgdat->kswapd)
2976                 return 0;
2977
2978         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2979         if (IS_ERR(pgdat->kswapd)) {
2980                 /* failure at boot is fatal */
2981                 BUG_ON(system_state == SYSTEM_BOOTING);
2982                 printk("Failed to start kswapd on node %d\n",nid);
2983                 ret = -1;
2984         }
2985         return ret;
2986 }
2987
2988 /*
2989  * Called by memory hotplug when all memory in a node is offlined.
2990  */
2991 void kswapd_stop(int nid)
2992 {
2993         struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
2994
2995         if (kswapd)
2996                 kthread_stop(kswapd);
2997 }
2998
2999 static int __init kswapd_init(void)
3000 {
3001         int nid;
3002
3003         swap_setup();
3004         for_each_node_state(nid, N_HIGH_MEMORY)
3005                 kswapd_run(nid);
3006         hotcpu_notifier(cpu_callback, 0);
3007         return 0;
3008 }
3009
3010 module_init(kswapd_init)
3011
3012 #ifdef CONFIG_NUMA
3013 /*
3014  * Zone reclaim mode
3015  *
3016  * If non-zero call zone_reclaim when the number of free pages falls below
3017  * the watermarks.
3018  */
3019 int zone_reclaim_mode __read_mostly;
3020
3021 #define RECLAIM_OFF 0
3022 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
3023 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
3024 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
3025
3026 /*
3027  * Priority for ZONE_RECLAIM. This determines the fraction of pages
3028  * of a node considered for each zone_reclaim. 4 scans 1/16th of
3029  * a zone.
3030  */
3031 #define ZONE_RECLAIM_PRIORITY 4
3032
3033 /*
3034  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3035  * occur.
3036  */
3037 int sysctl_min_unmapped_ratio = 1;
3038
3039 /*
3040  * If the number of slab pages in a zone grows beyond this percentage then
3041  * slab reclaim needs to occur.
3042  */
3043 int sysctl_min_slab_ratio = 5;
3044
3045 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3046 {
3047         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3048         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3049                 zone_page_state(zone, NR_ACTIVE_FILE);
3050
3051         /*
3052          * It's possible for there to be more file mapped pages than
3053          * accounted for by the pages on the file LRU lists because
3054          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3055          */
3056         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3057 }
3058
3059 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3060 static long zone_pagecache_reclaimable(struct zone *zone)
3061 {
3062         long nr_pagecache_reclaimable;
3063         long delta = 0;
3064
3065         /*
3066          * If RECLAIM_SWAP is set, then all file pages are considered
3067          * potentially reclaimable. Otherwise, we have to worry about
3068          * pages like swapcache and zone_unmapped_file_pages() provides
3069          * a better estimate
3070          */
3071         if (zone_reclaim_mode & RECLAIM_SWAP)
3072                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3073         else
3074                 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3075
3076         /* If we can't clean pages, remove dirty pages from consideration */
3077         if (!(zone_reclaim_mode & RECLAIM_WRITE))
3078                 delta += zone_page_state(zone, NR_FILE_DIRTY);
3079
3080         /* Watch for any possible underflows due to delta */
3081         if (unlikely(delta > nr_pagecache_reclaimable))
3082                 delta = nr_pagecache_reclaimable;
3083
3084         return nr_pagecache_reclaimable - delta;
3085 }
3086
3087 /*
3088  * Try to free up some pages from this zone through reclaim.
3089  */
3090 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3091 {
3092         /* Minimum pages needed in order to stay on node */
3093         const unsigned long nr_pages = 1 << order;
3094         struct task_struct *p = current;
3095         struct reclaim_state reclaim_state;
3096         int priority;
3097         struct scan_control sc = {
3098                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3099                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3100                 .may_swap = 1,
3101                 .nr_to_reclaim = max_t(unsigned long, nr_pages,
3102                                        SWAP_CLUSTER_MAX),
3103                 .gfp_mask = gfp_mask,
3104                 .swappiness = vm_swappiness,
3105                 .order = order,
3106         };
3107         struct shrink_control shrink = {
3108                 .gfp_mask = sc.gfp_mask,
3109         };
3110         unsigned long nr_slab_pages0, nr_slab_pages1;
3111
3112         cond_resched();
3113         /*
3114          * We need to be able to allocate from the reserves for RECLAIM_SWAP
3115          * and we also need to be able to write out pages for RECLAIM_WRITE
3116          * and RECLAIM_SWAP.
3117          */
3118         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3119         lockdep_set_current_reclaim_state(gfp_mask);
3120         reclaim_state.reclaimed_slab = 0;
3121         p->reclaim_state = &reclaim_state;
3122
3123         if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3124                 /*
3125                  * Free memory by calling shrink zone with increasing
3126                  * priorities until we have enough memory freed.
3127                  */
3128                 priority = ZONE_RECLAIM_PRIORITY;
3129                 do {
3130                         shrink_zone(priority, zone, &sc);
3131                         priority--;
3132                 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
3133         }
3134
3135         nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3136         if (nr_slab_pages0 > zone->min_slab_pages) {
3137                 /*
3138                  * shrink_slab() does not currently allow us to determine how
3139                  * many pages were freed in this zone. So we take the current
3140                  * number of slab pages and shake the slab until it is reduced
3141                  * by the same nr_pages that we used for reclaiming unmapped
3142                  * pages.
3143                  *
3144                  * Note that shrink_slab will free memory on all zones and may
3145                  * take a long time.
3146                  */
3147                 for (;;) {
3148                         unsigned long lru_pages = zone_reclaimable_pages(zone);
3149
3150                         /* No reclaimable slab or very low memory pressure */
3151                         if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3152                                 break;
3153
3154                         /* Freed enough memory */
3155                         nr_slab_pages1 = zone_page_state(zone,
3156                                                         NR_SLAB_RECLAIMABLE);
3157                         if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3158                                 break;
3159                 }
3160
3161                 /*
3162                  * Update nr_reclaimed by the number of slab pages we
3163                  * reclaimed from this zone.
3164                  */
3165                 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3166                 if (nr_slab_pages1 < nr_slab_pages0)
3167                         sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3168         }
3169
3170         p->reclaim_state = NULL;
3171         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3172         lockdep_clear_current_reclaim_state();
3173         return sc.nr_reclaimed >= nr_pages;
3174 }
3175
3176 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3177 {
3178         int node_id;
3179         int ret;
3180
3181         /*
3182          * Zone reclaim reclaims unmapped file backed pages and
3183          * slab pages if we are over the defined limits.
3184          *
3185          * A small portion of unmapped file backed pages is needed for
3186          * file I/O otherwise pages read by file I/O will be immediately
3187          * thrown out if the zone is overallocated. So we do not reclaim
3188          * if less than a specified percentage of the zone is used by
3189          * unmapped file backed pages.
3190          */
3191         if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3192             zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3193                 return ZONE_RECLAIM_FULL;
3194
3195         if (zone->all_unreclaimable)
3196                 return ZONE_RECLAIM_FULL;
3197
3198         /*
3199          * Do not scan if the allocation should not be delayed.
3200          */
3201         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3202                 return ZONE_RECLAIM_NOSCAN;
3203
3204         /*
3205          * Only run zone reclaim on the local zone or on zones that do not
3206          * have associated processors. This will favor the local processor
3207          * over remote processors and spread off node memory allocations
3208          * as wide as possible.
3209          */
3210         node_id = zone_to_nid(zone);
3211         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3212                 return ZONE_RECLAIM_NOSCAN;
3213
3214         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3215                 return ZONE_RECLAIM_NOSCAN;
3216
3217         ret = __zone_reclaim(zone, gfp_mask, order);
3218         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3219
3220         if (!ret)
3221                 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3222
3223         return ret;
3224 }
3225 #endif
3226
3227 /*
3228  * page_evictable - test whether a page is evictable
3229  * @page: the page to test
3230  * @vma: the VMA in which the page is or will be mapped, may be NULL
3231  *
3232  * Test whether page is evictable--i.e., should be placed on active/inactive
3233  * lists vs unevictable list.  The vma argument is !NULL when called from the
3234  * fault path to determine how to instantate a new page.
3235  *
3236  * Reasons page might not be evictable:
3237  * (1) page's mapping marked unevictable
3238  * (2) page is part of an mlocked VMA
3239  *
3240  */
3241 int page_evictable(struct page *page, struct vm_area_struct *vma)
3242 {
3243
3244         if (mapping_unevictable(page_mapping(page)))
3245                 return 0;
3246
3247         if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
3248                 return 0;
3249
3250         return 1;
3251 }
3252
3253 /**
3254  * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
3255  * @page: page to check evictability and move to appropriate lru list
3256  * @zone: zone page is in
3257  *
3258  * Checks a page for evictability and moves the page to the appropriate
3259  * zone lru list.
3260  *
3261  * Restrictions: zone->lru_lock must be held, page must be on LRU and must
3262  * have PageUnevictable set.
3263  */
3264 static void check_move_unevictable_page(struct page *page, struct zone *zone)
3265 {
3266         VM_BUG_ON(PageActive(page));
3267
3268 retry:
3269         ClearPageUnevictable(page);
3270         if (page_evictable(page, NULL)) {
3271                 enum lru_list l = page_lru_base_type(page);
3272
3273                 __dec_zone_state(zone, NR_UNEVICTABLE);
3274                 list_move(&page->lru, &zone->lru[l].list);
3275                 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
3276                 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
3277                 __count_vm_event(UNEVICTABLE_PGRESCUED);
3278         } else {
3279                 /*
3280                  * rotate unevictable list
3281                  */
3282                 SetPageUnevictable(page);
3283                 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
3284                 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
3285                 if (page_evictable(page, NULL))
3286                         goto retry;
3287         }
3288 }
3289
3290 /**
3291  * scan_mapping_unevictable_pages - scan an address space for evictable pages
3292  * @mapping: struct address_space to scan for evictable pages
3293  *
3294  * Scan all pages in mapping.  Check unevictable pages for
3295  * evictability and move them to the appropriate zone lru list.
3296  */
3297 void scan_mapping_unevictable_pages(struct address_space *mapping)
3298 {
3299         pgoff_t next = 0;
3300         pgoff_t end   = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
3301                          PAGE_CACHE_SHIFT;
3302         struct zone *zone;
3303         struct pagevec pvec;
3304
3305         if (mapping->nrpages == 0)
3306                 return;
3307
3308         pagevec_init(&pvec, 0);
3309         while (next < end &&
3310                 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
3311                 int i;
3312                 int pg_scanned = 0;
3313
3314                 zone = NULL;
3315
3316                 for (i = 0; i < pagevec_count(&pvec); i++) {
3317                         struct page *page = pvec.pages[i];
3318                         pgoff_t page_index = page->index;
3319                         struct zone *pagezone = page_zone(page);
3320
3321                         pg_scanned++;
3322                         if (page_index > next)
3323                                 next = page_index;
3324                         next++;
3325
3326                         if (pagezone != zone) {
3327                                 if (zone)
3328                                         spin_unlock_irq(&zone->lru_lock);
3329                                 zone = pagezone;
3330                                 spin_lock_irq(&zone->lru_lock);
3331                         }
3332
3333                         if (PageLRU(page) && PageUnevictable(page))
3334                                 check_move_unevictable_page(page, zone);
3335                 }
3336                 if (zone)
3337                         spin_unlock_irq(&zone->lru_lock);
3338                 pagevec_release(&pvec);
3339
3340                 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
3341         }
3342
3343 }
3344
3345 /**
3346  * scan_zone_unevictable_pages - check unevictable list for evictable pages
3347  * @zone - zone of which to scan the unevictable list
3348  *
3349  * Scan @zone's unevictable LRU lists to check for pages that have become
3350  * evictable.  Move those that have to @zone's inactive list where they
3351  * become candidates for reclaim, unless shrink_inactive_zone() decides
3352  * to reactivate them.  Pages that are still unevictable are rotated
3353  * back onto @zone's unevictable list.
3354  */
3355 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
3356 static void scan_zone_unevictable_pages(struct zone *zone)
3357 {
3358         struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
3359         unsigned long scan;
3360         unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
3361
3362         while (nr_to_scan > 0) {
3363                 unsigned long batch_size = min(nr_to_scan,
3364                                                 SCAN_UNEVICTABLE_BATCH_SIZE);
3365
3366                 spin_lock_irq(&zone->lru_lock);
3367                 for (scan = 0;  scan < batch_size; scan++) {
3368                         struct page *page = lru_to_page(l_unevictable);
3369
3370                         if (!trylock_page(page))
3371                                 continue;
3372
3373                         prefetchw_prev_lru_page(page, l_unevictable, flags);
3374
3375                         if (likely(PageLRU(page) && PageUnevictable(page)))
3376                                 check_move_unevictable_page(page, zone);
3377
3378                         unlock_page(page);
3379                 }
3380                 spin_unlock_irq(&zone->lru_lock);
3381
3382                 nr_to_scan -= batch_size;
3383         }
3384 }
3385
3386
3387 /**
3388  * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
3389  *
3390  * A really big hammer:  scan all zones' unevictable LRU lists to check for
3391  * pages that have become evictable.  Move those back to the zones'
3392  * inactive list where they become candidates for reclaim.
3393  * This occurs when, e.g., we have unswappable pages on the unevictable lists,
3394  * and we add swap to the system.  As such, it runs in the context of a task
3395  * that has possibly/probably made some previously unevictable pages
3396  * evictable.
3397  */
3398 static void scan_all_zones_unevictable_pages(void)
3399 {
3400         struct zone *zone;
3401
3402         for_each_zone(zone) {
3403                 scan_zone_unevictable_pages(zone);
3404         }
3405 }
3406
3407 /*
3408  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
3409  * all nodes' unevictable lists for evictable pages
3410  */
3411 unsigned long scan_unevictable_pages;
3412
3413 int scan_unevictable_handler(struct ctl_table *table, int write,
3414                            void __user *buffer,
3415                            size_t *length, loff_t *ppos)
3416 {
3417         proc_doulongvec_minmax(table, write, buffer, length, ppos);
3418
3419         if (write && *(unsigned long *)table->data)
3420                 scan_all_zones_unevictable_pages();
3421
3422         scan_unevictable_pages = 0;
3423         return 0;
3424 }
3425
3426 #ifdef CONFIG_NUMA
3427 /*
3428  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
3429  * a specified node's per zone unevictable lists for evictable pages.
3430  */
3431
3432 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
3433                                           struct sysdev_attribute *attr,
3434                                           char *buf)
3435 {
3436         return sprintf(buf, "0\n");     /* always zero; should fit... */
3437 }
3438
3439 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
3440                                            struct sysdev_attribute *attr,
3441                                         const char *buf, size_t count)
3442 {
3443         struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
3444         struct zone *zone;
3445         unsigned long res;
3446         unsigned long req = strict_strtoul(buf, 10, &res);
3447
3448         if (!req)
3449                 return 1;       /* zero is no-op */
3450
3451         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
3452                 if (!populated_zone(zone))
3453                         continue;
3454                 scan_zone_unevictable_pages(zone);
3455         }
3456         return 1;
3457 }
3458
3459
3460 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3461                         read_scan_unevictable_node,
3462                         write_scan_unevictable_node);
3463
3464 int scan_unevictable_register_node(struct node *node)
3465 {
3466         return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
3467 }
3468
3469 void scan_unevictable_unregister_node(struct node *node)
3470 {
3471         sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
3472 }
3473 #endif