vmscan: fix a livelock in kswapd
[linux-2.6.git] / mm / vmscan.c
1 /*
2  *  linux/mm/vmscan.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  *
6  *  Swap reorganised 29.12.95, Stephen Tweedie.
7  *  kswapd added: 7.1.96  sct
8  *  Removed kswapd_ctl limits, and swap out as many pages as needed
9  *  to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10  *  Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11  *  Multiqueue VM started 5.8.00, Rik van Riel.
12  */
13
14 #include <linux/mm.h>
15 #include <linux/module.h>
16 #include <linux/gfp.h>
17 #include <linux/kernel_stat.h>
18 #include <linux/swap.h>
19 #include <linux/pagemap.h>
20 #include <linux/init.h>
21 #include <linux/highmem.h>
22 #include <linux/vmstat.h>
23 #include <linux/file.h>
24 #include <linux/writeback.h>
25 #include <linux/blkdev.h>
26 #include <linux/buffer_head.h>  /* for try_to_release_page(),
27                                         buffer_heads_over_limit */
28 #include <linux/mm_inline.h>
29 #include <linux/pagevec.h>
30 #include <linux/backing-dev.h>
31 #include <linux/rmap.h>
32 #include <linux/topology.h>
33 #include <linux/cpu.h>
34 #include <linux/cpuset.h>
35 #include <linux/compaction.h>
36 #include <linux/notifier.h>
37 #include <linux/rwsem.h>
38 #include <linux/delay.h>
39 #include <linux/kthread.h>
40 #include <linux/freezer.h>
41 #include <linux/memcontrol.h>
42 #include <linux/delayacct.h>
43 #include <linux/sysctl.h>
44 #include <linux/oom.h>
45 #include <linux/prefetch.h>
46
47 #include <asm/tlbflush.h>
48 #include <asm/div64.h>
49
50 #include <linux/swapops.h>
51
52 #include "internal.h"
53
54 #define CREATE_TRACE_POINTS
55 #include <trace/events/vmscan.h>
56
57 /*
58  * reclaim_mode determines how the inactive list is shrunk
59  * RECLAIM_MODE_SINGLE: Reclaim only order-0 pages
60  * RECLAIM_MODE_ASYNC:  Do not block
61  * RECLAIM_MODE_SYNC:   Allow blocking e.g. call wait_on_page_writeback
62  * RECLAIM_MODE_LUMPYRECLAIM: For high-order allocations, take a reference
63  *                      page from the LRU and reclaim all pages within a
64  *                      naturally aligned range
65  * RECLAIM_MODE_COMPACTION: For high-order allocations, reclaim a number of
66  *                      order-0 pages and then compact the zone
67  */
68 typedef unsigned __bitwise__ reclaim_mode_t;
69 #define RECLAIM_MODE_SINGLE             ((__force reclaim_mode_t)0x01u)
70 #define RECLAIM_MODE_ASYNC              ((__force reclaim_mode_t)0x02u)
71 #define RECLAIM_MODE_SYNC               ((__force reclaim_mode_t)0x04u)
72 #define RECLAIM_MODE_LUMPYRECLAIM       ((__force reclaim_mode_t)0x08u)
73 #define RECLAIM_MODE_COMPACTION         ((__force reclaim_mode_t)0x10u)
74
75 struct scan_control {
76         /* Incremented by the number of inactive pages that were scanned */
77         unsigned long nr_scanned;
78
79         /* Number of pages freed so far during a call to shrink_zones() */
80         unsigned long nr_reclaimed;
81
82         /* How many pages shrink_list() should reclaim */
83         unsigned long nr_to_reclaim;
84
85         unsigned long hibernation_mode;
86
87         /* This context's GFP mask */
88         gfp_t gfp_mask;
89
90         int may_writepage;
91
92         /* Can mapped pages be reclaimed? */
93         int may_unmap;
94
95         /* Can pages be swapped as part of reclaim? */
96         int may_swap;
97
98         int swappiness;
99
100         int order;
101
102         /*
103          * Intend to reclaim enough continuous memory rather than reclaim
104          * enough amount of memory. i.e, mode for high order allocation.
105          */
106         reclaim_mode_t reclaim_mode;
107
108         /* Which cgroup do we reclaim from */
109         struct mem_cgroup *mem_cgroup;
110
111         /*
112          * Nodemask of nodes allowed by the caller. If NULL, all nodes
113          * are scanned.
114          */
115         nodemask_t      *nodemask;
116 };
117
118 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
119
120 #ifdef ARCH_HAS_PREFETCH
121 #define prefetch_prev_lru_page(_page, _base, _field)                    \
122         do {                                                            \
123                 if ((_page)->lru.prev != _base) {                       \
124                         struct page *prev;                              \
125                                                                         \
126                         prev = lru_to_page(&(_page->lru));              \
127                         prefetch(&prev->_field);                        \
128                 }                                                       \
129         } while (0)
130 #else
131 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
132 #endif
133
134 #ifdef ARCH_HAS_PREFETCHW
135 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
136         do {                                                            \
137                 if ((_page)->lru.prev != _base) {                       \
138                         struct page *prev;                              \
139                                                                         \
140                         prev = lru_to_page(&(_page->lru));              \
141                         prefetchw(&prev->_field);                       \
142                 }                                                       \
143         } while (0)
144 #else
145 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
146 #endif
147
148 /*
149  * From 0 .. 100.  Higher means more swappy.
150  */
151 int vm_swappiness = 60;
152 long vm_total_pages;    /* The total number of pages which the VM controls */
153
154 static LIST_HEAD(shrinker_list);
155 static DECLARE_RWSEM(shrinker_rwsem);
156
157 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
158 #define scanning_global_lru(sc) (!(sc)->mem_cgroup)
159 #else
160 #define scanning_global_lru(sc) (1)
161 #endif
162
163 static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
164                                                   struct scan_control *sc)
165 {
166         if (!scanning_global_lru(sc))
167                 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
168
169         return &zone->reclaim_stat;
170 }
171
172 static unsigned long zone_nr_lru_pages(struct zone *zone,
173                                 struct scan_control *sc, enum lru_list lru)
174 {
175         if (!scanning_global_lru(sc))
176                 return mem_cgroup_zone_nr_lru_pages(sc->mem_cgroup, zone, lru);
177
178         return zone_page_state(zone, NR_LRU_BASE + lru);
179 }
180
181
182 /*
183  * Add a shrinker callback to be called from the vm
184  */
185 void register_shrinker(struct shrinker *shrinker)
186 {
187         shrinker->nr = 0;
188         down_write(&shrinker_rwsem);
189         list_add_tail(&shrinker->list, &shrinker_list);
190         up_write(&shrinker_rwsem);
191 }
192 EXPORT_SYMBOL(register_shrinker);
193
194 /*
195  * Remove one
196  */
197 void unregister_shrinker(struct shrinker *shrinker)
198 {
199         down_write(&shrinker_rwsem);
200         list_del(&shrinker->list);
201         up_write(&shrinker_rwsem);
202 }
203 EXPORT_SYMBOL(unregister_shrinker);
204
205 static inline int do_shrinker_shrink(struct shrinker *shrinker,
206                                      struct shrink_control *sc,
207                                      unsigned long nr_to_scan)
208 {
209         sc->nr_to_scan = nr_to_scan;
210         return (*shrinker->shrink)(shrinker, sc);
211 }
212
213 #define SHRINK_BATCH 128
214 /*
215  * Call the shrink functions to age shrinkable caches
216  *
217  * Here we assume it costs one seek to replace a lru page and that it also
218  * takes a seek to recreate a cache object.  With this in mind we age equal
219  * percentages of the lru and ageable caches.  This should balance the seeks
220  * generated by these structures.
221  *
222  * If the vm encountered mapped pages on the LRU it increase the pressure on
223  * slab to avoid swapping.
224  *
225  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
226  *
227  * `lru_pages' represents the number of on-LRU pages in all the zones which
228  * are eligible for the caller's allocation attempt.  It is used for balancing
229  * slab reclaim versus page reclaim.
230  *
231  * Returns the number of slab objects which we shrunk.
232  */
233 unsigned long shrink_slab(struct shrink_control *shrink,
234                           unsigned long nr_pages_scanned,
235                           unsigned long lru_pages)
236 {
237         struct shrinker *shrinker;
238         unsigned long ret = 0;
239
240         if (nr_pages_scanned == 0)
241                 nr_pages_scanned = SWAP_CLUSTER_MAX;
242
243         if (!down_read_trylock(&shrinker_rwsem)) {
244                 /* Assume we'll be able to shrink next time */
245                 ret = 1;
246                 goto out;
247         }
248
249         list_for_each_entry(shrinker, &shrinker_list, list) {
250                 unsigned long long delta;
251                 unsigned long total_scan;
252                 unsigned long max_pass;
253
254                 max_pass = do_shrinker_shrink(shrinker, shrink, 0);
255                 delta = (4 * nr_pages_scanned) / shrinker->seeks;
256                 delta *= max_pass;
257                 do_div(delta, lru_pages + 1);
258                 shrinker->nr += delta;
259                 if (shrinker->nr < 0) {
260                         printk(KERN_ERR "shrink_slab: %pF negative objects to "
261                                "delete nr=%ld\n",
262                                shrinker->shrink, shrinker->nr);
263                         shrinker->nr = max_pass;
264                 }
265
266                 /*
267                  * Avoid risking looping forever due to too large nr value:
268                  * never try to free more than twice the estimate number of
269                  * freeable entries.
270                  */
271                 if (shrinker->nr > max_pass * 2)
272                         shrinker->nr = max_pass * 2;
273
274                 total_scan = shrinker->nr;
275                 shrinker->nr = 0;
276
277                 while (total_scan >= SHRINK_BATCH) {
278                         long this_scan = SHRINK_BATCH;
279                         int shrink_ret;
280                         int nr_before;
281
282                         nr_before = do_shrinker_shrink(shrinker, shrink, 0);
283                         shrink_ret = do_shrinker_shrink(shrinker, shrink,
284                                                         this_scan);
285                         if (shrink_ret == -1)
286                                 break;
287                         if (shrink_ret < nr_before)
288                                 ret += nr_before - shrink_ret;
289                         count_vm_events(SLABS_SCANNED, this_scan);
290                         total_scan -= this_scan;
291
292                         cond_resched();
293                 }
294
295                 shrinker->nr += total_scan;
296         }
297         up_read(&shrinker_rwsem);
298 out:
299         cond_resched();
300         return ret;
301 }
302
303 static void set_reclaim_mode(int priority, struct scan_control *sc,
304                                    bool sync)
305 {
306         reclaim_mode_t syncmode = sync ? RECLAIM_MODE_SYNC : RECLAIM_MODE_ASYNC;
307
308         /*
309          * Initially assume we are entering either lumpy reclaim or
310          * reclaim/compaction.Depending on the order, we will either set the
311          * sync mode or just reclaim order-0 pages later.
312          */
313         if (COMPACTION_BUILD)
314                 sc->reclaim_mode = RECLAIM_MODE_COMPACTION;
315         else
316                 sc->reclaim_mode = RECLAIM_MODE_LUMPYRECLAIM;
317
318         /*
319          * Avoid using lumpy reclaim or reclaim/compaction if possible by
320          * restricting when its set to either costly allocations or when
321          * under memory pressure
322          */
323         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
324                 sc->reclaim_mode |= syncmode;
325         else if (sc->order && priority < DEF_PRIORITY - 2)
326                 sc->reclaim_mode |= syncmode;
327         else
328                 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
329 }
330
331 static void reset_reclaim_mode(struct scan_control *sc)
332 {
333         sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
334 }
335
336 static inline int is_page_cache_freeable(struct page *page)
337 {
338         /*
339          * A freeable page cache page is referenced only by the caller
340          * that isolated the page, the page cache radix tree and
341          * optional buffer heads at page->private.
342          */
343         return page_count(page) - page_has_private(page) == 2;
344 }
345
346 static int may_write_to_queue(struct backing_dev_info *bdi,
347                               struct scan_control *sc)
348 {
349         if (current->flags & PF_SWAPWRITE)
350                 return 1;
351         if (!bdi_write_congested(bdi))
352                 return 1;
353         if (bdi == current->backing_dev_info)
354                 return 1;
355
356         /* lumpy reclaim for hugepage often need a lot of write */
357         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
358                 return 1;
359         return 0;
360 }
361
362 /*
363  * We detected a synchronous write error writing a page out.  Probably
364  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
365  * fsync(), msync() or close().
366  *
367  * The tricky part is that after writepage we cannot touch the mapping: nothing
368  * prevents it from being freed up.  But we have a ref on the page and once
369  * that page is locked, the mapping is pinned.
370  *
371  * We're allowed to run sleeping lock_page() here because we know the caller has
372  * __GFP_FS.
373  */
374 static void handle_write_error(struct address_space *mapping,
375                                 struct page *page, int error)
376 {
377         lock_page(page);
378         if (page_mapping(page) == mapping)
379                 mapping_set_error(mapping, error);
380         unlock_page(page);
381 }
382
383 /* possible outcome of pageout() */
384 typedef enum {
385         /* failed to write page out, page is locked */
386         PAGE_KEEP,
387         /* move page to the active list, page is locked */
388         PAGE_ACTIVATE,
389         /* page has been sent to the disk successfully, page is unlocked */
390         PAGE_SUCCESS,
391         /* page is clean and locked */
392         PAGE_CLEAN,
393 } pageout_t;
394
395 /*
396  * pageout is called by shrink_page_list() for each dirty page.
397  * Calls ->writepage().
398  */
399 static pageout_t pageout(struct page *page, struct address_space *mapping,
400                          struct scan_control *sc)
401 {
402         /*
403          * If the page is dirty, only perform writeback if that write
404          * will be non-blocking.  To prevent this allocation from being
405          * stalled by pagecache activity.  But note that there may be
406          * stalls if we need to run get_block().  We could test
407          * PagePrivate for that.
408          *
409          * If this process is currently in __generic_file_aio_write() against
410          * this page's queue, we can perform writeback even if that
411          * will block.
412          *
413          * If the page is swapcache, write it back even if that would
414          * block, for some throttling. This happens by accident, because
415          * swap_backing_dev_info is bust: it doesn't reflect the
416          * congestion state of the swapdevs.  Easy to fix, if needed.
417          */
418         if (!is_page_cache_freeable(page))
419                 return PAGE_KEEP;
420         if (!mapping) {
421                 /*
422                  * Some data journaling orphaned pages can have
423                  * page->mapping == NULL while being dirty with clean buffers.
424                  */
425                 if (page_has_private(page)) {
426                         if (try_to_free_buffers(page)) {
427                                 ClearPageDirty(page);
428                                 printk("%s: orphaned page\n", __func__);
429                                 return PAGE_CLEAN;
430                         }
431                 }
432                 return PAGE_KEEP;
433         }
434         if (mapping->a_ops->writepage == NULL)
435                 return PAGE_ACTIVATE;
436         if (!may_write_to_queue(mapping->backing_dev_info, sc))
437                 return PAGE_KEEP;
438
439         if (clear_page_dirty_for_io(page)) {
440                 int res;
441                 struct writeback_control wbc = {
442                         .sync_mode = WB_SYNC_NONE,
443                         .nr_to_write = SWAP_CLUSTER_MAX,
444                         .range_start = 0,
445                         .range_end = LLONG_MAX,
446                         .for_reclaim = 1,
447                 };
448
449                 SetPageReclaim(page);
450                 res = mapping->a_ops->writepage(page, &wbc);
451                 if (res < 0)
452                         handle_write_error(mapping, page, res);
453                 if (res == AOP_WRITEPAGE_ACTIVATE) {
454                         ClearPageReclaim(page);
455                         return PAGE_ACTIVATE;
456                 }
457
458                 /*
459                  * Wait on writeback if requested to. This happens when
460                  * direct reclaiming a large contiguous area and the
461                  * first attempt to free a range of pages fails.
462                  */
463                 if (PageWriteback(page) &&
464                     (sc->reclaim_mode & RECLAIM_MODE_SYNC))
465                         wait_on_page_writeback(page);
466
467                 if (!PageWriteback(page)) {
468                         /* synchronous write or broken a_ops? */
469                         ClearPageReclaim(page);
470                 }
471                 trace_mm_vmscan_writepage(page,
472                         trace_reclaim_flags(page, sc->reclaim_mode));
473                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
474                 return PAGE_SUCCESS;
475         }
476
477         return PAGE_CLEAN;
478 }
479
480 /*
481  * Same as remove_mapping, but if the page is removed from the mapping, it
482  * gets returned with a refcount of 0.
483  */
484 static int __remove_mapping(struct address_space *mapping, struct page *page)
485 {
486         BUG_ON(!PageLocked(page));
487         BUG_ON(mapping != page_mapping(page));
488
489         spin_lock_irq(&mapping->tree_lock);
490         /*
491          * The non racy check for a busy page.
492          *
493          * Must be careful with the order of the tests. When someone has
494          * a ref to the page, it may be possible that they dirty it then
495          * drop the reference. So if PageDirty is tested before page_count
496          * here, then the following race may occur:
497          *
498          * get_user_pages(&page);
499          * [user mapping goes away]
500          * write_to(page);
501          *                              !PageDirty(page)    [good]
502          * SetPageDirty(page);
503          * put_page(page);
504          *                              !page_count(page)   [good, discard it]
505          *
506          * [oops, our write_to data is lost]
507          *
508          * Reversing the order of the tests ensures such a situation cannot
509          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
510          * load is not satisfied before that of page->_count.
511          *
512          * Note that if SetPageDirty is always performed via set_page_dirty,
513          * and thus under tree_lock, then this ordering is not required.
514          */
515         if (!page_freeze_refs(page, 2))
516                 goto cannot_free;
517         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
518         if (unlikely(PageDirty(page))) {
519                 page_unfreeze_refs(page, 2);
520                 goto cannot_free;
521         }
522
523         if (PageSwapCache(page)) {
524                 swp_entry_t swap = { .val = page_private(page) };
525                 __delete_from_swap_cache(page);
526                 spin_unlock_irq(&mapping->tree_lock);
527                 swapcache_free(swap, page);
528         } else {
529                 void (*freepage)(struct page *);
530
531                 freepage = mapping->a_ops->freepage;
532
533                 __delete_from_page_cache(page);
534                 spin_unlock_irq(&mapping->tree_lock);
535                 mem_cgroup_uncharge_cache_page(page);
536
537                 if (freepage != NULL)
538                         freepage(page);
539         }
540
541         return 1;
542
543 cannot_free:
544         spin_unlock_irq(&mapping->tree_lock);
545         return 0;
546 }
547
548 /*
549  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
550  * someone else has a ref on the page, abort and return 0.  If it was
551  * successfully detached, return 1.  Assumes the caller has a single ref on
552  * this page.
553  */
554 int remove_mapping(struct address_space *mapping, struct page *page)
555 {
556         if (__remove_mapping(mapping, page)) {
557                 /*
558                  * Unfreezing the refcount with 1 rather than 2 effectively
559                  * drops the pagecache ref for us without requiring another
560                  * atomic operation.
561                  */
562                 page_unfreeze_refs(page, 1);
563                 return 1;
564         }
565         return 0;
566 }
567
568 /**
569  * putback_lru_page - put previously isolated page onto appropriate LRU list
570  * @page: page to be put back to appropriate lru list
571  *
572  * Add previously isolated @page to appropriate LRU list.
573  * Page may still be unevictable for other reasons.
574  *
575  * lru_lock must not be held, interrupts must be enabled.
576  */
577 void putback_lru_page(struct page *page)
578 {
579         int lru;
580         int active = !!TestClearPageActive(page);
581         int was_unevictable = PageUnevictable(page);
582
583         VM_BUG_ON(PageLRU(page));
584
585 redo:
586         ClearPageUnevictable(page);
587
588         if (page_evictable(page, NULL)) {
589                 /*
590                  * For evictable pages, we can use the cache.
591                  * In event of a race, worst case is we end up with an
592                  * unevictable page on [in]active list.
593                  * We know how to handle that.
594                  */
595                 lru = active + page_lru_base_type(page);
596                 lru_cache_add_lru(page, lru);
597         } else {
598                 /*
599                  * Put unevictable pages directly on zone's unevictable
600                  * list.
601                  */
602                 lru = LRU_UNEVICTABLE;
603                 add_page_to_unevictable_list(page);
604                 /*
605                  * When racing with an mlock clearing (page is
606                  * unlocked), make sure that if the other thread does
607                  * not observe our setting of PG_lru and fails
608                  * isolation, we see PG_mlocked cleared below and move
609                  * the page back to the evictable list.
610                  *
611                  * The other side is TestClearPageMlocked().
612                  */
613                 smp_mb();
614         }
615
616         /*
617          * page's status can change while we move it among lru. If an evictable
618          * page is on unevictable list, it never be freed. To avoid that,
619          * check after we added it to the list, again.
620          */
621         if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
622                 if (!isolate_lru_page(page)) {
623                         put_page(page);
624                         goto redo;
625                 }
626                 /* This means someone else dropped this page from LRU
627                  * So, it will be freed or putback to LRU again. There is
628                  * nothing to do here.
629                  */
630         }
631
632         if (was_unevictable && lru != LRU_UNEVICTABLE)
633                 count_vm_event(UNEVICTABLE_PGRESCUED);
634         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
635                 count_vm_event(UNEVICTABLE_PGCULLED);
636
637         put_page(page);         /* drop ref from isolate */
638 }
639
640 enum page_references {
641         PAGEREF_RECLAIM,
642         PAGEREF_RECLAIM_CLEAN,
643         PAGEREF_KEEP,
644         PAGEREF_ACTIVATE,
645 };
646
647 static enum page_references page_check_references(struct page *page,
648                                                   struct scan_control *sc)
649 {
650         int referenced_ptes, referenced_page;
651         unsigned long vm_flags;
652
653         referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
654         referenced_page = TestClearPageReferenced(page);
655
656         /* Lumpy reclaim - ignore references */
657         if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
658                 return PAGEREF_RECLAIM;
659
660         /*
661          * Mlock lost the isolation race with us.  Let try_to_unmap()
662          * move the page to the unevictable list.
663          */
664         if (vm_flags & VM_LOCKED)
665                 return PAGEREF_RECLAIM;
666
667         if (referenced_ptes) {
668                 if (PageAnon(page))
669                         return PAGEREF_ACTIVATE;
670                 /*
671                  * All mapped pages start out with page table
672                  * references from the instantiating fault, so we need
673                  * to look twice if a mapped file page is used more
674                  * than once.
675                  *
676                  * Mark it and spare it for another trip around the
677                  * inactive list.  Another page table reference will
678                  * lead to its activation.
679                  *
680                  * Note: the mark is set for activated pages as well
681                  * so that recently deactivated but used pages are
682                  * quickly recovered.
683                  */
684                 SetPageReferenced(page);
685
686                 if (referenced_page)
687                         return PAGEREF_ACTIVATE;
688
689                 return PAGEREF_KEEP;
690         }
691
692         /* Reclaim if clean, defer dirty pages to writeback */
693         if (referenced_page && !PageSwapBacked(page))
694                 return PAGEREF_RECLAIM_CLEAN;
695
696         return PAGEREF_RECLAIM;
697 }
698
699 static noinline_for_stack void free_page_list(struct list_head *free_pages)
700 {
701         struct pagevec freed_pvec;
702         struct page *page, *tmp;
703
704         pagevec_init(&freed_pvec, 1);
705
706         list_for_each_entry_safe(page, tmp, free_pages, lru) {
707                 list_del(&page->lru);
708                 if (!pagevec_add(&freed_pvec, page)) {
709                         __pagevec_free(&freed_pvec);
710                         pagevec_reinit(&freed_pvec);
711                 }
712         }
713
714         pagevec_free(&freed_pvec);
715 }
716
717 /*
718  * shrink_page_list() returns the number of reclaimed pages
719  */
720 static unsigned long shrink_page_list(struct list_head *page_list,
721                                       struct zone *zone,
722                                       struct scan_control *sc)
723 {
724         LIST_HEAD(ret_pages);
725         LIST_HEAD(free_pages);
726         int pgactivate = 0;
727         unsigned long nr_dirty = 0;
728         unsigned long nr_congested = 0;
729         unsigned long nr_reclaimed = 0;
730
731         cond_resched();
732
733         while (!list_empty(page_list)) {
734                 enum page_references references;
735                 struct address_space *mapping;
736                 struct page *page;
737                 int may_enter_fs;
738
739                 cond_resched();
740
741                 page = lru_to_page(page_list);
742                 list_del(&page->lru);
743
744                 if (!trylock_page(page))
745                         goto keep;
746
747                 VM_BUG_ON(PageActive(page));
748                 VM_BUG_ON(page_zone(page) != zone);
749
750                 sc->nr_scanned++;
751
752                 if (unlikely(!page_evictable(page, NULL)))
753                         goto cull_mlocked;
754
755                 if (!sc->may_unmap && page_mapped(page))
756                         goto keep_locked;
757
758                 /* Double the slab pressure for mapped and swapcache pages */
759                 if (page_mapped(page) || PageSwapCache(page))
760                         sc->nr_scanned++;
761
762                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
763                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
764
765                 if (PageWriteback(page)) {
766                         /*
767                          * Synchronous reclaim is performed in two passes,
768                          * first an asynchronous pass over the list to
769                          * start parallel writeback, and a second synchronous
770                          * pass to wait for the IO to complete.  Wait here
771                          * for any page for which writeback has already
772                          * started.
773                          */
774                         if ((sc->reclaim_mode & RECLAIM_MODE_SYNC) &&
775                             may_enter_fs)
776                                 wait_on_page_writeback(page);
777                         else {
778                                 unlock_page(page);
779                                 goto keep_lumpy;
780                         }
781                 }
782
783                 references = page_check_references(page, sc);
784                 switch (references) {
785                 case PAGEREF_ACTIVATE:
786                         goto activate_locked;
787                 case PAGEREF_KEEP:
788                         goto keep_locked;
789                 case PAGEREF_RECLAIM:
790                 case PAGEREF_RECLAIM_CLEAN:
791                         ; /* try to reclaim the page below */
792                 }
793
794                 /*
795                  * Anonymous process memory has backing store?
796                  * Try to allocate it some swap space here.
797                  */
798                 if (PageAnon(page) && !PageSwapCache(page)) {
799                         if (!(sc->gfp_mask & __GFP_IO))
800                                 goto keep_locked;
801                         if (!add_to_swap(page))
802                                 goto activate_locked;
803                         may_enter_fs = 1;
804                 }
805
806                 mapping = page_mapping(page);
807
808                 /*
809                  * The page is mapped into the page tables of one or more
810                  * processes. Try to unmap it here.
811                  */
812                 if (page_mapped(page) && mapping) {
813                         switch (try_to_unmap(page, TTU_UNMAP)) {
814                         case SWAP_FAIL:
815                                 goto activate_locked;
816                         case SWAP_AGAIN:
817                                 goto keep_locked;
818                         case SWAP_MLOCK:
819                                 goto cull_mlocked;
820                         case SWAP_SUCCESS:
821                                 ; /* try to free the page below */
822                         }
823                 }
824
825                 if (PageDirty(page)) {
826                         nr_dirty++;
827
828                         if (references == PAGEREF_RECLAIM_CLEAN)
829                                 goto keep_locked;
830                         if (!may_enter_fs)
831                                 goto keep_locked;
832                         if (!sc->may_writepage)
833                                 goto keep_locked;
834
835                         /* Page is dirty, try to write it out here */
836                         switch (pageout(page, mapping, sc)) {
837                         case PAGE_KEEP:
838                                 nr_congested++;
839                                 goto keep_locked;
840                         case PAGE_ACTIVATE:
841                                 goto activate_locked;
842                         case PAGE_SUCCESS:
843                                 if (PageWriteback(page))
844                                         goto keep_lumpy;
845                                 if (PageDirty(page))
846                                         goto keep;
847
848                                 /*
849                                  * A synchronous write - probably a ramdisk.  Go
850                                  * ahead and try to reclaim the page.
851                                  */
852                                 if (!trylock_page(page))
853                                         goto keep;
854                                 if (PageDirty(page) || PageWriteback(page))
855                                         goto keep_locked;
856                                 mapping = page_mapping(page);
857                         case PAGE_CLEAN:
858                                 ; /* try to free the page below */
859                         }
860                 }
861
862                 /*
863                  * If the page has buffers, try to free the buffer mappings
864                  * associated with this page. If we succeed we try to free
865                  * the page as well.
866                  *
867                  * We do this even if the page is PageDirty().
868                  * try_to_release_page() does not perform I/O, but it is
869                  * possible for a page to have PageDirty set, but it is actually
870                  * clean (all its buffers are clean).  This happens if the
871                  * buffers were written out directly, with submit_bh(). ext3
872                  * will do this, as well as the blockdev mapping.
873                  * try_to_release_page() will discover that cleanness and will
874                  * drop the buffers and mark the page clean - it can be freed.
875                  *
876                  * Rarely, pages can have buffers and no ->mapping.  These are
877                  * the pages which were not successfully invalidated in
878                  * truncate_complete_page().  We try to drop those buffers here
879                  * and if that worked, and the page is no longer mapped into
880                  * process address space (page_count == 1) it can be freed.
881                  * Otherwise, leave the page on the LRU so it is swappable.
882                  */
883                 if (page_has_private(page)) {
884                         if (!try_to_release_page(page, sc->gfp_mask))
885                                 goto activate_locked;
886                         if (!mapping && page_count(page) == 1) {
887                                 unlock_page(page);
888                                 if (put_page_testzero(page))
889                                         goto free_it;
890                                 else {
891                                         /*
892                                          * rare race with speculative reference.
893                                          * the speculative reference will free
894                                          * this page shortly, so we may
895                                          * increment nr_reclaimed here (and
896                                          * leave it off the LRU).
897                                          */
898                                         nr_reclaimed++;
899                                         continue;
900                                 }
901                         }
902                 }
903
904                 if (!mapping || !__remove_mapping(mapping, page))
905                         goto keep_locked;
906
907                 /*
908                  * At this point, we have no other references and there is
909                  * no way to pick any more up (removed from LRU, removed
910                  * from pagecache). Can use non-atomic bitops now (and
911                  * we obviously don't have to worry about waking up a process
912                  * waiting on the page lock, because there are no references.
913                  */
914                 __clear_page_locked(page);
915 free_it:
916                 nr_reclaimed++;
917
918                 /*
919                  * Is there need to periodically free_page_list? It would
920                  * appear not as the counts should be low
921                  */
922                 list_add(&page->lru, &free_pages);
923                 continue;
924
925 cull_mlocked:
926                 if (PageSwapCache(page))
927                         try_to_free_swap(page);
928                 unlock_page(page);
929                 putback_lru_page(page);
930                 reset_reclaim_mode(sc);
931                 continue;
932
933 activate_locked:
934                 /* Not a candidate for swapping, so reclaim swap space. */
935                 if (PageSwapCache(page) && vm_swap_full())
936                         try_to_free_swap(page);
937                 VM_BUG_ON(PageActive(page));
938                 SetPageActive(page);
939                 pgactivate++;
940 keep_locked:
941                 unlock_page(page);
942 keep:
943                 reset_reclaim_mode(sc);
944 keep_lumpy:
945                 list_add(&page->lru, &ret_pages);
946                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
947         }
948
949         /*
950          * Tag a zone as congested if all the dirty pages encountered were
951          * backed by a congested BDI. In this case, reclaimers should just
952          * back off and wait for congestion to clear because further reclaim
953          * will encounter the same problem
954          */
955         if (nr_dirty && nr_dirty == nr_congested && scanning_global_lru(sc))
956                 zone_set_flag(zone, ZONE_CONGESTED);
957
958         free_page_list(&free_pages);
959
960         list_splice(&ret_pages, page_list);
961         count_vm_events(PGACTIVATE, pgactivate);
962         return nr_reclaimed;
963 }
964
965 /*
966  * Attempt to remove the specified page from its LRU.  Only take this page
967  * if it is of the appropriate PageActive status.  Pages which are being
968  * freed elsewhere are also ignored.
969  *
970  * page:        page to consider
971  * mode:        one of the LRU isolation modes defined above
972  *
973  * returns 0 on success, -ve errno on failure.
974  */
975 int __isolate_lru_page(struct page *page, int mode, int file)
976 {
977         int ret = -EINVAL;
978
979         /* Only take pages on the LRU. */
980         if (!PageLRU(page))
981                 return ret;
982
983         /*
984          * When checking the active state, we need to be sure we are
985          * dealing with comparible boolean values.  Take the logical not
986          * of each.
987          */
988         if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
989                 return ret;
990
991         if (mode != ISOLATE_BOTH && page_is_file_cache(page) != file)
992                 return ret;
993
994         /*
995          * When this function is being called for lumpy reclaim, we
996          * initially look into all LRU pages, active, inactive and
997          * unevictable; only give shrink_page_list evictable pages.
998          */
999         if (PageUnevictable(page))
1000                 return ret;
1001
1002         ret = -EBUSY;
1003
1004         if (likely(get_page_unless_zero(page))) {
1005                 /*
1006                  * Be careful not to clear PageLRU until after we're
1007                  * sure the page is not being freed elsewhere -- the
1008                  * page release code relies on it.
1009                  */
1010                 ClearPageLRU(page);
1011                 ret = 0;
1012         }
1013
1014         return ret;
1015 }
1016
1017 /*
1018  * zone->lru_lock is heavily contended.  Some of the functions that
1019  * shrink the lists perform better by taking out a batch of pages
1020  * and working on them outside the LRU lock.
1021  *
1022  * For pagecache intensive workloads, this function is the hottest
1023  * spot in the kernel (apart from copy_*_user functions).
1024  *
1025  * Appropriate locks must be held before calling this function.
1026  *
1027  * @nr_to_scan: The number of pages to look through on the list.
1028  * @src:        The LRU list to pull pages off.
1029  * @dst:        The temp list to put pages on to.
1030  * @scanned:    The number of pages that were scanned.
1031  * @order:      The caller's attempted allocation order
1032  * @mode:       One of the LRU isolation modes
1033  * @file:       True [1] if isolating file [!anon] pages
1034  *
1035  * returns how many pages were moved onto *@dst.
1036  */
1037 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1038                 struct list_head *src, struct list_head *dst,
1039                 unsigned long *scanned, int order, int mode, int file)
1040 {
1041         unsigned long nr_taken = 0;
1042         unsigned long nr_lumpy_taken = 0;
1043         unsigned long nr_lumpy_dirty = 0;
1044         unsigned long nr_lumpy_failed = 0;
1045         unsigned long scan;
1046
1047         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1048                 struct page *page;
1049                 unsigned long pfn;
1050                 unsigned long end_pfn;
1051                 unsigned long page_pfn;
1052                 int zone_id;
1053
1054                 page = lru_to_page(src);
1055                 prefetchw_prev_lru_page(page, src, flags);
1056
1057                 VM_BUG_ON(!PageLRU(page));
1058
1059                 switch (__isolate_lru_page(page, mode, file)) {
1060                 case 0:
1061                         list_move(&page->lru, dst);
1062                         mem_cgroup_del_lru(page);
1063                         nr_taken += hpage_nr_pages(page);
1064                         break;
1065
1066                 case -EBUSY:
1067                         /* else it is being freed elsewhere */
1068                         list_move(&page->lru, src);
1069                         mem_cgroup_rotate_lru_list(page, page_lru(page));
1070                         continue;
1071
1072                 default:
1073                         BUG();
1074                 }
1075
1076                 if (!order)
1077                         continue;
1078
1079                 /*
1080                  * Attempt to take all pages in the order aligned region
1081                  * surrounding the tag page.  Only take those pages of
1082                  * the same active state as that tag page.  We may safely
1083                  * round the target page pfn down to the requested order
1084                  * as the mem_map is guaranteed valid out to MAX_ORDER,
1085                  * where that page is in a different zone we will detect
1086                  * it from its zone id and abort this block scan.
1087                  */
1088                 zone_id = page_zone_id(page);
1089                 page_pfn = page_to_pfn(page);
1090                 pfn = page_pfn & ~((1 << order) - 1);
1091                 end_pfn = pfn + (1 << order);
1092                 for (; pfn < end_pfn; pfn++) {
1093                         struct page *cursor_page;
1094
1095                         /* The target page is in the block, ignore it. */
1096                         if (unlikely(pfn == page_pfn))
1097                                 continue;
1098
1099                         /* Avoid holes within the zone. */
1100                         if (unlikely(!pfn_valid_within(pfn)))
1101                                 break;
1102
1103                         cursor_page = pfn_to_page(pfn);
1104
1105                         /* Check that we have not crossed a zone boundary. */
1106                         if (unlikely(page_zone_id(cursor_page) != zone_id))
1107                                 break;
1108
1109                         /*
1110                          * If we don't have enough swap space, reclaiming of
1111                          * anon page which don't already have a swap slot is
1112                          * pointless.
1113                          */
1114                         if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
1115                             !PageSwapCache(cursor_page))
1116                                 break;
1117
1118                         if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1119                                 list_move(&cursor_page->lru, dst);
1120                                 mem_cgroup_del_lru(cursor_page);
1121                                 nr_taken += hpage_nr_pages(page);
1122                                 nr_lumpy_taken++;
1123                                 if (PageDirty(cursor_page))
1124                                         nr_lumpy_dirty++;
1125                                 scan++;
1126                         } else {
1127                                 /*
1128                                  * Check if the page is freed already.
1129                                  *
1130                                  * We can't use page_count() as that
1131                                  * requires compound_head and we don't
1132                                  * have a pin on the page here. If a
1133                                  * page is tail, we may or may not
1134                                  * have isolated the head, so assume
1135                                  * it's not free, it'd be tricky to
1136                                  * track the head status without a
1137                                  * page pin.
1138                                  */
1139                                 if (!PageTail(cursor_page) &&
1140                                     !atomic_read(&cursor_page->_count))
1141                                         continue;
1142                                 break;
1143                         }
1144                 }
1145
1146                 /* If we break out of the loop above, lumpy reclaim failed */
1147                 if (pfn < end_pfn)
1148                         nr_lumpy_failed++;
1149         }
1150
1151         *scanned = scan;
1152
1153         trace_mm_vmscan_lru_isolate(order,
1154                         nr_to_scan, scan,
1155                         nr_taken,
1156                         nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1157                         mode);
1158         return nr_taken;
1159 }
1160
1161 static unsigned long isolate_pages_global(unsigned long nr,
1162                                         struct list_head *dst,
1163                                         unsigned long *scanned, int order,
1164                                         int mode, struct zone *z,
1165                                         int active, int file)
1166 {
1167         int lru = LRU_BASE;
1168         if (active)
1169                 lru += LRU_ACTIVE;
1170         if (file)
1171                 lru += LRU_FILE;
1172         return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1173                                                                 mode, file);
1174 }
1175
1176 /*
1177  * clear_active_flags() is a helper for shrink_active_list(), clearing
1178  * any active bits from the pages in the list.
1179  */
1180 static unsigned long clear_active_flags(struct list_head *page_list,
1181                                         unsigned int *count)
1182 {
1183         int nr_active = 0;
1184         int lru;
1185         struct page *page;
1186
1187         list_for_each_entry(page, page_list, lru) {
1188                 int numpages = hpage_nr_pages(page);
1189                 lru = page_lru_base_type(page);
1190                 if (PageActive(page)) {
1191                         lru += LRU_ACTIVE;
1192                         ClearPageActive(page);
1193                         nr_active += numpages;
1194                 }
1195                 if (count)
1196                         count[lru] += numpages;
1197         }
1198
1199         return nr_active;
1200 }
1201
1202 /**
1203  * isolate_lru_page - tries to isolate a page from its LRU list
1204  * @page: page to isolate from its LRU list
1205  *
1206  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1207  * vmstat statistic corresponding to whatever LRU list the page was on.
1208  *
1209  * Returns 0 if the page was removed from an LRU list.
1210  * Returns -EBUSY if the page was not on an LRU list.
1211  *
1212  * The returned page will have PageLRU() cleared.  If it was found on
1213  * the active list, it will have PageActive set.  If it was found on
1214  * the unevictable list, it will have the PageUnevictable bit set. That flag
1215  * may need to be cleared by the caller before letting the page go.
1216  *
1217  * The vmstat statistic corresponding to the list on which the page was
1218  * found will be decremented.
1219  *
1220  * Restrictions:
1221  * (1) Must be called with an elevated refcount on the page. This is a
1222  *     fundamentnal difference from isolate_lru_pages (which is called
1223  *     without a stable reference).
1224  * (2) the lru_lock must not be held.
1225  * (3) interrupts must be enabled.
1226  */
1227 int isolate_lru_page(struct page *page)
1228 {
1229         int ret = -EBUSY;
1230
1231         VM_BUG_ON(!page_count(page));
1232
1233         if (PageLRU(page)) {
1234                 struct zone *zone = page_zone(page);
1235
1236                 spin_lock_irq(&zone->lru_lock);
1237                 if (PageLRU(page)) {
1238                         int lru = page_lru(page);
1239                         ret = 0;
1240                         get_page(page);
1241                         ClearPageLRU(page);
1242
1243                         del_page_from_lru_list(zone, page, lru);
1244                 }
1245                 spin_unlock_irq(&zone->lru_lock);
1246         }
1247         return ret;
1248 }
1249
1250 /*
1251  * Are there way too many processes in the direct reclaim path already?
1252  */
1253 static int too_many_isolated(struct zone *zone, int file,
1254                 struct scan_control *sc)
1255 {
1256         unsigned long inactive, isolated;
1257
1258         if (current_is_kswapd())
1259                 return 0;
1260
1261         if (!scanning_global_lru(sc))
1262                 return 0;
1263
1264         if (file) {
1265                 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1266                 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1267         } else {
1268                 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1269                 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1270         }
1271
1272         return isolated > inactive;
1273 }
1274
1275 /*
1276  * TODO: Try merging with migrations version of putback_lru_pages
1277  */
1278 static noinline_for_stack void
1279 putback_lru_pages(struct zone *zone, struct scan_control *sc,
1280                                 unsigned long nr_anon, unsigned long nr_file,
1281                                 struct list_head *page_list)
1282 {
1283         struct page *page;
1284         struct pagevec pvec;
1285         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1286
1287         pagevec_init(&pvec, 1);
1288
1289         /*
1290          * Put back any unfreeable pages.
1291          */
1292         spin_lock(&zone->lru_lock);
1293         while (!list_empty(page_list)) {
1294                 int lru;
1295                 page = lru_to_page(page_list);
1296                 VM_BUG_ON(PageLRU(page));
1297                 list_del(&page->lru);
1298                 if (unlikely(!page_evictable(page, NULL))) {
1299                         spin_unlock_irq(&zone->lru_lock);
1300                         putback_lru_page(page);
1301                         spin_lock_irq(&zone->lru_lock);
1302                         continue;
1303                 }
1304                 SetPageLRU(page);
1305                 lru = page_lru(page);
1306                 add_page_to_lru_list(zone, page, lru);
1307                 if (is_active_lru(lru)) {
1308                         int file = is_file_lru(lru);
1309                         int numpages = hpage_nr_pages(page);
1310                         reclaim_stat->recent_rotated[file] += numpages;
1311                 }
1312                 if (!pagevec_add(&pvec, page)) {
1313                         spin_unlock_irq(&zone->lru_lock);
1314                         __pagevec_release(&pvec);
1315                         spin_lock_irq(&zone->lru_lock);
1316                 }
1317         }
1318         __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1319         __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1320
1321         spin_unlock_irq(&zone->lru_lock);
1322         pagevec_release(&pvec);
1323 }
1324
1325 static noinline_for_stack void update_isolated_counts(struct zone *zone,
1326                                         struct scan_control *sc,
1327                                         unsigned long *nr_anon,
1328                                         unsigned long *nr_file,
1329                                         struct list_head *isolated_list)
1330 {
1331         unsigned long nr_active;
1332         unsigned int count[NR_LRU_LISTS] = { 0, };
1333         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1334
1335         nr_active = clear_active_flags(isolated_list, count);
1336         __count_vm_events(PGDEACTIVATE, nr_active);
1337
1338         __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1339                               -count[LRU_ACTIVE_FILE]);
1340         __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1341                               -count[LRU_INACTIVE_FILE]);
1342         __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1343                               -count[LRU_ACTIVE_ANON]);
1344         __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1345                               -count[LRU_INACTIVE_ANON]);
1346
1347         *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1348         *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1349         __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1350         __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1351
1352         reclaim_stat->recent_scanned[0] += *nr_anon;
1353         reclaim_stat->recent_scanned[1] += *nr_file;
1354 }
1355
1356 /*
1357  * Returns true if the caller should wait to clean dirty/writeback pages.
1358  *
1359  * If we are direct reclaiming for contiguous pages and we do not reclaim
1360  * everything in the list, try again and wait for writeback IO to complete.
1361  * This will stall high-order allocations noticeably. Only do that when really
1362  * need to free the pages under high memory pressure.
1363  */
1364 static inline bool should_reclaim_stall(unsigned long nr_taken,
1365                                         unsigned long nr_freed,
1366                                         int priority,
1367                                         struct scan_control *sc)
1368 {
1369         int lumpy_stall_priority;
1370
1371         /* kswapd should not stall on sync IO */
1372         if (current_is_kswapd())
1373                 return false;
1374
1375         /* Only stall on lumpy reclaim */
1376         if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
1377                 return false;
1378
1379         /* If we have relaimed everything on the isolated list, no stall */
1380         if (nr_freed == nr_taken)
1381                 return false;
1382
1383         /*
1384          * For high-order allocations, there are two stall thresholds.
1385          * High-cost allocations stall immediately where as lower
1386          * order allocations such as stacks require the scanning
1387          * priority to be much higher before stalling.
1388          */
1389         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1390                 lumpy_stall_priority = DEF_PRIORITY;
1391         else
1392                 lumpy_stall_priority = DEF_PRIORITY / 3;
1393
1394         return priority <= lumpy_stall_priority;
1395 }
1396
1397 /*
1398  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1399  * of reclaimed pages
1400  */
1401 static noinline_for_stack unsigned long
1402 shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1403                         struct scan_control *sc, int priority, int file)
1404 {
1405         LIST_HEAD(page_list);
1406         unsigned long nr_scanned;
1407         unsigned long nr_reclaimed = 0;
1408         unsigned long nr_taken;
1409         unsigned long nr_anon;
1410         unsigned long nr_file;
1411
1412         while (unlikely(too_many_isolated(zone, file, sc))) {
1413                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1414
1415                 /* We are about to die and free our memory. Return now. */
1416                 if (fatal_signal_pending(current))
1417                         return SWAP_CLUSTER_MAX;
1418         }
1419
1420         set_reclaim_mode(priority, sc, false);
1421         lru_add_drain();
1422         spin_lock_irq(&zone->lru_lock);
1423
1424         if (scanning_global_lru(sc)) {
1425                 nr_taken = isolate_pages_global(nr_to_scan,
1426                         &page_list, &nr_scanned, sc->order,
1427                         sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM ?
1428                                         ISOLATE_BOTH : ISOLATE_INACTIVE,
1429                         zone, 0, file);
1430                 zone->pages_scanned += nr_scanned;
1431                 if (current_is_kswapd())
1432                         __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1433                                                nr_scanned);
1434                 else
1435                         __count_zone_vm_events(PGSCAN_DIRECT, zone,
1436                                                nr_scanned);
1437         } else {
1438                 nr_taken = mem_cgroup_isolate_pages(nr_to_scan,
1439                         &page_list, &nr_scanned, sc->order,
1440                         sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM ?
1441                                         ISOLATE_BOTH : ISOLATE_INACTIVE,
1442                         zone, sc->mem_cgroup,
1443                         0, file);
1444                 /*
1445                  * mem_cgroup_isolate_pages() keeps track of
1446                  * scanned pages on its own.
1447                  */
1448         }
1449
1450         if (nr_taken == 0) {
1451                 spin_unlock_irq(&zone->lru_lock);
1452                 return 0;
1453         }
1454
1455         update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
1456
1457         spin_unlock_irq(&zone->lru_lock);
1458
1459         nr_reclaimed = shrink_page_list(&page_list, zone, sc);
1460
1461         /* Check if we should syncronously wait for writeback */
1462         if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1463                 set_reclaim_mode(priority, sc, true);
1464                 nr_reclaimed += shrink_page_list(&page_list, zone, sc);
1465         }
1466
1467         local_irq_disable();
1468         if (current_is_kswapd())
1469                 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1470         __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1471
1472         putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
1473
1474         trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1475                 zone_idx(zone),
1476                 nr_scanned, nr_reclaimed,
1477                 priority,
1478                 trace_shrink_flags(file, sc->reclaim_mode));
1479         return nr_reclaimed;
1480 }
1481
1482 /*
1483  * This moves pages from the active list to the inactive list.
1484  *
1485  * We move them the other way if the page is referenced by one or more
1486  * processes, from rmap.
1487  *
1488  * If the pages are mostly unmapped, the processing is fast and it is
1489  * appropriate to hold zone->lru_lock across the whole operation.  But if
1490  * the pages are mapped, the processing is slow (page_referenced()) so we
1491  * should drop zone->lru_lock around each page.  It's impossible to balance
1492  * this, so instead we remove the pages from the LRU while processing them.
1493  * It is safe to rely on PG_active against the non-LRU pages in here because
1494  * nobody will play with that bit on a non-LRU page.
1495  *
1496  * The downside is that we have to touch page->_count against each page.
1497  * But we had to alter page->flags anyway.
1498  */
1499
1500 static void move_active_pages_to_lru(struct zone *zone,
1501                                      struct list_head *list,
1502                                      enum lru_list lru)
1503 {
1504         unsigned long pgmoved = 0;
1505         struct pagevec pvec;
1506         struct page *page;
1507
1508         pagevec_init(&pvec, 1);
1509
1510         while (!list_empty(list)) {
1511                 page = lru_to_page(list);
1512
1513                 VM_BUG_ON(PageLRU(page));
1514                 SetPageLRU(page);
1515
1516                 list_move(&page->lru, &zone->lru[lru].list);
1517                 mem_cgroup_add_lru_list(page, lru);
1518                 pgmoved += hpage_nr_pages(page);
1519
1520                 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1521                         spin_unlock_irq(&zone->lru_lock);
1522                         if (buffer_heads_over_limit)
1523                                 pagevec_strip(&pvec);
1524                         __pagevec_release(&pvec);
1525                         spin_lock_irq(&zone->lru_lock);
1526                 }
1527         }
1528         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1529         if (!is_active_lru(lru))
1530                 __count_vm_events(PGDEACTIVATE, pgmoved);
1531 }
1532
1533 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1534                         struct scan_control *sc, int priority, int file)
1535 {
1536         unsigned long nr_taken;
1537         unsigned long pgscanned;
1538         unsigned long vm_flags;
1539         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1540         LIST_HEAD(l_active);
1541         LIST_HEAD(l_inactive);
1542         struct page *page;
1543         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1544         unsigned long nr_rotated = 0;
1545
1546         lru_add_drain();
1547         spin_lock_irq(&zone->lru_lock);
1548         if (scanning_global_lru(sc)) {
1549                 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1550                                                 &pgscanned, sc->order,
1551                                                 ISOLATE_ACTIVE, zone,
1552                                                 1, file);
1553                 zone->pages_scanned += pgscanned;
1554         } else {
1555                 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1556                                                 &pgscanned, sc->order,
1557                                                 ISOLATE_ACTIVE, zone,
1558                                                 sc->mem_cgroup, 1, file);
1559                 /*
1560                  * mem_cgroup_isolate_pages() keeps track of
1561                  * scanned pages on its own.
1562                  */
1563         }
1564
1565         reclaim_stat->recent_scanned[file] += nr_taken;
1566
1567         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1568         if (file)
1569                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1570         else
1571                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1572         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1573         spin_unlock_irq(&zone->lru_lock);
1574
1575         while (!list_empty(&l_hold)) {
1576                 cond_resched();
1577                 page = lru_to_page(&l_hold);
1578                 list_del(&page->lru);
1579
1580                 if (unlikely(!page_evictable(page, NULL))) {
1581                         putback_lru_page(page);
1582                         continue;
1583                 }
1584
1585                 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1586                         nr_rotated += hpage_nr_pages(page);
1587                         /*
1588                          * Identify referenced, file-backed active pages and
1589                          * give them one more trip around the active list. So
1590                          * that executable code get better chances to stay in
1591                          * memory under moderate memory pressure.  Anon pages
1592                          * are not likely to be evicted by use-once streaming
1593                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1594                          * so we ignore them here.
1595                          */
1596                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1597                                 list_add(&page->lru, &l_active);
1598                                 continue;
1599                         }
1600                 }
1601
1602                 ClearPageActive(page);  /* we are de-activating */
1603                 list_add(&page->lru, &l_inactive);
1604         }
1605
1606         /*
1607          * Move pages back to the lru list.
1608          */
1609         spin_lock_irq(&zone->lru_lock);
1610         /*
1611          * Count referenced pages from currently used mappings as rotated,
1612          * even though only some of them are actually re-activated.  This
1613          * helps balance scan pressure between file and anonymous pages in
1614          * get_scan_ratio.
1615          */
1616         reclaim_stat->recent_rotated[file] += nr_rotated;
1617
1618         move_active_pages_to_lru(zone, &l_active,
1619                                                 LRU_ACTIVE + file * LRU_FILE);
1620         move_active_pages_to_lru(zone, &l_inactive,
1621                                                 LRU_BASE   + file * LRU_FILE);
1622         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1623         spin_unlock_irq(&zone->lru_lock);
1624 }
1625
1626 #ifdef CONFIG_SWAP
1627 static int inactive_anon_is_low_global(struct zone *zone)
1628 {
1629         unsigned long active, inactive;
1630
1631         active = zone_page_state(zone, NR_ACTIVE_ANON);
1632         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1633
1634         if (inactive * zone->inactive_ratio < active)
1635                 return 1;
1636
1637         return 0;
1638 }
1639
1640 /**
1641  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1642  * @zone: zone to check
1643  * @sc:   scan control of this context
1644  *
1645  * Returns true if the zone does not have enough inactive anon pages,
1646  * meaning some active anon pages need to be deactivated.
1647  */
1648 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1649 {
1650         int low;
1651
1652         /*
1653          * If we don't have swap space, anonymous page deactivation
1654          * is pointless.
1655          */
1656         if (!total_swap_pages)
1657                 return 0;
1658
1659         if (scanning_global_lru(sc))
1660                 low = inactive_anon_is_low_global(zone);
1661         else
1662                 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1663         return low;
1664 }
1665 #else
1666 static inline int inactive_anon_is_low(struct zone *zone,
1667                                         struct scan_control *sc)
1668 {
1669         return 0;
1670 }
1671 #endif
1672
1673 static int inactive_file_is_low_global(struct zone *zone)
1674 {
1675         unsigned long active, inactive;
1676
1677         active = zone_page_state(zone, NR_ACTIVE_FILE);
1678         inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1679
1680         return (active > inactive);
1681 }
1682
1683 /**
1684  * inactive_file_is_low - check if file pages need to be deactivated
1685  * @zone: zone to check
1686  * @sc:   scan control of this context
1687  *
1688  * When the system is doing streaming IO, memory pressure here
1689  * ensures that active file pages get deactivated, until more
1690  * than half of the file pages are on the inactive list.
1691  *
1692  * Once we get to that situation, protect the system's working
1693  * set from being evicted by disabling active file page aging.
1694  *
1695  * This uses a different ratio than the anonymous pages, because
1696  * the page cache uses a use-once replacement algorithm.
1697  */
1698 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1699 {
1700         int low;
1701
1702         if (scanning_global_lru(sc))
1703                 low = inactive_file_is_low_global(zone);
1704         else
1705                 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1706         return low;
1707 }
1708
1709 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1710                                 int file)
1711 {
1712         if (file)
1713                 return inactive_file_is_low(zone, sc);
1714         else
1715                 return inactive_anon_is_low(zone, sc);
1716 }
1717
1718 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1719         struct zone *zone, struct scan_control *sc, int priority)
1720 {
1721         int file = is_file_lru(lru);
1722
1723         if (is_active_lru(lru)) {
1724                 if (inactive_list_is_low(zone, sc, file))
1725                     shrink_active_list(nr_to_scan, zone, sc, priority, file);
1726                 return 0;
1727         }
1728
1729         return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1730 }
1731
1732 /*
1733  * Determine how aggressively the anon and file LRU lists should be
1734  * scanned.  The relative value of each set of LRU lists is determined
1735  * by looking at the fraction of the pages scanned we did rotate back
1736  * onto the active list instead of evict.
1737  *
1738  * nr[0] = anon pages to scan; nr[1] = file pages to scan
1739  */
1740 static void get_scan_count(struct zone *zone, struct scan_control *sc,
1741                                         unsigned long *nr, int priority)
1742 {
1743         unsigned long anon, file, free;
1744         unsigned long anon_prio, file_prio;
1745         unsigned long ap, fp;
1746         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1747         u64 fraction[2], denominator;
1748         enum lru_list l;
1749         int noswap = 0;
1750         int force_scan = 0;
1751
1752
1753         anon  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1754                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1755         file  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1756                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1757
1758         if (((anon + file) >> priority) < SWAP_CLUSTER_MAX) {
1759                 /* kswapd does zone balancing and need to scan this zone */
1760                 if (scanning_global_lru(sc) && current_is_kswapd())
1761                         force_scan = 1;
1762                 /* memcg may have small limit and need to avoid priority drop */
1763                 if (!scanning_global_lru(sc))
1764                         force_scan = 1;
1765         }
1766
1767         /* If we have no swap space, do not bother scanning anon pages. */
1768         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1769                 noswap = 1;
1770                 fraction[0] = 0;
1771                 fraction[1] = 1;
1772                 denominator = 1;
1773                 goto out;
1774         }
1775
1776         if (scanning_global_lru(sc)) {
1777                 free  = zone_page_state(zone, NR_FREE_PAGES);
1778                 /* If we have very few page cache pages,
1779                    force-scan anon pages. */
1780                 if (unlikely(file + free <= high_wmark_pages(zone))) {
1781                         fraction[0] = 1;
1782                         fraction[1] = 0;
1783                         denominator = 1;
1784                         goto out;
1785                 }
1786         }
1787
1788         /*
1789          * With swappiness at 100, anonymous and file have the same priority.
1790          * This scanning priority is essentially the inverse of IO cost.
1791          */
1792         anon_prio = sc->swappiness;
1793         file_prio = 200 - sc->swappiness;
1794
1795         /*
1796          * OK, so we have swap space and a fair amount of page cache
1797          * pages.  We use the recently rotated / recently scanned
1798          * ratios to determine how valuable each cache is.
1799          *
1800          * Because workloads change over time (and to avoid overflow)
1801          * we keep these statistics as a floating average, which ends
1802          * up weighing recent references more than old ones.
1803          *
1804          * anon in [0], file in [1]
1805          */
1806         spin_lock_irq(&zone->lru_lock);
1807         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1808                 reclaim_stat->recent_scanned[0] /= 2;
1809                 reclaim_stat->recent_rotated[0] /= 2;
1810         }
1811
1812         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1813                 reclaim_stat->recent_scanned[1] /= 2;
1814                 reclaim_stat->recent_rotated[1] /= 2;
1815         }
1816
1817         /*
1818          * The amount of pressure on anon vs file pages is inversely
1819          * proportional to the fraction of recently scanned pages on
1820          * each list that were recently referenced and in active use.
1821          */
1822         ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1823         ap /= reclaim_stat->recent_rotated[0] + 1;
1824
1825         fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1826         fp /= reclaim_stat->recent_rotated[1] + 1;
1827         spin_unlock_irq(&zone->lru_lock);
1828
1829         fraction[0] = ap;
1830         fraction[1] = fp;
1831         denominator = ap + fp + 1;
1832 out:
1833         for_each_evictable_lru(l) {
1834                 int file = is_file_lru(l);
1835                 unsigned long scan;
1836
1837                 scan = zone_nr_lru_pages(zone, sc, l);
1838                 if (priority || noswap) {
1839                         scan >>= priority;
1840                         scan = div64_u64(scan * fraction[file], denominator);
1841                 }
1842
1843                 /*
1844                  * If zone is small or memcg is small, nr[l] can be 0.
1845                  * This results no-scan on this priority and priority drop down.
1846                  * For global direct reclaim, it can visit next zone and tend
1847                  * not to have problems. For global kswapd, it's for zone
1848                  * balancing and it need to scan a small amounts. When using
1849                  * memcg, priority drop can cause big latency. So, it's better
1850                  * to scan small amount. See may_noscan above.
1851                  */
1852                 if (!scan && force_scan) {
1853                         if (file)
1854                                 scan = SWAP_CLUSTER_MAX;
1855                         else if (!noswap)
1856                                 scan = SWAP_CLUSTER_MAX;
1857                 }
1858                 nr[l] = scan;
1859         }
1860 }
1861
1862 /*
1863  * Reclaim/compaction depends on a number of pages being freed. To avoid
1864  * disruption to the system, a small number of order-0 pages continue to be
1865  * rotated and reclaimed in the normal fashion. However, by the time we get
1866  * back to the allocator and call try_to_compact_zone(), we ensure that
1867  * there are enough free pages for it to be likely successful
1868  */
1869 static inline bool should_continue_reclaim(struct zone *zone,
1870                                         unsigned long nr_reclaimed,
1871                                         unsigned long nr_scanned,
1872                                         struct scan_control *sc)
1873 {
1874         unsigned long pages_for_compaction;
1875         unsigned long inactive_lru_pages;
1876
1877         /* If not in reclaim/compaction mode, stop */
1878         if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
1879                 return false;
1880
1881         /* Consider stopping depending on scan and reclaim activity */
1882         if (sc->gfp_mask & __GFP_REPEAT) {
1883                 /*
1884                  * For __GFP_REPEAT allocations, stop reclaiming if the
1885                  * full LRU list has been scanned and we are still failing
1886                  * to reclaim pages. This full LRU scan is potentially
1887                  * expensive but a __GFP_REPEAT caller really wants to succeed
1888                  */
1889                 if (!nr_reclaimed && !nr_scanned)
1890                         return false;
1891         } else {
1892                 /*
1893                  * For non-__GFP_REPEAT allocations which can presumably
1894                  * fail without consequence, stop if we failed to reclaim
1895                  * any pages from the last SWAP_CLUSTER_MAX number of
1896                  * pages that were scanned. This will return to the
1897                  * caller faster at the risk reclaim/compaction and
1898                  * the resulting allocation attempt fails
1899                  */
1900                 if (!nr_reclaimed)
1901                         return false;
1902         }
1903
1904         /*
1905          * If we have not reclaimed enough pages for compaction and the
1906          * inactive lists are large enough, continue reclaiming
1907          */
1908         pages_for_compaction = (2UL << sc->order);
1909         inactive_lru_pages = zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON) +
1910                                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1911         if (sc->nr_reclaimed < pages_for_compaction &&
1912                         inactive_lru_pages > pages_for_compaction)
1913                 return true;
1914
1915         /* If compaction would go ahead or the allocation would succeed, stop */
1916         switch (compaction_suitable(zone, sc->order)) {
1917         case COMPACT_PARTIAL:
1918         case COMPACT_CONTINUE:
1919                 return false;
1920         default:
1921                 return true;
1922         }
1923 }
1924
1925 /*
1926  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1927  */
1928 static void shrink_zone(int priority, struct zone *zone,
1929                                 struct scan_control *sc)
1930 {
1931         unsigned long nr[NR_LRU_LISTS];
1932         unsigned long nr_to_scan;
1933         enum lru_list l;
1934         unsigned long nr_reclaimed, nr_scanned;
1935         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1936
1937 restart:
1938         nr_reclaimed = 0;
1939         nr_scanned = sc->nr_scanned;
1940         get_scan_count(zone, sc, nr, priority);
1941
1942         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1943                                         nr[LRU_INACTIVE_FILE]) {
1944                 for_each_evictable_lru(l) {
1945                         if (nr[l]) {
1946                                 nr_to_scan = min_t(unsigned long,
1947                                                    nr[l], SWAP_CLUSTER_MAX);
1948                                 nr[l] -= nr_to_scan;
1949
1950                                 nr_reclaimed += shrink_list(l, nr_to_scan,
1951                                                             zone, sc, priority);
1952                         }
1953                 }
1954                 /*
1955                  * On large memory systems, scan >> priority can become
1956                  * really large. This is fine for the starting priority;
1957                  * we want to put equal scanning pressure on each zone.
1958                  * However, if the VM has a harder time of freeing pages,
1959                  * with multiple processes reclaiming pages, the total
1960                  * freeing target can get unreasonably large.
1961                  */
1962                 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
1963                         break;
1964         }
1965         sc->nr_reclaimed += nr_reclaimed;
1966
1967         /*
1968          * Even if we did not try to evict anon pages at all, we want to
1969          * rebalance the anon lru active/inactive ratio.
1970          */
1971         if (inactive_anon_is_low(zone, sc))
1972                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1973
1974         /* reclaim/compaction might need reclaim to continue */
1975         if (should_continue_reclaim(zone, nr_reclaimed,
1976                                         sc->nr_scanned - nr_scanned, sc))
1977                 goto restart;
1978
1979         throttle_vm_writeout(sc->gfp_mask);
1980 }
1981
1982 /*
1983  * This is the direct reclaim path, for page-allocating processes.  We only
1984  * try to reclaim pages from zones which will satisfy the caller's allocation
1985  * request.
1986  *
1987  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1988  * Because:
1989  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1990  *    allocation or
1991  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1992  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1993  *    zone defense algorithm.
1994  *
1995  * If a zone is deemed to be full of pinned pages then just give it a light
1996  * scan then give up on it.
1997  */
1998 static void shrink_zones(int priority, struct zonelist *zonelist,
1999                                         struct scan_control *sc)
2000 {
2001         struct zoneref *z;
2002         struct zone *zone;
2003         unsigned long nr_soft_reclaimed;
2004         unsigned long nr_soft_scanned;
2005
2006         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2007                                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2008                 if (!populated_zone(zone))
2009                         continue;
2010                 /*
2011                  * Take care memory controller reclaiming has small influence
2012                  * to global LRU.
2013                  */
2014                 if (scanning_global_lru(sc)) {
2015                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2016                                 continue;
2017                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2018                                 continue;       /* Let kswapd poll it */
2019                         /*
2020                          * This steals pages from memory cgroups over softlimit
2021                          * and returns the number of reclaimed pages and
2022                          * scanned pages. This works for global memory pressure
2023                          * and balancing, not for a memcg's limit.
2024                          */
2025                         nr_soft_scanned = 0;
2026                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2027                                                 sc->order, sc->gfp_mask,
2028                                                 &nr_soft_scanned);
2029                         sc->nr_reclaimed += nr_soft_reclaimed;
2030                         sc->nr_scanned += nr_soft_scanned;
2031                         /* need some check for avoid more shrink_zone() */
2032                 }
2033
2034                 shrink_zone(priority, zone, sc);
2035         }
2036 }
2037
2038 static bool zone_reclaimable(struct zone *zone)
2039 {
2040         return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2041 }
2042
2043 /* All zones in zonelist are unreclaimable? */
2044 static bool all_unreclaimable(struct zonelist *zonelist,
2045                 struct scan_control *sc)
2046 {
2047         struct zoneref *z;
2048         struct zone *zone;
2049
2050         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2051                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2052                 if (!populated_zone(zone))
2053                         continue;
2054                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2055                         continue;
2056                 if (!zone->all_unreclaimable)
2057                         return false;
2058         }
2059
2060         return true;
2061 }
2062
2063 /*
2064  * This is the main entry point to direct page reclaim.
2065  *
2066  * If a full scan of the inactive list fails to free enough memory then we
2067  * are "out of memory" and something needs to be killed.
2068  *
2069  * If the caller is !__GFP_FS then the probability of a failure is reasonably
2070  * high - the zone may be full of dirty or under-writeback pages, which this
2071  * caller can't do much about.  We kick the writeback threads and take explicit
2072  * naps in the hope that some of these pages can be written.  But if the
2073  * allocating task holds filesystem locks which prevent writeout this might not
2074  * work, and the allocation attempt will fail.
2075  *
2076  * returns:     0, if no pages reclaimed
2077  *              else, the number of pages reclaimed
2078  */
2079 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2080                                         struct scan_control *sc,
2081                                         struct shrink_control *shrink)
2082 {
2083         int priority;
2084         unsigned long total_scanned = 0;
2085         struct reclaim_state *reclaim_state = current->reclaim_state;
2086         struct zoneref *z;
2087         struct zone *zone;
2088         unsigned long writeback_threshold;
2089
2090         get_mems_allowed();
2091         delayacct_freepages_start();
2092
2093         if (scanning_global_lru(sc))
2094                 count_vm_event(ALLOCSTALL);
2095
2096         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2097                 sc->nr_scanned = 0;
2098                 if (!priority)
2099                         disable_swap_token(sc->mem_cgroup);
2100                 shrink_zones(priority, zonelist, sc);
2101                 /*
2102                  * Don't shrink slabs when reclaiming memory from
2103                  * over limit cgroups
2104                  */
2105                 if (scanning_global_lru(sc)) {
2106                         unsigned long lru_pages = 0;
2107                         for_each_zone_zonelist(zone, z, zonelist,
2108                                         gfp_zone(sc->gfp_mask)) {
2109                                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2110                                         continue;
2111
2112                                 lru_pages += zone_reclaimable_pages(zone);
2113                         }
2114
2115                         shrink_slab(shrink, sc->nr_scanned, lru_pages);
2116                         if (reclaim_state) {
2117                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2118                                 reclaim_state->reclaimed_slab = 0;
2119                         }
2120                 }
2121                 total_scanned += sc->nr_scanned;
2122                 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2123                         goto out;
2124
2125                 /*
2126                  * Try to write back as many pages as we just scanned.  This
2127                  * tends to cause slow streaming writers to write data to the
2128                  * disk smoothly, at the dirtying rate, which is nice.   But
2129                  * that's undesirable in laptop mode, where we *want* lumpy
2130                  * writeout.  So in laptop mode, write out the whole world.
2131                  */
2132                 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2133                 if (total_scanned > writeback_threshold) {
2134                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
2135                         sc->may_writepage = 1;
2136                 }
2137
2138                 /* Take a nap, wait for some writeback to complete */
2139                 if (!sc->hibernation_mode && sc->nr_scanned &&
2140                     priority < DEF_PRIORITY - 2) {
2141                         struct zone *preferred_zone;
2142
2143                         first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2144                                                 &cpuset_current_mems_allowed,
2145                                                 &preferred_zone);
2146                         wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2147                 }
2148         }
2149
2150 out:
2151         delayacct_freepages_end();
2152         put_mems_allowed();
2153
2154         if (sc->nr_reclaimed)
2155                 return sc->nr_reclaimed;
2156
2157         /*
2158          * As hibernation is going on, kswapd is freezed so that it can't mark
2159          * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2160          * check.
2161          */
2162         if (oom_killer_disabled)
2163                 return 0;
2164
2165         /* top priority shrink_zones still had more to do? don't OOM, then */
2166         if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
2167                 return 1;
2168
2169         return 0;
2170 }
2171
2172 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2173                                 gfp_t gfp_mask, nodemask_t *nodemask)
2174 {
2175         unsigned long nr_reclaimed;
2176         struct scan_control sc = {
2177                 .gfp_mask = gfp_mask,
2178                 .may_writepage = !laptop_mode,
2179                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2180                 .may_unmap = 1,
2181                 .may_swap = 1,
2182                 .swappiness = vm_swappiness,
2183                 .order = order,
2184                 .mem_cgroup = NULL,
2185                 .nodemask = nodemask,
2186         };
2187         struct shrink_control shrink = {
2188                 .gfp_mask = sc.gfp_mask,
2189         };
2190
2191         trace_mm_vmscan_direct_reclaim_begin(order,
2192                                 sc.may_writepage,
2193                                 gfp_mask);
2194
2195         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2196
2197         trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2198
2199         return nr_reclaimed;
2200 }
2201
2202 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2203
2204 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2205                                                 gfp_t gfp_mask, bool noswap,
2206                                                 unsigned int swappiness,
2207                                                 struct zone *zone,
2208                                                 unsigned long *nr_scanned)
2209 {
2210         struct scan_control sc = {
2211                 .nr_scanned = 0,
2212                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2213                 .may_writepage = !laptop_mode,
2214                 .may_unmap = 1,
2215                 .may_swap = !noswap,
2216                 .swappiness = swappiness,
2217                 .order = 0,
2218                 .mem_cgroup = mem,
2219         };
2220
2221         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2222                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2223
2224         trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2225                                                       sc.may_writepage,
2226                                                       sc.gfp_mask);
2227
2228         /*
2229          * NOTE: Although we can get the priority field, using it
2230          * here is not a good idea, since it limits the pages we can scan.
2231          * if we don't reclaim here, the shrink_zone from balance_pgdat
2232          * will pick up pages from other mem cgroup's as well. We hack
2233          * the priority and make it zero.
2234          */
2235         shrink_zone(0, zone, &sc);
2236
2237         trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2238
2239         *nr_scanned = sc.nr_scanned;
2240         return sc.nr_reclaimed;
2241 }
2242
2243 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
2244                                            gfp_t gfp_mask,
2245                                            bool noswap,
2246                                            unsigned int swappiness)
2247 {
2248         struct zonelist *zonelist;
2249         unsigned long nr_reclaimed;
2250         int nid;
2251         struct scan_control sc = {
2252                 .may_writepage = !laptop_mode,
2253                 .may_unmap = 1,
2254                 .may_swap = !noswap,
2255                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2256                 .swappiness = swappiness,
2257                 .order = 0,
2258                 .mem_cgroup = mem_cont,
2259                 .nodemask = NULL, /* we don't care the placement */
2260                 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2261                                 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2262         };
2263         struct shrink_control shrink = {
2264                 .gfp_mask = sc.gfp_mask,
2265         };
2266
2267         /*
2268          * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2269          * take care of from where we get pages. So the node where we start the
2270          * scan does not need to be the current node.
2271          */
2272         nid = mem_cgroup_select_victim_node(mem_cont);
2273
2274         zonelist = NODE_DATA(nid)->node_zonelists;
2275
2276         trace_mm_vmscan_memcg_reclaim_begin(0,
2277                                             sc.may_writepage,
2278                                             sc.gfp_mask);
2279
2280         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2281
2282         trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2283
2284         return nr_reclaimed;
2285 }
2286 #endif
2287
2288 /*
2289  * pgdat_balanced is used when checking if a node is balanced for high-order
2290  * allocations. Only zones that meet watermarks and are in a zone allowed
2291  * by the callers classzone_idx are added to balanced_pages. The total of
2292  * balanced pages must be at least 25% of the zones allowed by classzone_idx
2293  * for the node to be considered balanced. Forcing all zones to be balanced
2294  * for high orders can cause excessive reclaim when there are imbalanced zones.
2295  * The choice of 25% is due to
2296  *   o a 16M DMA zone that is balanced will not balance a zone on any
2297  *     reasonable sized machine
2298  *   o On all other machines, the top zone must be at least a reasonable
2299  *     percentage of the middle zones. For example, on 32-bit x86, highmem
2300  *     would need to be at least 256M for it to be balance a whole node.
2301  *     Similarly, on x86-64 the Normal zone would need to be at least 1G
2302  *     to balance a node on its own. These seemed like reasonable ratios.
2303  */
2304 static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2305                                                 int classzone_idx)
2306 {
2307         unsigned long present_pages = 0;
2308         int i;
2309
2310         for (i = 0; i <= classzone_idx; i++)
2311                 present_pages += pgdat->node_zones[i].present_pages;
2312
2313         /* A special case here: if zone has no page, we think it's balanced */
2314         return balanced_pages >= (present_pages >> 2);
2315 }
2316
2317 /* is kswapd sleeping prematurely? */
2318 static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2319                                         int classzone_idx)
2320 {
2321         int i;
2322         unsigned long balanced = 0;
2323         bool all_zones_ok = true;
2324
2325         /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2326         if (remaining)
2327                 return true;
2328
2329         /* Check the watermark levels */
2330         for (i = 0; i <= classzone_idx; i++) {
2331                 struct zone *zone = pgdat->node_zones + i;
2332
2333                 if (!populated_zone(zone))
2334                         continue;
2335
2336                 /*
2337                  * balance_pgdat() skips over all_unreclaimable after
2338                  * DEF_PRIORITY. Effectively, it considers them balanced so
2339                  * they must be considered balanced here as well if kswapd
2340                  * is to sleep
2341                  */
2342                 if (zone->all_unreclaimable) {
2343                         balanced += zone->present_pages;
2344                         continue;
2345                 }
2346
2347                 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2348                                                         i, 0))
2349                         all_zones_ok = false;
2350                 else
2351                         balanced += zone->present_pages;
2352         }
2353
2354         /*
2355          * For high-order requests, the balanced zones must contain at least
2356          * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2357          * must be balanced
2358          */
2359         if (order)
2360                 return !pgdat_balanced(pgdat, balanced, classzone_idx);
2361         else
2362                 return !all_zones_ok;
2363 }
2364
2365 /*
2366  * For kswapd, balance_pgdat() will work across all this node's zones until
2367  * they are all at high_wmark_pages(zone).
2368  *
2369  * Returns the final order kswapd was reclaiming at
2370  *
2371  * There is special handling here for zones which are full of pinned pages.
2372  * This can happen if the pages are all mlocked, or if they are all used by
2373  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
2374  * What we do is to detect the case where all pages in the zone have been
2375  * scanned twice and there has been zero successful reclaim.  Mark the zone as
2376  * dead and from now on, only perform a short scan.  Basically we're polling
2377  * the zone for when the problem goes away.
2378  *
2379  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
2380  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2381  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2382  * lower zones regardless of the number of free pages in the lower zones. This
2383  * interoperates with the page allocator fallback scheme to ensure that aging
2384  * of pages is balanced across the zones.
2385  */
2386 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2387                                                         int *classzone_idx)
2388 {
2389         int all_zones_ok;
2390         unsigned long balanced;
2391         int priority;
2392         int i;
2393         int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
2394         unsigned long total_scanned;
2395         struct reclaim_state *reclaim_state = current->reclaim_state;
2396         unsigned long nr_soft_reclaimed;
2397         unsigned long nr_soft_scanned;
2398         struct scan_control sc = {
2399                 .gfp_mask = GFP_KERNEL,
2400                 .may_unmap = 1,
2401                 .may_swap = 1,
2402                 /*
2403                  * kswapd doesn't want to be bailed out while reclaim. because
2404                  * we want to put equal scanning pressure on each zone.
2405                  */
2406                 .nr_to_reclaim = ULONG_MAX,
2407                 .swappiness = vm_swappiness,
2408                 .order = order,
2409                 .mem_cgroup = NULL,
2410         };
2411         struct shrink_control shrink = {
2412                 .gfp_mask = sc.gfp_mask,
2413         };
2414 loop_again:
2415         total_scanned = 0;
2416         sc.nr_reclaimed = 0;
2417         sc.may_writepage = !laptop_mode;
2418         count_vm_event(PAGEOUTRUN);
2419
2420         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2421                 unsigned long lru_pages = 0;
2422                 int has_under_min_watermark_zone = 0;
2423
2424                 /* The swap token gets in the way of swapout... */
2425                 if (!priority)
2426                         disable_swap_token(NULL);
2427
2428                 all_zones_ok = 1;
2429                 balanced = 0;
2430
2431                 /*
2432                  * Scan in the highmem->dma direction for the highest
2433                  * zone which needs scanning
2434                  */
2435                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2436                         struct zone *zone = pgdat->node_zones + i;
2437
2438                         if (!populated_zone(zone))
2439                                 continue;
2440
2441                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2442                                 continue;
2443
2444                         /*
2445                          * Do some background aging of the anon list, to give
2446                          * pages a chance to be referenced before reclaiming.
2447                          */
2448                         if (inactive_anon_is_low(zone, &sc))
2449                                 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2450                                                         &sc, priority, 0);
2451
2452                         if (!zone_watermark_ok_safe(zone, order,
2453                                         high_wmark_pages(zone), 0, 0)) {
2454                                 end_zone = i;
2455                                 break;
2456                         }
2457                 }
2458                 if (i < 0)
2459                         goto out;
2460
2461                 for (i = 0; i <= end_zone; i++) {
2462                         struct zone *zone = pgdat->node_zones + i;
2463
2464                         lru_pages += zone_reclaimable_pages(zone);
2465                 }
2466
2467                 /*
2468                  * Now scan the zone in the dma->highmem direction, stopping
2469                  * at the last zone which needs scanning.
2470                  *
2471                  * We do this because the page allocator works in the opposite
2472                  * direction.  This prevents the page allocator from allocating
2473                  * pages behind kswapd's direction of progress, which would
2474                  * cause too much scanning of the lower zones.
2475                  */
2476                 for (i = 0; i <= end_zone; i++) {
2477                         struct zone *zone = pgdat->node_zones + i;
2478                         int nr_slab;
2479                         unsigned long balance_gap;
2480
2481                         if (!populated_zone(zone))
2482                                 continue;
2483
2484                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2485                                 continue;
2486
2487                         sc.nr_scanned = 0;
2488
2489                         nr_soft_scanned = 0;
2490                         /*
2491                          * Call soft limit reclaim before calling shrink_zone.
2492                          */
2493                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2494                                                         order, sc.gfp_mask,
2495                                                         &nr_soft_scanned);
2496                         sc.nr_reclaimed += nr_soft_reclaimed;
2497                         total_scanned += nr_soft_scanned;
2498
2499                         /*
2500                          * We put equal pressure on every zone, unless
2501                          * one zone has way too many pages free
2502                          * already. The "too many pages" is defined
2503                          * as the high wmark plus a "gap" where the
2504                          * gap is either the low watermark or 1%
2505                          * of the zone, whichever is smaller.
2506                          */
2507                         balance_gap = min(low_wmark_pages(zone),
2508                                 (zone->present_pages +
2509                                         KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2510                                 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2511                         if (!zone_watermark_ok_safe(zone, order,
2512                                         high_wmark_pages(zone) + balance_gap,
2513                                         end_zone, 0)) {
2514                                 shrink_zone(priority, zone, &sc);
2515
2516                                 reclaim_state->reclaimed_slab = 0;
2517                                 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2518                                 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2519                                 total_scanned += sc.nr_scanned;
2520
2521                                 if (nr_slab == 0 && !zone_reclaimable(zone))
2522                                         zone->all_unreclaimable = 1;
2523                         }
2524
2525                         /*
2526                          * If we've done a decent amount of scanning and
2527                          * the reclaim ratio is low, start doing writepage
2528                          * even in laptop mode
2529                          */
2530                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2531                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2532                                 sc.may_writepage = 1;
2533
2534                         if (zone->all_unreclaimable) {
2535                                 if (end_zone && end_zone == i)
2536                                         end_zone--;
2537                                 continue;
2538                         }
2539
2540                         if (!zone_watermark_ok_safe(zone, order,
2541                                         high_wmark_pages(zone), end_zone, 0)) {
2542                                 all_zones_ok = 0;
2543                                 /*
2544                                  * We are still under min water mark.  This
2545                                  * means that we have a GFP_ATOMIC allocation
2546                                  * failure risk. Hurry up!
2547                                  */
2548                                 if (!zone_watermark_ok_safe(zone, order,
2549                                             min_wmark_pages(zone), end_zone, 0))
2550                                         has_under_min_watermark_zone = 1;
2551                         } else {
2552                                 /*
2553                                  * If a zone reaches its high watermark,
2554                                  * consider it to be no longer congested. It's
2555                                  * possible there are dirty pages backed by
2556                                  * congested BDIs but as pressure is relieved,
2557                                  * spectulatively avoid congestion waits
2558                                  */
2559                                 zone_clear_flag(zone, ZONE_CONGESTED);
2560                                 if (i <= *classzone_idx)
2561                                         balanced += zone->present_pages;
2562                         }
2563
2564                 }
2565                 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2566                         break;          /* kswapd: all done */
2567                 /*
2568                  * OK, kswapd is getting into trouble.  Take a nap, then take
2569                  * another pass across the zones.
2570                  */
2571                 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2572                         if (has_under_min_watermark_zone)
2573                                 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2574                         else
2575                                 congestion_wait(BLK_RW_ASYNC, HZ/10);
2576                 }
2577
2578                 /*
2579                  * We do this so kswapd doesn't build up large priorities for
2580                  * example when it is freeing in parallel with allocators. It
2581                  * matches the direct reclaim path behaviour in terms of impact
2582                  * on zone->*_priority.
2583                  */
2584                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2585                         break;
2586         }
2587 out:
2588
2589         /*
2590          * order-0: All zones must meet high watermark for a balanced node
2591          * high-order: Balanced zones must make up at least 25% of the node
2592          *             for the node to be balanced
2593          */
2594         if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2595                 cond_resched();
2596
2597                 try_to_freeze();
2598
2599                 /*
2600                  * Fragmentation may mean that the system cannot be
2601                  * rebalanced for high-order allocations in all zones.
2602                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2603                  * it means the zones have been fully scanned and are still
2604                  * not balanced. For high-order allocations, there is
2605                  * little point trying all over again as kswapd may
2606                  * infinite loop.
2607                  *
2608                  * Instead, recheck all watermarks at order-0 as they
2609                  * are the most important. If watermarks are ok, kswapd will go
2610                  * back to sleep. High-order users can still perform direct
2611                  * reclaim if they wish.
2612                  */
2613                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2614                         order = sc.order = 0;
2615
2616                 goto loop_again;
2617         }
2618
2619         /*
2620          * If kswapd was reclaiming at a higher order, it has the option of
2621          * sleeping without all zones being balanced. Before it does, it must
2622          * ensure that the watermarks for order-0 on *all* zones are met and
2623          * that the congestion flags are cleared. The congestion flag must
2624          * be cleared as kswapd is the only mechanism that clears the flag
2625          * and it is potentially going to sleep here.
2626          */
2627         if (order) {
2628                 for (i = 0; i <= end_zone; i++) {
2629                         struct zone *zone = pgdat->node_zones + i;
2630
2631                         if (!populated_zone(zone))
2632                                 continue;
2633
2634                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2635                                 continue;
2636
2637                         /* Confirm the zone is balanced for order-0 */
2638                         if (!zone_watermark_ok(zone, 0,
2639                                         high_wmark_pages(zone), 0, 0)) {
2640                                 order = sc.order = 0;
2641                                 goto loop_again;
2642                         }
2643
2644                         /* If balanced, clear the congested flag */
2645                         zone_clear_flag(zone, ZONE_CONGESTED);
2646                 }
2647         }
2648
2649         /*
2650          * Return the order we were reclaiming at so sleeping_prematurely()
2651          * makes a decision on the order we were last reclaiming at. However,
2652          * if another caller entered the allocator slow path while kswapd
2653          * was awake, order will remain at the higher level
2654          */
2655         *classzone_idx = end_zone;
2656         return order;
2657 }
2658
2659 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2660 {
2661         long remaining = 0;
2662         DEFINE_WAIT(wait);
2663
2664         if (freezing(current) || kthread_should_stop())
2665                 return;
2666
2667         prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2668
2669         /* Try to sleep for a short interval */
2670         if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2671                 remaining = schedule_timeout(HZ/10);
2672                 finish_wait(&pgdat->kswapd_wait, &wait);
2673                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2674         }
2675
2676         /*
2677          * After a short sleep, check if it was a premature sleep. If not, then
2678          * go fully to sleep until explicitly woken up.
2679          */
2680         if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2681                 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2682
2683                 /*
2684                  * vmstat counters are not perfectly accurate and the estimated
2685                  * value for counters such as NR_FREE_PAGES can deviate from the
2686                  * true value by nr_online_cpus * threshold. To avoid the zone
2687                  * watermarks being breached while under pressure, we reduce the
2688                  * per-cpu vmstat threshold while kswapd is awake and restore
2689                  * them before going back to sleep.
2690                  */
2691                 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2692                 schedule();
2693                 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2694         } else {
2695                 if (remaining)
2696                         count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2697                 else
2698                         count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2699         }
2700         finish_wait(&pgdat->kswapd_wait, &wait);
2701 }
2702
2703 /*
2704  * The background pageout daemon, started as a kernel thread
2705  * from the init process.
2706  *
2707  * This basically trickles out pages so that we have _some_
2708  * free memory available even if there is no other activity
2709  * that frees anything up. This is needed for things like routing
2710  * etc, where we otherwise might have all activity going on in
2711  * asynchronous contexts that cannot page things out.
2712  *
2713  * If there are applications that are active memory-allocators
2714  * (most normal use), this basically shouldn't matter.
2715  */
2716 static int kswapd(void *p)
2717 {
2718         unsigned long order, new_order;
2719         int classzone_idx, new_classzone_idx;
2720         pg_data_t *pgdat = (pg_data_t*)p;
2721         struct task_struct *tsk = current;
2722
2723         struct reclaim_state reclaim_state = {
2724                 .reclaimed_slab = 0,
2725         };
2726         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2727
2728         lockdep_set_current_reclaim_state(GFP_KERNEL);
2729
2730         if (!cpumask_empty(cpumask))
2731                 set_cpus_allowed_ptr(tsk, cpumask);
2732         current->reclaim_state = &reclaim_state;
2733
2734         /*
2735          * Tell the memory management that we're a "memory allocator",
2736          * and that if we need more memory we should get access to it
2737          * regardless (see "__alloc_pages()"). "kswapd" should
2738          * never get caught in the normal page freeing logic.
2739          *
2740          * (Kswapd normally doesn't need memory anyway, but sometimes
2741          * you need a small amount of memory in order to be able to
2742          * page out something else, and this flag essentially protects
2743          * us from recursively trying to free more memory as we're
2744          * trying to free the first piece of memory in the first place).
2745          */
2746         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2747         set_freezable();
2748
2749         order = new_order = 0;
2750         classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
2751         for ( ; ; ) {
2752                 int ret;
2753
2754                 /*
2755                  * If the last balance_pgdat was unsuccessful it's unlikely a
2756                  * new request of a similar or harder type will succeed soon
2757                  * so consider going to sleep on the basis we reclaimed at
2758                  */
2759                 if (classzone_idx >= new_classzone_idx && order == new_order) {
2760                         new_order = pgdat->kswapd_max_order;
2761                         new_classzone_idx = pgdat->classzone_idx;
2762                         pgdat->kswapd_max_order =  0;
2763                         pgdat->classzone_idx = pgdat->nr_zones - 1;
2764                 }
2765
2766                 if (order < new_order || classzone_idx > new_classzone_idx) {
2767                         /*
2768                          * Don't sleep if someone wants a larger 'order'
2769                          * allocation or has tigher zone constraints
2770                          */
2771                         order = new_order;
2772                         classzone_idx = new_classzone_idx;
2773                 } else {
2774                         kswapd_try_to_sleep(pgdat, order, classzone_idx);
2775                         order = pgdat->kswapd_max_order;
2776                         classzone_idx = pgdat->classzone_idx;
2777                         pgdat->kswapd_max_order = 0;
2778                         pgdat->classzone_idx = pgdat->nr_zones - 1;
2779                 }
2780
2781                 ret = try_to_freeze();
2782                 if (kthread_should_stop())
2783                         break;
2784
2785                 /*
2786                  * We can speed up thawing tasks if we don't call balance_pgdat
2787                  * after returning from the refrigerator
2788                  */
2789                 if (!ret) {
2790                         trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2791                         order = balance_pgdat(pgdat, order, &classzone_idx);
2792                 }
2793         }
2794         return 0;
2795 }
2796
2797 /*
2798  * A zone is low on free memory, so wake its kswapd task to service it.
2799  */
2800 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
2801 {
2802         pg_data_t *pgdat;
2803
2804         if (!populated_zone(zone))
2805                 return;
2806
2807         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2808                 return;
2809         pgdat = zone->zone_pgdat;
2810         if (pgdat->kswapd_max_order < order) {
2811                 pgdat->kswapd_max_order = order;
2812                 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
2813         }
2814         if (!waitqueue_active(&pgdat->kswapd_wait))
2815                 return;
2816         if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
2817                 return;
2818
2819         trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
2820         wake_up_interruptible(&pgdat->kswapd_wait);
2821 }
2822
2823 /*
2824  * The reclaimable count would be mostly accurate.
2825  * The less reclaimable pages may be
2826  * - mlocked pages, which will be moved to unevictable list when encountered
2827  * - mapped pages, which may require several travels to be reclaimed
2828  * - dirty pages, which is not "instantly" reclaimable
2829  */
2830 unsigned long global_reclaimable_pages(void)
2831 {
2832         int nr;
2833
2834         nr = global_page_state(NR_ACTIVE_FILE) +
2835              global_page_state(NR_INACTIVE_FILE);
2836
2837         if (nr_swap_pages > 0)
2838                 nr += global_page_state(NR_ACTIVE_ANON) +
2839                       global_page_state(NR_INACTIVE_ANON);
2840
2841         return nr;
2842 }
2843
2844 unsigned long zone_reclaimable_pages(struct zone *zone)
2845 {
2846         int nr;
2847
2848         nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2849              zone_page_state(zone, NR_INACTIVE_FILE);
2850
2851         if (nr_swap_pages > 0)
2852                 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2853                       zone_page_state(zone, NR_INACTIVE_ANON);
2854
2855         return nr;
2856 }
2857
2858 #ifdef CONFIG_HIBERNATION
2859 /*
2860  * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2861  * freed pages.
2862  *
2863  * Rather than trying to age LRUs the aim is to preserve the overall
2864  * LRU order by reclaiming preferentially
2865  * inactive > active > active referenced > active mapped
2866  */
2867 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
2868 {
2869         struct reclaim_state reclaim_state;
2870         struct scan_control sc = {
2871                 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2872                 .may_swap = 1,
2873                 .may_unmap = 1,
2874                 .may_writepage = 1,
2875                 .nr_to_reclaim = nr_to_reclaim,
2876                 .hibernation_mode = 1,
2877                 .swappiness = vm_swappiness,
2878                 .order = 0,
2879         };
2880         struct shrink_control shrink = {
2881                 .gfp_mask = sc.gfp_mask,
2882         };
2883         struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2884         struct task_struct *p = current;
2885         unsigned long nr_reclaimed;
2886
2887         p->flags |= PF_MEMALLOC;
2888         lockdep_set_current_reclaim_state(sc.gfp_mask);
2889         reclaim_state.reclaimed_slab = 0;
2890         p->reclaim_state = &reclaim_state;
2891
2892         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2893
2894         p->reclaim_state = NULL;
2895         lockdep_clear_current_reclaim_state();
2896         p->flags &= ~PF_MEMALLOC;
2897
2898         return nr_reclaimed;
2899 }
2900 #endif /* CONFIG_HIBERNATION */
2901
2902 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2903    not required for correctness.  So if the last cpu in a node goes
2904    away, we get changed to run anywhere: as the first one comes back,
2905    restore their cpu bindings. */
2906 static int __devinit cpu_callback(struct notifier_block *nfb,
2907                                   unsigned long action, void *hcpu)
2908 {
2909         int nid;
2910
2911         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2912                 for_each_node_state(nid, N_HIGH_MEMORY) {
2913                         pg_data_t *pgdat = NODE_DATA(nid);
2914                         const struct cpumask *mask;
2915
2916                         mask = cpumask_of_node(pgdat->node_id);
2917
2918                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2919                                 /* One of our CPUs online: restore mask */
2920                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2921                 }
2922         }
2923         return NOTIFY_OK;
2924 }
2925
2926 /*
2927  * This kswapd start function will be called by init and node-hot-add.
2928  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2929  */
2930 int kswapd_run(int nid)
2931 {
2932         pg_data_t *pgdat = NODE_DATA(nid);
2933         int ret = 0;
2934
2935         if (pgdat->kswapd)
2936                 return 0;
2937
2938         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2939         if (IS_ERR(pgdat->kswapd)) {
2940                 /* failure at boot is fatal */
2941                 BUG_ON(system_state == SYSTEM_BOOTING);
2942                 printk("Failed to start kswapd on node %d\n",nid);
2943                 ret = -1;
2944         }
2945         return ret;
2946 }
2947
2948 /*
2949  * Called by memory hotplug when all memory in a node is offlined.
2950  */
2951 void kswapd_stop(int nid)
2952 {
2953         struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
2954
2955         if (kswapd)
2956                 kthread_stop(kswapd);
2957 }
2958
2959 static int __init kswapd_init(void)
2960 {
2961         int nid;
2962
2963         swap_setup();
2964         for_each_node_state(nid, N_HIGH_MEMORY)
2965                 kswapd_run(nid);
2966         hotcpu_notifier(cpu_callback, 0);
2967         return 0;
2968 }
2969
2970 module_init(kswapd_init)
2971
2972 #ifdef CONFIG_NUMA
2973 /*
2974  * Zone reclaim mode
2975  *
2976  * If non-zero call zone_reclaim when the number of free pages falls below
2977  * the watermarks.
2978  */
2979 int zone_reclaim_mode __read_mostly;
2980
2981 #define RECLAIM_OFF 0
2982 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
2983 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
2984 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
2985
2986 /*
2987  * Priority for ZONE_RECLAIM. This determines the fraction of pages
2988  * of a node considered for each zone_reclaim. 4 scans 1/16th of
2989  * a zone.
2990  */
2991 #define ZONE_RECLAIM_PRIORITY 4
2992
2993 /*
2994  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2995  * occur.
2996  */
2997 int sysctl_min_unmapped_ratio = 1;
2998
2999 /*
3000  * If the number of slab pages in a zone grows beyond this percentage then
3001  * slab reclaim needs to occur.
3002  */
3003 int sysctl_min_slab_ratio = 5;
3004
3005 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3006 {
3007         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3008         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3009                 zone_page_state(zone, NR_ACTIVE_FILE);
3010
3011         /*
3012          * It's possible for there to be more file mapped pages than
3013          * accounted for by the pages on the file LRU lists because
3014          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3015          */
3016         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3017 }
3018
3019 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3020 static long zone_pagecache_reclaimable(struct zone *zone)
3021 {
3022         long nr_pagecache_reclaimable;
3023         long delta = 0;
3024
3025         /*
3026          * If RECLAIM_SWAP is set, then all file pages are considered
3027          * potentially reclaimable. Otherwise, we have to worry about
3028          * pages like swapcache and zone_unmapped_file_pages() provides
3029          * a better estimate
3030          */
3031         if (zone_reclaim_mode & RECLAIM_SWAP)
3032                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3033         else
3034                 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3035
3036         /* If we can't clean pages, remove dirty pages from consideration */
3037         if (!(zone_reclaim_mode & RECLAIM_WRITE))
3038                 delta += zone_page_state(zone, NR_FILE_DIRTY);
3039
3040         /* Watch for any possible underflows due to delta */
3041         if (unlikely(delta > nr_pagecache_reclaimable))
3042                 delta = nr_pagecache_reclaimable;
3043
3044         return nr_pagecache_reclaimable - delta;
3045 }
3046
3047 /*
3048  * Try to free up some pages from this zone through reclaim.
3049  */
3050 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3051 {
3052         /* Minimum pages needed in order to stay on node */
3053         const unsigned long nr_pages = 1 << order;
3054         struct task_struct *p = current;
3055         struct reclaim_state reclaim_state;
3056         int priority;
3057         struct scan_control sc = {
3058                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3059                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3060                 .may_swap = 1,
3061                 .nr_to_reclaim = max_t(unsigned long, nr_pages,
3062                                        SWAP_CLUSTER_MAX),
3063                 .gfp_mask = gfp_mask,
3064                 .swappiness = vm_swappiness,
3065                 .order = order,
3066         };
3067         struct shrink_control shrink = {
3068                 .gfp_mask = sc.gfp_mask,
3069         };
3070         unsigned long nr_slab_pages0, nr_slab_pages1;
3071
3072         cond_resched();
3073         /*
3074          * We need to be able to allocate from the reserves for RECLAIM_SWAP
3075          * and we also need to be able to write out pages for RECLAIM_WRITE
3076          * and RECLAIM_SWAP.
3077          */
3078         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3079         lockdep_set_current_reclaim_state(gfp_mask);
3080         reclaim_state.reclaimed_slab = 0;
3081         p->reclaim_state = &reclaim_state;
3082
3083         if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3084                 /*
3085                  * Free memory by calling shrink zone with increasing
3086                  * priorities until we have enough memory freed.
3087                  */
3088                 priority = ZONE_RECLAIM_PRIORITY;
3089                 do {
3090                         shrink_zone(priority, zone, &sc);
3091                         priority--;
3092                 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
3093         }
3094
3095         nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3096         if (nr_slab_pages0 > zone->min_slab_pages) {
3097                 /*
3098                  * shrink_slab() does not currently allow us to determine how
3099                  * many pages were freed in this zone. So we take the current
3100                  * number of slab pages and shake the slab until it is reduced
3101                  * by the same nr_pages that we used for reclaiming unmapped
3102                  * pages.
3103                  *
3104                  * Note that shrink_slab will free memory on all zones and may
3105                  * take a long time.
3106                  */
3107                 for (;;) {
3108                         unsigned long lru_pages = zone_reclaimable_pages(zone);
3109
3110                         /* No reclaimable slab or very low memory pressure */
3111                         if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3112                                 break;
3113
3114                         /* Freed enough memory */
3115                         nr_slab_pages1 = zone_page_state(zone,
3116                                                         NR_SLAB_RECLAIMABLE);
3117                         if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3118                                 break;
3119                 }
3120
3121                 /*
3122                  * Update nr_reclaimed by the number of slab pages we
3123                  * reclaimed from this zone.
3124                  */
3125                 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3126                 if (nr_slab_pages1 < nr_slab_pages0)
3127                         sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3128         }
3129
3130         p->reclaim_state = NULL;
3131         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3132         lockdep_clear_current_reclaim_state();
3133         return sc.nr_reclaimed >= nr_pages;
3134 }
3135
3136 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3137 {
3138         int node_id;
3139         int ret;
3140
3141         /*
3142          * Zone reclaim reclaims unmapped file backed pages and
3143          * slab pages if we are over the defined limits.
3144          *
3145          * A small portion of unmapped file backed pages is needed for
3146          * file I/O otherwise pages read by file I/O will be immediately
3147          * thrown out if the zone is overallocated. So we do not reclaim
3148          * if less than a specified percentage of the zone is used by
3149          * unmapped file backed pages.
3150          */
3151         if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3152             zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3153                 return ZONE_RECLAIM_FULL;
3154
3155         if (zone->all_unreclaimable)
3156                 return ZONE_RECLAIM_FULL;
3157
3158         /*
3159          * Do not scan if the allocation should not be delayed.
3160          */
3161         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3162                 return ZONE_RECLAIM_NOSCAN;
3163
3164         /*
3165          * Only run zone reclaim on the local zone or on zones that do not
3166          * have associated processors. This will favor the local processor
3167          * over remote processors and spread off node memory allocations
3168          * as wide as possible.
3169          */
3170         node_id = zone_to_nid(zone);
3171         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3172                 return ZONE_RECLAIM_NOSCAN;
3173
3174         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3175                 return ZONE_RECLAIM_NOSCAN;
3176
3177         ret = __zone_reclaim(zone, gfp_mask, order);
3178         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3179
3180         if (!ret)
3181                 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3182
3183         return ret;
3184 }
3185 #endif
3186
3187 /*
3188  * page_evictable - test whether a page is evictable
3189  * @page: the page to test
3190  * @vma: the VMA in which the page is or will be mapped, may be NULL
3191  *
3192  * Test whether page is evictable--i.e., should be placed on active/inactive
3193  * lists vs unevictable list.  The vma argument is !NULL when called from the
3194  * fault path to determine how to instantate a new page.
3195  *
3196  * Reasons page might not be evictable:
3197  * (1) page's mapping marked unevictable
3198  * (2) page is part of an mlocked VMA
3199  *
3200  */
3201 int page_evictable(struct page *page, struct vm_area_struct *vma)
3202 {
3203
3204         if (mapping_unevictable(page_mapping(page)))
3205                 return 0;
3206
3207         if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
3208                 return 0;
3209
3210         return 1;
3211 }
3212
3213 /**
3214  * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
3215  * @page: page to check evictability and move to appropriate lru list
3216  * @zone: zone page is in
3217  *
3218  * Checks a page for evictability and moves the page to the appropriate
3219  * zone lru list.
3220  *
3221  * Restrictions: zone->lru_lock must be held, page must be on LRU and must
3222  * have PageUnevictable set.
3223  */
3224 static void check_move_unevictable_page(struct page *page, struct zone *zone)
3225 {
3226         VM_BUG_ON(PageActive(page));
3227
3228 retry:
3229         ClearPageUnevictable(page);
3230         if (page_evictable(page, NULL)) {
3231                 enum lru_list l = page_lru_base_type(page);
3232
3233                 __dec_zone_state(zone, NR_UNEVICTABLE);
3234                 list_move(&page->lru, &zone->lru[l].list);
3235                 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
3236                 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
3237                 __count_vm_event(UNEVICTABLE_PGRESCUED);
3238         } else {
3239                 /*
3240                  * rotate unevictable list
3241                  */
3242                 SetPageUnevictable(page);
3243                 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
3244                 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
3245                 if (page_evictable(page, NULL))
3246                         goto retry;
3247         }
3248 }
3249
3250 /**
3251  * scan_mapping_unevictable_pages - scan an address space for evictable pages
3252  * @mapping: struct address_space to scan for evictable pages
3253  *
3254  * Scan all pages in mapping.  Check unevictable pages for
3255  * evictability and move them to the appropriate zone lru list.
3256  */
3257 void scan_mapping_unevictable_pages(struct address_space *mapping)
3258 {
3259         pgoff_t next = 0;
3260         pgoff_t end   = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
3261                          PAGE_CACHE_SHIFT;
3262         struct zone *zone;
3263         struct pagevec pvec;
3264
3265         if (mapping->nrpages == 0)
3266                 return;
3267
3268         pagevec_init(&pvec, 0);
3269         while (next < end &&
3270                 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
3271                 int i;
3272                 int pg_scanned = 0;
3273
3274                 zone = NULL;
3275
3276                 for (i = 0; i < pagevec_count(&pvec); i++) {
3277                         struct page *page = pvec.pages[i];
3278                         pgoff_t page_index = page->index;
3279                         struct zone *pagezone = page_zone(page);
3280
3281                         pg_scanned++;
3282                         if (page_index > next)
3283                                 next = page_index;
3284                         next++;
3285
3286                         if (pagezone != zone) {
3287                                 if (zone)
3288                                         spin_unlock_irq(&zone->lru_lock);
3289                                 zone = pagezone;
3290                                 spin_lock_irq(&zone->lru_lock);
3291                         }
3292
3293                         if (PageLRU(page) && PageUnevictable(page))
3294                                 check_move_unevictable_page(page, zone);
3295                 }
3296                 if (zone)
3297                         spin_unlock_irq(&zone->lru_lock);
3298                 pagevec_release(&pvec);
3299
3300                 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
3301         }
3302
3303 }
3304
3305 /**
3306  * scan_zone_unevictable_pages - check unevictable list for evictable pages
3307  * @zone - zone of which to scan the unevictable list
3308  *
3309  * Scan @zone's unevictable LRU lists to check for pages that have become
3310  * evictable.  Move those that have to @zone's inactive list where they
3311  * become candidates for reclaim, unless shrink_inactive_zone() decides
3312  * to reactivate them.  Pages that are still unevictable are rotated
3313  * back onto @zone's unevictable list.
3314  */
3315 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
3316 static void scan_zone_unevictable_pages(struct zone *zone)
3317 {
3318         struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
3319         unsigned long scan;
3320         unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
3321
3322         while (nr_to_scan > 0) {
3323                 unsigned long batch_size = min(nr_to_scan,
3324                                                 SCAN_UNEVICTABLE_BATCH_SIZE);
3325
3326                 spin_lock_irq(&zone->lru_lock);
3327                 for (scan = 0;  scan < batch_size; scan++) {
3328                         struct page *page = lru_to_page(l_unevictable);
3329
3330                         if (!trylock_page(page))
3331                                 continue;
3332
3333                         prefetchw_prev_lru_page(page, l_unevictable, flags);
3334
3335                         if (likely(PageLRU(page) && PageUnevictable(page)))
3336                                 check_move_unevictable_page(page, zone);
3337
3338                         unlock_page(page);
3339                 }
3340                 spin_unlock_irq(&zone->lru_lock);
3341
3342                 nr_to_scan -= batch_size;
3343         }
3344 }
3345
3346
3347 /**
3348  * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
3349  *
3350  * A really big hammer:  scan all zones' unevictable LRU lists to check for
3351  * pages that have become evictable.  Move those back to the zones'
3352  * inactive list where they become candidates for reclaim.
3353  * This occurs when, e.g., we have unswappable pages on the unevictable lists,
3354  * and we add swap to the system.  As such, it runs in the context of a task
3355  * that has possibly/probably made some previously unevictable pages
3356  * evictable.
3357  */
3358 static void scan_all_zones_unevictable_pages(void)
3359 {
3360         struct zone *zone;
3361
3362         for_each_zone(zone) {
3363                 scan_zone_unevictable_pages(zone);
3364         }
3365 }
3366
3367 /*
3368  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
3369  * all nodes' unevictable lists for evictable pages
3370  */
3371 unsigned long scan_unevictable_pages;
3372
3373 int scan_unevictable_handler(struct ctl_table *table, int write,
3374                            void __user *buffer,
3375                            size_t *length, loff_t *ppos)
3376 {
3377         proc_doulongvec_minmax(table, write, buffer, length, ppos);
3378
3379         if (write && *(unsigned long *)table->data)
3380                 scan_all_zones_unevictable_pages();
3381
3382         scan_unevictable_pages = 0;
3383         return 0;
3384 }
3385
3386 #ifdef CONFIG_NUMA
3387 /*
3388  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
3389  * a specified node's per zone unevictable lists for evictable pages.
3390  */
3391
3392 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
3393                                           struct sysdev_attribute *attr,
3394                                           char *buf)
3395 {
3396         return sprintf(buf, "0\n");     /* always zero; should fit... */
3397 }
3398
3399 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
3400                                            struct sysdev_attribute *attr,
3401                                         const char *buf, size_t count)
3402 {
3403         struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
3404         struct zone *zone;
3405         unsigned long res;
3406         unsigned long req = strict_strtoul(buf, 10, &res);
3407
3408         if (!req)
3409                 return 1;       /* zero is no-op */
3410
3411         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
3412                 if (!populated_zone(zone))
3413                         continue;
3414                 scan_zone_unevictable_pages(zone);
3415         }
3416         return 1;
3417 }
3418
3419
3420 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3421                         read_scan_unevictable_node,
3422                         write_scan_unevictable_node);
3423
3424 int scan_unevictable_register_node(struct node *node)
3425 {
3426         return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
3427 }
3428
3429 void scan_unevictable_unregister_node(struct node *node)
3430 {
3431         sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
3432 }
3433 #endif