Merge branch 'for-2.6.39/stack-plug' into for-2.6.39/core
[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
45 #include <asm/tlbflush.h>
46 #include <asm/div64.h>
47
48 #include <linux/swapops.h>
49
50 #include "internal.h"
51
52 #define CREATE_TRACE_POINTS
53 #include <trace/events/vmscan.h>
54
55 /*
56  * reclaim_mode determines how the inactive list is shrunk
57  * RECLAIM_MODE_SINGLE: Reclaim only order-0 pages
58  * RECLAIM_MODE_ASYNC:  Do not block
59  * RECLAIM_MODE_SYNC:   Allow blocking e.g. call wait_on_page_writeback
60  * RECLAIM_MODE_LUMPYRECLAIM: For high-order allocations, take a reference
61  *                      page from the LRU and reclaim all pages within a
62  *                      naturally aligned range
63  * RECLAIM_MODE_COMPACTION: For high-order allocations, reclaim a number of
64  *                      order-0 pages and then compact the zone
65  */
66 typedef unsigned __bitwise__ reclaim_mode_t;
67 #define RECLAIM_MODE_SINGLE             ((__force reclaim_mode_t)0x01u)
68 #define RECLAIM_MODE_ASYNC              ((__force reclaim_mode_t)0x02u)
69 #define RECLAIM_MODE_SYNC               ((__force reclaim_mode_t)0x04u)
70 #define RECLAIM_MODE_LUMPYRECLAIM       ((__force reclaim_mode_t)0x08u)
71 #define RECLAIM_MODE_COMPACTION         ((__force reclaim_mode_t)0x10u)
72
73 struct scan_control {
74         /* Incremented by the number of inactive pages that were scanned */
75         unsigned long nr_scanned;
76
77         /* Number of pages freed so far during a call to shrink_zones() */
78         unsigned long nr_reclaimed;
79
80         /* How many pages shrink_list() should reclaim */
81         unsigned long nr_to_reclaim;
82
83         unsigned long hibernation_mode;
84
85         /* This context's GFP mask */
86         gfp_t gfp_mask;
87
88         int may_writepage;
89
90         /* Can mapped pages be reclaimed? */
91         int may_unmap;
92
93         /* Can pages be swapped as part of reclaim? */
94         int may_swap;
95
96         int swappiness;
97
98         int order;
99
100         /*
101          * Intend to reclaim enough continuous memory rather than reclaim
102          * enough amount of memory. i.e, mode for high order allocation.
103          */
104         reclaim_mode_t reclaim_mode;
105
106         /* Which cgroup do we reclaim from */
107         struct mem_cgroup *mem_cgroup;
108
109         /*
110          * Nodemask of nodes allowed by the caller. If NULL, all nodes
111          * are scanned.
112          */
113         nodemask_t      *nodemask;
114 };
115
116 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
117
118 #ifdef ARCH_HAS_PREFETCH
119 #define prefetch_prev_lru_page(_page, _base, _field)                    \
120         do {                                                            \
121                 if ((_page)->lru.prev != _base) {                       \
122                         struct page *prev;                              \
123                                                                         \
124                         prev = lru_to_page(&(_page->lru));              \
125                         prefetch(&prev->_field);                        \
126                 }                                                       \
127         } while (0)
128 #else
129 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
130 #endif
131
132 #ifdef ARCH_HAS_PREFETCHW
133 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
134         do {                                                            \
135                 if ((_page)->lru.prev != _base) {                       \
136                         struct page *prev;                              \
137                                                                         \
138                         prev = lru_to_page(&(_page->lru));              \
139                         prefetchw(&prev->_field);                       \
140                 }                                                       \
141         } while (0)
142 #else
143 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
144 #endif
145
146 /*
147  * From 0 .. 100.  Higher means more swappy.
148  */
149 int vm_swappiness = 60;
150 long vm_total_pages;    /* The total number of pages which the VM controls */
151
152 static LIST_HEAD(shrinker_list);
153 static DECLARE_RWSEM(shrinker_rwsem);
154
155 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
156 #define scanning_global_lru(sc) (!(sc)->mem_cgroup)
157 #else
158 #define scanning_global_lru(sc) (1)
159 #endif
160
161 static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
162                                                   struct scan_control *sc)
163 {
164         if (!scanning_global_lru(sc))
165                 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
166
167         return &zone->reclaim_stat;
168 }
169
170 static unsigned long zone_nr_lru_pages(struct zone *zone,
171                                 struct scan_control *sc, enum lru_list lru)
172 {
173         if (!scanning_global_lru(sc))
174                 return mem_cgroup_zone_nr_pages(sc->mem_cgroup, zone, lru);
175
176         return zone_page_state(zone, NR_LRU_BASE + lru);
177 }
178
179
180 /*
181  * Add a shrinker callback to be called from the vm
182  */
183 void register_shrinker(struct shrinker *shrinker)
184 {
185         shrinker->nr = 0;
186         down_write(&shrinker_rwsem);
187         list_add_tail(&shrinker->list, &shrinker_list);
188         up_write(&shrinker_rwsem);
189 }
190 EXPORT_SYMBOL(register_shrinker);
191
192 /*
193  * Remove one
194  */
195 void unregister_shrinker(struct shrinker *shrinker)
196 {
197         down_write(&shrinker_rwsem);
198         list_del(&shrinker->list);
199         up_write(&shrinker_rwsem);
200 }
201 EXPORT_SYMBOL(unregister_shrinker);
202
203 #define SHRINK_BATCH 128
204 /*
205  * Call the shrink functions to age shrinkable caches
206  *
207  * Here we assume it costs one seek to replace a lru page and that it also
208  * takes a seek to recreate a cache object.  With this in mind we age equal
209  * percentages of the lru and ageable caches.  This should balance the seeks
210  * generated by these structures.
211  *
212  * If the vm encountered mapped pages on the LRU it increase the pressure on
213  * slab to avoid swapping.
214  *
215  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
216  *
217  * `lru_pages' represents the number of on-LRU pages in all the zones which
218  * are eligible for the caller's allocation attempt.  It is used for balancing
219  * slab reclaim versus page reclaim.
220  *
221  * Returns the number of slab objects which we shrunk.
222  */
223 unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
224                         unsigned long lru_pages)
225 {
226         struct shrinker *shrinker;
227         unsigned long ret = 0;
228
229         if (scanned == 0)
230                 scanned = SWAP_CLUSTER_MAX;
231
232         if (!down_read_trylock(&shrinker_rwsem))
233                 return 1;       /* Assume we'll be able to shrink next time */
234
235         list_for_each_entry(shrinker, &shrinker_list, list) {
236                 unsigned long long delta;
237                 unsigned long total_scan;
238                 unsigned long max_pass;
239
240                 max_pass = (*shrinker->shrink)(shrinker, 0, gfp_mask);
241                 delta = (4 * scanned) / shrinker->seeks;
242                 delta *= max_pass;
243                 do_div(delta, lru_pages + 1);
244                 shrinker->nr += delta;
245                 if (shrinker->nr < 0) {
246                         printk(KERN_ERR "shrink_slab: %pF negative objects to "
247                                "delete nr=%ld\n",
248                                shrinker->shrink, shrinker->nr);
249                         shrinker->nr = max_pass;
250                 }
251
252                 /*
253                  * Avoid risking looping forever due to too large nr value:
254                  * never try to free more than twice the estimate number of
255                  * freeable entries.
256                  */
257                 if (shrinker->nr > max_pass * 2)
258                         shrinker->nr = max_pass * 2;
259
260                 total_scan = shrinker->nr;
261                 shrinker->nr = 0;
262
263                 while (total_scan >= SHRINK_BATCH) {
264                         long this_scan = SHRINK_BATCH;
265                         int shrink_ret;
266                         int nr_before;
267
268                         nr_before = (*shrinker->shrink)(shrinker, 0, gfp_mask);
269                         shrink_ret = (*shrinker->shrink)(shrinker, this_scan,
270                                                                 gfp_mask);
271                         if (shrink_ret == -1)
272                                 break;
273                         if (shrink_ret < nr_before)
274                                 ret += nr_before - shrink_ret;
275                         count_vm_events(SLABS_SCANNED, this_scan);
276                         total_scan -= this_scan;
277
278                         cond_resched();
279                 }
280
281                 shrinker->nr += total_scan;
282         }
283         up_read(&shrinker_rwsem);
284         return ret;
285 }
286
287 static void set_reclaim_mode(int priority, struct scan_control *sc,
288                                    bool sync)
289 {
290         reclaim_mode_t syncmode = sync ? RECLAIM_MODE_SYNC : RECLAIM_MODE_ASYNC;
291
292         /*
293          * Initially assume we are entering either lumpy reclaim or
294          * reclaim/compaction.Depending on the order, we will either set the
295          * sync mode or just reclaim order-0 pages later.
296          */
297         if (COMPACTION_BUILD)
298                 sc->reclaim_mode = RECLAIM_MODE_COMPACTION;
299         else
300                 sc->reclaim_mode = RECLAIM_MODE_LUMPYRECLAIM;
301
302         /*
303          * Avoid using lumpy reclaim or reclaim/compaction if possible by
304          * restricting when its set to either costly allocations or when
305          * under memory pressure
306          */
307         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
308                 sc->reclaim_mode |= syncmode;
309         else if (sc->order && priority < DEF_PRIORITY - 2)
310                 sc->reclaim_mode |= syncmode;
311         else
312                 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
313 }
314
315 static void reset_reclaim_mode(struct scan_control *sc)
316 {
317         sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
318 }
319
320 static inline int is_page_cache_freeable(struct page *page)
321 {
322         /*
323          * A freeable page cache page is referenced only by the caller
324          * that isolated the page, the page cache radix tree and
325          * optional buffer heads at page->private.
326          */
327         return page_count(page) - page_has_private(page) == 2;
328 }
329
330 static int may_write_to_queue(struct backing_dev_info *bdi,
331                               struct scan_control *sc)
332 {
333         if (current->flags & PF_SWAPWRITE)
334                 return 1;
335         if (!bdi_write_congested(bdi))
336                 return 1;
337         if (bdi == current->backing_dev_info)
338                 return 1;
339
340         /* lumpy reclaim for hugepage often need a lot of write */
341         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
342                 return 1;
343         return 0;
344 }
345
346 /*
347  * We detected a synchronous write error writing a page out.  Probably
348  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
349  * fsync(), msync() or close().
350  *
351  * The tricky part is that after writepage we cannot touch the mapping: nothing
352  * prevents it from being freed up.  But we have a ref on the page and once
353  * that page is locked, the mapping is pinned.
354  *
355  * We're allowed to run sleeping lock_page() here because we know the caller has
356  * __GFP_FS.
357  */
358 static void handle_write_error(struct address_space *mapping,
359                                 struct page *page, int error)
360 {
361         lock_page(page);
362         if (page_mapping(page) == mapping)
363                 mapping_set_error(mapping, error);
364         unlock_page(page);
365 }
366
367 /* possible outcome of pageout() */
368 typedef enum {
369         /* failed to write page out, page is locked */
370         PAGE_KEEP,
371         /* move page to the active list, page is locked */
372         PAGE_ACTIVATE,
373         /* page has been sent to the disk successfully, page is unlocked */
374         PAGE_SUCCESS,
375         /* page is clean and locked */
376         PAGE_CLEAN,
377 } pageout_t;
378
379 /*
380  * pageout is called by shrink_page_list() for each dirty page.
381  * Calls ->writepage().
382  */
383 static pageout_t pageout(struct page *page, struct address_space *mapping,
384                          struct scan_control *sc)
385 {
386         /*
387          * If the page is dirty, only perform writeback if that write
388          * will be non-blocking.  To prevent this allocation from being
389          * stalled by pagecache activity.  But note that there may be
390          * stalls if we need to run get_block().  We could test
391          * PagePrivate for that.
392          *
393          * If this process is currently in __generic_file_aio_write() against
394          * this page's queue, we can perform writeback even if that
395          * will block.
396          *
397          * If the page is swapcache, write it back even if that would
398          * block, for some throttling. This happens by accident, because
399          * swap_backing_dev_info is bust: it doesn't reflect the
400          * congestion state of the swapdevs.  Easy to fix, if needed.
401          */
402         if (!is_page_cache_freeable(page))
403                 return PAGE_KEEP;
404         if (!mapping) {
405                 /*
406                  * Some data journaling orphaned pages can have
407                  * page->mapping == NULL while being dirty with clean buffers.
408                  */
409                 if (page_has_private(page)) {
410                         if (try_to_free_buffers(page)) {
411                                 ClearPageDirty(page);
412                                 printk("%s: orphaned page\n", __func__);
413                                 return PAGE_CLEAN;
414                         }
415                 }
416                 return PAGE_KEEP;
417         }
418         if (mapping->a_ops->writepage == NULL)
419                 return PAGE_ACTIVATE;
420         if (!may_write_to_queue(mapping->backing_dev_info, sc))
421                 return PAGE_KEEP;
422
423         if (clear_page_dirty_for_io(page)) {
424                 int res;
425                 struct writeback_control wbc = {
426                         .sync_mode = WB_SYNC_NONE,
427                         .nr_to_write = SWAP_CLUSTER_MAX,
428                         .range_start = 0,
429                         .range_end = LLONG_MAX,
430                         .for_reclaim = 1,
431                 };
432
433                 SetPageReclaim(page);
434                 res = mapping->a_ops->writepage(page, &wbc);
435                 if (res < 0)
436                         handle_write_error(mapping, page, res);
437                 if (res == AOP_WRITEPAGE_ACTIVATE) {
438                         ClearPageReclaim(page);
439                         return PAGE_ACTIVATE;
440                 }
441
442                 /*
443                  * Wait on writeback if requested to. This happens when
444                  * direct reclaiming a large contiguous area and the
445                  * first attempt to free a range of pages fails.
446                  */
447                 if (PageWriteback(page) &&
448                     (sc->reclaim_mode & RECLAIM_MODE_SYNC))
449                         wait_on_page_writeback(page);
450
451                 if (!PageWriteback(page)) {
452                         /* synchronous write or broken a_ops? */
453                         ClearPageReclaim(page);
454                 }
455                 trace_mm_vmscan_writepage(page,
456                         trace_reclaim_flags(page, sc->reclaim_mode));
457                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
458                 return PAGE_SUCCESS;
459         }
460
461         return PAGE_CLEAN;
462 }
463
464 /*
465  * Same as remove_mapping, but if the page is removed from the mapping, it
466  * gets returned with a refcount of 0.
467  */
468 static int __remove_mapping(struct address_space *mapping, struct page *page)
469 {
470         BUG_ON(!PageLocked(page));
471         BUG_ON(mapping != page_mapping(page));
472
473         spin_lock_irq(&mapping->tree_lock);
474         /*
475          * The non racy check for a busy page.
476          *
477          * Must be careful with the order of the tests. When someone has
478          * a ref to the page, it may be possible that they dirty it then
479          * drop the reference. So if PageDirty is tested before page_count
480          * here, then the following race may occur:
481          *
482          * get_user_pages(&page);
483          * [user mapping goes away]
484          * write_to(page);
485          *                              !PageDirty(page)    [good]
486          * SetPageDirty(page);
487          * put_page(page);
488          *                              !page_count(page)   [good, discard it]
489          *
490          * [oops, our write_to data is lost]
491          *
492          * Reversing the order of the tests ensures such a situation cannot
493          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
494          * load is not satisfied before that of page->_count.
495          *
496          * Note that if SetPageDirty is always performed via set_page_dirty,
497          * and thus under tree_lock, then this ordering is not required.
498          */
499         if (!page_freeze_refs(page, 2))
500                 goto cannot_free;
501         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
502         if (unlikely(PageDirty(page))) {
503                 page_unfreeze_refs(page, 2);
504                 goto cannot_free;
505         }
506
507         if (PageSwapCache(page)) {
508                 swp_entry_t swap = { .val = page_private(page) };
509                 __delete_from_swap_cache(page);
510                 spin_unlock_irq(&mapping->tree_lock);
511                 swapcache_free(swap, page);
512         } else {
513                 void (*freepage)(struct page *);
514
515                 freepage = mapping->a_ops->freepage;
516
517                 __remove_from_page_cache(page);
518                 spin_unlock_irq(&mapping->tree_lock);
519                 mem_cgroup_uncharge_cache_page(page);
520
521                 if (freepage != NULL)
522                         freepage(page);
523         }
524
525         return 1;
526
527 cannot_free:
528         spin_unlock_irq(&mapping->tree_lock);
529         return 0;
530 }
531
532 /*
533  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
534  * someone else has a ref on the page, abort and return 0.  If it was
535  * successfully detached, return 1.  Assumes the caller has a single ref on
536  * this page.
537  */
538 int remove_mapping(struct address_space *mapping, struct page *page)
539 {
540         if (__remove_mapping(mapping, page)) {
541                 /*
542                  * Unfreezing the refcount with 1 rather than 2 effectively
543                  * drops the pagecache ref for us without requiring another
544                  * atomic operation.
545                  */
546                 page_unfreeze_refs(page, 1);
547                 return 1;
548         }
549         return 0;
550 }
551
552 /**
553  * putback_lru_page - put previously isolated page onto appropriate LRU list
554  * @page: page to be put back to appropriate lru list
555  *
556  * Add previously isolated @page to appropriate LRU list.
557  * Page may still be unevictable for other reasons.
558  *
559  * lru_lock must not be held, interrupts must be enabled.
560  */
561 void putback_lru_page(struct page *page)
562 {
563         int lru;
564         int active = !!TestClearPageActive(page);
565         int was_unevictable = PageUnevictable(page);
566
567         VM_BUG_ON(PageLRU(page));
568
569 redo:
570         ClearPageUnevictable(page);
571
572         if (page_evictable(page, NULL)) {
573                 /*
574                  * For evictable pages, we can use the cache.
575                  * In event of a race, worst case is we end up with an
576                  * unevictable page on [in]active list.
577                  * We know how to handle that.
578                  */
579                 lru = active + page_lru_base_type(page);
580                 lru_cache_add_lru(page, lru);
581         } else {
582                 /*
583                  * Put unevictable pages directly on zone's unevictable
584                  * list.
585                  */
586                 lru = LRU_UNEVICTABLE;
587                 add_page_to_unevictable_list(page);
588                 /*
589                  * When racing with an mlock clearing (page is
590                  * unlocked), make sure that if the other thread does
591                  * not observe our setting of PG_lru and fails
592                  * isolation, we see PG_mlocked cleared below and move
593                  * the page back to the evictable list.
594                  *
595                  * The other side is TestClearPageMlocked().
596                  */
597                 smp_mb();
598         }
599
600         /*
601          * page's status can change while we move it among lru. If an evictable
602          * page is on unevictable list, it never be freed. To avoid that,
603          * check after we added it to the list, again.
604          */
605         if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
606                 if (!isolate_lru_page(page)) {
607                         put_page(page);
608                         goto redo;
609                 }
610                 /* This means someone else dropped this page from LRU
611                  * So, it will be freed or putback to LRU again. There is
612                  * nothing to do here.
613                  */
614         }
615
616         if (was_unevictable && lru != LRU_UNEVICTABLE)
617                 count_vm_event(UNEVICTABLE_PGRESCUED);
618         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
619                 count_vm_event(UNEVICTABLE_PGCULLED);
620
621         put_page(page);         /* drop ref from isolate */
622 }
623
624 enum page_references {
625         PAGEREF_RECLAIM,
626         PAGEREF_RECLAIM_CLEAN,
627         PAGEREF_KEEP,
628         PAGEREF_ACTIVATE,
629 };
630
631 static enum page_references page_check_references(struct page *page,
632                                                   struct scan_control *sc)
633 {
634         int referenced_ptes, referenced_page;
635         unsigned long vm_flags;
636
637         referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
638         referenced_page = TestClearPageReferenced(page);
639
640         /* Lumpy reclaim - ignore references */
641         if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
642                 return PAGEREF_RECLAIM;
643
644         /*
645          * Mlock lost the isolation race with us.  Let try_to_unmap()
646          * move the page to the unevictable list.
647          */
648         if (vm_flags & VM_LOCKED)
649                 return PAGEREF_RECLAIM;
650
651         if (referenced_ptes) {
652                 if (PageAnon(page))
653                         return PAGEREF_ACTIVATE;
654                 /*
655                  * All mapped pages start out with page table
656                  * references from the instantiating fault, so we need
657                  * to look twice if a mapped file page is used more
658                  * than once.
659                  *
660                  * Mark it and spare it for another trip around the
661                  * inactive list.  Another page table reference will
662                  * lead to its activation.
663                  *
664                  * Note: the mark is set for activated pages as well
665                  * so that recently deactivated but used pages are
666                  * quickly recovered.
667                  */
668                 SetPageReferenced(page);
669
670                 if (referenced_page)
671                         return PAGEREF_ACTIVATE;
672
673                 return PAGEREF_KEEP;
674         }
675
676         /* Reclaim if clean, defer dirty pages to writeback */
677         if (referenced_page && !PageSwapBacked(page))
678                 return PAGEREF_RECLAIM_CLEAN;
679
680         return PAGEREF_RECLAIM;
681 }
682
683 static noinline_for_stack void free_page_list(struct list_head *free_pages)
684 {
685         struct pagevec freed_pvec;
686         struct page *page, *tmp;
687
688         pagevec_init(&freed_pvec, 1);
689
690         list_for_each_entry_safe(page, tmp, free_pages, lru) {
691                 list_del(&page->lru);
692                 if (!pagevec_add(&freed_pvec, page)) {
693                         __pagevec_free(&freed_pvec);
694                         pagevec_reinit(&freed_pvec);
695                 }
696         }
697
698         pagevec_free(&freed_pvec);
699 }
700
701 /*
702  * shrink_page_list() returns the number of reclaimed pages
703  */
704 static unsigned long shrink_page_list(struct list_head *page_list,
705                                       struct zone *zone,
706                                       struct scan_control *sc)
707 {
708         LIST_HEAD(ret_pages);
709         LIST_HEAD(free_pages);
710         int pgactivate = 0;
711         unsigned long nr_dirty = 0;
712         unsigned long nr_congested = 0;
713         unsigned long nr_reclaimed = 0;
714
715         cond_resched();
716
717         while (!list_empty(page_list)) {
718                 enum page_references references;
719                 struct address_space *mapping;
720                 struct page *page;
721                 int may_enter_fs;
722
723                 cond_resched();
724
725                 page = lru_to_page(page_list);
726                 list_del(&page->lru);
727
728                 if (!trylock_page(page))
729                         goto keep;
730
731                 VM_BUG_ON(PageActive(page));
732                 VM_BUG_ON(page_zone(page) != zone);
733
734                 sc->nr_scanned++;
735
736                 if (unlikely(!page_evictable(page, NULL)))
737                         goto cull_mlocked;
738
739                 if (!sc->may_unmap && page_mapped(page))
740                         goto keep_locked;
741
742                 /* Double the slab pressure for mapped and swapcache pages */
743                 if (page_mapped(page) || PageSwapCache(page))
744                         sc->nr_scanned++;
745
746                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
747                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
748
749                 if (PageWriteback(page)) {
750                         /*
751                          * Synchronous reclaim is performed in two passes,
752                          * first an asynchronous pass over the list to
753                          * start parallel writeback, and a second synchronous
754                          * pass to wait for the IO to complete.  Wait here
755                          * for any page for which writeback has already
756                          * started.
757                          */
758                         if ((sc->reclaim_mode & RECLAIM_MODE_SYNC) &&
759                             may_enter_fs)
760                                 wait_on_page_writeback(page);
761                         else {
762                                 unlock_page(page);
763                                 goto keep_lumpy;
764                         }
765                 }
766
767                 references = page_check_references(page, sc);
768                 switch (references) {
769                 case PAGEREF_ACTIVATE:
770                         goto activate_locked;
771                 case PAGEREF_KEEP:
772                         goto keep_locked;
773                 case PAGEREF_RECLAIM:
774                 case PAGEREF_RECLAIM_CLEAN:
775                         ; /* try to reclaim the page below */
776                 }
777
778                 /*
779                  * Anonymous process memory has backing store?
780                  * Try to allocate it some swap space here.
781                  */
782                 if (PageAnon(page) && !PageSwapCache(page)) {
783                         if (!(sc->gfp_mask & __GFP_IO))
784                                 goto keep_locked;
785                         if (!add_to_swap(page))
786                                 goto activate_locked;
787                         may_enter_fs = 1;
788                 }
789
790                 mapping = page_mapping(page);
791
792                 /*
793                  * The page is mapped into the page tables of one or more
794                  * processes. Try to unmap it here.
795                  */
796                 if (page_mapped(page) && mapping) {
797                         switch (try_to_unmap(page, TTU_UNMAP)) {
798                         case SWAP_FAIL:
799                                 goto activate_locked;
800                         case SWAP_AGAIN:
801                                 goto keep_locked;
802                         case SWAP_MLOCK:
803                                 goto cull_mlocked;
804                         case SWAP_SUCCESS:
805                                 ; /* try to free the page below */
806                         }
807                 }
808
809                 if (PageDirty(page)) {
810                         nr_dirty++;
811
812                         if (references == PAGEREF_RECLAIM_CLEAN)
813                                 goto keep_locked;
814                         if (!may_enter_fs)
815                                 goto keep_locked;
816                         if (!sc->may_writepage)
817                                 goto keep_locked;
818
819                         /* Page is dirty, try to write it out here */
820                         switch (pageout(page, mapping, sc)) {
821                         case PAGE_KEEP:
822                                 nr_congested++;
823                                 goto keep_locked;
824                         case PAGE_ACTIVATE:
825                                 goto activate_locked;
826                         case PAGE_SUCCESS:
827                                 if (PageWriteback(page))
828                                         goto keep_lumpy;
829                                 if (PageDirty(page))
830                                         goto keep;
831
832                                 /*
833                                  * A synchronous write - probably a ramdisk.  Go
834                                  * ahead and try to reclaim the page.
835                                  */
836                                 if (!trylock_page(page))
837                                         goto keep;
838                                 if (PageDirty(page) || PageWriteback(page))
839                                         goto keep_locked;
840                                 mapping = page_mapping(page);
841                         case PAGE_CLEAN:
842                                 ; /* try to free the page below */
843                         }
844                 }
845
846                 /*
847                  * If the page has buffers, try to free the buffer mappings
848                  * associated with this page. If we succeed we try to free
849                  * the page as well.
850                  *
851                  * We do this even if the page is PageDirty().
852                  * try_to_release_page() does not perform I/O, but it is
853                  * possible for a page to have PageDirty set, but it is actually
854                  * clean (all its buffers are clean).  This happens if the
855                  * buffers were written out directly, with submit_bh(). ext3
856                  * will do this, as well as the blockdev mapping.
857                  * try_to_release_page() will discover that cleanness and will
858                  * drop the buffers and mark the page clean - it can be freed.
859                  *
860                  * Rarely, pages can have buffers and no ->mapping.  These are
861                  * the pages which were not successfully invalidated in
862                  * truncate_complete_page().  We try to drop those buffers here
863                  * and if that worked, and the page is no longer mapped into
864                  * process address space (page_count == 1) it can be freed.
865                  * Otherwise, leave the page on the LRU so it is swappable.
866                  */
867                 if (page_has_private(page)) {
868                         if (!try_to_release_page(page, sc->gfp_mask))
869                                 goto activate_locked;
870                         if (!mapping && page_count(page) == 1) {
871                                 unlock_page(page);
872                                 if (put_page_testzero(page))
873                                         goto free_it;
874                                 else {
875                                         /*
876                                          * rare race with speculative reference.
877                                          * the speculative reference will free
878                                          * this page shortly, so we may
879                                          * increment nr_reclaimed here (and
880                                          * leave it off the LRU).
881                                          */
882                                         nr_reclaimed++;
883                                         continue;
884                                 }
885                         }
886                 }
887
888                 if (!mapping || !__remove_mapping(mapping, page))
889                         goto keep_locked;
890
891                 /*
892                  * At this point, we have no other references and there is
893                  * no way to pick any more up (removed from LRU, removed
894                  * from pagecache). Can use non-atomic bitops now (and
895                  * we obviously don't have to worry about waking up a process
896                  * waiting on the page lock, because there are no references.
897                  */
898                 __clear_page_locked(page);
899 free_it:
900                 nr_reclaimed++;
901
902                 /*
903                  * Is there need to periodically free_page_list? It would
904                  * appear not as the counts should be low
905                  */
906                 list_add(&page->lru, &free_pages);
907                 continue;
908
909 cull_mlocked:
910                 if (PageSwapCache(page))
911                         try_to_free_swap(page);
912                 unlock_page(page);
913                 putback_lru_page(page);
914                 reset_reclaim_mode(sc);
915                 continue;
916
917 activate_locked:
918                 /* Not a candidate for swapping, so reclaim swap space. */
919                 if (PageSwapCache(page) && vm_swap_full())
920                         try_to_free_swap(page);
921                 VM_BUG_ON(PageActive(page));
922                 SetPageActive(page);
923                 pgactivate++;
924 keep_locked:
925                 unlock_page(page);
926 keep:
927                 reset_reclaim_mode(sc);
928 keep_lumpy:
929                 list_add(&page->lru, &ret_pages);
930                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
931         }
932
933         /*
934          * Tag a zone as congested if all the dirty pages encountered were
935          * backed by a congested BDI. In this case, reclaimers should just
936          * back off and wait for congestion to clear because further reclaim
937          * will encounter the same problem
938          */
939         if (nr_dirty == nr_congested && nr_dirty != 0)
940                 zone_set_flag(zone, ZONE_CONGESTED);
941
942         free_page_list(&free_pages);
943
944         list_splice(&ret_pages, page_list);
945         count_vm_events(PGACTIVATE, pgactivate);
946         return nr_reclaimed;
947 }
948
949 /*
950  * Attempt to remove the specified page from its LRU.  Only take this page
951  * if it is of the appropriate PageActive status.  Pages which are being
952  * freed elsewhere are also ignored.
953  *
954  * page:        page to consider
955  * mode:        one of the LRU isolation modes defined above
956  *
957  * returns 0 on success, -ve errno on failure.
958  */
959 int __isolate_lru_page(struct page *page, int mode, int file)
960 {
961         int ret = -EINVAL;
962
963         /* Only take pages on the LRU. */
964         if (!PageLRU(page))
965                 return ret;
966
967         /*
968          * When checking the active state, we need to be sure we are
969          * dealing with comparible boolean values.  Take the logical not
970          * of each.
971          */
972         if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
973                 return ret;
974
975         if (mode != ISOLATE_BOTH && page_is_file_cache(page) != file)
976                 return ret;
977
978         /*
979          * When this function is being called for lumpy reclaim, we
980          * initially look into all LRU pages, active, inactive and
981          * unevictable; only give shrink_page_list evictable pages.
982          */
983         if (PageUnevictable(page))
984                 return ret;
985
986         ret = -EBUSY;
987
988         if (likely(get_page_unless_zero(page))) {
989                 /*
990                  * Be careful not to clear PageLRU until after we're
991                  * sure the page is not being freed elsewhere -- the
992                  * page release code relies on it.
993                  */
994                 ClearPageLRU(page);
995                 ret = 0;
996         }
997
998         return ret;
999 }
1000
1001 /*
1002  * zone->lru_lock is heavily contended.  Some of the functions that
1003  * shrink the lists perform better by taking out a batch of pages
1004  * and working on them outside the LRU lock.
1005  *
1006  * For pagecache intensive workloads, this function is the hottest
1007  * spot in the kernel (apart from copy_*_user functions).
1008  *
1009  * Appropriate locks must be held before calling this function.
1010  *
1011  * @nr_to_scan: The number of pages to look through on the list.
1012  * @src:        The LRU list to pull pages off.
1013  * @dst:        The temp list to put pages on to.
1014  * @scanned:    The number of pages that were scanned.
1015  * @order:      The caller's attempted allocation order
1016  * @mode:       One of the LRU isolation modes
1017  * @file:       True [1] if isolating file [!anon] pages
1018  *
1019  * returns how many pages were moved onto *@dst.
1020  */
1021 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1022                 struct list_head *src, struct list_head *dst,
1023                 unsigned long *scanned, int order, int mode, int file)
1024 {
1025         unsigned long nr_taken = 0;
1026         unsigned long nr_lumpy_taken = 0;
1027         unsigned long nr_lumpy_dirty = 0;
1028         unsigned long nr_lumpy_failed = 0;
1029         unsigned long scan;
1030
1031         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1032                 struct page *page;
1033                 unsigned long pfn;
1034                 unsigned long end_pfn;
1035                 unsigned long page_pfn;
1036                 int zone_id;
1037
1038                 page = lru_to_page(src);
1039                 prefetchw_prev_lru_page(page, src, flags);
1040
1041                 VM_BUG_ON(!PageLRU(page));
1042
1043                 switch (__isolate_lru_page(page, mode, file)) {
1044                 case 0:
1045                         list_move(&page->lru, dst);
1046                         mem_cgroup_del_lru(page);
1047                         nr_taken += hpage_nr_pages(page);
1048                         break;
1049
1050                 case -EBUSY:
1051                         /* else it is being freed elsewhere */
1052                         list_move(&page->lru, src);
1053                         mem_cgroup_rotate_lru_list(page, page_lru(page));
1054                         continue;
1055
1056                 default:
1057                         BUG();
1058                 }
1059
1060                 if (!order)
1061                         continue;
1062
1063                 /*
1064                  * Attempt to take all pages in the order aligned region
1065                  * surrounding the tag page.  Only take those pages of
1066                  * the same active state as that tag page.  We may safely
1067                  * round the target page pfn down to the requested order
1068                  * as the mem_map is guarenteed valid out to MAX_ORDER,
1069                  * where that page is in a different zone we will detect
1070                  * it from its zone id and abort this block scan.
1071                  */
1072                 zone_id = page_zone_id(page);
1073                 page_pfn = page_to_pfn(page);
1074                 pfn = page_pfn & ~((1 << order) - 1);
1075                 end_pfn = pfn + (1 << order);
1076                 for (; pfn < end_pfn; pfn++) {
1077                         struct page *cursor_page;
1078
1079                         /* The target page is in the block, ignore it. */
1080                         if (unlikely(pfn == page_pfn))
1081                                 continue;
1082
1083                         /* Avoid holes within the zone. */
1084                         if (unlikely(!pfn_valid_within(pfn)))
1085                                 break;
1086
1087                         cursor_page = pfn_to_page(pfn);
1088
1089                         /* Check that we have not crossed a zone boundary. */
1090                         if (unlikely(page_zone_id(cursor_page) != zone_id))
1091                                 break;
1092
1093                         /*
1094                          * If we don't have enough swap space, reclaiming of
1095                          * anon page which don't already have a swap slot is
1096                          * pointless.
1097                          */
1098                         if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
1099                             !PageSwapCache(cursor_page))
1100                                 break;
1101
1102                         if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1103                                 list_move(&cursor_page->lru, dst);
1104                                 mem_cgroup_del_lru(cursor_page);
1105                                 nr_taken += hpage_nr_pages(page);
1106                                 nr_lumpy_taken++;
1107                                 if (PageDirty(cursor_page))
1108                                         nr_lumpy_dirty++;
1109                                 scan++;
1110                         } else {
1111                                 /* the page is freed already. */
1112                                 if (!page_count(cursor_page))
1113                                         continue;
1114                                 break;
1115                         }
1116                 }
1117
1118                 /* If we break out of the loop above, lumpy reclaim failed */
1119                 if (pfn < end_pfn)
1120                         nr_lumpy_failed++;
1121         }
1122
1123         *scanned = scan;
1124
1125         trace_mm_vmscan_lru_isolate(order,
1126                         nr_to_scan, scan,
1127                         nr_taken,
1128                         nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1129                         mode);
1130         return nr_taken;
1131 }
1132
1133 static unsigned long isolate_pages_global(unsigned long nr,
1134                                         struct list_head *dst,
1135                                         unsigned long *scanned, int order,
1136                                         int mode, struct zone *z,
1137                                         int active, int file)
1138 {
1139         int lru = LRU_BASE;
1140         if (active)
1141                 lru += LRU_ACTIVE;
1142         if (file)
1143                 lru += LRU_FILE;
1144         return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1145                                                                 mode, file);
1146 }
1147
1148 /*
1149  * clear_active_flags() is a helper for shrink_active_list(), clearing
1150  * any active bits from the pages in the list.
1151  */
1152 static unsigned long clear_active_flags(struct list_head *page_list,
1153                                         unsigned int *count)
1154 {
1155         int nr_active = 0;
1156         int lru;
1157         struct page *page;
1158
1159         list_for_each_entry(page, page_list, lru) {
1160                 int numpages = hpage_nr_pages(page);
1161                 lru = page_lru_base_type(page);
1162                 if (PageActive(page)) {
1163                         lru += LRU_ACTIVE;
1164                         ClearPageActive(page);
1165                         nr_active += numpages;
1166                 }
1167                 if (count)
1168                         count[lru] += numpages;
1169         }
1170
1171         return nr_active;
1172 }
1173
1174 /**
1175  * isolate_lru_page - tries to isolate a page from its LRU list
1176  * @page: page to isolate from its LRU list
1177  *
1178  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1179  * vmstat statistic corresponding to whatever LRU list the page was on.
1180  *
1181  * Returns 0 if the page was removed from an LRU list.
1182  * Returns -EBUSY if the page was not on an LRU list.
1183  *
1184  * The returned page will have PageLRU() cleared.  If it was found on
1185  * the active list, it will have PageActive set.  If it was found on
1186  * the unevictable list, it will have the PageUnevictable bit set. That flag
1187  * may need to be cleared by the caller before letting the page go.
1188  *
1189  * The vmstat statistic corresponding to the list on which the page was
1190  * found will be decremented.
1191  *
1192  * Restrictions:
1193  * (1) Must be called with an elevated refcount on the page. This is a
1194  *     fundamentnal difference from isolate_lru_pages (which is called
1195  *     without a stable reference).
1196  * (2) the lru_lock must not be held.
1197  * (3) interrupts must be enabled.
1198  */
1199 int isolate_lru_page(struct page *page)
1200 {
1201         int ret = -EBUSY;
1202
1203         if (PageLRU(page)) {
1204                 struct zone *zone = page_zone(page);
1205
1206                 spin_lock_irq(&zone->lru_lock);
1207                 if (PageLRU(page) && get_page_unless_zero(page)) {
1208                         int lru = page_lru(page);
1209                         ret = 0;
1210                         ClearPageLRU(page);
1211
1212                         del_page_from_lru_list(zone, page, lru);
1213                 }
1214                 spin_unlock_irq(&zone->lru_lock);
1215         }
1216         return ret;
1217 }
1218
1219 /*
1220  * Are there way too many processes in the direct reclaim path already?
1221  */
1222 static int too_many_isolated(struct zone *zone, int file,
1223                 struct scan_control *sc)
1224 {
1225         unsigned long inactive, isolated;
1226
1227         if (current_is_kswapd())
1228                 return 0;
1229
1230         if (!scanning_global_lru(sc))
1231                 return 0;
1232
1233         if (file) {
1234                 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1235                 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1236         } else {
1237                 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1238                 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1239         }
1240
1241         return isolated > inactive;
1242 }
1243
1244 /*
1245  * TODO: Try merging with migrations version of putback_lru_pages
1246  */
1247 static noinline_for_stack void
1248 putback_lru_pages(struct zone *zone, struct scan_control *sc,
1249                                 unsigned long nr_anon, unsigned long nr_file,
1250                                 struct list_head *page_list)
1251 {
1252         struct page *page;
1253         struct pagevec pvec;
1254         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1255
1256         pagevec_init(&pvec, 1);
1257
1258         /*
1259          * Put back any unfreeable pages.
1260          */
1261         spin_lock(&zone->lru_lock);
1262         while (!list_empty(page_list)) {
1263                 int lru;
1264                 page = lru_to_page(page_list);
1265                 VM_BUG_ON(PageLRU(page));
1266                 list_del(&page->lru);
1267                 if (unlikely(!page_evictable(page, NULL))) {
1268                         spin_unlock_irq(&zone->lru_lock);
1269                         putback_lru_page(page);
1270                         spin_lock_irq(&zone->lru_lock);
1271                         continue;
1272                 }
1273                 SetPageLRU(page);
1274                 lru = page_lru(page);
1275                 add_page_to_lru_list(zone, page, lru);
1276                 if (is_active_lru(lru)) {
1277                         int file = is_file_lru(lru);
1278                         int numpages = hpage_nr_pages(page);
1279                         reclaim_stat->recent_rotated[file] += numpages;
1280                 }
1281                 if (!pagevec_add(&pvec, page)) {
1282                         spin_unlock_irq(&zone->lru_lock);
1283                         __pagevec_release(&pvec);
1284                         spin_lock_irq(&zone->lru_lock);
1285                 }
1286         }
1287         __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1288         __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1289
1290         spin_unlock_irq(&zone->lru_lock);
1291         pagevec_release(&pvec);
1292 }
1293
1294 static noinline_for_stack void update_isolated_counts(struct zone *zone,
1295                                         struct scan_control *sc,
1296                                         unsigned long *nr_anon,
1297                                         unsigned long *nr_file,
1298                                         struct list_head *isolated_list)
1299 {
1300         unsigned long nr_active;
1301         unsigned int count[NR_LRU_LISTS] = { 0, };
1302         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1303
1304         nr_active = clear_active_flags(isolated_list, count);
1305         __count_vm_events(PGDEACTIVATE, nr_active);
1306
1307         __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1308                               -count[LRU_ACTIVE_FILE]);
1309         __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1310                               -count[LRU_INACTIVE_FILE]);
1311         __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1312                               -count[LRU_ACTIVE_ANON]);
1313         __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1314                               -count[LRU_INACTIVE_ANON]);
1315
1316         *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1317         *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1318         __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1319         __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1320
1321         reclaim_stat->recent_scanned[0] += *nr_anon;
1322         reclaim_stat->recent_scanned[1] += *nr_file;
1323 }
1324
1325 /*
1326  * Returns true if the caller should wait to clean dirty/writeback pages.
1327  *
1328  * If we are direct reclaiming for contiguous pages and we do not reclaim
1329  * everything in the list, try again and wait for writeback IO to complete.
1330  * This will stall high-order allocations noticeably. Only do that when really
1331  * need to free the pages under high memory pressure.
1332  */
1333 static inline bool should_reclaim_stall(unsigned long nr_taken,
1334                                         unsigned long nr_freed,
1335                                         int priority,
1336                                         struct scan_control *sc)
1337 {
1338         int lumpy_stall_priority;
1339
1340         /* kswapd should not stall on sync IO */
1341         if (current_is_kswapd())
1342                 return false;
1343
1344         /* Only stall on lumpy reclaim */
1345         if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
1346                 return false;
1347
1348         /* If we have relaimed everything on the isolated list, no stall */
1349         if (nr_freed == nr_taken)
1350                 return false;
1351
1352         /*
1353          * For high-order allocations, there are two stall thresholds.
1354          * High-cost allocations stall immediately where as lower
1355          * order allocations such as stacks require the scanning
1356          * priority to be much higher before stalling.
1357          */
1358         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1359                 lumpy_stall_priority = DEF_PRIORITY;
1360         else
1361                 lumpy_stall_priority = DEF_PRIORITY / 3;
1362
1363         return priority <= lumpy_stall_priority;
1364 }
1365
1366 /*
1367  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1368  * of reclaimed pages
1369  */
1370 static noinline_for_stack unsigned long
1371 shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1372                         struct scan_control *sc, int priority, int file)
1373 {
1374         LIST_HEAD(page_list);
1375         unsigned long nr_scanned;
1376         unsigned long nr_reclaimed = 0;
1377         unsigned long nr_taken;
1378         unsigned long nr_anon;
1379         unsigned long nr_file;
1380
1381         while (unlikely(too_many_isolated(zone, file, sc))) {
1382                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1383
1384                 /* We are about to die and free our memory. Return now. */
1385                 if (fatal_signal_pending(current))
1386                         return SWAP_CLUSTER_MAX;
1387         }
1388
1389         set_reclaim_mode(priority, sc, false);
1390         lru_add_drain();
1391         spin_lock_irq(&zone->lru_lock);
1392
1393         if (scanning_global_lru(sc)) {
1394                 nr_taken = isolate_pages_global(nr_to_scan,
1395                         &page_list, &nr_scanned, sc->order,
1396                         sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM ?
1397                                         ISOLATE_BOTH : ISOLATE_INACTIVE,
1398                         zone, 0, file);
1399                 zone->pages_scanned += nr_scanned;
1400                 if (current_is_kswapd())
1401                         __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1402                                                nr_scanned);
1403                 else
1404                         __count_zone_vm_events(PGSCAN_DIRECT, zone,
1405                                                nr_scanned);
1406         } else {
1407                 nr_taken = mem_cgroup_isolate_pages(nr_to_scan,
1408                         &page_list, &nr_scanned, sc->order,
1409                         sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM ?
1410                                         ISOLATE_BOTH : ISOLATE_INACTIVE,
1411                         zone, sc->mem_cgroup,
1412                         0, file);
1413                 /*
1414                  * mem_cgroup_isolate_pages() keeps track of
1415                  * scanned pages on its own.
1416                  */
1417         }
1418
1419         if (nr_taken == 0) {
1420                 spin_unlock_irq(&zone->lru_lock);
1421                 return 0;
1422         }
1423
1424         update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
1425
1426         spin_unlock_irq(&zone->lru_lock);
1427
1428         nr_reclaimed = shrink_page_list(&page_list, zone, sc);
1429
1430         /* Check if we should syncronously wait for writeback */
1431         if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1432                 set_reclaim_mode(priority, sc, true);
1433                 nr_reclaimed += shrink_page_list(&page_list, zone, sc);
1434         }
1435
1436         local_irq_disable();
1437         if (current_is_kswapd())
1438                 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1439         __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1440
1441         putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
1442
1443         trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1444                 zone_idx(zone),
1445                 nr_scanned, nr_reclaimed,
1446                 priority,
1447                 trace_shrink_flags(file, sc->reclaim_mode));
1448         return nr_reclaimed;
1449 }
1450
1451 /*
1452  * This moves pages from the active list to the inactive list.
1453  *
1454  * We move them the other way if the page is referenced by one or more
1455  * processes, from rmap.
1456  *
1457  * If the pages are mostly unmapped, the processing is fast and it is
1458  * appropriate to hold zone->lru_lock across the whole operation.  But if
1459  * the pages are mapped, the processing is slow (page_referenced()) so we
1460  * should drop zone->lru_lock around each page.  It's impossible to balance
1461  * this, so instead we remove the pages from the LRU while processing them.
1462  * It is safe to rely on PG_active against the non-LRU pages in here because
1463  * nobody will play with that bit on a non-LRU page.
1464  *
1465  * The downside is that we have to touch page->_count against each page.
1466  * But we had to alter page->flags anyway.
1467  */
1468
1469 static void move_active_pages_to_lru(struct zone *zone,
1470                                      struct list_head *list,
1471                                      enum lru_list lru)
1472 {
1473         unsigned long pgmoved = 0;
1474         struct pagevec pvec;
1475         struct page *page;
1476
1477         pagevec_init(&pvec, 1);
1478
1479         while (!list_empty(list)) {
1480                 page = lru_to_page(list);
1481
1482                 VM_BUG_ON(PageLRU(page));
1483                 SetPageLRU(page);
1484
1485                 list_move(&page->lru, &zone->lru[lru].list);
1486                 mem_cgroup_add_lru_list(page, lru);
1487                 pgmoved += hpage_nr_pages(page);
1488
1489                 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1490                         spin_unlock_irq(&zone->lru_lock);
1491                         if (buffer_heads_over_limit)
1492                                 pagevec_strip(&pvec);
1493                         __pagevec_release(&pvec);
1494                         spin_lock_irq(&zone->lru_lock);
1495                 }
1496         }
1497         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1498         if (!is_active_lru(lru))
1499                 __count_vm_events(PGDEACTIVATE, pgmoved);
1500 }
1501
1502 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1503                         struct scan_control *sc, int priority, int file)
1504 {
1505         unsigned long nr_taken;
1506         unsigned long pgscanned;
1507         unsigned long vm_flags;
1508         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1509         LIST_HEAD(l_active);
1510         LIST_HEAD(l_inactive);
1511         struct page *page;
1512         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1513         unsigned long nr_rotated = 0;
1514
1515         lru_add_drain();
1516         spin_lock_irq(&zone->lru_lock);
1517         if (scanning_global_lru(sc)) {
1518                 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1519                                                 &pgscanned, sc->order,
1520                                                 ISOLATE_ACTIVE, zone,
1521                                                 1, file);
1522                 zone->pages_scanned += pgscanned;
1523         } else {
1524                 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1525                                                 &pgscanned, sc->order,
1526                                                 ISOLATE_ACTIVE, zone,
1527                                                 sc->mem_cgroup, 1, file);
1528                 /*
1529                  * mem_cgroup_isolate_pages() keeps track of
1530                  * scanned pages on its own.
1531                  */
1532         }
1533
1534         reclaim_stat->recent_scanned[file] += nr_taken;
1535
1536         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1537         if (file)
1538                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1539         else
1540                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1541         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1542         spin_unlock_irq(&zone->lru_lock);
1543
1544         while (!list_empty(&l_hold)) {
1545                 cond_resched();
1546                 page = lru_to_page(&l_hold);
1547                 list_del(&page->lru);
1548
1549                 if (unlikely(!page_evictable(page, NULL))) {
1550                         putback_lru_page(page);
1551                         continue;
1552                 }
1553
1554                 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1555                         nr_rotated += hpage_nr_pages(page);
1556                         /*
1557                          * Identify referenced, file-backed active pages and
1558                          * give them one more trip around the active list. So
1559                          * that executable code get better chances to stay in
1560                          * memory under moderate memory pressure.  Anon pages
1561                          * are not likely to be evicted by use-once streaming
1562                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1563                          * so we ignore them here.
1564                          */
1565                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1566                                 list_add(&page->lru, &l_active);
1567                                 continue;
1568                         }
1569                 }
1570
1571                 ClearPageActive(page);  /* we are de-activating */
1572                 list_add(&page->lru, &l_inactive);
1573         }
1574
1575         /*
1576          * Move pages back to the lru list.
1577          */
1578         spin_lock_irq(&zone->lru_lock);
1579         /*
1580          * Count referenced pages from currently used mappings as rotated,
1581          * even though only some of them are actually re-activated.  This
1582          * helps balance scan pressure between file and anonymous pages in
1583          * get_scan_ratio.
1584          */
1585         reclaim_stat->recent_rotated[file] += nr_rotated;
1586
1587         move_active_pages_to_lru(zone, &l_active,
1588                                                 LRU_ACTIVE + file * LRU_FILE);
1589         move_active_pages_to_lru(zone, &l_inactive,
1590                                                 LRU_BASE   + file * LRU_FILE);
1591         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1592         spin_unlock_irq(&zone->lru_lock);
1593 }
1594
1595 #ifdef CONFIG_SWAP
1596 static int inactive_anon_is_low_global(struct zone *zone)
1597 {
1598         unsigned long active, inactive;
1599
1600         active = zone_page_state(zone, NR_ACTIVE_ANON);
1601         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1602
1603         if (inactive * zone->inactive_ratio < active)
1604                 return 1;
1605
1606         return 0;
1607 }
1608
1609 /**
1610  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1611  * @zone: zone to check
1612  * @sc:   scan control of this context
1613  *
1614  * Returns true if the zone does not have enough inactive anon pages,
1615  * meaning some active anon pages need to be deactivated.
1616  */
1617 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1618 {
1619         int low;
1620
1621         /*
1622          * If we don't have swap space, anonymous page deactivation
1623          * is pointless.
1624          */
1625         if (!total_swap_pages)
1626                 return 0;
1627
1628         if (scanning_global_lru(sc))
1629                 low = inactive_anon_is_low_global(zone);
1630         else
1631                 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1632         return low;
1633 }
1634 #else
1635 static inline int inactive_anon_is_low(struct zone *zone,
1636                                         struct scan_control *sc)
1637 {
1638         return 0;
1639 }
1640 #endif
1641
1642 static int inactive_file_is_low_global(struct zone *zone)
1643 {
1644         unsigned long active, inactive;
1645
1646         active = zone_page_state(zone, NR_ACTIVE_FILE);
1647         inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1648
1649         return (active > inactive);
1650 }
1651
1652 /**
1653  * inactive_file_is_low - check if file pages need to be deactivated
1654  * @zone: zone to check
1655  * @sc:   scan control of this context
1656  *
1657  * When the system is doing streaming IO, memory pressure here
1658  * ensures that active file pages get deactivated, until more
1659  * than half of the file pages are on the inactive list.
1660  *
1661  * Once we get to that situation, protect the system's working
1662  * set from being evicted by disabling active file page aging.
1663  *
1664  * This uses a different ratio than the anonymous pages, because
1665  * the page cache uses a use-once replacement algorithm.
1666  */
1667 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1668 {
1669         int low;
1670
1671         if (scanning_global_lru(sc))
1672                 low = inactive_file_is_low_global(zone);
1673         else
1674                 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1675         return low;
1676 }
1677
1678 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1679                                 int file)
1680 {
1681         if (file)
1682                 return inactive_file_is_low(zone, sc);
1683         else
1684                 return inactive_anon_is_low(zone, sc);
1685 }
1686
1687 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1688         struct zone *zone, struct scan_control *sc, int priority)
1689 {
1690         int file = is_file_lru(lru);
1691
1692         if (is_active_lru(lru)) {
1693                 if (inactive_list_is_low(zone, sc, file))
1694                     shrink_active_list(nr_to_scan, zone, sc, priority, file);
1695                 return 0;
1696         }
1697
1698         return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1699 }
1700
1701 /*
1702  * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1703  * until we collected @swap_cluster_max pages to scan.
1704  */
1705 static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
1706                                        unsigned long *nr_saved_scan)
1707 {
1708         unsigned long nr;
1709
1710         *nr_saved_scan += nr_to_scan;
1711         nr = *nr_saved_scan;
1712
1713         if (nr >= SWAP_CLUSTER_MAX)
1714                 *nr_saved_scan = 0;
1715         else
1716                 nr = 0;
1717
1718         return nr;
1719 }
1720
1721 /*
1722  * Determine how aggressively the anon and file LRU lists should be
1723  * scanned.  The relative value of each set of LRU lists is determined
1724  * by looking at the fraction of the pages scanned we did rotate back
1725  * onto the active list instead of evict.
1726  *
1727  * nr[0] = anon pages to scan; nr[1] = file pages to scan
1728  */
1729 static void get_scan_count(struct zone *zone, struct scan_control *sc,
1730                                         unsigned long *nr, int priority)
1731 {
1732         unsigned long anon, file, free;
1733         unsigned long anon_prio, file_prio;
1734         unsigned long ap, fp;
1735         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1736         u64 fraction[2], denominator;
1737         enum lru_list l;
1738         int noswap = 0;
1739
1740         /* If we have no swap space, do not bother scanning anon pages. */
1741         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1742                 noswap = 1;
1743                 fraction[0] = 0;
1744                 fraction[1] = 1;
1745                 denominator = 1;
1746                 goto out;
1747         }
1748
1749         anon  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1750                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1751         file  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1752                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1753
1754         if (scanning_global_lru(sc)) {
1755                 free  = zone_page_state(zone, NR_FREE_PAGES);
1756                 /* If we have very few page cache pages,
1757                    force-scan anon pages. */
1758                 if (unlikely(file + free <= high_wmark_pages(zone))) {
1759                         fraction[0] = 1;
1760                         fraction[1] = 0;
1761                         denominator = 1;
1762                         goto out;
1763                 }
1764         }
1765
1766         /*
1767          * With swappiness at 100, anonymous and file have the same priority.
1768          * This scanning priority is essentially the inverse of IO cost.
1769          */
1770         anon_prio = sc->swappiness;
1771         file_prio = 200 - sc->swappiness;
1772
1773         /*
1774          * OK, so we have swap space and a fair amount of page cache
1775          * pages.  We use the recently rotated / recently scanned
1776          * ratios to determine how valuable each cache is.
1777          *
1778          * Because workloads change over time (and to avoid overflow)
1779          * we keep these statistics as a floating average, which ends
1780          * up weighing recent references more than old ones.
1781          *
1782          * anon in [0], file in [1]
1783          */
1784         spin_lock_irq(&zone->lru_lock);
1785         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1786                 reclaim_stat->recent_scanned[0] /= 2;
1787                 reclaim_stat->recent_rotated[0] /= 2;
1788         }
1789
1790         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1791                 reclaim_stat->recent_scanned[1] /= 2;
1792                 reclaim_stat->recent_rotated[1] /= 2;
1793         }
1794
1795         /*
1796          * The amount of pressure on anon vs file pages is inversely
1797          * proportional to the fraction of recently scanned pages on
1798          * each list that were recently referenced and in active use.
1799          */
1800         ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1801         ap /= reclaim_stat->recent_rotated[0] + 1;
1802
1803         fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1804         fp /= reclaim_stat->recent_rotated[1] + 1;
1805         spin_unlock_irq(&zone->lru_lock);
1806
1807         fraction[0] = ap;
1808         fraction[1] = fp;
1809         denominator = ap + fp + 1;
1810 out:
1811         for_each_evictable_lru(l) {
1812                 int file = is_file_lru(l);
1813                 unsigned long scan;
1814
1815                 scan = zone_nr_lru_pages(zone, sc, l);
1816                 if (priority || noswap) {
1817                         scan >>= priority;
1818                         scan = div64_u64(scan * fraction[file], denominator);
1819                 }
1820                 nr[l] = nr_scan_try_batch(scan,
1821                                           &reclaim_stat->nr_saved_scan[l]);
1822         }
1823 }
1824
1825 /*
1826  * Reclaim/compaction depends on a number of pages being freed. To avoid
1827  * disruption to the system, a small number of order-0 pages continue to be
1828  * rotated and reclaimed in the normal fashion. However, by the time we get
1829  * back to the allocator and call try_to_compact_zone(), we ensure that
1830  * there are enough free pages for it to be likely successful
1831  */
1832 static inline bool should_continue_reclaim(struct zone *zone,
1833                                         unsigned long nr_reclaimed,
1834                                         unsigned long nr_scanned,
1835                                         struct scan_control *sc)
1836 {
1837         unsigned long pages_for_compaction;
1838         unsigned long inactive_lru_pages;
1839
1840         /* If not in reclaim/compaction mode, stop */
1841         if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
1842                 return false;
1843
1844         /* Consider stopping depending on scan and reclaim activity */
1845         if (sc->gfp_mask & __GFP_REPEAT) {
1846                 /*
1847                  * For __GFP_REPEAT allocations, stop reclaiming if the
1848                  * full LRU list has been scanned and we are still failing
1849                  * to reclaim pages. This full LRU scan is potentially
1850                  * expensive but a __GFP_REPEAT caller really wants to succeed
1851                  */
1852                 if (!nr_reclaimed && !nr_scanned)
1853                         return false;
1854         } else {
1855                 /*
1856                  * For non-__GFP_REPEAT allocations which can presumably
1857                  * fail without consequence, stop if we failed to reclaim
1858                  * any pages from the last SWAP_CLUSTER_MAX number of
1859                  * pages that were scanned. This will return to the
1860                  * caller faster at the risk reclaim/compaction and
1861                  * the resulting allocation attempt fails
1862                  */
1863                 if (!nr_reclaimed)
1864                         return false;
1865         }
1866
1867         /*
1868          * If we have not reclaimed enough pages for compaction and the
1869          * inactive lists are large enough, continue reclaiming
1870          */
1871         pages_for_compaction = (2UL << sc->order);
1872         inactive_lru_pages = zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON) +
1873                                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1874         if (sc->nr_reclaimed < pages_for_compaction &&
1875                         inactive_lru_pages > pages_for_compaction)
1876                 return true;
1877
1878         /* If compaction would go ahead or the allocation would succeed, stop */
1879         switch (compaction_suitable(zone, sc->order)) {
1880         case COMPACT_PARTIAL:
1881         case COMPACT_CONTINUE:
1882                 return false;
1883         default:
1884                 return true;
1885         }
1886 }
1887
1888 /*
1889  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1890  */
1891 static void shrink_zone(int priority, struct zone *zone,
1892                                 struct scan_control *sc)
1893 {
1894         unsigned long nr[NR_LRU_LISTS];
1895         unsigned long nr_to_scan;
1896         enum lru_list l;
1897         unsigned long nr_reclaimed, nr_scanned;
1898         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1899
1900 restart:
1901         nr_reclaimed = 0;
1902         nr_scanned = sc->nr_scanned;
1903         get_scan_count(zone, sc, nr, priority);
1904
1905         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1906                                         nr[LRU_INACTIVE_FILE]) {
1907                 for_each_evictable_lru(l) {
1908                         if (nr[l]) {
1909                                 nr_to_scan = min_t(unsigned long,
1910                                                    nr[l], SWAP_CLUSTER_MAX);
1911                                 nr[l] -= nr_to_scan;
1912
1913                                 nr_reclaimed += shrink_list(l, nr_to_scan,
1914                                                             zone, sc, priority);
1915                         }
1916                 }
1917                 /*
1918                  * On large memory systems, scan >> priority can become
1919                  * really large. This is fine for the starting priority;
1920                  * we want to put equal scanning pressure on each zone.
1921                  * However, if the VM has a harder time of freeing pages,
1922                  * with multiple processes reclaiming pages, the total
1923                  * freeing target can get unreasonably large.
1924                  */
1925                 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
1926                         break;
1927         }
1928         sc->nr_reclaimed += nr_reclaimed;
1929
1930         /*
1931          * Even if we did not try to evict anon pages at all, we want to
1932          * rebalance the anon lru active/inactive ratio.
1933          */
1934         if (inactive_anon_is_low(zone, sc))
1935                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1936
1937         /* reclaim/compaction might need reclaim to continue */
1938         if (should_continue_reclaim(zone, nr_reclaimed,
1939                                         sc->nr_scanned - nr_scanned, sc))
1940                 goto restart;
1941
1942         throttle_vm_writeout(sc->gfp_mask);
1943 }
1944
1945 /*
1946  * This is the direct reclaim path, for page-allocating processes.  We only
1947  * try to reclaim pages from zones which will satisfy the caller's allocation
1948  * request.
1949  *
1950  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1951  * Because:
1952  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1953  *    allocation or
1954  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1955  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1956  *    zone defense algorithm.
1957  *
1958  * If a zone is deemed to be full of pinned pages then just give it a light
1959  * scan then give up on it.
1960  */
1961 static void shrink_zones(int priority, struct zonelist *zonelist,
1962                                         struct scan_control *sc)
1963 {
1964         struct zoneref *z;
1965         struct zone *zone;
1966
1967         for_each_zone_zonelist_nodemask(zone, z, zonelist,
1968                                         gfp_zone(sc->gfp_mask), sc->nodemask) {
1969                 if (!populated_zone(zone))
1970                         continue;
1971                 /*
1972                  * Take care memory controller reclaiming has small influence
1973                  * to global LRU.
1974                  */
1975                 if (scanning_global_lru(sc)) {
1976                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1977                                 continue;
1978                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1979                                 continue;       /* Let kswapd poll it */
1980                 }
1981
1982                 shrink_zone(priority, zone, sc);
1983         }
1984 }
1985
1986 static bool zone_reclaimable(struct zone *zone)
1987 {
1988         return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
1989 }
1990
1991 /*
1992  * As hibernation is going on, kswapd is freezed so that it can't mark
1993  * the zone into all_unreclaimable. It can't handle OOM during hibernation.
1994  * So let's check zone's unreclaimable in direct reclaim as well as kswapd.
1995  */
1996 static bool all_unreclaimable(struct zonelist *zonelist,
1997                 struct scan_control *sc)
1998 {
1999         struct zoneref *z;
2000         struct zone *zone;
2001         bool all_unreclaimable = true;
2002
2003         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2004                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2005                 if (!populated_zone(zone))
2006                         continue;
2007                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2008                         continue;
2009                 if (zone_reclaimable(zone)) {
2010                         all_unreclaimable = false;
2011                         break;
2012                 }
2013         }
2014
2015         return all_unreclaimable;
2016 }
2017
2018 /*
2019  * This is the main entry point to direct page reclaim.
2020  *
2021  * If a full scan of the inactive list fails to free enough memory then we
2022  * are "out of memory" and something needs to be killed.
2023  *
2024  * If the caller is !__GFP_FS then the probability of a failure is reasonably
2025  * high - the zone may be full of dirty or under-writeback pages, which this
2026  * caller can't do much about.  We kick the writeback threads and take explicit
2027  * naps in the hope that some of these pages can be written.  But if the
2028  * allocating task holds filesystem locks which prevent writeout this might not
2029  * work, and the allocation attempt will fail.
2030  *
2031  * returns:     0, if no pages reclaimed
2032  *              else, the number of pages reclaimed
2033  */
2034 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2035                                         struct scan_control *sc)
2036 {
2037         int priority;
2038         unsigned long total_scanned = 0;
2039         struct reclaim_state *reclaim_state = current->reclaim_state;
2040         struct zoneref *z;
2041         struct zone *zone;
2042         unsigned long writeback_threshold;
2043
2044         get_mems_allowed();
2045         delayacct_freepages_start();
2046
2047         if (scanning_global_lru(sc))
2048                 count_vm_event(ALLOCSTALL);
2049
2050         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2051                 sc->nr_scanned = 0;
2052                 if (!priority)
2053                         disable_swap_token();
2054                 shrink_zones(priority, zonelist, sc);
2055                 /*
2056                  * Don't shrink slabs when reclaiming memory from
2057                  * over limit cgroups
2058                  */
2059                 if (scanning_global_lru(sc)) {
2060                         unsigned long lru_pages = 0;
2061                         for_each_zone_zonelist(zone, z, zonelist,
2062                                         gfp_zone(sc->gfp_mask)) {
2063                                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2064                                         continue;
2065
2066                                 lru_pages += zone_reclaimable_pages(zone);
2067                         }
2068
2069                         shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
2070                         if (reclaim_state) {
2071                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2072                                 reclaim_state->reclaimed_slab = 0;
2073                         }
2074                 }
2075                 total_scanned += sc->nr_scanned;
2076                 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2077                         goto out;
2078
2079                 /*
2080                  * Try to write back as many pages as we just scanned.  This
2081                  * tends to cause slow streaming writers to write data to the
2082                  * disk smoothly, at the dirtying rate, which is nice.   But
2083                  * that's undesirable in laptop mode, where we *want* lumpy
2084                  * writeout.  So in laptop mode, write out the whole world.
2085                  */
2086                 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2087                 if (total_scanned > writeback_threshold) {
2088                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
2089                         sc->may_writepage = 1;
2090                 }
2091
2092                 /* Take a nap, wait for some writeback to complete */
2093                 if (!sc->hibernation_mode && sc->nr_scanned &&
2094                     priority < DEF_PRIORITY - 2) {
2095                         struct zone *preferred_zone;
2096
2097                         first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2098                                                 &cpuset_current_mems_allowed,
2099                                                 &preferred_zone);
2100                         wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2101                 }
2102         }
2103
2104 out:
2105         delayacct_freepages_end();
2106         put_mems_allowed();
2107
2108         if (sc->nr_reclaimed)
2109                 return sc->nr_reclaimed;
2110
2111         /* top priority shrink_zones still had more to do? don't OOM, then */
2112         if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
2113                 return 1;
2114
2115         return 0;
2116 }
2117
2118 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2119                                 gfp_t gfp_mask, nodemask_t *nodemask)
2120 {
2121         unsigned long nr_reclaimed;
2122         struct scan_control sc = {
2123                 .gfp_mask = gfp_mask,
2124                 .may_writepage = !laptop_mode,
2125                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2126                 .may_unmap = 1,
2127                 .may_swap = 1,
2128                 .swappiness = vm_swappiness,
2129                 .order = order,
2130                 .mem_cgroup = NULL,
2131                 .nodemask = nodemask,
2132         };
2133
2134         trace_mm_vmscan_direct_reclaim_begin(order,
2135                                 sc.may_writepage,
2136                                 gfp_mask);
2137
2138         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2139
2140         trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2141
2142         return nr_reclaimed;
2143 }
2144
2145 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2146
2147 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2148                                                 gfp_t gfp_mask, bool noswap,
2149                                                 unsigned int swappiness,
2150                                                 struct zone *zone)
2151 {
2152         struct scan_control sc = {
2153                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2154                 .may_writepage = !laptop_mode,
2155                 .may_unmap = 1,
2156                 .may_swap = !noswap,
2157                 .swappiness = swappiness,
2158                 .order = 0,
2159                 .mem_cgroup = mem,
2160         };
2161         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2162                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2163
2164         trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2165                                                       sc.may_writepage,
2166                                                       sc.gfp_mask);
2167
2168         /*
2169          * NOTE: Although we can get the priority field, using it
2170          * here is not a good idea, since it limits the pages we can scan.
2171          * if we don't reclaim here, the shrink_zone from balance_pgdat
2172          * will pick up pages from other mem cgroup's as well. We hack
2173          * the priority and make it zero.
2174          */
2175         shrink_zone(0, zone, &sc);
2176
2177         trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2178
2179         return sc.nr_reclaimed;
2180 }
2181
2182 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
2183                                            gfp_t gfp_mask,
2184                                            bool noswap,
2185                                            unsigned int swappiness)
2186 {
2187         struct zonelist *zonelist;
2188         unsigned long nr_reclaimed;
2189         struct scan_control sc = {
2190                 .may_writepage = !laptop_mode,
2191                 .may_unmap = 1,
2192                 .may_swap = !noswap,
2193                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2194                 .swappiness = swappiness,
2195                 .order = 0,
2196                 .mem_cgroup = mem_cont,
2197                 .nodemask = NULL, /* we don't care the placement */
2198         };
2199
2200         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2201                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2202         zonelist = NODE_DATA(numa_node_id())->node_zonelists;
2203
2204         trace_mm_vmscan_memcg_reclaim_begin(0,
2205                                             sc.may_writepage,
2206                                             sc.gfp_mask);
2207
2208         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2209
2210         trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2211
2212         return nr_reclaimed;
2213 }
2214 #endif
2215
2216 /*
2217  * pgdat_balanced is used when checking if a node is balanced for high-order
2218  * allocations. Only zones that meet watermarks and are in a zone allowed
2219  * by the callers classzone_idx are added to balanced_pages. The total of
2220  * balanced pages must be at least 25% of the zones allowed by classzone_idx
2221  * for the node to be considered balanced. Forcing all zones to be balanced
2222  * for high orders can cause excessive reclaim when there are imbalanced zones.
2223  * The choice of 25% is due to
2224  *   o a 16M DMA zone that is balanced will not balance a zone on any
2225  *     reasonable sized machine
2226  *   o On all other machines, the top zone must be at least a reasonable
2227  *     precentage of the middle zones. For example, on 32-bit x86, highmem
2228  *     would need to be at least 256M for it to be balance a whole node.
2229  *     Similarly, on x86-64 the Normal zone would need to be at least 1G
2230  *     to balance a node on its own. These seemed like reasonable ratios.
2231  */
2232 static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2233                                                 int classzone_idx)
2234 {
2235         unsigned long present_pages = 0;
2236         int i;
2237
2238         for (i = 0; i <= classzone_idx; i++)
2239                 present_pages += pgdat->node_zones[i].present_pages;
2240
2241         return balanced_pages > (present_pages >> 2);
2242 }
2243
2244 /* is kswapd sleeping prematurely? */
2245 static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2246                                         int classzone_idx)
2247 {
2248         int i;
2249         unsigned long balanced = 0;
2250         bool all_zones_ok = true;
2251
2252         /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2253         if (remaining)
2254                 return true;
2255
2256         /* Check the watermark levels */
2257         for (i = 0; i < pgdat->nr_zones; i++) {
2258                 struct zone *zone = pgdat->node_zones + i;
2259
2260                 if (!populated_zone(zone))
2261                         continue;
2262
2263                 /*
2264                  * balance_pgdat() skips over all_unreclaimable after
2265                  * DEF_PRIORITY. Effectively, it considers them balanced so
2266                  * they must be considered balanced here as well if kswapd
2267                  * is to sleep
2268                  */
2269                 if (zone->all_unreclaimable) {
2270                         balanced += zone->present_pages;
2271                         continue;
2272                 }
2273
2274                 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2275                                                         classzone_idx, 0))
2276                         all_zones_ok = false;
2277                 else
2278                         balanced += zone->present_pages;
2279         }
2280
2281         /*
2282          * For high-order requests, the balanced zones must contain at least
2283          * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2284          * must be balanced
2285          */
2286         if (order)
2287                 return pgdat_balanced(pgdat, balanced, classzone_idx);
2288         else
2289                 return !all_zones_ok;
2290 }
2291
2292 /*
2293  * For kswapd, balance_pgdat() will work across all this node's zones until
2294  * they are all at high_wmark_pages(zone).
2295  *
2296  * Returns the final order kswapd was reclaiming at
2297  *
2298  * There is special handling here for zones which are full of pinned pages.
2299  * This can happen if the pages are all mlocked, or if they are all used by
2300  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
2301  * What we do is to detect the case where all pages in the zone have been
2302  * scanned twice and there has been zero successful reclaim.  Mark the zone as
2303  * dead and from now on, only perform a short scan.  Basically we're polling
2304  * the zone for when the problem goes away.
2305  *
2306  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
2307  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2308  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2309  * lower zones regardless of the number of free pages in the lower zones. This
2310  * interoperates with the page allocator fallback scheme to ensure that aging
2311  * of pages is balanced across the zones.
2312  */
2313 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2314                                                         int *classzone_idx)
2315 {
2316         int all_zones_ok;
2317         unsigned long balanced;
2318         int priority;
2319         int i;
2320         int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
2321         unsigned long total_scanned;
2322         struct reclaim_state *reclaim_state = current->reclaim_state;
2323         struct scan_control sc = {
2324                 .gfp_mask = GFP_KERNEL,
2325                 .may_unmap = 1,
2326                 .may_swap = 1,
2327                 /*
2328                  * kswapd doesn't want to be bailed out while reclaim. because
2329                  * we want to put equal scanning pressure on each zone.
2330                  */
2331                 .nr_to_reclaim = ULONG_MAX,
2332                 .swappiness = vm_swappiness,
2333                 .order = order,
2334                 .mem_cgroup = NULL,
2335         };
2336 loop_again:
2337         total_scanned = 0;
2338         sc.nr_reclaimed = 0;
2339         sc.may_writepage = !laptop_mode;
2340         count_vm_event(PAGEOUTRUN);
2341
2342         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2343                 unsigned long lru_pages = 0;
2344                 int has_under_min_watermark_zone = 0;
2345
2346                 /* The swap token gets in the way of swapout... */
2347                 if (!priority)
2348                         disable_swap_token();
2349
2350                 all_zones_ok = 1;
2351                 balanced = 0;
2352
2353                 /*
2354                  * Scan in the highmem->dma direction for the highest
2355                  * zone which needs scanning
2356                  */
2357                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2358                         struct zone *zone = pgdat->node_zones + i;
2359
2360                         if (!populated_zone(zone))
2361                                 continue;
2362
2363                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2364                                 continue;
2365
2366                         /*
2367                          * Do some background aging of the anon list, to give
2368                          * pages a chance to be referenced before reclaiming.
2369                          */
2370                         if (inactive_anon_is_low(zone, &sc))
2371                                 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2372                                                         &sc, priority, 0);
2373
2374                         if (!zone_watermark_ok_safe(zone, order,
2375                                         high_wmark_pages(zone), 0, 0)) {
2376                                 end_zone = i;
2377                                 *classzone_idx = i;
2378                                 break;
2379                         }
2380                 }
2381                 if (i < 0)
2382                         goto out;
2383
2384                 for (i = 0; i <= end_zone; i++) {
2385                         struct zone *zone = pgdat->node_zones + i;
2386
2387                         lru_pages += zone_reclaimable_pages(zone);
2388                 }
2389
2390                 /*
2391                  * Now scan the zone in the dma->highmem direction, stopping
2392                  * at the last zone which needs scanning.
2393                  *
2394                  * We do this because the page allocator works in the opposite
2395                  * direction.  This prevents the page allocator from allocating
2396                  * pages behind kswapd's direction of progress, which would
2397                  * cause too much scanning of the lower zones.
2398                  */
2399                 for (i = 0; i <= end_zone; i++) {
2400                         int compaction;
2401                         struct zone *zone = pgdat->node_zones + i;
2402                         int nr_slab;
2403
2404                         if (!populated_zone(zone))
2405                                 continue;
2406
2407                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2408                                 continue;
2409
2410                         sc.nr_scanned = 0;
2411
2412                         /*
2413                          * Call soft limit reclaim before calling shrink_zone.
2414                          * For now we ignore the return value
2415                          */
2416                         mem_cgroup_soft_limit_reclaim(zone, order, sc.gfp_mask);
2417
2418                         /*
2419                          * We put equal pressure on every zone, unless one
2420                          * zone has way too many pages free already.
2421                          */
2422                         if (!zone_watermark_ok_safe(zone, order,
2423                                         8*high_wmark_pages(zone), end_zone, 0))
2424                                 shrink_zone(priority, zone, &sc);
2425                         reclaim_state->reclaimed_slab = 0;
2426                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
2427                                                 lru_pages);
2428                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2429                         total_scanned += sc.nr_scanned;
2430
2431                         compaction = 0;
2432                         if (order &&
2433                             zone_watermark_ok(zone, 0,
2434                                                high_wmark_pages(zone),
2435                                               end_zone, 0) &&
2436                             !zone_watermark_ok(zone, order,
2437                                                high_wmark_pages(zone),
2438                                                end_zone, 0)) {
2439                                 compact_zone_order(zone,
2440                                                    order,
2441                                                    sc.gfp_mask, false,
2442                                                    COMPACT_MODE_KSWAPD);
2443                                 compaction = 1;
2444                         }
2445
2446                         if (zone->all_unreclaimable)
2447                                 continue;
2448                         if (!compaction && nr_slab == 0 &&
2449                             !zone_reclaimable(zone))
2450                                 zone->all_unreclaimable = 1;
2451                         /*
2452                          * If we've done a decent amount of scanning and
2453                          * the reclaim ratio is low, start doing writepage
2454                          * even in laptop mode
2455                          */
2456                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2457                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2458                                 sc.may_writepage = 1;
2459
2460                         if (!zone_watermark_ok_safe(zone, order,
2461                                         high_wmark_pages(zone), end_zone, 0)) {
2462                                 all_zones_ok = 0;
2463                                 /*
2464                                  * We are still under min water mark.  This
2465                                  * means that we have a GFP_ATOMIC allocation
2466                                  * failure risk. Hurry up!
2467                                  */
2468                                 if (!zone_watermark_ok_safe(zone, order,
2469                                             min_wmark_pages(zone), end_zone, 0))
2470                                         has_under_min_watermark_zone = 1;
2471                         } else {
2472                                 /*
2473                                  * If a zone reaches its high watermark,
2474                                  * consider it to be no longer congested. It's
2475                                  * possible there are dirty pages backed by
2476                                  * congested BDIs but as pressure is relieved,
2477                                  * spectulatively avoid congestion waits
2478                                  */
2479                                 zone_clear_flag(zone, ZONE_CONGESTED);
2480                                 if (i <= *classzone_idx)
2481                                         balanced += zone->present_pages;
2482                         }
2483
2484                 }
2485                 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2486                         break;          /* kswapd: all done */
2487                 /*
2488                  * OK, kswapd is getting into trouble.  Take a nap, then take
2489                  * another pass across the zones.
2490                  */
2491                 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2492                         if (has_under_min_watermark_zone)
2493                                 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2494                         else
2495                                 congestion_wait(BLK_RW_ASYNC, HZ/10);
2496                 }
2497
2498                 /*
2499                  * We do this so kswapd doesn't build up large priorities for
2500                  * example when it is freeing in parallel with allocators. It
2501                  * matches the direct reclaim path behaviour in terms of impact
2502                  * on zone->*_priority.
2503                  */
2504                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2505                         break;
2506         }
2507 out:
2508
2509         /*
2510          * order-0: All zones must meet high watermark for a balanced node
2511          * high-order: Balanced zones must make up at least 25% of the node
2512          *             for the node to be balanced
2513          */
2514         if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2515                 cond_resched();
2516
2517                 try_to_freeze();
2518
2519                 /*
2520                  * Fragmentation may mean that the system cannot be
2521                  * rebalanced for high-order allocations in all zones.
2522                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2523                  * it means the zones have been fully scanned and are still
2524                  * not balanced. For high-order allocations, there is
2525                  * little point trying all over again as kswapd may
2526                  * infinite loop.
2527                  *
2528                  * Instead, recheck all watermarks at order-0 as they
2529                  * are the most important. If watermarks are ok, kswapd will go
2530                  * back to sleep. High-order users can still perform direct
2531                  * reclaim if they wish.
2532                  */
2533                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2534                         order = sc.order = 0;
2535
2536                 goto loop_again;
2537         }
2538
2539         /*
2540          * If kswapd was reclaiming at a higher order, it has the option of
2541          * sleeping without all zones being balanced. Before it does, it must
2542          * ensure that the watermarks for order-0 on *all* zones are met and
2543          * that the congestion flags are cleared. The congestion flag must
2544          * be cleared as kswapd is the only mechanism that clears the flag
2545          * and it is potentially going to sleep here.
2546          */
2547         if (order) {
2548                 for (i = 0; i <= end_zone; i++) {
2549                         struct zone *zone = pgdat->node_zones + i;
2550
2551                         if (!populated_zone(zone))
2552                                 continue;
2553
2554                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2555                                 continue;
2556
2557                         /* Confirm the zone is balanced for order-0 */
2558                         if (!zone_watermark_ok(zone, 0,
2559                                         high_wmark_pages(zone), 0, 0)) {
2560                                 order = sc.order = 0;
2561                                 goto loop_again;
2562                         }
2563
2564                         /* If balanced, clear the congested flag */
2565                         zone_clear_flag(zone, ZONE_CONGESTED);
2566                 }
2567         }
2568
2569         /*
2570          * Return the order we were reclaiming at so sleeping_prematurely()
2571          * makes a decision on the order we were last reclaiming at. However,
2572          * if another caller entered the allocator slow path while kswapd
2573          * was awake, order will remain at the higher level
2574          */
2575         *classzone_idx = end_zone;
2576         return order;
2577 }
2578
2579 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2580 {
2581         long remaining = 0;
2582         DEFINE_WAIT(wait);
2583
2584         if (freezing(current) || kthread_should_stop())
2585                 return;
2586
2587         prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2588
2589         /* Try to sleep for a short interval */
2590         if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2591                 remaining = schedule_timeout(HZ/10);
2592                 finish_wait(&pgdat->kswapd_wait, &wait);
2593                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2594         }
2595
2596         /*
2597          * After a short sleep, check if it was a premature sleep. If not, then
2598          * go fully to sleep until explicitly woken up.
2599          */
2600         if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2601                 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2602
2603                 /*
2604                  * vmstat counters are not perfectly accurate and the estimated
2605                  * value for counters such as NR_FREE_PAGES can deviate from the
2606                  * true value by nr_online_cpus * threshold. To avoid the zone
2607                  * watermarks being breached while under pressure, we reduce the
2608                  * per-cpu vmstat threshold while kswapd is awake and restore
2609                  * them before going back to sleep.
2610                  */
2611                 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2612                 schedule();
2613                 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2614         } else {
2615                 if (remaining)
2616                         count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2617                 else
2618                         count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2619         }
2620         finish_wait(&pgdat->kswapd_wait, &wait);
2621 }
2622
2623 /*
2624  * The background pageout daemon, started as a kernel thread
2625  * from the init process.
2626  *
2627  * This basically trickles out pages so that we have _some_
2628  * free memory available even if there is no other activity
2629  * that frees anything up. This is needed for things like routing
2630  * etc, where we otherwise might have all activity going on in
2631  * asynchronous contexts that cannot page things out.
2632  *
2633  * If there are applications that are active memory-allocators
2634  * (most normal use), this basically shouldn't matter.
2635  */
2636 static int kswapd(void *p)
2637 {
2638         unsigned long order;
2639         int classzone_idx;
2640         pg_data_t *pgdat = (pg_data_t*)p;
2641         struct task_struct *tsk = current;
2642
2643         struct reclaim_state reclaim_state = {
2644                 .reclaimed_slab = 0,
2645         };
2646         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2647
2648         lockdep_set_current_reclaim_state(GFP_KERNEL);
2649
2650         if (!cpumask_empty(cpumask))
2651                 set_cpus_allowed_ptr(tsk, cpumask);
2652         current->reclaim_state = &reclaim_state;
2653
2654         /*
2655          * Tell the memory management that we're a "memory allocator",
2656          * and that if we need more memory we should get access to it
2657          * regardless (see "__alloc_pages()"). "kswapd" should
2658          * never get caught in the normal page freeing logic.
2659          *
2660          * (Kswapd normally doesn't need memory anyway, but sometimes
2661          * you need a small amount of memory in order to be able to
2662          * page out something else, and this flag essentially protects
2663          * us from recursively trying to free more memory as we're
2664          * trying to free the first piece of memory in the first place).
2665          */
2666         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2667         set_freezable();
2668
2669         order = 0;
2670         classzone_idx = MAX_NR_ZONES - 1;
2671         for ( ; ; ) {
2672                 unsigned long new_order;
2673                 int new_classzone_idx;
2674                 int ret;
2675
2676                 new_order = pgdat->kswapd_max_order;
2677                 new_classzone_idx = pgdat->classzone_idx;
2678                 pgdat->kswapd_max_order = 0;
2679                 pgdat->classzone_idx = MAX_NR_ZONES - 1;
2680                 if (order < new_order || classzone_idx > new_classzone_idx) {
2681                         /*
2682                          * Don't sleep if someone wants a larger 'order'
2683                          * allocation or has tigher zone constraints
2684                          */
2685                         order = new_order;
2686                         classzone_idx = new_classzone_idx;
2687                 } else {
2688                         kswapd_try_to_sleep(pgdat, order, classzone_idx);
2689                         order = pgdat->kswapd_max_order;
2690                         classzone_idx = pgdat->classzone_idx;
2691                         pgdat->kswapd_max_order = 0;
2692                         pgdat->classzone_idx = MAX_NR_ZONES - 1;
2693                 }
2694
2695                 ret = try_to_freeze();
2696                 if (kthread_should_stop())
2697                         break;
2698
2699                 /*
2700                  * We can speed up thawing tasks if we don't call balance_pgdat
2701                  * after returning from the refrigerator
2702                  */
2703                 if (!ret) {
2704                         trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2705                         order = balance_pgdat(pgdat, order, &classzone_idx);
2706                 }
2707         }
2708         return 0;
2709 }
2710
2711 /*
2712  * A zone is low on free memory, so wake its kswapd task to service it.
2713  */
2714 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
2715 {
2716         pg_data_t *pgdat;
2717
2718         if (!populated_zone(zone))
2719                 return;
2720
2721         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2722                 return;
2723         pgdat = zone->zone_pgdat;
2724         if (pgdat->kswapd_max_order < order) {
2725                 pgdat->kswapd_max_order = order;
2726                 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
2727         }
2728         if (!waitqueue_active(&pgdat->kswapd_wait))
2729                 return;
2730         if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
2731                 return;
2732
2733         trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
2734         wake_up_interruptible(&pgdat->kswapd_wait);
2735 }
2736
2737 /*
2738  * The reclaimable count would be mostly accurate.
2739  * The less reclaimable pages may be
2740  * - mlocked pages, which will be moved to unevictable list when encountered
2741  * - mapped pages, which may require several travels to be reclaimed
2742  * - dirty pages, which is not "instantly" reclaimable
2743  */
2744 unsigned long global_reclaimable_pages(void)
2745 {
2746         int nr;
2747
2748         nr = global_page_state(NR_ACTIVE_FILE) +
2749              global_page_state(NR_INACTIVE_FILE);
2750
2751         if (nr_swap_pages > 0)
2752                 nr += global_page_state(NR_ACTIVE_ANON) +
2753                       global_page_state(NR_INACTIVE_ANON);
2754
2755         return nr;
2756 }
2757
2758 unsigned long zone_reclaimable_pages(struct zone *zone)
2759 {
2760         int nr;
2761
2762         nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2763              zone_page_state(zone, NR_INACTIVE_FILE);
2764
2765         if (nr_swap_pages > 0)
2766                 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2767                       zone_page_state(zone, NR_INACTIVE_ANON);
2768
2769         return nr;
2770 }
2771
2772 #ifdef CONFIG_HIBERNATION
2773 /*
2774  * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2775  * freed pages.
2776  *
2777  * Rather than trying to age LRUs the aim is to preserve the overall
2778  * LRU order by reclaiming preferentially
2779  * inactive > active > active referenced > active mapped
2780  */
2781 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
2782 {
2783         struct reclaim_state reclaim_state;
2784         struct scan_control sc = {
2785                 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2786                 .may_swap = 1,
2787                 .may_unmap = 1,
2788                 .may_writepage = 1,
2789                 .nr_to_reclaim = nr_to_reclaim,
2790                 .hibernation_mode = 1,
2791                 .swappiness = vm_swappiness,
2792                 .order = 0,
2793         };
2794         struct zonelist * zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2795         struct task_struct *p = current;
2796         unsigned long nr_reclaimed;
2797
2798         p->flags |= PF_MEMALLOC;
2799         lockdep_set_current_reclaim_state(sc.gfp_mask);
2800         reclaim_state.reclaimed_slab = 0;
2801         p->reclaim_state = &reclaim_state;
2802
2803         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2804
2805         p->reclaim_state = NULL;
2806         lockdep_clear_current_reclaim_state();
2807         p->flags &= ~PF_MEMALLOC;
2808
2809         return nr_reclaimed;
2810 }
2811 #endif /* CONFIG_HIBERNATION */
2812
2813 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2814    not required for correctness.  So if the last cpu in a node goes
2815    away, we get changed to run anywhere: as the first one comes back,
2816    restore their cpu bindings. */
2817 static int __devinit cpu_callback(struct notifier_block *nfb,
2818                                   unsigned long action, void *hcpu)
2819 {
2820         int nid;
2821
2822         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2823                 for_each_node_state(nid, N_HIGH_MEMORY) {
2824                         pg_data_t *pgdat = NODE_DATA(nid);
2825                         const struct cpumask *mask;
2826
2827                         mask = cpumask_of_node(pgdat->node_id);
2828
2829                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2830                                 /* One of our CPUs online: restore mask */
2831                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2832                 }
2833         }
2834         return NOTIFY_OK;
2835 }
2836
2837 /*
2838  * This kswapd start function will be called by init and node-hot-add.
2839  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2840  */
2841 int kswapd_run(int nid)
2842 {
2843         pg_data_t *pgdat = NODE_DATA(nid);
2844         int ret = 0;
2845
2846         if (pgdat->kswapd)
2847                 return 0;
2848
2849         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2850         if (IS_ERR(pgdat->kswapd)) {
2851                 /* failure at boot is fatal */
2852                 BUG_ON(system_state == SYSTEM_BOOTING);
2853                 printk("Failed to start kswapd on node %d\n",nid);
2854                 ret = -1;
2855         }
2856         return ret;
2857 }
2858
2859 /*
2860  * Called by memory hotplug when all memory in a node is offlined.
2861  */
2862 void kswapd_stop(int nid)
2863 {
2864         struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
2865
2866         if (kswapd)
2867                 kthread_stop(kswapd);
2868 }
2869
2870 static int __init kswapd_init(void)
2871 {
2872         int nid;
2873
2874         swap_setup();
2875         for_each_node_state(nid, N_HIGH_MEMORY)
2876                 kswapd_run(nid);
2877         hotcpu_notifier(cpu_callback, 0);
2878         return 0;
2879 }
2880
2881 module_init(kswapd_init)
2882
2883 #ifdef CONFIG_NUMA
2884 /*
2885  * Zone reclaim mode
2886  *
2887  * If non-zero call zone_reclaim when the number of free pages falls below
2888  * the watermarks.
2889  */
2890 int zone_reclaim_mode __read_mostly;
2891
2892 #define RECLAIM_OFF 0
2893 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
2894 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
2895 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
2896
2897 /*
2898  * Priority for ZONE_RECLAIM. This determines the fraction of pages
2899  * of a node considered for each zone_reclaim. 4 scans 1/16th of
2900  * a zone.
2901  */
2902 #define ZONE_RECLAIM_PRIORITY 4
2903
2904 /*
2905  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2906  * occur.
2907  */
2908 int sysctl_min_unmapped_ratio = 1;
2909
2910 /*
2911  * If the number of slab pages in a zone grows beyond this percentage then
2912  * slab reclaim needs to occur.
2913  */
2914 int sysctl_min_slab_ratio = 5;
2915
2916 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
2917 {
2918         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
2919         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
2920                 zone_page_state(zone, NR_ACTIVE_FILE);
2921
2922         /*
2923          * It's possible for there to be more file mapped pages than
2924          * accounted for by the pages on the file LRU lists because
2925          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2926          */
2927         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
2928 }
2929
2930 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
2931 static long zone_pagecache_reclaimable(struct zone *zone)
2932 {
2933         long nr_pagecache_reclaimable;
2934         long delta = 0;
2935
2936         /*
2937          * If RECLAIM_SWAP is set, then all file pages are considered
2938          * potentially reclaimable. Otherwise, we have to worry about
2939          * pages like swapcache and zone_unmapped_file_pages() provides
2940          * a better estimate
2941          */
2942         if (zone_reclaim_mode & RECLAIM_SWAP)
2943                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
2944         else
2945                 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
2946
2947         /* If we can't clean pages, remove dirty pages from consideration */
2948         if (!(zone_reclaim_mode & RECLAIM_WRITE))
2949                 delta += zone_page_state(zone, NR_FILE_DIRTY);
2950
2951         /* Watch for any possible underflows due to delta */
2952         if (unlikely(delta > nr_pagecache_reclaimable))
2953                 delta = nr_pagecache_reclaimable;
2954
2955         return nr_pagecache_reclaimable - delta;
2956 }
2957
2958 /*
2959  * Try to free up some pages from this zone through reclaim.
2960  */
2961 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2962 {
2963         /* Minimum pages needed in order to stay on node */
2964         const unsigned long nr_pages = 1 << order;
2965         struct task_struct *p = current;
2966         struct reclaim_state reclaim_state;
2967         int priority;
2968         struct scan_control sc = {
2969                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2970                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2971                 .may_swap = 1,
2972                 .nr_to_reclaim = max_t(unsigned long, nr_pages,
2973                                        SWAP_CLUSTER_MAX),
2974                 .gfp_mask = gfp_mask,
2975                 .swappiness = vm_swappiness,
2976                 .order = order,
2977         };
2978         unsigned long nr_slab_pages0, nr_slab_pages1;
2979
2980         cond_resched();
2981         /*
2982          * We need to be able to allocate from the reserves for RECLAIM_SWAP
2983          * and we also need to be able to write out pages for RECLAIM_WRITE
2984          * and RECLAIM_SWAP.
2985          */
2986         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2987         lockdep_set_current_reclaim_state(gfp_mask);
2988         reclaim_state.reclaimed_slab = 0;
2989         p->reclaim_state = &reclaim_state;
2990
2991         if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
2992                 /*
2993                  * Free memory by calling shrink zone with increasing
2994                  * priorities until we have enough memory freed.
2995                  */
2996                 priority = ZONE_RECLAIM_PRIORITY;
2997                 do {
2998                         shrink_zone(priority, zone, &sc);
2999                         priority--;
3000                 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
3001         }
3002
3003         nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3004         if (nr_slab_pages0 > zone->min_slab_pages) {
3005                 /*
3006                  * shrink_slab() does not currently allow us to determine how
3007                  * many pages were freed in this zone. So we take the current
3008                  * number of slab pages and shake the slab until it is reduced
3009                  * by the same nr_pages that we used for reclaiming unmapped
3010                  * pages.
3011                  *
3012                  * Note that shrink_slab will free memory on all zones and may
3013                  * take a long time.
3014                  */
3015                 for (;;) {
3016                         unsigned long lru_pages = zone_reclaimable_pages(zone);
3017
3018                         /* No reclaimable slab or very low memory pressure */
3019                         if (!shrink_slab(sc.nr_scanned, gfp_mask, lru_pages))
3020                                 break;
3021
3022                         /* Freed enough memory */
3023                         nr_slab_pages1 = zone_page_state(zone,
3024                                                         NR_SLAB_RECLAIMABLE);
3025                         if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3026                                 break;
3027                 }
3028
3029                 /*
3030                  * Update nr_reclaimed by the number of slab pages we
3031                  * reclaimed from this zone.
3032                  */
3033                 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3034                 if (nr_slab_pages1 < nr_slab_pages0)
3035                         sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3036         }
3037
3038         p->reclaim_state = NULL;
3039         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3040         lockdep_clear_current_reclaim_state();
3041         return sc.nr_reclaimed >= nr_pages;
3042 }
3043
3044 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3045 {
3046         int node_id;
3047         int ret;
3048
3049         /*
3050          * Zone reclaim reclaims unmapped file backed pages and
3051          * slab pages if we are over the defined limits.
3052          *
3053          * A small portion of unmapped file backed pages is needed for
3054          * file I/O otherwise pages read by file I/O will be immediately
3055          * thrown out if the zone is overallocated. So we do not reclaim
3056          * if less than a specified percentage of the zone is used by
3057          * unmapped file backed pages.
3058          */
3059         if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3060             zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3061                 return ZONE_RECLAIM_FULL;
3062
3063         if (zone->all_unreclaimable)
3064                 return ZONE_RECLAIM_FULL;
3065
3066         /*
3067          * Do not scan if the allocation should not be delayed.
3068          */
3069         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3070                 return ZONE_RECLAIM_NOSCAN;
3071
3072         /*
3073          * Only run zone reclaim on the local zone or on zones that do not
3074          * have associated processors. This will favor the local processor
3075          * over remote processors and spread off node memory allocations
3076          * as wide as possible.
3077          */
3078         node_id = zone_to_nid(zone);
3079         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3080                 return ZONE_RECLAIM_NOSCAN;
3081
3082         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3083                 return ZONE_RECLAIM_NOSCAN;
3084
3085         ret = __zone_reclaim(zone, gfp_mask, order);
3086         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3087
3088         if (!ret)
3089                 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3090
3091         return ret;
3092 }
3093 #endif
3094
3095 /*
3096  * page_evictable - test whether a page is evictable
3097  * @page: the page to test
3098  * @vma: the VMA in which the page is or will be mapped, may be NULL
3099  *
3100  * Test whether page is evictable--i.e., should be placed on active/inactive
3101  * lists vs unevictable list.  The vma argument is !NULL when called from the
3102  * fault path to determine how to instantate a new page.
3103  *
3104  * Reasons page might not be evictable:
3105  * (1) page's mapping marked unevictable
3106  * (2) page is part of an mlocked VMA
3107  *
3108  */
3109 int page_evictable(struct page *page, struct vm_area_struct *vma)
3110 {
3111
3112         if (mapping_unevictable(page_mapping(page)))
3113                 return 0;
3114
3115         if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
3116                 return 0;
3117
3118         return 1;
3119 }
3120
3121 /**
3122  * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
3123  * @page: page to check evictability and move to appropriate lru list
3124  * @zone: zone page is in
3125  *
3126  * Checks a page for evictability and moves the page to the appropriate
3127  * zone lru list.
3128  *
3129  * Restrictions: zone->lru_lock must be held, page must be on LRU and must
3130  * have PageUnevictable set.
3131  */
3132 static void check_move_unevictable_page(struct page *page, struct zone *zone)
3133 {
3134         VM_BUG_ON(PageActive(page));
3135
3136 retry:
3137         ClearPageUnevictable(page);
3138         if (page_evictable(page, NULL)) {
3139                 enum lru_list l = page_lru_base_type(page);
3140
3141                 __dec_zone_state(zone, NR_UNEVICTABLE);
3142                 list_move(&page->lru, &zone->lru[l].list);
3143                 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
3144                 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
3145                 __count_vm_event(UNEVICTABLE_PGRESCUED);
3146         } else {
3147                 /*
3148                  * rotate unevictable list
3149                  */
3150                 SetPageUnevictable(page);
3151                 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
3152                 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
3153                 if (page_evictable(page, NULL))
3154                         goto retry;
3155         }
3156 }
3157
3158 /**
3159  * scan_mapping_unevictable_pages - scan an address space for evictable pages
3160  * @mapping: struct address_space to scan for evictable pages
3161  *
3162  * Scan all pages in mapping.  Check unevictable pages for
3163  * evictability and move them to the appropriate zone lru list.
3164  */
3165 void scan_mapping_unevictable_pages(struct address_space *mapping)
3166 {
3167         pgoff_t next = 0;
3168         pgoff_t end   = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
3169                          PAGE_CACHE_SHIFT;
3170         struct zone *zone;
3171         struct pagevec pvec;
3172
3173         if (mapping->nrpages == 0)
3174                 return;
3175
3176         pagevec_init(&pvec, 0);
3177         while (next < end &&
3178                 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
3179                 int i;
3180                 int pg_scanned = 0;
3181
3182                 zone = NULL;
3183
3184                 for (i = 0; i < pagevec_count(&pvec); i++) {
3185                         struct page *page = pvec.pages[i];
3186                         pgoff_t page_index = page->index;
3187                         struct zone *pagezone = page_zone(page);
3188
3189                         pg_scanned++;
3190                         if (page_index > next)
3191                                 next = page_index;
3192                         next++;
3193
3194                         if (pagezone != zone) {
3195                                 if (zone)
3196                                         spin_unlock_irq(&zone->lru_lock);
3197                                 zone = pagezone;
3198                                 spin_lock_irq(&zone->lru_lock);
3199                         }
3200
3201                         if (PageLRU(page) && PageUnevictable(page))
3202                                 check_move_unevictable_page(page, zone);
3203                 }
3204                 if (zone)
3205                         spin_unlock_irq(&zone->lru_lock);
3206                 pagevec_release(&pvec);
3207
3208                 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
3209         }
3210
3211 }
3212
3213 /**
3214  * scan_zone_unevictable_pages - check unevictable list for evictable pages
3215  * @zone - zone of which to scan the unevictable list
3216  *
3217  * Scan @zone's unevictable LRU lists to check for pages that have become
3218  * evictable.  Move those that have to @zone's inactive list where they
3219  * become candidates for reclaim, unless shrink_inactive_zone() decides
3220  * to reactivate them.  Pages that are still unevictable are rotated
3221  * back onto @zone's unevictable list.
3222  */
3223 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
3224 static void scan_zone_unevictable_pages(struct zone *zone)
3225 {
3226         struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
3227         unsigned long scan;
3228         unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
3229
3230         while (nr_to_scan > 0) {
3231                 unsigned long batch_size = min(nr_to_scan,
3232                                                 SCAN_UNEVICTABLE_BATCH_SIZE);
3233
3234                 spin_lock_irq(&zone->lru_lock);
3235                 for (scan = 0;  scan < batch_size; scan++) {
3236                         struct page *page = lru_to_page(l_unevictable);
3237
3238                         if (!trylock_page(page))
3239                                 continue;
3240
3241                         prefetchw_prev_lru_page(page, l_unevictable, flags);
3242
3243                         if (likely(PageLRU(page) && PageUnevictable(page)))
3244                                 check_move_unevictable_page(page, zone);
3245
3246                         unlock_page(page);
3247                 }
3248                 spin_unlock_irq(&zone->lru_lock);
3249
3250                 nr_to_scan -= batch_size;
3251         }
3252 }
3253
3254
3255 /**
3256  * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
3257  *
3258  * A really big hammer:  scan all zones' unevictable LRU lists to check for
3259  * pages that have become evictable.  Move those back to the zones'
3260  * inactive list where they become candidates for reclaim.
3261  * This occurs when, e.g., we have unswappable pages on the unevictable lists,
3262  * and we add swap to the system.  As such, it runs in the context of a task
3263  * that has possibly/probably made some previously unevictable pages
3264  * evictable.
3265  */
3266 static void scan_all_zones_unevictable_pages(void)
3267 {
3268         struct zone *zone;
3269
3270         for_each_zone(zone) {
3271                 scan_zone_unevictable_pages(zone);
3272         }
3273 }
3274
3275 /*
3276  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
3277  * all nodes' unevictable lists for evictable pages
3278  */
3279 unsigned long scan_unevictable_pages;
3280
3281 int scan_unevictable_handler(struct ctl_table *table, int write,
3282                            void __user *buffer,
3283                            size_t *length, loff_t *ppos)
3284 {
3285         proc_doulongvec_minmax(table, write, buffer, length, ppos);
3286
3287         if (write && *(unsigned long *)table->data)
3288                 scan_all_zones_unevictable_pages();
3289
3290         scan_unevictable_pages = 0;
3291         return 0;
3292 }
3293
3294 #ifdef CONFIG_NUMA
3295 /*
3296  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
3297  * a specified node's per zone unevictable lists for evictable pages.
3298  */
3299
3300 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
3301                                           struct sysdev_attribute *attr,
3302                                           char *buf)
3303 {
3304         return sprintf(buf, "0\n");     /* always zero; should fit... */
3305 }
3306
3307 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
3308                                            struct sysdev_attribute *attr,
3309                                         const char *buf, size_t count)
3310 {
3311         struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
3312         struct zone *zone;
3313         unsigned long res;
3314         unsigned long req = strict_strtoul(buf, 10, &res);
3315
3316         if (!req)
3317                 return 1;       /* zero is no-op */
3318
3319         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
3320                 if (!populated_zone(zone))
3321                         continue;
3322                 scan_zone_unevictable_pages(zone);
3323         }
3324         return 1;
3325 }
3326
3327
3328 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3329                         read_scan_unevictable_node,
3330                         write_scan_unevictable_node);
3331
3332 int scan_unevictable_register_node(struct node *node)
3333 {
3334         return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
3335 }
3336
3337 void scan_unevictable_unregister_node(struct node *node)
3338 {
3339         sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
3340 }
3341 #endif