mm: vmscan: fix typo in isolating lru pages
[linux-2.6.git] / mm / vmscan.c
1 /*
2  *  linux/mm/vmscan.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  *
6  *  Swap reorganised 29.12.95, Stephen Tweedie.
7  *  kswapd added: 7.1.96  sct
8  *  Removed kswapd_ctl limits, and swap out as many pages as needed
9  *  to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10  *  Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11  *  Multiqueue VM started 5.8.00, Rik van Riel.
12  */
13
14 #include <linux/mm.h>
15 #include <linux/module.h>
16 #include <linux/gfp.h>
17 #include <linux/kernel_stat.h>
18 #include <linux/swap.h>
19 #include <linux/pagemap.h>
20 #include <linux/init.h>
21 #include <linux/highmem.h>
22 #include <linux/vmstat.h>
23 #include <linux/file.h>
24 #include <linux/writeback.h>
25 #include <linux/blkdev.h>
26 #include <linux/buffer_head.h>  /* for try_to_release_page(),
27                                         buffer_heads_over_limit */
28 #include <linux/mm_inline.h>
29 #include <linux/pagevec.h>
30 #include <linux/backing-dev.h>
31 #include <linux/rmap.h>
32 #include <linux/topology.h>
33 #include <linux/cpu.h>
34 #include <linux/cpuset.h>
35 #include <linux/compaction.h>
36 #include <linux/notifier.h>
37 #include <linux/rwsem.h>
38 #include <linux/delay.h>
39 #include <linux/kthread.h>
40 #include <linux/freezer.h>
41 #include <linux/memcontrol.h>
42 #include <linux/delayacct.h>
43 #include <linux/sysctl.h>
44 #include <linux/oom.h>
45 #include <linux/prefetch.h>
46
47 #include <asm/tlbflush.h>
48 #include <asm/div64.h>
49
50 #include <linux/swapops.h>
51
52 #include "internal.h"
53
54 #define CREATE_TRACE_POINTS
55 #include <trace/events/vmscan.h>
56
57 /*
58  * reclaim_mode determines how the inactive list is shrunk
59  * RECLAIM_MODE_SINGLE: Reclaim only order-0 pages
60  * RECLAIM_MODE_ASYNC:  Do not block
61  * RECLAIM_MODE_SYNC:   Allow blocking e.g. call wait_on_page_writeback
62  * RECLAIM_MODE_LUMPYRECLAIM: For high-order allocations, take a reference
63  *                      page from the LRU and reclaim all pages within a
64  *                      naturally aligned range
65  * RECLAIM_MODE_COMPACTION: For high-order allocations, reclaim a number of
66  *                      order-0 pages and then compact the zone
67  */
68 typedef unsigned __bitwise__ reclaim_mode_t;
69 #define RECLAIM_MODE_SINGLE             ((__force reclaim_mode_t)0x01u)
70 #define RECLAIM_MODE_ASYNC              ((__force reclaim_mode_t)0x02u)
71 #define RECLAIM_MODE_SYNC               ((__force reclaim_mode_t)0x04u)
72 #define RECLAIM_MODE_LUMPYRECLAIM       ((__force reclaim_mode_t)0x08u)
73 #define RECLAIM_MODE_COMPACTION         ((__force reclaim_mode_t)0x10u)
74
75 struct scan_control {
76         /* Incremented by the number of inactive pages that were scanned */
77         unsigned long nr_scanned;
78
79         /* Number of pages freed so far during a call to shrink_zones() */
80         unsigned long nr_reclaimed;
81
82         /* How many pages shrink_list() should reclaim */
83         unsigned long nr_to_reclaim;
84
85         unsigned long hibernation_mode;
86
87         /* This context's GFP mask */
88         gfp_t gfp_mask;
89
90         int may_writepage;
91
92         /* Can mapped pages be reclaimed? */
93         int may_unmap;
94
95         /* Can pages be swapped as part of reclaim? */
96         int may_swap;
97
98         int order;
99
100         /*
101          * Intend to reclaim enough continuous memory rather than reclaim
102          * enough amount of memory. i.e, mode for high order allocation.
103          */
104         reclaim_mode_t reclaim_mode;
105
106         /* Which cgroup do we reclaim from */
107         struct mem_cgroup *mem_cgroup;
108
109         /*
110          * Nodemask of nodes allowed by the caller. If NULL, all nodes
111          * are scanned.
112          */
113         nodemask_t      *nodemask;
114 };
115
116 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
117
118 #ifdef ARCH_HAS_PREFETCH
119 #define prefetch_prev_lru_page(_page, _base, _field)                    \
120         do {                                                            \
121                 if ((_page)->lru.prev != _base) {                       \
122                         struct page *prev;                              \
123                                                                         \
124                         prev = lru_to_page(&(_page->lru));              \
125                         prefetch(&prev->_field);                        \
126                 }                                                       \
127         } while (0)
128 #else
129 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
130 #endif
131
132 #ifdef ARCH_HAS_PREFETCHW
133 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
134         do {                                                            \
135                 if ((_page)->lru.prev != _base) {                       \
136                         struct page *prev;                              \
137                                                                         \
138                         prev = lru_to_page(&(_page->lru));              \
139                         prefetchw(&prev->_field);                       \
140                 }                                                       \
141         } while (0)
142 #else
143 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
144 #endif
145
146 /*
147  * From 0 .. 100.  Higher means more swappy.
148  */
149 int vm_swappiness = 60;
150 long vm_total_pages;    /* The total number of pages which the VM controls */
151
152 static LIST_HEAD(shrinker_list);
153 static DECLARE_RWSEM(shrinker_rwsem);
154
155 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
156 #define scanning_global_lru(sc) (!(sc)->mem_cgroup)
157 #else
158 #define scanning_global_lru(sc) (1)
159 #endif
160
161 static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
162                                                   struct scan_control *sc)
163 {
164         if (!scanning_global_lru(sc))
165                 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
166
167         return &zone->reclaim_stat;
168 }
169
170 static unsigned long zone_nr_lru_pages(struct zone *zone,
171                                 struct scan_control *sc, enum lru_list lru)
172 {
173         if (!scanning_global_lru(sc))
174                 return mem_cgroup_zone_nr_lru_pages(sc->mem_cgroup,
175                                 zone_to_nid(zone), zone_idx(zone), BIT(lru));
176
177         return zone_page_state(zone, NR_LRU_BASE + lru);
178 }
179
180
181 /*
182  * Add a shrinker callback to be called from the vm
183  */
184 void register_shrinker(struct shrinker *shrinker)
185 {
186         atomic_long_set(&shrinker->nr_in_batch, 0);
187         down_write(&shrinker_rwsem);
188         list_add_tail(&shrinker->list, &shrinker_list);
189         up_write(&shrinker_rwsem);
190 }
191 EXPORT_SYMBOL(register_shrinker);
192
193 /*
194  * Remove one
195  */
196 void unregister_shrinker(struct shrinker *shrinker)
197 {
198         down_write(&shrinker_rwsem);
199         list_del(&shrinker->list);
200         up_write(&shrinker_rwsem);
201 }
202 EXPORT_SYMBOL(unregister_shrinker);
203
204 static inline int do_shrinker_shrink(struct shrinker *shrinker,
205                                      struct shrink_control *sc,
206                                      unsigned long nr_to_scan)
207 {
208         sc->nr_to_scan = nr_to_scan;
209         return (*shrinker->shrink)(shrinker, sc);
210 }
211
212 #define SHRINK_BATCH 128
213 /*
214  * Call the shrink functions to age shrinkable caches
215  *
216  * Here we assume it costs one seek to replace a lru page and that it also
217  * takes a seek to recreate a cache object.  With this in mind we age equal
218  * percentages of the lru and ageable caches.  This should balance the seeks
219  * generated by these structures.
220  *
221  * If the vm encountered mapped pages on the LRU it increase the pressure on
222  * slab to avoid swapping.
223  *
224  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
225  *
226  * `lru_pages' represents the number of on-LRU pages in all the zones which
227  * are eligible for the caller's allocation attempt.  It is used for balancing
228  * slab reclaim versus page reclaim.
229  *
230  * Returns the number of slab objects which we shrunk.
231  */
232 unsigned long shrink_slab(struct shrink_control *shrink,
233                           unsigned long nr_pages_scanned,
234                           unsigned long lru_pages)
235 {
236         struct shrinker *shrinker;
237         unsigned long ret = 0;
238
239         if (nr_pages_scanned == 0)
240                 nr_pages_scanned = SWAP_CLUSTER_MAX;
241
242         if (!down_read_trylock(&shrinker_rwsem)) {
243                 /* Assume we'll be able to shrink next time */
244                 ret = 1;
245                 goto out;
246         }
247
248         list_for_each_entry(shrinker, &shrinker_list, list) {
249                 unsigned long long delta;
250                 long total_scan;
251                 long max_pass;
252                 int shrink_ret = 0;
253                 long nr;
254                 long new_nr;
255                 long batch_size = shrinker->batch ? shrinker->batch
256                                                   : SHRINK_BATCH;
257
258                 max_pass = do_shrinker_shrink(shrinker, shrink, 0);
259                 if (max_pass <= 0)
260                         continue;
261
262                 /*
263                  * copy the current shrinker scan count into a local variable
264                  * and zero it so that other concurrent shrinker invocations
265                  * don't also do this scanning work.
266                  */
267                 nr = atomic_long_xchg(&shrinker->nr_in_batch, 0);
268
269                 total_scan = nr;
270                 delta = (4 * nr_pages_scanned) / shrinker->seeks;
271                 delta *= max_pass;
272                 do_div(delta, lru_pages + 1);
273                 total_scan += delta;
274                 if (total_scan < 0) {
275                         printk(KERN_ERR "shrink_slab: %pF negative objects to "
276                                "delete nr=%ld\n",
277                                shrinker->shrink, total_scan);
278                         total_scan = max_pass;
279                 }
280
281                 /*
282                  * We need to avoid excessive windup on filesystem shrinkers
283                  * due to large numbers of GFP_NOFS allocations causing the
284                  * shrinkers to return -1 all the time. This results in a large
285                  * nr being built up so when a shrink that can do some work
286                  * comes along it empties the entire cache due to nr >>>
287                  * max_pass.  This is bad for sustaining a working set in
288                  * memory.
289                  *
290                  * Hence only allow the shrinker to scan the entire cache when
291                  * a large delta change is calculated directly.
292                  */
293                 if (delta < max_pass / 4)
294                         total_scan = min(total_scan, max_pass / 2);
295
296                 /*
297                  * Avoid risking looping forever due to too large nr value:
298                  * never try to free more than twice the estimate number of
299                  * freeable entries.
300                  */
301                 if (total_scan > max_pass * 2)
302                         total_scan = max_pass * 2;
303
304                 trace_mm_shrink_slab_start(shrinker, shrink, nr,
305                                         nr_pages_scanned, lru_pages,
306                                         max_pass, delta, total_scan);
307
308                 while (total_scan >= batch_size) {
309                         int nr_before;
310
311                         nr_before = do_shrinker_shrink(shrinker, shrink, 0);
312                         shrink_ret = do_shrinker_shrink(shrinker, shrink,
313                                                         batch_size);
314                         if (shrink_ret == -1)
315                                 break;
316                         if (shrink_ret < nr_before)
317                                 ret += nr_before - shrink_ret;
318                         count_vm_events(SLABS_SCANNED, batch_size);
319                         total_scan -= batch_size;
320
321                         cond_resched();
322                 }
323
324                 /*
325                  * move the unused scan count back into the shrinker in a
326                  * manner that handles concurrent updates. If we exhausted the
327                  * scan, there is no need to do an update.
328                  */
329                 if (total_scan > 0)
330                         new_nr = atomic_long_add_return(total_scan,
331                                         &shrinker->nr_in_batch);
332                 else
333                         new_nr = atomic_long_read(&shrinker->nr_in_batch);
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                 if (!PageWriteback(page)) {
499                         /* synchronous write or broken a_ops? */
500                         ClearPageReclaim(page);
501                 }
502                 trace_mm_vmscan_writepage(page,
503                         trace_reclaim_flags(page, sc->reclaim_mode));
504                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
505                 return PAGE_SUCCESS;
506         }
507
508         return PAGE_CLEAN;
509 }
510
511 /*
512  * Same as remove_mapping, but if the page is removed from the mapping, it
513  * gets returned with a refcount of 0.
514  */
515 static int __remove_mapping(struct address_space *mapping, struct page *page)
516 {
517         BUG_ON(!PageLocked(page));
518         BUG_ON(mapping != page_mapping(page));
519
520         spin_lock_irq(&mapping->tree_lock);
521         /*
522          * The non racy check for a busy page.
523          *
524          * Must be careful with the order of the tests. When someone has
525          * a ref to the page, it may be possible that they dirty it then
526          * drop the reference. So if PageDirty is tested before page_count
527          * here, then the following race may occur:
528          *
529          * get_user_pages(&page);
530          * [user mapping goes away]
531          * write_to(page);
532          *                              !PageDirty(page)    [good]
533          * SetPageDirty(page);
534          * put_page(page);
535          *                              !page_count(page)   [good, discard it]
536          *
537          * [oops, our write_to data is lost]
538          *
539          * Reversing the order of the tests ensures such a situation cannot
540          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
541          * load is not satisfied before that of page->_count.
542          *
543          * Note that if SetPageDirty is always performed via set_page_dirty,
544          * and thus under tree_lock, then this ordering is not required.
545          */
546         if (!page_freeze_refs(page, 2))
547                 goto cannot_free;
548         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
549         if (unlikely(PageDirty(page))) {
550                 page_unfreeze_refs(page, 2);
551                 goto cannot_free;
552         }
553
554         if (PageSwapCache(page)) {
555                 swp_entry_t swap = { .val = page_private(page) };
556                 __delete_from_swap_cache(page);
557                 spin_unlock_irq(&mapping->tree_lock);
558                 swapcache_free(swap, page);
559         } else {
560                 void (*freepage)(struct page *);
561
562                 freepage = mapping->a_ops->freepage;
563
564                 __delete_from_page_cache(page);
565                 spin_unlock_irq(&mapping->tree_lock);
566                 mem_cgroup_uncharge_cache_page(page);
567
568                 if (freepage != NULL)
569                         freepage(page);
570         }
571
572         return 1;
573
574 cannot_free:
575         spin_unlock_irq(&mapping->tree_lock);
576         return 0;
577 }
578
579 /*
580  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
581  * someone else has a ref on the page, abort and return 0.  If it was
582  * successfully detached, return 1.  Assumes the caller has a single ref on
583  * this page.
584  */
585 int remove_mapping(struct address_space *mapping, struct page *page)
586 {
587         if (__remove_mapping(mapping, page)) {
588                 /*
589                  * Unfreezing the refcount with 1 rather than 2 effectively
590                  * drops the pagecache ref for us without requiring another
591                  * atomic operation.
592                  */
593                 page_unfreeze_refs(page, 1);
594                 return 1;
595         }
596         return 0;
597 }
598
599 /**
600  * putback_lru_page - put previously isolated page onto appropriate LRU list
601  * @page: page to be put back to appropriate lru list
602  *
603  * Add previously isolated @page to appropriate LRU list.
604  * Page may still be unevictable for other reasons.
605  *
606  * lru_lock must not be held, interrupts must be enabled.
607  */
608 void putback_lru_page(struct page *page)
609 {
610         int lru;
611         int active = !!TestClearPageActive(page);
612         int was_unevictable = PageUnevictable(page);
613
614         VM_BUG_ON(PageLRU(page));
615
616 redo:
617         ClearPageUnevictable(page);
618
619         if (page_evictable(page, NULL)) {
620                 /*
621                  * For evictable pages, we can use the cache.
622                  * In event of a race, worst case is we end up with an
623                  * unevictable page on [in]active list.
624                  * We know how to handle that.
625                  */
626                 lru = active + page_lru_base_type(page);
627                 lru_cache_add_lru(page, lru);
628         } else {
629                 /*
630                  * Put unevictable pages directly on zone's unevictable
631                  * list.
632                  */
633                 lru = LRU_UNEVICTABLE;
634                 add_page_to_unevictable_list(page);
635                 /*
636                  * When racing with an mlock or AS_UNEVICTABLE clearing
637                  * (page is unlocked) make sure that if the other thread
638                  * does not observe our setting of PG_lru and fails
639                  * isolation/check_move_unevictable_page,
640                  * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
641                  * the page back to the evictable list.
642                  *
643                  * The other side is TestClearPageMlocked() or shmem_lock().
644                  */
645                 smp_mb();
646         }
647
648         /*
649          * page's status can change while we move it among lru. If an evictable
650          * page is on unevictable list, it never be freed. To avoid that,
651          * check after we added it to the list, again.
652          */
653         if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
654                 if (!isolate_lru_page(page)) {
655                         put_page(page);
656                         goto redo;
657                 }
658                 /* This means someone else dropped this page from LRU
659                  * So, it will be freed or putback to LRU again. There is
660                  * nothing to do here.
661                  */
662         }
663
664         if (was_unevictable && lru != LRU_UNEVICTABLE)
665                 count_vm_event(UNEVICTABLE_PGRESCUED);
666         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
667                 count_vm_event(UNEVICTABLE_PGCULLED);
668
669         put_page(page);         /* drop ref from isolate */
670 }
671
672 enum page_references {
673         PAGEREF_RECLAIM,
674         PAGEREF_RECLAIM_CLEAN,
675         PAGEREF_KEEP,
676         PAGEREF_ACTIVATE,
677 };
678
679 static enum page_references page_check_references(struct page *page,
680                                                   struct scan_control *sc)
681 {
682         int referenced_ptes, referenced_page;
683         unsigned long vm_flags;
684
685         referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
686         referenced_page = TestClearPageReferenced(page);
687
688         /* Lumpy reclaim - ignore references */
689         if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
690                 return PAGEREF_RECLAIM;
691
692         /*
693          * Mlock lost the isolation race with us.  Let try_to_unmap()
694          * move the page to the unevictable list.
695          */
696         if (vm_flags & VM_LOCKED)
697                 return PAGEREF_RECLAIM;
698
699         if (referenced_ptes) {
700                 if (PageAnon(page))
701                         return PAGEREF_ACTIVATE;
702                 /*
703                  * All mapped pages start out with page table
704                  * references from the instantiating fault, so we need
705                  * to look twice if a mapped file page is used more
706                  * than once.
707                  *
708                  * Mark it and spare it for another trip around the
709                  * inactive list.  Another page table reference will
710                  * lead to its activation.
711                  *
712                  * Note: the mark is set for activated pages as well
713                  * so that recently deactivated but used pages are
714                  * quickly recovered.
715                  */
716                 SetPageReferenced(page);
717
718                 if (referenced_page || referenced_ptes > 1)
719                         return PAGEREF_ACTIVATE;
720
721                 /*
722                  * Activate file-backed executable pages after first usage.
723                  */
724                 if (vm_flags & VM_EXEC)
725                         return PAGEREF_ACTIVATE;
726
727                 return PAGEREF_KEEP;
728         }
729
730         /* Reclaim if clean, defer dirty pages to writeback */
731         if (referenced_page && !PageSwapBacked(page))
732                 return PAGEREF_RECLAIM_CLEAN;
733
734         return PAGEREF_RECLAIM;
735 }
736
737 /*
738  * shrink_page_list() returns the number of reclaimed pages
739  */
740 static unsigned long shrink_page_list(struct list_head *page_list,
741                                       struct zone *zone,
742                                       struct scan_control *sc,
743                                       int priority,
744                                       unsigned long *ret_nr_dirty,
745                                       unsigned long *ret_nr_writeback)
746 {
747         LIST_HEAD(ret_pages);
748         LIST_HEAD(free_pages);
749         int pgactivate = 0;
750         unsigned long nr_dirty = 0;
751         unsigned long nr_congested = 0;
752         unsigned long nr_reclaimed = 0;
753         unsigned long nr_writeback = 0;
754
755         cond_resched();
756
757         while (!list_empty(page_list)) {
758                 enum page_references references;
759                 struct address_space *mapping;
760                 struct page *page;
761                 int may_enter_fs;
762
763                 cond_resched();
764
765                 page = lru_to_page(page_list);
766                 list_del(&page->lru);
767
768                 if (!trylock_page(page))
769                         goto keep;
770
771                 VM_BUG_ON(PageActive(page));
772                 VM_BUG_ON(page_zone(page) != zone);
773
774                 sc->nr_scanned++;
775
776                 if (unlikely(!page_evictable(page, NULL)))
777                         goto cull_mlocked;
778
779                 if (!sc->may_unmap && page_mapped(page))
780                         goto keep_locked;
781
782                 /* Double the slab pressure for mapped and swapcache pages */
783                 if (page_mapped(page) || PageSwapCache(page))
784                         sc->nr_scanned++;
785
786                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
787                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
788
789                 if (PageWriteback(page)) {
790                         nr_writeback++;
791                         /*
792                          * Synchronous reclaim cannot queue pages for
793                          * writeback due to the possibility of stack overflow
794                          * but if it encounters a page under writeback, wait
795                          * for the IO to complete.
796                          */
797                         if ((sc->reclaim_mode & RECLAIM_MODE_SYNC) &&
798                             may_enter_fs)
799                                 wait_on_page_writeback(page);
800                         else {
801                                 unlock_page(page);
802                                 goto keep_lumpy;
803                         }
804                 }
805
806                 references = page_check_references(page, sc);
807                 switch (references) {
808                 case PAGEREF_ACTIVATE:
809                         goto activate_locked;
810                 case PAGEREF_KEEP:
811                         goto keep_locked;
812                 case PAGEREF_RECLAIM:
813                 case PAGEREF_RECLAIM_CLEAN:
814                         ; /* try to reclaim the page below */
815                 }
816
817                 /*
818                  * Anonymous process memory has backing store?
819                  * Try to allocate it some swap space here.
820                  */
821                 if (PageAnon(page) && !PageSwapCache(page)) {
822                         if (!(sc->gfp_mask & __GFP_IO))
823                                 goto keep_locked;
824                         if (!add_to_swap(page))
825                                 goto activate_locked;
826                         may_enter_fs = 1;
827                 }
828
829                 mapping = page_mapping(page);
830
831                 /*
832                  * The page is mapped into the page tables of one or more
833                  * processes. Try to unmap it here.
834                  */
835                 if (page_mapped(page) && mapping) {
836                         switch (try_to_unmap(page, TTU_UNMAP)) {
837                         case SWAP_FAIL:
838                                 goto activate_locked;
839                         case SWAP_AGAIN:
840                                 goto keep_locked;
841                         case SWAP_MLOCK:
842                                 goto cull_mlocked;
843                         case SWAP_SUCCESS:
844                                 ; /* try to free the page below */
845                         }
846                 }
847
848                 if (PageDirty(page)) {
849                         nr_dirty++;
850
851                         /*
852                          * Only kswapd can writeback filesystem pages to
853                          * avoid risk of stack overflow but do not writeback
854                          * unless under significant pressure.
855                          */
856                         if (page_is_file_cache(page) &&
857                                         (!current_is_kswapd() || priority >= DEF_PRIORITY - 2)) {
858                                 /*
859                                  * Immediately reclaim when written back.
860                                  * Similar in principal to deactivate_page()
861                                  * except we already have the page isolated
862                                  * and know it's dirty
863                                  */
864                                 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
865                                 SetPageReclaim(page);
866
867                                 goto keep_locked;
868                         }
869
870                         if (references == PAGEREF_RECLAIM_CLEAN)
871                                 goto keep_locked;
872                         if (!may_enter_fs)
873                                 goto keep_locked;
874                         if (!sc->may_writepage)
875                                 goto keep_locked;
876
877                         /* Page is dirty, try to write it out here */
878                         switch (pageout(page, mapping, sc)) {
879                         case PAGE_KEEP:
880                                 nr_congested++;
881                                 goto keep_locked;
882                         case PAGE_ACTIVATE:
883                                 goto activate_locked;
884                         case PAGE_SUCCESS:
885                                 if (PageWriteback(page))
886                                         goto keep_lumpy;
887                                 if (PageDirty(page))
888                                         goto keep;
889
890                                 /*
891                                  * A synchronous write - probably a ramdisk.  Go
892                                  * ahead and try to reclaim the page.
893                                  */
894                                 if (!trylock_page(page))
895                                         goto keep;
896                                 if (PageDirty(page) || PageWriteback(page))
897                                         goto keep_locked;
898                                 mapping = page_mapping(page);
899                         case PAGE_CLEAN:
900                                 ; /* try to free the page below */
901                         }
902                 }
903
904                 /*
905                  * If the page has buffers, try to free the buffer mappings
906                  * associated with this page. If we succeed we try to free
907                  * the page as well.
908                  *
909                  * We do this even if the page is PageDirty().
910                  * try_to_release_page() does not perform I/O, but it is
911                  * possible for a page to have PageDirty set, but it is actually
912                  * clean (all its buffers are clean).  This happens if the
913                  * buffers were written out directly, with submit_bh(). ext3
914                  * will do this, as well as the blockdev mapping.
915                  * try_to_release_page() will discover that cleanness and will
916                  * drop the buffers and mark the page clean - it can be freed.
917                  *
918                  * Rarely, pages can have buffers and no ->mapping.  These are
919                  * the pages which were not successfully invalidated in
920                  * truncate_complete_page().  We try to drop those buffers here
921                  * and if that worked, and the page is no longer mapped into
922                  * process address space (page_count == 1) it can be freed.
923                  * Otherwise, leave the page on the LRU so it is swappable.
924                  */
925                 if (page_has_private(page)) {
926                         if (!try_to_release_page(page, sc->gfp_mask))
927                                 goto activate_locked;
928                         if (!mapping && page_count(page) == 1) {
929                                 unlock_page(page);
930                                 if (put_page_testzero(page))
931                                         goto free_it;
932                                 else {
933                                         /*
934                                          * rare race with speculative reference.
935                                          * the speculative reference will free
936                                          * this page shortly, so we may
937                                          * increment nr_reclaimed here (and
938                                          * leave it off the LRU).
939                                          */
940                                         nr_reclaimed++;
941                                         continue;
942                                 }
943                         }
944                 }
945
946                 if (!mapping || !__remove_mapping(mapping, page))
947                         goto keep_locked;
948
949                 /*
950                  * At this point, we have no other references and there is
951                  * no way to pick any more up (removed from LRU, removed
952                  * from pagecache). Can use non-atomic bitops now (and
953                  * we obviously don't have to worry about waking up a process
954                  * waiting on the page lock, because there are no references.
955                  */
956                 __clear_page_locked(page);
957 free_it:
958                 nr_reclaimed++;
959
960                 /*
961                  * Is there need to periodically free_page_list? It would
962                  * appear not as the counts should be low
963                  */
964                 list_add(&page->lru, &free_pages);
965                 continue;
966
967 cull_mlocked:
968                 if (PageSwapCache(page))
969                         try_to_free_swap(page);
970                 unlock_page(page);
971                 putback_lru_page(page);
972                 reset_reclaim_mode(sc);
973                 continue;
974
975 activate_locked:
976                 /* Not a candidate for swapping, so reclaim swap space. */
977                 if (PageSwapCache(page) && vm_swap_full())
978                         try_to_free_swap(page);
979                 VM_BUG_ON(PageActive(page));
980                 SetPageActive(page);
981                 pgactivate++;
982 keep_locked:
983                 unlock_page(page);
984 keep:
985                 reset_reclaim_mode(sc);
986 keep_lumpy:
987                 list_add(&page->lru, &ret_pages);
988                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
989         }
990
991         /*
992          * Tag a zone as congested if all the dirty pages encountered were
993          * backed by a congested BDI. In this case, reclaimers should just
994          * back off and wait for congestion to clear because further reclaim
995          * will encounter the same problem
996          */
997         if (nr_dirty && nr_dirty == nr_congested && scanning_global_lru(sc))
998                 zone_set_flag(zone, ZONE_CONGESTED);
999
1000         free_hot_cold_page_list(&free_pages, 1);
1001
1002         list_splice(&ret_pages, page_list);
1003         count_vm_events(PGACTIVATE, pgactivate);
1004         *ret_nr_dirty += nr_dirty;
1005         *ret_nr_writeback += nr_writeback;
1006         return nr_reclaimed;
1007 }
1008
1009 /*
1010  * Attempt to remove the specified page from its LRU.  Only take this page
1011  * if it is of the appropriate PageActive status.  Pages which are being
1012  * freed elsewhere are also ignored.
1013  *
1014  * page:        page to consider
1015  * mode:        one of the LRU isolation modes defined above
1016  *
1017  * returns 0 on success, -ve errno on failure.
1018  */
1019 int __isolate_lru_page(struct page *page, isolate_mode_t mode, int file)
1020 {
1021         bool all_lru_mode;
1022         int ret = -EINVAL;
1023
1024         /* Only take pages on the LRU. */
1025         if (!PageLRU(page))
1026                 return ret;
1027
1028         all_lru_mode = (mode & (ISOLATE_ACTIVE|ISOLATE_INACTIVE)) ==
1029                 (ISOLATE_ACTIVE|ISOLATE_INACTIVE);
1030
1031         /*
1032          * When checking the active state, we need to be sure we are
1033          * dealing with comparible boolean values.  Take the logical not
1034          * of each.
1035          */
1036         if (!all_lru_mode && !PageActive(page) != !(mode & ISOLATE_ACTIVE))
1037                 return ret;
1038
1039         if (!all_lru_mode && !!page_is_file_cache(page) != file)
1040                 return ret;
1041
1042         /*
1043          * When this function is being called for lumpy reclaim, we
1044          * initially look into all LRU pages, active, inactive and
1045          * unevictable; only give shrink_page_list evictable pages.
1046          */
1047         if (PageUnevictable(page))
1048                 return ret;
1049
1050         ret = -EBUSY;
1051
1052         if ((mode & ISOLATE_CLEAN) && (PageDirty(page) || PageWriteback(page)))
1053                 return ret;
1054
1055         if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1056                 return ret;
1057
1058         if (likely(get_page_unless_zero(page))) {
1059                 /*
1060                  * Be careful not to clear PageLRU until after we're
1061                  * sure the page is not being freed elsewhere -- the
1062                  * page release code relies on it.
1063                  */
1064                 ClearPageLRU(page);
1065                 ret = 0;
1066         }
1067
1068         return ret;
1069 }
1070
1071 /*
1072  * zone->lru_lock is heavily contended.  Some of the functions that
1073  * shrink the lists perform better by taking out a batch of pages
1074  * and working on them outside the LRU lock.
1075  *
1076  * For pagecache intensive workloads, this function is the hottest
1077  * spot in the kernel (apart from copy_*_user functions).
1078  *
1079  * Appropriate locks must be held before calling this function.
1080  *
1081  * @nr_to_scan: The number of pages to look through on the list.
1082  * @src:        The LRU list to pull pages off.
1083  * @dst:        The temp list to put pages on to.
1084  * @scanned:    The number of pages that were scanned.
1085  * @order:      The caller's attempted allocation order
1086  * @mode:       One of the LRU isolation modes
1087  * @file:       True [1] if isolating file [!anon] pages
1088  *
1089  * returns how many pages were moved onto *@dst.
1090  */
1091 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1092                 struct list_head *src, struct list_head *dst,
1093                 unsigned long *scanned, int order, isolate_mode_t mode,
1094                 int file)
1095 {
1096         unsigned long nr_taken = 0;
1097         unsigned long nr_lumpy_taken = 0;
1098         unsigned long nr_lumpy_dirty = 0;
1099         unsigned long nr_lumpy_failed = 0;
1100         unsigned long scan;
1101
1102         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1103                 struct page *page;
1104                 unsigned long pfn;
1105                 unsigned long end_pfn;
1106                 unsigned long page_pfn;
1107                 int zone_id;
1108
1109                 page = lru_to_page(src);
1110                 prefetchw_prev_lru_page(page, src, flags);
1111
1112                 VM_BUG_ON(!PageLRU(page));
1113
1114                 switch (__isolate_lru_page(page, mode, file)) {
1115                 case 0:
1116                         list_move(&page->lru, dst);
1117                         mem_cgroup_del_lru(page);
1118                         nr_taken += hpage_nr_pages(page);
1119                         break;
1120
1121                 case -EBUSY:
1122                         /* else it is being freed elsewhere */
1123                         list_move(&page->lru, src);
1124                         mem_cgroup_rotate_lru_list(page, page_lru(page));
1125                         continue;
1126
1127                 default:
1128                         BUG();
1129                 }
1130
1131                 if (!order)
1132                         continue;
1133
1134                 /*
1135                  * Attempt to take all pages in the order aligned region
1136                  * surrounding the tag page.  Only take those pages of
1137                  * the same active state as that tag page.  We may safely
1138                  * round the target page pfn down to the requested order
1139                  * as the mem_map is guaranteed valid out to MAX_ORDER,
1140                  * where that page is in a different zone we will detect
1141                  * it from its zone id and abort this block scan.
1142                  */
1143                 zone_id = page_zone_id(page);
1144                 page_pfn = page_to_pfn(page);
1145                 pfn = page_pfn & ~((1 << order) - 1);
1146                 end_pfn = pfn + (1 << order);
1147                 for (; pfn < end_pfn; pfn++) {
1148                         struct page *cursor_page;
1149
1150                         /* The target page is in the block, ignore it. */
1151                         if (unlikely(pfn == page_pfn))
1152                                 continue;
1153
1154                         /* Avoid holes within the zone. */
1155                         if (unlikely(!pfn_valid_within(pfn)))
1156                                 break;
1157
1158                         cursor_page = pfn_to_page(pfn);
1159
1160                         /* Check that we have not crossed a zone boundary. */
1161                         if (unlikely(page_zone_id(cursor_page) != zone_id))
1162                                 break;
1163
1164                         /*
1165                          * If we don't have enough swap space, reclaiming of
1166                          * anon page which don't already have a swap slot is
1167                          * pointless.
1168                          */
1169                         if (nr_swap_pages <= 0 && PageSwapBacked(cursor_page) &&
1170                             !PageSwapCache(cursor_page))
1171                                 break;
1172
1173                         if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1174                                 list_move(&cursor_page->lru, dst);
1175                                 mem_cgroup_del_lru(cursor_page);
1176                                 nr_taken += hpage_nr_pages(cursor_page);
1177                                 nr_lumpy_taken++;
1178                                 if (PageDirty(cursor_page))
1179                                         nr_lumpy_dirty++;
1180                                 scan++;
1181                         } else {
1182                                 /*
1183                                  * Check if the page is freed already.
1184                                  *
1185                                  * We can't use page_count() as that
1186                                  * requires compound_head and we don't
1187                                  * have a pin on the page here. If a
1188                                  * page is tail, we may or may not
1189                                  * have isolated the head, so assume
1190                                  * it's not free, it'd be tricky to
1191                                  * track the head status without a
1192                                  * page pin.
1193                                  */
1194                                 if (!PageTail(cursor_page) &&
1195                                     !atomic_read(&cursor_page->_count))
1196                                         continue;
1197                                 break;
1198                         }
1199                 }
1200
1201                 /* If we break out of the loop above, lumpy reclaim failed */
1202                 if (pfn < end_pfn)
1203                         nr_lumpy_failed++;
1204         }
1205
1206         *scanned = scan;
1207
1208         trace_mm_vmscan_lru_isolate(order,
1209                         nr_to_scan, scan,
1210                         nr_taken,
1211                         nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1212                         mode);
1213         return nr_taken;
1214 }
1215
1216 static unsigned long isolate_pages_global(unsigned long nr,
1217                                         struct list_head *dst,
1218                                         unsigned long *scanned, int order,
1219                                         isolate_mode_t mode,
1220                                         struct zone *z, int active, int file)
1221 {
1222         int lru = LRU_BASE;
1223         if (active)
1224                 lru += LRU_ACTIVE;
1225         if (file)
1226                 lru += LRU_FILE;
1227         return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1228                                                                 mode, file);
1229 }
1230
1231 /*
1232  * clear_active_flags() is a helper for shrink_active_list(), clearing
1233  * any active bits from the pages in the list.
1234  */
1235 static unsigned long clear_active_flags(struct list_head *page_list,
1236                                         unsigned int *count)
1237 {
1238         int nr_active = 0;
1239         int lru;
1240         struct page *page;
1241
1242         list_for_each_entry(page, page_list, lru) {
1243                 int numpages = hpage_nr_pages(page);
1244                 lru = page_lru_base_type(page);
1245                 if (PageActive(page)) {
1246                         lru += LRU_ACTIVE;
1247                         ClearPageActive(page);
1248                         nr_active += numpages;
1249                 }
1250                 if (count)
1251                         count[lru] += numpages;
1252         }
1253
1254         return nr_active;
1255 }
1256
1257 /**
1258  * isolate_lru_page - tries to isolate a page from its LRU list
1259  * @page: page to isolate from its LRU list
1260  *
1261  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1262  * vmstat statistic corresponding to whatever LRU list the page was on.
1263  *
1264  * Returns 0 if the page was removed from an LRU list.
1265  * Returns -EBUSY if the page was not on an LRU list.
1266  *
1267  * The returned page will have PageLRU() cleared.  If it was found on
1268  * the active list, it will have PageActive set.  If it was found on
1269  * the unevictable list, it will have the PageUnevictable bit set. That flag
1270  * may need to be cleared by the caller before letting the page go.
1271  *
1272  * The vmstat statistic corresponding to the list on which the page was
1273  * found will be decremented.
1274  *
1275  * Restrictions:
1276  * (1) Must be called with an elevated refcount on the page. This is a
1277  *     fundamentnal difference from isolate_lru_pages (which is called
1278  *     without a stable reference).
1279  * (2) the lru_lock must not be held.
1280  * (3) interrupts must be enabled.
1281  */
1282 int isolate_lru_page(struct page *page)
1283 {
1284         int ret = -EBUSY;
1285
1286         VM_BUG_ON(!page_count(page));
1287
1288         if (PageLRU(page)) {
1289                 struct zone *zone = page_zone(page);
1290
1291                 spin_lock_irq(&zone->lru_lock);
1292                 if (PageLRU(page)) {
1293                         int lru = page_lru(page);
1294                         ret = 0;
1295                         get_page(page);
1296                         ClearPageLRU(page);
1297
1298                         del_page_from_lru_list(zone, page, lru);
1299                 }
1300                 spin_unlock_irq(&zone->lru_lock);
1301         }
1302         return ret;
1303 }
1304
1305 /*
1306  * Are there way too many processes in the direct reclaim path already?
1307  */
1308 static int too_many_isolated(struct zone *zone, int file,
1309                 struct scan_control *sc)
1310 {
1311         unsigned long inactive, isolated;
1312
1313         if (current_is_kswapd())
1314                 return 0;
1315
1316         if (!scanning_global_lru(sc))
1317                 return 0;
1318
1319         if (file) {
1320                 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1321                 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1322         } else {
1323                 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1324                 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1325         }
1326
1327         return isolated > inactive;
1328 }
1329
1330 /*
1331  * TODO: Try merging with migrations version of putback_lru_pages
1332  */
1333 static noinline_for_stack void
1334 putback_lru_pages(struct zone *zone, struct scan_control *sc,
1335                                 unsigned long nr_anon, unsigned long nr_file,
1336                                 struct list_head *page_list)
1337 {
1338         struct page *page;
1339         struct pagevec pvec;
1340         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1341
1342         pagevec_init(&pvec, 1);
1343
1344         /*
1345          * Put back any unfreeable pages.
1346          */
1347         spin_lock(&zone->lru_lock);
1348         while (!list_empty(page_list)) {
1349                 int lru;
1350                 page = lru_to_page(page_list);
1351                 VM_BUG_ON(PageLRU(page));
1352                 list_del(&page->lru);
1353                 if (unlikely(!page_evictable(page, NULL))) {
1354                         spin_unlock_irq(&zone->lru_lock);
1355                         putback_lru_page(page);
1356                         spin_lock_irq(&zone->lru_lock);
1357                         continue;
1358                 }
1359                 SetPageLRU(page);
1360                 lru = page_lru(page);
1361                 add_page_to_lru_list(zone, page, lru);
1362                 if (is_active_lru(lru)) {
1363                         int file = is_file_lru(lru);
1364                         int numpages = hpage_nr_pages(page);
1365                         reclaim_stat->recent_rotated[file] += numpages;
1366                 }
1367                 if (!pagevec_add(&pvec, page)) {
1368                         spin_unlock_irq(&zone->lru_lock);
1369                         __pagevec_release(&pvec);
1370                         spin_lock_irq(&zone->lru_lock);
1371                 }
1372         }
1373         __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1374         __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1375
1376         spin_unlock_irq(&zone->lru_lock);
1377         pagevec_release(&pvec);
1378 }
1379
1380 static noinline_for_stack void update_isolated_counts(struct zone *zone,
1381                                         struct scan_control *sc,
1382                                         unsigned long *nr_anon,
1383                                         unsigned long *nr_file,
1384                                         struct list_head *isolated_list)
1385 {
1386         unsigned long nr_active;
1387         unsigned int count[NR_LRU_LISTS] = { 0, };
1388         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1389
1390         nr_active = clear_active_flags(isolated_list, count);
1391         __count_vm_events(PGDEACTIVATE, nr_active);
1392
1393         __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1394                               -count[LRU_ACTIVE_FILE]);
1395         __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1396                               -count[LRU_INACTIVE_FILE]);
1397         __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1398                               -count[LRU_ACTIVE_ANON]);
1399         __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1400                               -count[LRU_INACTIVE_ANON]);
1401
1402         *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1403         *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1404         __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1405         __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1406
1407         reclaim_stat->recent_scanned[0] += *nr_anon;
1408         reclaim_stat->recent_scanned[1] += *nr_file;
1409 }
1410
1411 /*
1412  * Returns true if a direct reclaim should wait on pages under writeback.
1413  *
1414  * If we are direct reclaiming for contiguous pages and we do not reclaim
1415  * everything in the list, try again and wait for writeback IO to complete.
1416  * This will stall high-order allocations noticeably. Only do that when really
1417  * need to free the pages under high memory pressure.
1418  */
1419 static inline bool should_reclaim_stall(unsigned long nr_taken,
1420                                         unsigned long nr_freed,
1421                                         int priority,
1422                                         struct scan_control *sc)
1423 {
1424         int lumpy_stall_priority;
1425
1426         /* kswapd should not stall on sync IO */
1427         if (current_is_kswapd())
1428                 return false;
1429
1430         /* Only stall on lumpy reclaim */
1431         if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
1432                 return false;
1433
1434         /* If we have reclaimed everything on the isolated list, no stall */
1435         if (nr_freed == nr_taken)
1436                 return false;
1437
1438         /*
1439          * For high-order allocations, there are two stall thresholds.
1440          * High-cost allocations stall immediately where as lower
1441          * order allocations such as stacks require the scanning
1442          * priority to be much higher before stalling.
1443          */
1444         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1445                 lumpy_stall_priority = DEF_PRIORITY;
1446         else
1447                 lumpy_stall_priority = DEF_PRIORITY / 3;
1448
1449         return priority <= lumpy_stall_priority;
1450 }
1451
1452 /*
1453  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1454  * of reclaimed pages
1455  */
1456 static noinline_for_stack unsigned long
1457 shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1458                         struct scan_control *sc, int priority, int file)
1459 {
1460         LIST_HEAD(page_list);
1461         unsigned long nr_scanned;
1462         unsigned long nr_reclaimed = 0;
1463         unsigned long nr_taken;
1464         unsigned long nr_anon;
1465         unsigned long nr_file;
1466         unsigned long nr_dirty = 0;
1467         unsigned long nr_writeback = 0;
1468         isolate_mode_t reclaim_mode = ISOLATE_INACTIVE;
1469
1470         while (unlikely(too_many_isolated(zone, file, sc))) {
1471                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1472
1473                 /* We are about to die and free our memory. Return now. */
1474                 if (fatal_signal_pending(current))
1475                         return SWAP_CLUSTER_MAX;
1476         }
1477
1478         set_reclaim_mode(priority, sc, false);
1479         if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
1480                 reclaim_mode |= ISOLATE_ACTIVE;
1481
1482         lru_add_drain();
1483
1484         if (!sc->may_unmap)
1485                 reclaim_mode |= ISOLATE_UNMAPPED;
1486         if (!sc->may_writepage)
1487                 reclaim_mode |= ISOLATE_CLEAN;
1488
1489         spin_lock_irq(&zone->lru_lock);
1490
1491         if (scanning_global_lru(sc)) {
1492                 nr_taken = isolate_pages_global(nr_to_scan, &page_list,
1493                         &nr_scanned, sc->order, reclaim_mode, zone, 0, file);
1494                 zone->pages_scanned += nr_scanned;
1495                 if (current_is_kswapd())
1496                         __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1497                                                nr_scanned);
1498                 else
1499                         __count_zone_vm_events(PGSCAN_DIRECT, zone,
1500                                                nr_scanned);
1501         } else {
1502                 nr_taken = mem_cgroup_isolate_pages(nr_to_scan, &page_list,
1503                         &nr_scanned, sc->order, reclaim_mode, zone,
1504                         sc->mem_cgroup, 0, file);
1505                 /*
1506                  * mem_cgroup_isolate_pages() keeps track of
1507                  * scanned pages on its own.
1508                  */
1509         }
1510
1511         if (nr_taken == 0) {
1512                 spin_unlock_irq(&zone->lru_lock);
1513                 return 0;
1514         }
1515
1516         update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
1517
1518         spin_unlock_irq(&zone->lru_lock);
1519
1520         nr_reclaimed = shrink_page_list(&page_list, zone, sc, priority,
1521                                                 &nr_dirty, &nr_writeback);
1522
1523         /* Check if we should syncronously wait for writeback */
1524         if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1525                 set_reclaim_mode(priority, sc, true);
1526                 nr_reclaimed += shrink_page_list(&page_list, zone, sc,
1527                                         priority, &nr_dirty, &nr_writeback);
1528         }
1529
1530         local_irq_disable();
1531         if (current_is_kswapd())
1532                 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1533         __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1534
1535         putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
1536
1537         /*
1538          * If reclaim is isolating dirty pages under writeback, it implies
1539          * that the long-lived page allocation rate is exceeding the page
1540          * laundering rate. Either the global limits are not being effective
1541          * at throttling processes due to the page distribution throughout
1542          * zones or there is heavy usage of a slow backing device. The
1543          * only option is to throttle from reclaim context which is not ideal
1544          * as there is no guarantee the dirtying process is throttled in the
1545          * same way balance_dirty_pages() manages.
1546          *
1547          * This scales the number of dirty pages that must be under writeback
1548          * before throttling depending on priority. It is a simple backoff
1549          * function that has the most effect in the range DEF_PRIORITY to
1550          * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1551          * in trouble and reclaim is considered to be in trouble.
1552          *
1553          * DEF_PRIORITY   100% isolated pages must be PageWriteback to throttle
1554          * DEF_PRIORITY-1  50% must be PageWriteback
1555          * DEF_PRIORITY-2  25% must be PageWriteback, kswapd in trouble
1556          * ...
1557          * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1558          *                     isolated page is PageWriteback
1559          */
1560         if (nr_writeback && nr_writeback >= (nr_taken >> (DEF_PRIORITY-priority)))
1561                 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1562
1563         trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1564                 zone_idx(zone),
1565                 nr_scanned, nr_reclaimed,
1566                 priority,
1567                 trace_shrink_flags(file, sc->reclaim_mode));
1568         return nr_reclaimed;
1569 }
1570
1571 /*
1572  * This moves pages from the active list to the inactive list.
1573  *
1574  * We move them the other way if the page is referenced by one or more
1575  * processes, from rmap.
1576  *
1577  * If the pages are mostly unmapped, the processing is fast and it is
1578  * appropriate to hold zone->lru_lock across the whole operation.  But if
1579  * the pages are mapped, the processing is slow (page_referenced()) so we
1580  * should drop zone->lru_lock around each page.  It's impossible to balance
1581  * this, so instead we remove the pages from the LRU while processing them.
1582  * It is safe to rely on PG_active against the non-LRU pages in here because
1583  * nobody will play with that bit on a non-LRU page.
1584  *
1585  * The downside is that we have to touch page->_count against each page.
1586  * But we had to alter page->flags anyway.
1587  */
1588
1589 static void move_active_pages_to_lru(struct zone *zone,
1590                                      struct list_head *list,
1591                                      enum lru_list lru)
1592 {
1593         unsigned long pgmoved = 0;
1594         struct pagevec pvec;
1595         struct page *page;
1596
1597         pagevec_init(&pvec, 1);
1598
1599         while (!list_empty(list)) {
1600                 page = lru_to_page(list);
1601
1602                 VM_BUG_ON(PageLRU(page));
1603                 SetPageLRU(page);
1604
1605                 list_move(&page->lru, &zone->lru[lru].list);
1606                 mem_cgroup_add_lru_list(page, lru);
1607                 pgmoved += hpage_nr_pages(page);
1608
1609                 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1610                         spin_unlock_irq(&zone->lru_lock);
1611                         if (buffer_heads_over_limit)
1612                                 pagevec_strip(&pvec);
1613                         __pagevec_release(&pvec);
1614                         spin_lock_irq(&zone->lru_lock);
1615                 }
1616         }
1617         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1618         if (!is_active_lru(lru))
1619                 __count_vm_events(PGDEACTIVATE, pgmoved);
1620 }
1621
1622 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1623                         struct scan_control *sc, int priority, int file)
1624 {
1625         unsigned long nr_taken;
1626         unsigned long pgscanned;
1627         unsigned long vm_flags;
1628         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1629         LIST_HEAD(l_active);
1630         LIST_HEAD(l_inactive);
1631         struct page *page;
1632         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1633         unsigned long nr_rotated = 0;
1634         isolate_mode_t reclaim_mode = ISOLATE_ACTIVE;
1635
1636         lru_add_drain();
1637
1638         if (!sc->may_unmap)
1639                 reclaim_mode |= ISOLATE_UNMAPPED;
1640         if (!sc->may_writepage)
1641                 reclaim_mode |= ISOLATE_CLEAN;
1642
1643         spin_lock_irq(&zone->lru_lock);
1644         if (scanning_global_lru(sc)) {
1645                 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1646                                                 &pgscanned, sc->order,
1647                                                 reclaim_mode, zone,
1648                                                 1, file);
1649                 zone->pages_scanned += pgscanned;
1650         } else {
1651                 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1652                                                 &pgscanned, sc->order,
1653                                                 reclaim_mode, zone,
1654                                                 sc->mem_cgroup, 1, file);
1655                 /*
1656                  * mem_cgroup_isolate_pages() keeps track of
1657                  * scanned pages on its own.
1658                  */
1659         }
1660
1661         reclaim_stat->recent_scanned[file] += nr_taken;
1662
1663         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1664         if (file)
1665                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1666         else
1667                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1668         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1669         spin_unlock_irq(&zone->lru_lock);
1670
1671         while (!list_empty(&l_hold)) {
1672                 cond_resched();
1673                 page = lru_to_page(&l_hold);
1674                 list_del(&page->lru);
1675
1676                 if (unlikely(!page_evictable(page, NULL))) {
1677                         putback_lru_page(page);
1678                         continue;
1679                 }
1680
1681                 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1682                         nr_rotated += hpage_nr_pages(page);
1683                         /*
1684                          * Identify referenced, file-backed active pages and
1685                          * give them one more trip around the active list. So
1686                          * that executable code get better chances to stay in
1687                          * memory under moderate memory pressure.  Anon pages
1688                          * are not likely to be evicted by use-once streaming
1689                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1690                          * so we ignore them here.
1691                          */
1692                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1693                                 list_add(&page->lru, &l_active);
1694                                 continue;
1695                         }
1696                 }
1697
1698                 ClearPageActive(page);  /* we are de-activating */
1699                 list_add(&page->lru, &l_inactive);
1700         }
1701
1702         /*
1703          * Move pages back to the lru list.
1704          */
1705         spin_lock_irq(&zone->lru_lock);
1706         /*
1707          * Count referenced pages from currently used mappings as rotated,
1708          * even though only some of them are actually re-activated.  This
1709          * helps balance scan pressure between file and anonymous pages in
1710          * get_scan_ratio.
1711          */
1712         reclaim_stat->recent_rotated[file] += nr_rotated;
1713
1714         move_active_pages_to_lru(zone, &l_active,
1715                                                 LRU_ACTIVE + file * LRU_FILE);
1716         move_active_pages_to_lru(zone, &l_inactive,
1717                                                 LRU_BASE   + file * LRU_FILE);
1718         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1719         spin_unlock_irq(&zone->lru_lock);
1720 }
1721
1722 #ifdef CONFIG_SWAP
1723 static int inactive_anon_is_low_global(struct zone *zone)
1724 {
1725         unsigned long active, inactive;
1726
1727         active = zone_page_state(zone, NR_ACTIVE_ANON);
1728         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1729
1730         if (inactive * zone->inactive_ratio < active)
1731                 return 1;
1732
1733         return 0;
1734 }
1735
1736 /**
1737  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1738  * @zone: zone to check
1739  * @sc:   scan control of this context
1740  *
1741  * Returns true if the zone does not have enough inactive anon pages,
1742  * meaning some active anon pages need to be deactivated.
1743  */
1744 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1745 {
1746         int low;
1747
1748         /*
1749          * If we don't have swap space, anonymous page deactivation
1750          * is pointless.
1751          */
1752         if (!total_swap_pages)
1753                 return 0;
1754
1755         if (scanning_global_lru(sc))
1756                 low = inactive_anon_is_low_global(zone);
1757         else
1758                 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup, zone);
1759         return low;
1760 }
1761 #else
1762 static inline int inactive_anon_is_low(struct zone *zone,
1763                                         struct scan_control *sc)
1764 {
1765         return 0;
1766 }
1767 #endif
1768
1769 static int inactive_file_is_low_global(struct zone *zone)
1770 {
1771         unsigned long active, inactive;
1772
1773         active = zone_page_state(zone, NR_ACTIVE_FILE);
1774         inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1775
1776         return (active > inactive);
1777 }
1778
1779 /**
1780  * inactive_file_is_low - check if file pages need to be deactivated
1781  * @zone: zone to check
1782  * @sc:   scan control of this context
1783  *
1784  * When the system is doing streaming IO, memory pressure here
1785  * ensures that active file pages get deactivated, until more
1786  * than half of the file pages are on the inactive list.
1787  *
1788  * Once we get to that situation, protect the system's working
1789  * set from being evicted by disabling active file page aging.
1790  *
1791  * This uses a different ratio than the anonymous pages, because
1792  * the page cache uses a use-once replacement algorithm.
1793  */
1794 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1795 {
1796         int low;
1797
1798         if (scanning_global_lru(sc))
1799                 low = inactive_file_is_low_global(zone);
1800         else
1801                 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup, zone);
1802         return low;
1803 }
1804
1805 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1806                                 int file)
1807 {
1808         if (file)
1809                 return inactive_file_is_low(zone, sc);
1810         else
1811                 return inactive_anon_is_low(zone, sc);
1812 }
1813
1814 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1815         struct zone *zone, struct scan_control *sc, int priority)
1816 {
1817         int file = is_file_lru(lru);
1818
1819         if (is_active_lru(lru)) {
1820                 if (inactive_list_is_low(zone, sc, file))
1821                     shrink_active_list(nr_to_scan, zone, sc, priority, file);
1822                 return 0;
1823         }
1824
1825         return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1826 }
1827
1828 static int vmscan_swappiness(struct scan_control *sc)
1829 {
1830         if (scanning_global_lru(sc))
1831                 return vm_swappiness;
1832         return mem_cgroup_swappiness(sc->mem_cgroup);
1833 }
1834
1835 /*
1836  * Determine how aggressively the anon and file LRU lists should be
1837  * scanned.  The relative value of each set of LRU lists is determined
1838  * by looking at the fraction of the pages scanned we did rotate back
1839  * onto the active list instead of evict.
1840  *
1841  * nr[0] = anon pages to scan; nr[1] = file pages to scan
1842  */
1843 static void get_scan_count(struct zone *zone, struct scan_control *sc,
1844                                         unsigned long *nr, int priority)
1845 {
1846         unsigned long anon, file, free;
1847         unsigned long anon_prio, file_prio;
1848         unsigned long ap, fp;
1849         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1850         u64 fraction[2], denominator;
1851         enum lru_list l;
1852         int noswap = 0;
1853         bool force_scan = false;
1854
1855         /*
1856          * If the zone or memcg is small, nr[l] can be 0.  This
1857          * results in no scanning on this priority and a potential
1858          * priority drop.  Global direct reclaim can go to the next
1859          * zone and tends to have no problems. Global kswapd is for
1860          * zone balancing and it needs to scan a minimum amount. When
1861          * reclaiming for a memcg, a priority drop can cause high
1862          * latencies, so it's better to scan a minimum amount there as
1863          * well.
1864          */
1865         if (scanning_global_lru(sc) && current_is_kswapd())
1866                 force_scan = true;
1867         if (!scanning_global_lru(sc))
1868                 force_scan = true;
1869
1870         /* If we have no swap space, do not bother scanning anon pages. */
1871         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1872                 noswap = 1;
1873                 fraction[0] = 0;
1874                 fraction[1] = 1;
1875                 denominator = 1;
1876                 goto out;
1877         }
1878
1879         anon  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1880                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1881         file  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1882                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1883
1884         if (scanning_global_lru(sc)) {
1885                 free  = zone_page_state(zone, NR_FREE_PAGES);
1886                 /* If we have very few page cache pages,
1887                    force-scan anon pages. */
1888                 if (unlikely(file + free <= high_wmark_pages(zone))) {
1889                         fraction[0] = 1;
1890                         fraction[1] = 0;
1891                         denominator = 1;
1892                         goto out;
1893                 }
1894         }
1895
1896         /*
1897          * With swappiness at 100, anonymous and file have the same priority.
1898          * This scanning priority is essentially the inverse of IO cost.
1899          */
1900         anon_prio = vmscan_swappiness(sc);
1901         file_prio = 200 - vmscan_swappiness(sc);
1902
1903         /*
1904          * OK, so we have swap space and a fair amount of page cache
1905          * pages.  We use the recently rotated / recently scanned
1906          * ratios to determine how valuable each cache is.
1907          *
1908          * Because workloads change over time (and to avoid overflow)
1909          * we keep these statistics as a floating average, which ends
1910          * up weighing recent references more than old ones.
1911          *
1912          * anon in [0], file in [1]
1913          */
1914         spin_lock_irq(&zone->lru_lock);
1915         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1916                 reclaim_stat->recent_scanned[0] /= 2;
1917                 reclaim_stat->recent_rotated[0] /= 2;
1918         }
1919
1920         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1921                 reclaim_stat->recent_scanned[1] /= 2;
1922                 reclaim_stat->recent_rotated[1] /= 2;
1923         }
1924
1925         /*
1926          * The amount of pressure on anon vs file pages is inversely
1927          * proportional to the fraction of recently scanned pages on
1928          * each list that were recently referenced and in active use.
1929          */
1930         ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1931         ap /= reclaim_stat->recent_rotated[0] + 1;
1932
1933         fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1934         fp /= reclaim_stat->recent_rotated[1] + 1;
1935         spin_unlock_irq(&zone->lru_lock);
1936
1937         fraction[0] = ap;
1938         fraction[1] = fp;
1939         denominator = ap + fp + 1;
1940 out:
1941         for_each_evictable_lru(l) {
1942                 int file = is_file_lru(l);
1943                 unsigned long scan;
1944
1945                 scan = zone_nr_lru_pages(zone, sc, l);
1946                 if (priority || noswap) {
1947                         scan >>= priority;
1948                         if (!scan && force_scan)
1949                                 scan = SWAP_CLUSTER_MAX;
1950                         scan = div64_u64(scan * fraction[file], denominator);
1951                 }
1952                 nr[l] = scan;
1953         }
1954 }
1955
1956 /*
1957  * Reclaim/compaction depends on a number of pages being freed. To avoid
1958  * disruption to the system, a small number of order-0 pages continue to be
1959  * rotated and reclaimed in the normal fashion. However, by the time we get
1960  * back to the allocator and call try_to_compact_zone(), we ensure that
1961  * there are enough free pages for it to be likely successful
1962  */
1963 static inline bool should_continue_reclaim(struct zone *zone,
1964                                         unsigned long nr_reclaimed,
1965                                         unsigned long nr_scanned,
1966                                         struct scan_control *sc)
1967 {
1968         unsigned long pages_for_compaction;
1969         unsigned long inactive_lru_pages;
1970
1971         /* If not in reclaim/compaction mode, stop */
1972         if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
1973                 return false;
1974
1975         /* Consider stopping depending on scan and reclaim activity */
1976         if (sc->gfp_mask & __GFP_REPEAT) {
1977                 /*
1978                  * For __GFP_REPEAT allocations, stop reclaiming if the
1979                  * full LRU list has been scanned and we are still failing
1980                  * to reclaim pages. This full LRU scan is potentially
1981                  * expensive but a __GFP_REPEAT caller really wants to succeed
1982                  */
1983                 if (!nr_reclaimed && !nr_scanned)
1984                         return false;
1985         } else {
1986                 /*
1987                  * For non-__GFP_REPEAT allocations which can presumably
1988                  * fail without consequence, stop if we failed to reclaim
1989                  * any pages from the last SWAP_CLUSTER_MAX number of
1990                  * pages that were scanned. This will return to the
1991                  * caller faster at the risk reclaim/compaction and
1992                  * the resulting allocation attempt fails
1993                  */
1994                 if (!nr_reclaimed)
1995                         return false;
1996         }
1997
1998         /*
1999          * If we have not reclaimed enough pages for compaction and the
2000          * inactive lists are large enough, continue reclaiming
2001          */
2002         pages_for_compaction = (2UL << sc->order);
2003         inactive_lru_pages = zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
2004         if (nr_swap_pages > 0)
2005                 inactive_lru_pages += zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
2006         if (sc->nr_reclaimed < pages_for_compaction &&
2007                         inactive_lru_pages > pages_for_compaction)
2008                 return true;
2009
2010         /* If compaction would go ahead or the allocation would succeed, stop */
2011         switch (compaction_suitable(zone, sc->order)) {
2012         case COMPACT_PARTIAL:
2013         case COMPACT_CONTINUE:
2014                 return false;
2015         default:
2016                 return true;
2017         }
2018 }
2019
2020 /*
2021  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
2022  */
2023 static void shrink_zone(int priority, struct zone *zone,
2024                                 struct scan_control *sc)
2025 {
2026         unsigned long nr[NR_LRU_LISTS];
2027         unsigned long nr_to_scan;
2028         enum lru_list l;
2029         unsigned long nr_reclaimed, nr_scanned;
2030         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2031         struct blk_plug plug;
2032
2033 restart:
2034         nr_reclaimed = 0;
2035         nr_scanned = sc->nr_scanned;
2036         get_scan_count(zone, sc, nr, priority);
2037
2038         blk_start_plug(&plug);
2039         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2040                                         nr[LRU_INACTIVE_FILE]) {
2041                 for_each_evictable_lru(l) {
2042                         if (nr[l]) {
2043                                 nr_to_scan = min_t(unsigned long,
2044                                                    nr[l], SWAP_CLUSTER_MAX);
2045                                 nr[l] -= nr_to_scan;
2046
2047                                 nr_reclaimed += shrink_list(l, nr_to_scan,
2048                                                             zone, sc, priority);
2049                         }
2050                 }
2051                 /*
2052                  * On large memory systems, scan >> priority can become
2053                  * really large. This is fine for the starting priority;
2054                  * we want to put equal scanning pressure on each zone.
2055                  * However, if the VM has a harder time of freeing pages,
2056                  * with multiple processes reclaiming pages, the total
2057                  * freeing target can get unreasonably large.
2058                  */
2059                 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
2060                         break;
2061         }
2062         blk_finish_plug(&plug);
2063         sc->nr_reclaimed += nr_reclaimed;
2064
2065         /*
2066          * Even if we did not try to evict anon pages at all, we want to
2067          * rebalance the anon lru active/inactive ratio.
2068          */
2069         if (inactive_anon_is_low(zone, sc))
2070                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
2071
2072         /* reclaim/compaction might need reclaim to continue */
2073         if (should_continue_reclaim(zone, nr_reclaimed,
2074                                         sc->nr_scanned - nr_scanned, sc))
2075                 goto restart;
2076
2077         throttle_vm_writeout(sc->gfp_mask);
2078 }
2079
2080 /*
2081  * This is the direct reclaim path, for page-allocating processes.  We only
2082  * try to reclaim pages from zones which will satisfy the caller's allocation
2083  * request.
2084  *
2085  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2086  * Because:
2087  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2088  *    allocation or
2089  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2090  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2091  *    zone defense algorithm.
2092  *
2093  * If a zone is deemed to be full of pinned pages then just give it a light
2094  * scan then give up on it.
2095  *
2096  * This function returns true if a zone is being reclaimed for a costly
2097  * high-order allocation and compaction is either ready to begin or deferred.
2098  * This indicates to the caller that it should retry the allocation or fail.
2099  */
2100 static bool shrink_zones(int priority, struct zonelist *zonelist,
2101                                         struct scan_control *sc)
2102 {
2103         struct zoneref *z;
2104         struct zone *zone;
2105         unsigned long nr_soft_reclaimed;
2106         unsigned long nr_soft_scanned;
2107         bool should_abort_reclaim = false;
2108
2109         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2110                                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2111                 if (!populated_zone(zone))
2112                         continue;
2113                 /*
2114                  * Take care memory controller reclaiming has small influence
2115                  * to global LRU.
2116                  */
2117                 if (scanning_global_lru(sc)) {
2118                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2119                                 continue;
2120                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2121                                 continue;       /* Let kswapd poll it */
2122                         if (COMPACTION_BUILD) {
2123                                 /*
2124                                  * If we already have plenty of memory free for
2125                                  * compaction in this zone, don't free any more.
2126                                  * Even though compaction is invoked for any
2127                                  * non-zero order, only frequent costly order
2128                                  * reclamation is disruptive enough to become a
2129                                  * noticable problem, like transparent huge page
2130                                  * allocations.
2131                                  */
2132                                 if (sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2133                                         (compaction_suitable(zone, sc->order) ||
2134                                          compaction_deferred(zone))) {
2135                                         should_abort_reclaim = true;
2136                                         continue;
2137                                 }
2138                         }
2139                         /*
2140                          * This steals pages from memory cgroups over softlimit
2141                          * and returns the number of reclaimed pages and
2142                          * scanned pages. This works for global memory pressure
2143                          * and balancing, not for a memcg's limit.
2144                          */
2145                         nr_soft_scanned = 0;
2146                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2147                                                 sc->order, sc->gfp_mask,
2148                                                 &nr_soft_scanned);
2149                         sc->nr_reclaimed += nr_soft_reclaimed;
2150                         sc->nr_scanned += nr_soft_scanned;
2151                         /* need some check for avoid more shrink_zone() */
2152                 }
2153
2154                 shrink_zone(priority, zone, sc);
2155         }
2156
2157         return should_abort_reclaim;
2158 }
2159
2160 static bool zone_reclaimable(struct zone *zone)
2161 {
2162         return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2163 }
2164
2165 /* All zones in zonelist are unreclaimable? */
2166 static bool all_unreclaimable(struct zonelist *zonelist,
2167                 struct scan_control *sc)
2168 {
2169         struct zoneref *z;
2170         struct zone *zone;
2171
2172         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2173                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2174                 if (!populated_zone(zone))
2175                         continue;
2176                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2177                         continue;
2178                 if (!zone->all_unreclaimable)
2179                         return false;
2180         }
2181
2182         return true;
2183 }
2184
2185 /*
2186  * This is the main entry point to direct page reclaim.
2187  *
2188  * If a full scan of the inactive list fails to free enough memory then we
2189  * are "out of memory" and something needs to be killed.
2190  *
2191  * If the caller is !__GFP_FS then the probability of a failure is reasonably
2192  * high - the zone may be full of dirty or under-writeback pages, which this
2193  * caller can't do much about.  We kick the writeback threads and take explicit
2194  * naps in the hope that some of these pages can be written.  But if the
2195  * allocating task holds filesystem locks which prevent writeout this might not
2196  * work, and the allocation attempt will fail.
2197  *
2198  * returns:     0, if no pages reclaimed
2199  *              else, the number of pages reclaimed
2200  */
2201 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2202                                         struct scan_control *sc,
2203                                         struct shrink_control *shrink)
2204 {
2205         int priority;
2206         unsigned long total_scanned = 0;
2207         struct reclaim_state *reclaim_state = current->reclaim_state;
2208         struct zoneref *z;
2209         struct zone *zone;
2210         unsigned long writeback_threshold;
2211
2212         get_mems_allowed();
2213         delayacct_freepages_start();
2214
2215         if (scanning_global_lru(sc))
2216                 count_vm_event(ALLOCSTALL);
2217
2218         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2219                 sc->nr_scanned = 0;
2220                 if (!priority)
2221                         disable_swap_token(sc->mem_cgroup);
2222                 if (shrink_zones(priority, zonelist, sc))
2223                         break;
2224
2225                 /*
2226                  * Don't shrink slabs when reclaiming memory from
2227                  * over limit cgroups
2228                  */
2229                 if (scanning_global_lru(sc)) {
2230                         unsigned long lru_pages = 0;
2231                         for_each_zone_zonelist(zone, z, zonelist,
2232                                         gfp_zone(sc->gfp_mask)) {
2233                                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2234                                         continue;
2235
2236                                 lru_pages += zone_reclaimable_pages(zone);
2237                         }
2238
2239                         shrink_slab(shrink, sc->nr_scanned, lru_pages);
2240                         if (reclaim_state) {
2241                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2242                                 reclaim_state->reclaimed_slab = 0;
2243                         }
2244                 }
2245                 total_scanned += sc->nr_scanned;
2246                 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2247                         goto out;
2248
2249                 /*
2250                  * Try to write back as many pages as we just scanned.  This
2251                  * tends to cause slow streaming writers to write data to the
2252                  * disk smoothly, at the dirtying rate, which is nice.   But
2253                  * that's undesirable in laptop mode, where we *want* lumpy
2254                  * writeout.  So in laptop mode, write out the whole world.
2255                  */
2256                 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2257                 if (total_scanned > writeback_threshold) {
2258                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2259                                                 WB_REASON_TRY_TO_FREE_PAGES);
2260                         sc->may_writepage = 1;
2261                 }
2262
2263                 /* Take a nap, wait for some writeback to complete */
2264                 if (!sc->hibernation_mode && sc->nr_scanned &&
2265                     priority < DEF_PRIORITY - 2) {
2266                         struct zone *preferred_zone;
2267
2268                         first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2269                                                 &cpuset_current_mems_allowed,
2270                                                 &preferred_zone);
2271                         wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2272                 }
2273         }
2274
2275 out:
2276         delayacct_freepages_end();
2277         put_mems_allowed();
2278
2279         if (sc->nr_reclaimed)
2280                 return sc->nr_reclaimed;
2281
2282         /*
2283          * As hibernation is going on, kswapd is freezed so that it can't mark
2284          * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2285          * check.
2286          */
2287         if (oom_killer_disabled)
2288                 return 0;
2289
2290         /* top priority shrink_zones still had more to do? don't OOM, then */
2291         if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
2292                 return 1;
2293
2294         return 0;
2295 }
2296
2297 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2298                                 gfp_t gfp_mask, nodemask_t *nodemask)
2299 {
2300         unsigned long nr_reclaimed;
2301         struct scan_control sc = {
2302                 .gfp_mask = gfp_mask,
2303                 .may_writepage = !laptop_mode,
2304                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2305                 .may_unmap = 1,
2306                 .may_swap = 1,
2307                 .order = order,
2308                 .mem_cgroup = NULL,
2309                 .nodemask = nodemask,
2310         };
2311         struct shrink_control shrink = {
2312                 .gfp_mask = sc.gfp_mask,
2313         };
2314
2315         trace_mm_vmscan_direct_reclaim_begin(order,
2316                                 sc.may_writepage,
2317                                 gfp_mask);
2318
2319         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2320
2321         trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2322
2323         return nr_reclaimed;
2324 }
2325
2326 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2327
2328 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2329                                                 gfp_t gfp_mask, bool noswap,
2330                                                 struct zone *zone,
2331                                                 unsigned long *nr_scanned)
2332 {
2333         struct scan_control sc = {
2334                 .nr_scanned = 0,
2335                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2336                 .may_writepage = !laptop_mode,
2337                 .may_unmap = 1,
2338                 .may_swap = !noswap,
2339                 .order = 0,
2340                 .mem_cgroup = mem,
2341         };
2342
2343         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2344                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2345
2346         trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2347                                                       sc.may_writepage,
2348                                                       sc.gfp_mask);
2349
2350         /*
2351          * NOTE: Although we can get the priority field, using it
2352          * here is not a good idea, since it limits the pages we can scan.
2353          * if we don't reclaim here, the shrink_zone from balance_pgdat
2354          * will pick up pages from other mem cgroup's as well. We hack
2355          * the priority and make it zero.
2356          */
2357         shrink_zone(0, zone, &sc);
2358
2359         trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2360
2361         *nr_scanned = sc.nr_scanned;
2362         return sc.nr_reclaimed;
2363 }
2364
2365 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
2366                                            gfp_t gfp_mask,
2367                                            bool noswap)
2368 {
2369         struct zonelist *zonelist;
2370         unsigned long nr_reclaimed;
2371         int nid;
2372         struct scan_control sc = {
2373                 .may_writepage = !laptop_mode,
2374                 .may_unmap = 1,
2375                 .may_swap = !noswap,
2376                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2377                 .order = 0,
2378                 .mem_cgroup = mem_cont,
2379                 .nodemask = NULL, /* we don't care the placement */
2380                 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2381                                 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2382         };
2383         struct shrink_control shrink = {
2384                 .gfp_mask = sc.gfp_mask,
2385         };
2386
2387         /*
2388          * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2389          * take care of from where we get pages. So the node where we start the
2390          * scan does not need to be the current node.
2391          */
2392         nid = mem_cgroup_select_victim_node(mem_cont);
2393
2394         zonelist = NODE_DATA(nid)->node_zonelists;
2395
2396         trace_mm_vmscan_memcg_reclaim_begin(0,
2397                                             sc.may_writepage,
2398                                             sc.gfp_mask);
2399
2400         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2401
2402         trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2403
2404         return nr_reclaimed;
2405 }
2406 #endif
2407
2408 /*
2409  * pgdat_balanced is used when checking if a node is balanced for high-order
2410  * allocations. Only zones that meet watermarks and are in a zone allowed
2411  * by the callers classzone_idx are added to balanced_pages. The total of
2412  * balanced pages must be at least 25% of the zones allowed by classzone_idx
2413  * for the node to be considered balanced. Forcing all zones to be balanced
2414  * for high orders can cause excessive reclaim when there are imbalanced zones.
2415  * The choice of 25% is due to
2416  *   o a 16M DMA zone that is balanced will not balance a zone on any
2417  *     reasonable sized machine
2418  *   o On all other machines, the top zone must be at least a reasonable
2419  *     percentage of the middle zones. For example, on 32-bit x86, highmem
2420  *     would need to be at least 256M for it to be balance a whole node.
2421  *     Similarly, on x86-64 the Normal zone would need to be at least 1G
2422  *     to balance a node on its own. These seemed like reasonable ratios.
2423  */
2424 static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2425                                                 int classzone_idx)
2426 {
2427         unsigned long present_pages = 0;
2428         int i;
2429
2430         for (i = 0; i <= classzone_idx; i++)
2431                 present_pages += pgdat->node_zones[i].present_pages;
2432
2433         /* A special case here: if zone has no page, we think it's balanced */
2434         return balanced_pages >= (present_pages >> 2);
2435 }
2436
2437 /* is kswapd sleeping prematurely? */
2438 static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2439                                         int classzone_idx)
2440 {
2441         int i;
2442         unsigned long balanced = 0;
2443         bool all_zones_ok = true;
2444
2445         /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2446         if (remaining)
2447                 return true;
2448
2449         /* Check the watermark levels */
2450         for (i = 0; i <= classzone_idx; i++) {
2451                 struct zone *zone = pgdat->node_zones + i;
2452
2453                 if (!populated_zone(zone))
2454                         continue;
2455
2456                 /*
2457                  * balance_pgdat() skips over all_unreclaimable after
2458                  * DEF_PRIORITY. Effectively, it considers them balanced so
2459                  * they must be considered balanced here as well if kswapd
2460                  * is to sleep
2461                  */
2462                 if (zone->all_unreclaimable) {
2463                         balanced += zone->present_pages;
2464                         continue;
2465                 }
2466
2467                 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2468                                                         i, 0))
2469                         all_zones_ok = false;
2470                 else
2471                         balanced += zone->present_pages;
2472         }
2473
2474         /*
2475          * For high-order requests, the balanced zones must contain at least
2476          * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2477          * must be balanced
2478          */
2479         if (order)
2480                 return !pgdat_balanced(pgdat, balanced, classzone_idx);
2481         else
2482                 return !all_zones_ok;
2483 }
2484
2485 /*
2486  * For kswapd, balance_pgdat() will work across all this node's zones until
2487  * they are all at high_wmark_pages(zone).
2488  *
2489  * Returns the final order kswapd was reclaiming at
2490  *
2491  * There is special handling here for zones which are full of pinned pages.
2492  * This can happen if the pages are all mlocked, or if they are all used by
2493  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
2494  * What we do is to detect the case where all pages in the zone have been
2495  * scanned twice and there has been zero successful reclaim.  Mark the zone as
2496  * dead and from now on, only perform a short scan.  Basically we're polling
2497  * the zone for when the problem goes away.
2498  *
2499  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
2500  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2501  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2502  * lower zones regardless of the number of free pages in the lower zones. This
2503  * interoperates with the page allocator fallback scheme to ensure that aging
2504  * of pages is balanced across the zones.
2505  */
2506 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2507                                                         int *classzone_idx)
2508 {
2509         int all_zones_ok;
2510         unsigned long balanced;
2511         int priority;
2512         int i;
2513         int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
2514         unsigned long total_scanned;
2515         struct reclaim_state *reclaim_state = current->reclaim_state;
2516         unsigned long nr_soft_reclaimed;
2517         unsigned long nr_soft_scanned;
2518         struct scan_control sc = {
2519                 .gfp_mask = GFP_KERNEL,
2520                 .may_unmap = 1,
2521                 .may_swap = 1,
2522                 /*
2523                  * kswapd doesn't want to be bailed out while reclaim. because
2524                  * we want to put equal scanning pressure on each zone.
2525                  */
2526                 .nr_to_reclaim = ULONG_MAX,
2527                 .order = order,
2528                 .mem_cgroup = NULL,
2529         };
2530         struct shrink_control shrink = {
2531                 .gfp_mask = sc.gfp_mask,
2532         };
2533 loop_again:
2534         total_scanned = 0;
2535         sc.nr_reclaimed = 0;
2536         sc.may_writepage = !laptop_mode;
2537         count_vm_event(PAGEOUTRUN);
2538
2539         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2540                 unsigned long lru_pages = 0;
2541                 int has_under_min_watermark_zone = 0;
2542
2543                 /* The swap token gets in the way of swapout... */
2544                 if (!priority)
2545                         disable_swap_token(NULL);
2546
2547                 all_zones_ok = 1;
2548                 balanced = 0;
2549
2550                 /*
2551                  * Scan in the highmem->dma direction for the highest
2552                  * zone which needs scanning
2553                  */
2554                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2555                         struct zone *zone = pgdat->node_zones + i;
2556
2557                         if (!populated_zone(zone))
2558                                 continue;
2559
2560                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2561                                 continue;
2562
2563                         /*
2564                          * Do some background aging of the anon list, to give
2565                          * pages a chance to be referenced before reclaiming.
2566                          */
2567                         if (inactive_anon_is_low(zone, &sc))
2568                                 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2569                                                         &sc, priority, 0);
2570
2571                         if (!zone_watermark_ok_safe(zone, order,
2572                                         high_wmark_pages(zone), 0, 0)) {
2573                                 end_zone = i;
2574                                 break;
2575                         } else {
2576                                 /* If balanced, clear the congested flag */
2577                                 zone_clear_flag(zone, ZONE_CONGESTED);
2578                         }
2579                 }
2580                 if (i < 0)
2581                         goto out;
2582
2583                 for (i = 0; i <= end_zone; i++) {
2584                         struct zone *zone = pgdat->node_zones + i;
2585
2586                         lru_pages += zone_reclaimable_pages(zone);
2587                 }
2588
2589                 /*
2590                  * Now scan the zone in the dma->highmem direction, stopping
2591                  * at the last zone which needs scanning.
2592                  *
2593                  * We do this because the page allocator works in the opposite
2594                  * direction.  This prevents the page allocator from allocating
2595                  * pages behind kswapd's direction of progress, which would
2596                  * cause too much scanning of the lower zones.
2597                  */
2598                 for (i = 0; i <= end_zone; i++) {
2599                         struct zone *zone = pgdat->node_zones + i;
2600                         int nr_slab;
2601                         unsigned long balance_gap;
2602
2603                         if (!populated_zone(zone))
2604                                 continue;
2605
2606                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2607                                 continue;
2608
2609                         sc.nr_scanned = 0;
2610
2611                         nr_soft_scanned = 0;
2612                         /*
2613                          * Call soft limit reclaim before calling shrink_zone.
2614                          */
2615                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2616                                                         order, sc.gfp_mask,
2617                                                         &nr_soft_scanned);
2618                         sc.nr_reclaimed += nr_soft_reclaimed;
2619                         total_scanned += nr_soft_scanned;
2620
2621                         /*
2622                          * We put equal pressure on every zone, unless
2623                          * one zone has way too many pages free
2624                          * already. The "too many pages" is defined
2625                          * as the high wmark plus a "gap" where the
2626                          * gap is either the low watermark or 1%
2627                          * of the zone, whichever is smaller.
2628                          */
2629                         balance_gap = min(low_wmark_pages(zone),
2630                                 (zone->present_pages +
2631                                         KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2632                                 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2633                         if (!zone_watermark_ok_safe(zone, order,
2634                                         high_wmark_pages(zone) + balance_gap,
2635                                         end_zone, 0)) {
2636                                 shrink_zone(priority, zone, &sc);
2637
2638                                 reclaim_state->reclaimed_slab = 0;
2639                                 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2640                                 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2641                                 total_scanned += sc.nr_scanned;
2642
2643                                 if (nr_slab == 0 && !zone_reclaimable(zone))
2644                                         zone->all_unreclaimable = 1;
2645                         }
2646
2647                         /*
2648                          * If we've done a decent amount of scanning and
2649                          * the reclaim ratio is low, start doing writepage
2650                          * even in laptop mode
2651                          */
2652                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2653                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2654                                 sc.may_writepage = 1;
2655
2656                         if (zone->all_unreclaimable) {
2657                                 if (end_zone && end_zone == i)
2658                                         end_zone--;
2659                                 continue;
2660                         }
2661
2662                         if (!zone_watermark_ok_safe(zone, order,
2663                                         high_wmark_pages(zone), end_zone, 0)) {
2664                                 all_zones_ok = 0;
2665                                 /*
2666                                  * We are still under min water mark.  This
2667                                  * means that we have a GFP_ATOMIC allocation
2668                                  * failure risk. Hurry up!
2669                                  */
2670                                 if (!zone_watermark_ok_safe(zone, order,
2671                                             min_wmark_pages(zone), end_zone, 0))
2672                                         has_under_min_watermark_zone = 1;
2673                         } else {
2674                                 /*
2675                                  * If a zone reaches its high watermark,
2676                                  * consider it to be no longer congested. It's
2677                                  * possible there are dirty pages backed by
2678                                  * congested BDIs but as pressure is relieved,
2679                                  * spectulatively avoid congestion waits
2680                                  */
2681                                 zone_clear_flag(zone, ZONE_CONGESTED);
2682                                 if (i <= *classzone_idx)
2683                                         balanced += zone->present_pages;
2684                         }
2685
2686                 }
2687                 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2688                         break;          /* kswapd: all done */
2689                 /*
2690                  * OK, kswapd is getting into trouble.  Take a nap, then take
2691                  * another pass across the zones.
2692                  */
2693                 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2694                         if (has_under_min_watermark_zone)
2695                                 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2696                         else
2697                                 congestion_wait(BLK_RW_ASYNC, HZ/10);
2698                 }
2699
2700                 /*
2701                  * We do this so kswapd doesn't build up large priorities for
2702                  * example when it is freeing in parallel with allocators. It
2703                  * matches the direct reclaim path behaviour in terms of impact
2704                  * on zone->*_priority.
2705                  */
2706                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2707                         break;
2708         }
2709 out:
2710
2711         /*
2712          * order-0: All zones must meet high watermark for a balanced node
2713          * high-order: Balanced zones must make up at least 25% of the node
2714          *             for the node to be balanced
2715          */
2716         if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2717                 cond_resched();
2718
2719                 try_to_freeze();
2720
2721                 /*
2722                  * Fragmentation may mean that the system cannot be
2723                  * rebalanced for high-order allocations in all zones.
2724                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2725                  * it means the zones have been fully scanned and are still
2726                  * not balanced. For high-order allocations, there is
2727                  * little point trying all over again as kswapd may
2728                  * infinite loop.
2729                  *
2730                  * Instead, recheck all watermarks at order-0 as they
2731                  * are the most important. If watermarks are ok, kswapd will go
2732                  * back to sleep. High-order users can still perform direct
2733                  * reclaim if they wish.
2734                  */
2735                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2736                         order = sc.order = 0;
2737
2738                 goto loop_again;
2739         }
2740
2741         /*
2742          * If kswapd was reclaiming at a higher order, it has the option of
2743          * sleeping without all zones being balanced. Before it does, it must
2744          * ensure that the watermarks for order-0 on *all* zones are met and
2745          * that the congestion flags are cleared. The congestion flag must
2746          * be cleared as kswapd is the only mechanism that clears the flag
2747          * and it is potentially going to sleep here.
2748          */
2749         if (order) {
2750                 for (i = 0; i <= end_zone; i++) {
2751                         struct zone *zone = pgdat->node_zones + i;
2752
2753                         if (!populated_zone(zone))
2754                                 continue;
2755
2756                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2757                                 continue;
2758
2759                         /* Confirm the zone is balanced for order-0 */
2760                         if (!zone_watermark_ok(zone, 0,
2761                                         high_wmark_pages(zone), 0, 0)) {
2762                                 order = sc.order = 0;
2763                                 goto loop_again;
2764                         }
2765
2766                         /* If balanced, clear the congested flag */
2767                         zone_clear_flag(zone, ZONE_CONGESTED);
2768                         if (i <= *classzone_idx)
2769                                 balanced += zone->present_pages;
2770                 }
2771         }
2772
2773         /*
2774          * Return the order we were reclaiming at so sleeping_prematurely()
2775          * makes a decision on the order we were last reclaiming at. However,
2776          * if another caller entered the allocator slow path while kswapd
2777          * was awake, order will remain at the higher level
2778          */
2779         *classzone_idx = end_zone;
2780         return order;
2781 }
2782
2783 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2784 {
2785         long remaining = 0;
2786         DEFINE_WAIT(wait);
2787
2788         if (freezing(current) || kthread_should_stop())
2789                 return;
2790
2791         prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2792
2793         /* Try to sleep for a short interval */
2794         if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2795                 remaining = schedule_timeout(HZ/10);
2796                 finish_wait(&pgdat->kswapd_wait, &wait);
2797                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2798         }
2799
2800         /*
2801          * After a short sleep, check if it was a premature sleep. If not, then
2802          * go fully to sleep until explicitly woken up.
2803          */
2804         if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2805                 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2806
2807                 /*
2808                  * vmstat counters are not perfectly accurate and the estimated
2809                  * value for counters such as NR_FREE_PAGES can deviate from the
2810                  * true value by nr_online_cpus * threshold. To avoid the zone
2811                  * watermarks being breached while under pressure, we reduce the
2812                  * per-cpu vmstat threshold while kswapd is awake and restore
2813                  * them before going back to sleep.
2814                  */
2815                 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2816                 schedule();
2817                 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2818         } else {
2819                 if (remaining)
2820                         count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2821                 else
2822                         count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2823         }
2824         finish_wait(&pgdat->kswapd_wait, &wait);
2825 }
2826
2827 /*
2828  * The background pageout daemon, started as a kernel thread
2829  * from the init process.
2830  *
2831  * This basically trickles out pages so that we have _some_
2832  * free memory available even if there is no other activity
2833  * that frees anything up. This is needed for things like routing
2834  * etc, where we otherwise might have all activity going on in
2835  * asynchronous contexts that cannot page things out.
2836  *
2837  * If there are applications that are active memory-allocators
2838  * (most normal use), this basically shouldn't matter.
2839  */
2840 static int kswapd(void *p)
2841 {
2842         unsigned long order, new_order;
2843         unsigned balanced_order;
2844         int classzone_idx, new_classzone_idx;
2845         int balanced_classzone_idx;
2846         pg_data_t *pgdat = (pg_data_t*)p;
2847         struct task_struct *tsk = current;
2848
2849         struct reclaim_state reclaim_state = {
2850                 .reclaimed_slab = 0,
2851         };
2852         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2853
2854         lockdep_set_current_reclaim_state(GFP_KERNEL);
2855
2856         if (!cpumask_empty(cpumask))
2857                 set_cpus_allowed_ptr(tsk, cpumask);
2858         current->reclaim_state = &reclaim_state;
2859
2860         /*
2861          * Tell the memory management that we're a "memory allocator",
2862          * and that if we need more memory we should get access to it
2863          * regardless (see "__alloc_pages()"). "kswapd" should
2864          * never get caught in the normal page freeing logic.
2865          *
2866          * (Kswapd normally doesn't need memory anyway, but sometimes
2867          * you need a small amount of memory in order to be able to
2868          * page out something else, and this flag essentially protects
2869          * us from recursively trying to free more memory as we're
2870          * trying to free the first piece of memory in the first place).
2871          */
2872         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2873         set_freezable();
2874
2875         order = new_order = 0;
2876         balanced_order = 0;
2877         classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
2878         balanced_classzone_idx = classzone_idx;
2879         for ( ; ; ) {
2880                 int ret;
2881
2882                 /*
2883                  * If the last balance_pgdat was unsuccessful it's unlikely a
2884                  * new request of a similar or harder type will succeed soon
2885                  * so consider going to sleep on the basis we reclaimed at
2886                  */
2887                 if (balanced_classzone_idx >= new_classzone_idx &&
2888                                         balanced_order == new_order) {
2889                         new_order = pgdat->kswapd_max_order;
2890                         new_classzone_idx = pgdat->classzone_idx;
2891                         pgdat->kswapd_max_order =  0;
2892                         pgdat->classzone_idx = pgdat->nr_zones - 1;
2893                 }
2894
2895                 if (order < new_order || classzone_idx > new_classzone_idx) {
2896                         /*
2897                          * Don't sleep if someone wants a larger 'order'
2898                          * allocation or has tigher zone constraints
2899                          */
2900                         order = new_order;
2901                         classzone_idx = new_classzone_idx;
2902                 } else {
2903                         kswapd_try_to_sleep(pgdat, balanced_order,
2904                                                 balanced_classzone_idx);
2905                         order = pgdat->kswapd_max_order;
2906                         classzone_idx = pgdat->classzone_idx;
2907                         new_order = order;
2908                         new_classzone_idx = classzone_idx;
2909                         pgdat->kswapd_max_order = 0;
2910                         pgdat->classzone_idx = pgdat->nr_zones - 1;
2911                 }
2912
2913                 ret = try_to_freeze();
2914                 if (kthread_should_stop())
2915                         break;
2916
2917                 /*
2918                  * We can speed up thawing tasks if we don't call balance_pgdat
2919                  * after returning from the refrigerator
2920                  */
2921                 if (!ret) {
2922                         trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2923                         balanced_classzone_idx = classzone_idx;
2924                         balanced_order = balance_pgdat(pgdat, order,
2925                                                 &balanced_classzone_idx);
2926                 }
2927         }
2928         return 0;
2929 }
2930
2931 /*
2932  * A zone is low on free memory, so wake its kswapd task to service it.
2933  */
2934 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
2935 {
2936         pg_data_t *pgdat;
2937
2938         if (!populated_zone(zone))
2939                 return;
2940
2941         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2942                 return;
2943         pgdat = zone->zone_pgdat;
2944         if (pgdat->kswapd_max_order < order) {
2945                 pgdat->kswapd_max_order = order;
2946                 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
2947         }
2948         if (!waitqueue_active(&pgdat->kswapd_wait))
2949                 return;
2950         if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
2951                 return;
2952
2953         trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
2954         wake_up_interruptible(&pgdat->kswapd_wait);
2955 }
2956
2957 /*
2958  * The reclaimable count would be mostly accurate.
2959  * The less reclaimable pages may be
2960  * - mlocked pages, which will be moved to unevictable list when encountered
2961  * - mapped pages, which may require several travels to be reclaimed
2962  * - dirty pages, which is not "instantly" reclaimable
2963  */
2964 unsigned long global_reclaimable_pages(void)
2965 {
2966         int nr;
2967
2968         nr = global_page_state(NR_ACTIVE_FILE) +
2969              global_page_state(NR_INACTIVE_FILE);
2970
2971         if (nr_swap_pages > 0)
2972                 nr += global_page_state(NR_ACTIVE_ANON) +
2973                       global_page_state(NR_INACTIVE_ANON);
2974
2975         return nr;
2976 }
2977
2978 unsigned long zone_reclaimable_pages(struct zone *zone)
2979 {
2980         int nr;
2981
2982         nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2983              zone_page_state(zone, NR_INACTIVE_FILE);
2984
2985         if (nr_swap_pages > 0)
2986                 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2987                       zone_page_state(zone, NR_INACTIVE_ANON);
2988
2989         return nr;
2990 }
2991
2992 #ifdef CONFIG_HIBERNATION
2993 /*
2994  * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2995  * freed pages.
2996  *
2997  * Rather than trying to age LRUs the aim is to preserve the overall
2998  * LRU order by reclaiming preferentially
2999  * inactive > active > active referenced > active mapped
3000  */
3001 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3002 {
3003         struct reclaim_state reclaim_state;
3004         struct scan_control sc = {
3005                 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3006                 .may_swap = 1,
3007                 .may_unmap = 1,
3008                 .may_writepage = 1,
3009                 .nr_to_reclaim = nr_to_reclaim,
3010                 .hibernation_mode = 1,
3011                 .order = 0,
3012         };
3013         struct shrink_control shrink = {
3014                 .gfp_mask = sc.gfp_mask,
3015         };
3016         struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3017         struct task_struct *p = current;
3018         unsigned long nr_reclaimed;
3019
3020         p->flags |= PF_MEMALLOC;
3021         lockdep_set_current_reclaim_state(sc.gfp_mask);
3022         reclaim_state.reclaimed_slab = 0;
3023         p->reclaim_state = &reclaim_state;
3024
3025         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
3026
3027         p->reclaim_state = NULL;
3028         lockdep_clear_current_reclaim_state();
3029         p->flags &= ~PF_MEMALLOC;
3030
3031         return nr_reclaimed;
3032 }
3033 #endif /* CONFIG_HIBERNATION */
3034
3035 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3036    not required for correctness.  So if the last cpu in a node goes
3037    away, we get changed to run anywhere: as the first one comes back,
3038    restore their cpu bindings. */
3039 static int __devinit cpu_callback(struct notifier_block *nfb,
3040                                   unsigned long action, void *hcpu)
3041 {
3042         int nid;
3043
3044         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3045                 for_each_node_state(nid, N_HIGH_MEMORY) {
3046                         pg_data_t *pgdat = NODE_DATA(nid);
3047                         const struct cpumask *mask;
3048
3049                         mask = cpumask_of_node(pgdat->node_id);
3050
3051                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3052                                 /* One of our CPUs online: restore mask */
3053                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3054                 }
3055         }
3056         return NOTIFY_OK;
3057 }
3058
3059 /*
3060  * This kswapd start function will be called by init and node-hot-add.
3061  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3062  */
3063 int kswapd_run(int nid)
3064 {
3065         pg_data_t *pgdat = NODE_DATA(nid);
3066         int ret = 0;
3067
3068         if (pgdat->kswapd)
3069                 return 0;
3070
3071         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3072         if (IS_ERR(pgdat->kswapd)) {
3073                 /* failure at boot is fatal */
3074                 BUG_ON(system_state == SYSTEM_BOOTING);
3075                 printk("Failed to start kswapd on node %d\n",nid);
3076                 ret = -1;
3077         }
3078         return ret;
3079 }
3080
3081 /*
3082  * Called by memory hotplug when all memory in a node is offlined.
3083  */
3084 void kswapd_stop(int nid)
3085 {
3086         struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3087
3088         if (kswapd)
3089                 kthread_stop(kswapd);
3090 }
3091
3092 static int __init kswapd_init(void)
3093 {
3094         int nid;
3095
3096         swap_setup();
3097         for_each_node_state(nid, N_HIGH_MEMORY)
3098                 kswapd_run(nid);
3099         hotcpu_notifier(cpu_callback, 0);
3100         return 0;
3101 }
3102
3103 module_init(kswapd_init)
3104
3105 #ifdef CONFIG_NUMA
3106 /*
3107  * Zone reclaim mode
3108  *
3109  * If non-zero call zone_reclaim when the number of free pages falls below
3110  * the watermarks.
3111  */
3112 int zone_reclaim_mode __read_mostly;
3113
3114 #define RECLAIM_OFF 0
3115 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
3116 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
3117 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
3118
3119 /*
3120  * Priority for ZONE_RECLAIM. This determines the fraction of pages
3121  * of a node considered for each zone_reclaim. 4 scans 1/16th of
3122  * a zone.
3123  */
3124 #define ZONE_RECLAIM_PRIORITY 4
3125
3126 /*
3127  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3128  * occur.
3129  */
3130 int sysctl_min_unmapped_ratio = 1;
3131
3132 /*
3133  * If the number of slab pages in a zone grows beyond this percentage then
3134  * slab reclaim needs to occur.
3135  */
3136 int sysctl_min_slab_ratio = 5;
3137
3138 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3139 {
3140         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3141         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3142                 zone_page_state(zone, NR_ACTIVE_FILE);
3143
3144         /*
3145          * It's possible for there to be more file mapped pages than
3146          * accounted for by the pages on the file LRU lists because
3147          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3148          */
3149         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3150 }
3151
3152 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3153 static long zone_pagecache_reclaimable(struct zone *zone)
3154 {
3155         long nr_pagecache_reclaimable;
3156         long delta = 0;
3157
3158         /*
3159          * If RECLAIM_SWAP is set, then all file pages are considered
3160          * potentially reclaimable. Otherwise, we have to worry about
3161          * pages like swapcache and zone_unmapped_file_pages() provides
3162          * a better estimate
3163          */
3164         if (zone_reclaim_mode & RECLAIM_SWAP)
3165                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3166         else
3167                 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3168
3169         /* If we can't clean pages, remove dirty pages from consideration */
3170         if (!(zone_reclaim_mode & RECLAIM_WRITE))
3171                 delta += zone_page_state(zone, NR_FILE_DIRTY);
3172
3173         /* Watch for any possible underflows due to delta */
3174         if (unlikely(delta > nr_pagecache_reclaimable))
3175                 delta = nr_pagecache_reclaimable;
3176
3177         return nr_pagecache_reclaimable - delta;
3178 }
3179
3180 /*
3181  * Try to free up some pages from this zone through reclaim.
3182  */
3183 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3184 {
3185         /* Minimum pages needed in order to stay on node */
3186         const unsigned long nr_pages = 1 << order;
3187         struct task_struct *p = current;
3188         struct reclaim_state reclaim_state;
3189         int priority;
3190         struct scan_control sc = {
3191                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3192                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3193                 .may_swap = 1,
3194                 .nr_to_reclaim = max_t(unsigned long, nr_pages,
3195                                        SWAP_CLUSTER_MAX),
3196                 .gfp_mask = gfp_mask,
3197                 .order = order,
3198         };
3199         struct shrink_control shrink = {
3200                 .gfp_mask = sc.gfp_mask,
3201         };
3202         unsigned long nr_slab_pages0, nr_slab_pages1;
3203
3204         cond_resched();
3205         /*
3206          * We need to be able to allocate from the reserves for RECLAIM_SWAP
3207          * and we also need to be able to write out pages for RECLAIM_WRITE
3208          * and RECLAIM_SWAP.
3209          */
3210         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3211         lockdep_set_current_reclaim_state(gfp_mask);
3212         reclaim_state.reclaimed_slab = 0;
3213         p->reclaim_state = &reclaim_state;
3214
3215         if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3216                 /*
3217                  * Free memory by calling shrink zone with increasing
3218                  * priorities until we have enough memory freed.
3219                  */
3220                 priority = ZONE_RECLAIM_PRIORITY;
3221                 do {
3222                         shrink_zone(priority, zone, &sc);
3223                         priority--;
3224                 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
3225         }
3226
3227         nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3228         if (nr_slab_pages0 > zone->min_slab_pages) {
3229                 /*
3230                  * shrink_slab() does not currently allow us to determine how
3231                  * many pages were freed in this zone. So we take the current
3232                  * number of slab pages and shake the slab until it is reduced
3233                  * by the same nr_pages that we used for reclaiming unmapped
3234                  * pages.
3235                  *
3236                  * Note that shrink_slab will free memory on all zones and may
3237                  * take a long time.
3238                  */
3239                 for (;;) {
3240                         unsigned long lru_pages = zone_reclaimable_pages(zone);
3241
3242                         /* No reclaimable slab or very low memory pressure */
3243                         if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3244                                 break;
3245
3246                         /* Freed enough memory */
3247                         nr_slab_pages1 = zone_page_state(zone,
3248                                                         NR_SLAB_RECLAIMABLE);
3249                         if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3250                                 break;
3251                 }
3252
3253                 /*
3254                  * Update nr_reclaimed by the number of slab pages we
3255                  * reclaimed from this zone.
3256                  */
3257                 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3258                 if (nr_slab_pages1 < nr_slab_pages0)
3259                         sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3260         }
3261
3262         p->reclaim_state = NULL;
3263         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3264         lockdep_clear_current_reclaim_state();
3265         return sc.nr_reclaimed >= nr_pages;
3266 }
3267
3268 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3269 {
3270         int node_id;
3271         int ret;
3272
3273         /*
3274          * Zone reclaim reclaims unmapped file backed pages and
3275          * slab pages if we are over the defined limits.
3276          *
3277          * A small portion of unmapped file backed pages is needed for
3278          * file I/O otherwise pages read by file I/O will be immediately
3279          * thrown out if the zone is overallocated. So we do not reclaim
3280          * if less than a specified percentage of the zone is used by
3281          * unmapped file backed pages.
3282          */
3283         if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3284             zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3285                 return ZONE_RECLAIM_FULL;
3286
3287         if (zone->all_unreclaimable)
3288                 return ZONE_RECLAIM_FULL;
3289
3290         /*
3291          * Do not scan if the allocation should not be delayed.
3292          */
3293         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3294                 return ZONE_RECLAIM_NOSCAN;
3295
3296         /*
3297          * Only run zone reclaim on the local zone or on zones that do not
3298          * have associated processors. This will favor the local processor
3299          * over remote processors and spread off node memory allocations
3300          * as wide as possible.
3301          */
3302         node_id = zone_to_nid(zone);
3303         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3304                 return ZONE_RECLAIM_NOSCAN;
3305
3306         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3307                 return ZONE_RECLAIM_NOSCAN;
3308
3309         ret = __zone_reclaim(zone, gfp_mask, order);
3310         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3311
3312         if (!ret)
3313                 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3314
3315         return ret;
3316 }
3317 #endif
3318
3319 /*
3320  * page_evictable - test whether a page is evictable
3321  * @page: the page to test
3322  * @vma: the VMA in which the page is or will be mapped, may be NULL
3323  *
3324  * Test whether page is evictable--i.e., should be placed on active/inactive
3325  * lists vs unevictable list.  The vma argument is !NULL when called from the
3326  * fault path to determine how to instantate a new page.
3327  *
3328  * Reasons page might not be evictable:
3329  * (1) page's mapping marked unevictable
3330  * (2) page is part of an mlocked VMA
3331  *
3332  */
3333 int page_evictable(struct page *page, struct vm_area_struct *vma)
3334 {
3335
3336         if (mapping_unevictable(page_mapping(page)))
3337                 return 0;
3338
3339         if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
3340                 return 0;
3341
3342         return 1;
3343 }
3344
3345 /**
3346  * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
3347  * @page: page to check evictability and move to appropriate lru list
3348  * @zone: zone page is in
3349  *
3350  * Checks a page for evictability and moves the page to the appropriate
3351  * zone lru list.
3352  *
3353  * Restrictions: zone->lru_lock must be held, page must be on LRU and must
3354  * have PageUnevictable set.
3355  */
3356 static void check_move_unevictable_page(struct page *page, struct zone *zone)
3357 {
3358         VM_BUG_ON(PageActive(page));
3359
3360 retry:
3361         ClearPageUnevictable(page);
3362         if (page_evictable(page, NULL)) {
3363                 enum lru_list l = page_lru_base_type(page);
3364
3365                 __dec_zone_state(zone, NR_UNEVICTABLE);
3366                 list_move(&page->lru, &zone->lru[l].list);
3367                 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
3368                 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
3369                 __count_vm_event(UNEVICTABLE_PGRESCUED);
3370         } else {
3371                 /*
3372                  * rotate unevictable list
3373                  */
3374                 SetPageUnevictable(page);
3375                 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
3376                 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
3377                 if (page_evictable(page, NULL))
3378                         goto retry;
3379         }
3380 }
3381
3382 /**
3383  * scan_mapping_unevictable_pages - scan an address space for evictable pages
3384  * @mapping: struct address_space to scan for evictable pages
3385  *
3386  * Scan all pages in mapping.  Check unevictable pages for
3387  * evictability and move them to the appropriate zone lru list.
3388  */
3389 void scan_mapping_unevictable_pages(struct address_space *mapping)
3390 {
3391         pgoff_t next = 0;
3392         pgoff_t end   = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
3393                          PAGE_CACHE_SHIFT;
3394         struct zone *zone;
3395         struct pagevec pvec;
3396
3397         if (mapping->nrpages == 0)
3398                 return;
3399
3400         pagevec_init(&pvec, 0);
3401         while (next < end &&
3402                 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
3403                 int i;
3404                 int pg_scanned = 0;
3405
3406                 zone = NULL;
3407
3408                 for (i = 0; i < pagevec_count(&pvec); i++) {
3409                         struct page *page = pvec.pages[i];
3410                         pgoff_t page_index = page->index;
3411                         struct zone *pagezone = page_zone(page);
3412
3413                         pg_scanned++;
3414                         if (page_index > next)
3415                                 next = page_index;
3416                         next++;
3417
3418                         if (pagezone != zone) {
3419                                 if (zone)
3420                                         spin_unlock_irq(&zone->lru_lock);
3421                                 zone = pagezone;
3422                                 spin_lock_irq(&zone->lru_lock);
3423                         }
3424
3425                         if (PageLRU(page) && PageUnevictable(page))
3426                                 check_move_unevictable_page(page, zone);
3427                 }
3428                 if (zone)
3429                         spin_unlock_irq(&zone->lru_lock);
3430                 pagevec_release(&pvec);
3431
3432                 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
3433         }
3434
3435 }
3436
3437 static void warn_scan_unevictable_pages(void)
3438 {
3439         printk_once(KERN_WARNING
3440                     "%s: The scan_unevictable_pages sysctl/node-interface has been "
3441                     "disabled for lack of a legitimate use case.  If you have "
3442                     "one, please send an email to linux-mm@kvack.org.\n",
3443                     current->comm);
3444 }
3445
3446 /*
3447  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
3448  * all nodes' unevictable lists for evictable pages
3449  */
3450 unsigned long scan_unevictable_pages;
3451
3452 int scan_unevictable_handler(struct ctl_table *table, int write,
3453                            void __user *buffer,
3454                            size_t *length, loff_t *ppos)
3455 {
3456         warn_scan_unevictable_pages();
3457         proc_doulongvec_minmax(table, write, buffer, length, ppos);
3458         scan_unevictable_pages = 0;
3459         return 0;
3460 }
3461
3462 #ifdef CONFIG_NUMA
3463 /*
3464  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
3465  * a specified node's per zone unevictable lists for evictable pages.
3466  */
3467
3468 static ssize_t read_scan_unevictable_node(struct device *dev,
3469                                           struct device_attribute *attr,
3470                                           char *buf)
3471 {
3472         warn_scan_unevictable_pages();
3473         return sprintf(buf, "0\n");     /* always zero; should fit... */
3474 }
3475
3476 static ssize_t write_scan_unevictable_node(struct device *dev,
3477                                            struct device_attribute *attr,
3478                                         const char *buf, size_t count)
3479 {
3480         warn_scan_unevictable_pages();
3481         return 1;
3482 }
3483
3484
3485 static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3486                         read_scan_unevictable_node,
3487                         write_scan_unevictable_node);
3488
3489 int scan_unevictable_register_node(struct node *node)
3490 {
3491         return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
3492 }
3493
3494 void scan_unevictable_unregister_node(struct node *node)
3495 {
3496         device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);
3497 }
3498 #endif