mm: account for the number of times direct reclaimers get throttled
[linux-3.10.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/backing-dev.h>
30 #include <linux/rmap.h>
31 #include <linux/topology.h>
32 #include <linux/cpu.h>
33 #include <linux/cpuset.h>
34 #include <linux/compaction.h>
35 #include <linux/notifier.h>
36 #include <linux/rwsem.h>
37 #include <linux/delay.h>
38 #include <linux/kthread.h>
39 #include <linux/freezer.h>
40 #include <linux/memcontrol.h>
41 #include <linux/delayacct.h>
42 #include <linux/sysctl.h>
43 #include <linux/oom.h>
44 #include <linux/prefetch.h>
45
46 #include <asm/tlbflush.h>
47 #include <asm/div64.h>
48
49 #include <linux/swapops.h>
50
51 #include "internal.h"
52
53 #define CREATE_TRACE_POINTS
54 #include <trace/events/vmscan.h>
55
56 struct scan_control {
57         /* Incremented by the number of inactive pages that were scanned */
58         unsigned long nr_scanned;
59
60         /* Number of pages freed so far during a call to shrink_zones() */
61         unsigned long nr_reclaimed;
62
63         /* How many pages shrink_list() should reclaim */
64         unsigned long nr_to_reclaim;
65
66         unsigned long hibernation_mode;
67
68         /* This context's GFP mask */
69         gfp_t gfp_mask;
70
71         int may_writepage;
72
73         /* Can mapped pages be reclaimed? */
74         int may_unmap;
75
76         /* Can pages be swapped as part of reclaim? */
77         int may_swap;
78
79         int order;
80
81         /* Scan (total_size >> priority) pages at once */
82         int priority;
83
84         /*
85          * The memory cgroup that hit its limit and as a result is the
86          * primary target of this reclaim invocation.
87          */
88         struct mem_cgroup *target_mem_cgroup;
89
90         /*
91          * Nodemask of nodes allowed by the caller. If NULL, all nodes
92          * are scanned.
93          */
94         nodemask_t      *nodemask;
95 };
96
97 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
98
99 #ifdef ARCH_HAS_PREFETCH
100 #define prefetch_prev_lru_page(_page, _base, _field)                    \
101         do {                                                            \
102                 if ((_page)->lru.prev != _base) {                       \
103                         struct page *prev;                              \
104                                                                         \
105                         prev = lru_to_page(&(_page->lru));              \
106                         prefetch(&prev->_field);                        \
107                 }                                                       \
108         } while (0)
109 #else
110 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
111 #endif
112
113 #ifdef ARCH_HAS_PREFETCHW
114 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
115         do {                                                            \
116                 if ((_page)->lru.prev != _base) {                       \
117                         struct page *prev;                              \
118                                                                         \
119                         prev = lru_to_page(&(_page->lru));              \
120                         prefetchw(&prev->_field);                       \
121                 }                                                       \
122         } while (0)
123 #else
124 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
125 #endif
126
127 /*
128  * From 0 .. 100.  Higher means more swappy.
129  */
130 int vm_swappiness = 60;
131 long vm_total_pages;    /* The total number of pages which the VM controls */
132
133 static LIST_HEAD(shrinker_list);
134 static DECLARE_RWSEM(shrinker_rwsem);
135
136 #ifdef CONFIG_MEMCG
137 static bool global_reclaim(struct scan_control *sc)
138 {
139         return !sc->target_mem_cgroup;
140 }
141 #else
142 static bool global_reclaim(struct scan_control *sc)
143 {
144         return true;
145 }
146 #endif
147
148 static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru)
149 {
150         if (!mem_cgroup_disabled())
151                 return mem_cgroup_get_lru_size(lruvec, lru);
152
153         return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru);
154 }
155
156 /*
157  * Add a shrinker callback to be called from the vm
158  */
159 void register_shrinker(struct shrinker *shrinker)
160 {
161         atomic_long_set(&shrinker->nr_in_batch, 0);
162         down_write(&shrinker_rwsem);
163         list_add_tail(&shrinker->list, &shrinker_list);
164         up_write(&shrinker_rwsem);
165 }
166 EXPORT_SYMBOL(register_shrinker);
167
168 /*
169  * Remove one
170  */
171 void unregister_shrinker(struct shrinker *shrinker)
172 {
173         down_write(&shrinker_rwsem);
174         list_del(&shrinker->list);
175         up_write(&shrinker_rwsem);
176 }
177 EXPORT_SYMBOL(unregister_shrinker);
178
179 static inline int do_shrinker_shrink(struct shrinker *shrinker,
180                                      struct shrink_control *sc,
181                                      unsigned long nr_to_scan)
182 {
183         sc->nr_to_scan = nr_to_scan;
184         return (*shrinker->shrink)(shrinker, sc);
185 }
186
187 #define SHRINK_BATCH 128
188 /*
189  * Call the shrink functions to age shrinkable caches
190  *
191  * Here we assume it costs one seek to replace a lru page and that it also
192  * takes a seek to recreate a cache object.  With this in mind we age equal
193  * percentages of the lru and ageable caches.  This should balance the seeks
194  * generated by these structures.
195  *
196  * If the vm encountered mapped pages on the LRU it increase the pressure on
197  * slab to avoid swapping.
198  *
199  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
200  *
201  * `lru_pages' represents the number of on-LRU pages in all the zones which
202  * are eligible for the caller's allocation attempt.  It is used for balancing
203  * slab reclaim versus page reclaim.
204  *
205  * Returns the number of slab objects which we shrunk.
206  */
207 unsigned long shrink_slab(struct shrink_control *shrink,
208                           unsigned long nr_pages_scanned,
209                           unsigned long lru_pages)
210 {
211         struct shrinker *shrinker;
212         unsigned long ret = 0;
213
214         if (nr_pages_scanned == 0)
215                 nr_pages_scanned = SWAP_CLUSTER_MAX;
216
217         if (!down_read_trylock(&shrinker_rwsem)) {
218                 /* Assume we'll be able to shrink next time */
219                 ret = 1;
220                 goto out;
221         }
222
223         list_for_each_entry(shrinker, &shrinker_list, list) {
224                 unsigned long long delta;
225                 long total_scan;
226                 long max_pass;
227                 int shrink_ret = 0;
228                 long nr;
229                 long new_nr;
230                 long batch_size = shrinker->batch ? shrinker->batch
231                                                   : SHRINK_BATCH;
232
233                 max_pass = do_shrinker_shrink(shrinker, shrink, 0);
234                 if (max_pass <= 0)
235                         continue;
236
237                 /*
238                  * copy the current shrinker scan count into a local variable
239                  * and zero it so that other concurrent shrinker invocations
240                  * don't also do this scanning work.
241                  */
242                 nr = atomic_long_xchg(&shrinker->nr_in_batch, 0);
243
244                 total_scan = nr;
245                 delta = (4 * nr_pages_scanned) / shrinker->seeks;
246                 delta *= max_pass;
247                 do_div(delta, lru_pages + 1);
248                 total_scan += delta;
249                 if (total_scan < 0) {
250                         printk(KERN_ERR "shrink_slab: %pF negative objects to "
251                                "delete nr=%ld\n",
252                                shrinker->shrink, total_scan);
253                         total_scan = max_pass;
254                 }
255
256                 /*
257                  * We need to avoid excessive windup on filesystem shrinkers
258                  * due to large numbers of GFP_NOFS allocations causing the
259                  * shrinkers to return -1 all the time. This results in a large
260                  * nr being built up so when a shrink that can do some work
261                  * comes along it empties the entire cache due to nr >>>
262                  * max_pass.  This is bad for sustaining a working set in
263                  * memory.
264                  *
265                  * Hence only allow the shrinker to scan the entire cache when
266                  * a large delta change is calculated directly.
267                  */
268                 if (delta < max_pass / 4)
269                         total_scan = min(total_scan, max_pass / 2);
270
271                 /*
272                  * Avoid risking looping forever due to too large nr value:
273                  * never try to free more than twice the estimate number of
274                  * freeable entries.
275                  */
276                 if (total_scan > max_pass * 2)
277                         total_scan = max_pass * 2;
278
279                 trace_mm_shrink_slab_start(shrinker, shrink, nr,
280                                         nr_pages_scanned, lru_pages,
281                                         max_pass, delta, total_scan);
282
283                 while (total_scan >= batch_size) {
284                         int nr_before;
285
286                         nr_before = do_shrinker_shrink(shrinker, shrink, 0);
287                         shrink_ret = do_shrinker_shrink(shrinker, shrink,
288                                                         batch_size);
289                         if (shrink_ret == -1)
290                                 break;
291                         if (shrink_ret < nr_before)
292                                 ret += nr_before - shrink_ret;
293                         count_vm_events(SLABS_SCANNED, batch_size);
294                         total_scan -= batch_size;
295
296                         cond_resched();
297                 }
298
299                 /*
300                  * move the unused scan count back into the shrinker in a
301                  * manner that handles concurrent updates. If we exhausted the
302                  * scan, there is no need to do an update.
303                  */
304                 if (total_scan > 0)
305                         new_nr = atomic_long_add_return(total_scan,
306                                         &shrinker->nr_in_batch);
307                 else
308                         new_nr = atomic_long_read(&shrinker->nr_in_batch);
309
310                 trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr);
311         }
312         up_read(&shrinker_rwsem);
313 out:
314         cond_resched();
315         return ret;
316 }
317
318 static inline int is_page_cache_freeable(struct page *page)
319 {
320         /*
321          * A freeable page cache page is referenced only by the caller
322          * that isolated the page, the page cache radix tree and
323          * optional buffer heads at page->private.
324          */
325         return page_count(page) - page_has_private(page) == 2;
326 }
327
328 static int may_write_to_queue(struct backing_dev_info *bdi,
329                               struct scan_control *sc)
330 {
331         if (current->flags & PF_SWAPWRITE)
332                 return 1;
333         if (!bdi_write_congested(bdi))
334                 return 1;
335         if (bdi == current->backing_dev_info)
336                 return 1;
337         return 0;
338 }
339
340 /*
341  * We detected a synchronous write error writing a page out.  Probably
342  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
343  * fsync(), msync() or close().
344  *
345  * The tricky part is that after writepage we cannot touch the mapping: nothing
346  * prevents it from being freed up.  But we have a ref on the page and once
347  * that page is locked, the mapping is pinned.
348  *
349  * We're allowed to run sleeping lock_page() here because we know the caller has
350  * __GFP_FS.
351  */
352 static void handle_write_error(struct address_space *mapping,
353                                 struct page *page, int error)
354 {
355         lock_page(page);
356         if (page_mapping(page) == mapping)
357                 mapping_set_error(mapping, error);
358         unlock_page(page);
359 }
360
361 /* possible outcome of pageout() */
362 typedef enum {
363         /* failed to write page out, page is locked */
364         PAGE_KEEP,
365         /* move page to the active list, page is locked */
366         PAGE_ACTIVATE,
367         /* page has been sent to the disk successfully, page is unlocked */
368         PAGE_SUCCESS,
369         /* page is clean and locked */
370         PAGE_CLEAN,
371 } pageout_t;
372
373 /*
374  * pageout is called by shrink_page_list() for each dirty page.
375  * Calls ->writepage().
376  */
377 static pageout_t pageout(struct page *page, struct address_space *mapping,
378                          struct scan_control *sc)
379 {
380         /*
381          * If the page is dirty, only perform writeback if that write
382          * will be non-blocking.  To prevent this allocation from being
383          * stalled by pagecache activity.  But note that there may be
384          * stalls if we need to run get_block().  We could test
385          * PagePrivate for that.
386          *
387          * If this process is currently in __generic_file_aio_write() against
388          * this page's queue, we can perform writeback even if that
389          * will block.
390          *
391          * If the page is swapcache, write it back even if that would
392          * block, for some throttling. This happens by accident, because
393          * swap_backing_dev_info is bust: it doesn't reflect the
394          * congestion state of the swapdevs.  Easy to fix, if needed.
395          */
396         if (!is_page_cache_freeable(page))
397                 return PAGE_KEEP;
398         if (!mapping) {
399                 /*
400                  * Some data journaling orphaned pages can have
401                  * page->mapping == NULL while being dirty with clean buffers.
402                  */
403                 if (page_has_private(page)) {
404                         if (try_to_free_buffers(page)) {
405                                 ClearPageDirty(page);
406                                 printk("%s: orphaned page\n", __func__);
407                                 return PAGE_CLEAN;
408                         }
409                 }
410                 return PAGE_KEEP;
411         }
412         if (mapping->a_ops->writepage == NULL)
413                 return PAGE_ACTIVATE;
414         if (!may_write_to_queue(mapping->backing_dev_info, sc))
415                 return PAGE_KEEP;
416
417         if (clear_page_dirty_for_io(page)) {
418                 int res;
419                 struct writeback_control wbc = {
420                         .sync_mode = WB_SYNC_NONE,
421                         .nr_to_write = SWAP_CLUSTER_MAX,
422                         .range_start = 0,
423                         .range_end = LLONG_MAX,
424                         .for_reclaim = 1,
425                 };
426
427                 SetPageReclaim(page);
428                 res = mapping->a_ops->writepage(page, &wbc);
429                 if (res < 0)
430                         handle_write_error(mapping, page, res);
431                 if (res == AOP_WRITEPAGE_ACTIVATE) {
432                         ClearPageReclaim(page);
433                         return PAGE_ACTIVATE;
434                 }
435
436                 if (!PageWriteback(page)) {
437                         /* synchronous write or broken a_ops? */
438                         ClearPageReclaim(page);
439                 }
440                 trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
441                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
442                 return PAGE_SUCCESS;
443         }
444
445         return PAGE_CLEAN;
446 }
447
448 /*
449  * Same as remove_mapping, but if the page is removed from the mapping, it
450  * gets returned with a refcount of 0.
451  */
452 static int __remove_mapping(struct address_space *mapping, struct page *page)
453 {
454         BUG_ON(!PageLocked(page));
455         BUG_ON(mapping != page_mapping(page));
456
457         spin_lock_irq(&mapping->tree_lock);
458         /*
459          * The non racy check for a busy page.
460          *
461          * Must be careful with the order of the tests. When someone has
462          * a ref to the page, it may be possible that they dirty it then
463          * drop the reference. So if PageDirty is tested before page_count
464          * here, then the following race may occur:
465          *
466          * get_user_pages(&page);
467          * [user mapping goes away]
468          * write_to(page);
469          *                              !PageDirty(page)    [good]
470          * SetPageDirty(page);
471          * put_page(page);
472          *                              !page_count(page)   [good, discard it]
473          *
474          * [oops, our write_to data is lost]
475          *
476          * Reversing the order of the tests ensures such a situation cannot
477          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
478          * load is not satisfied before that of page->_count.
479          *
480          * Note that if SetPageDirty is always performed via set_page_dirty,
481          * and thus under tree_lock, then this ordering is not required.
482          */
483         if (!page_freeze_refs(page, 2))
484                 goto cannot_free;
485         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
486         if (unlikely(PageDirty(page))) {
487                 page_unfreeze_refs(page, 2);
488                 goto cannot_free;
489         }
490
491         if (PageSwapCache(page)) {
492                 swp_entry_t swap = { .val = page_private(page) };
493                 __delete_from_swap_cache(page);
494                 spin_unlock_irq(&mapping->tree_lock);
495                 swapcache_free(swap, page);
496         } else {
497                 void (*freepage)(struct page *);
498
499                 freepage = mapping->a_ops->freepage;
500
501                 __delete_from_page_cache(page);
502                 spin_unlock_irq(&mapping->tree_lock);
503                 mem_cgroup_uncharge_cache_page(page);
504
505                 if (freepage != NULL)
506                         freepage(page);
507         }
508
509         return 1;
510
511 cannot_free:
512         spin_unlock_irq(&mapping->tree_lock);
513         return 0;
514 }
515
516 /*
517  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
518  * someone else has a ref on the page, abort and return 0.  If it was
519  * successfully detached, return 1.  Assumes the caller has a single ref on
520  * this page.
521  */
522 int remove_mapping(struct address_space *mapping, struct page *page)
523 {
524         if (__remove_mapping(mapping, page)) {
525                 /*
526                  * Unfreezing the refcount with 1 rather than 2 effectively
527                  * drops the pagecache ref for us without requiring another
528                  * atomic operation.
529                  */
530                 page_unfreeze_refs(page, 1);
531                 return 1;
532         }
533         return 0;
534 }
535
536 /**
537  * putback_lru_page - put previously isolated page onto appropriate LRU list
538  * @page: page to be put back to appropriate lru list
539  *
540  * Add previously isolated @page to appropriate LRU list.
541  * Page may still be unevictable for other reasons.
542  *
543  * lru_lock must not be held, interrupts must be enabled.
544  */
545 void putback_lru_page(struct page *page)
546 {
547         int lru;
548         int active = !!TestClearPageActive(page);
549         int was_unevictable = PageUnevictable(page);
550
551         VM_BUG_ON(PageLRU(page));
552
553 redo:
554         ClearPageUnevictable(page);
555
556         if (page_evictable(page, NULL)) {
557                 /*
558                  * For evictable pages, we can use the cache.
559                  * In event of a race, worst case is we end up with an
560                  * unevictable page on [in]active list.
561                  * We know how to handle that.
562                  */
563                 lru = active + page_lru_base_type(page);
564                 lru_cache_add_lru(page, lru);
565         } else {
566                 /*
567                  * Put unevictable pages directly on zone's unevictable
568                  * list.
569                  */
570                 lru = LRU_UNEVICTABLE;
571                 add_page_to_unevictable_list(page);
572                 /*
573                  * When racing with an mlock or AS_UNEVICTABLE clearing
574                  * (page is unlocked) make sure that if the other thread
575                  * does not observe our setting of PG_lru and fails
576                  * isolation/check_move_unevictable_pages,
577                  * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
578                  * the page back to the evictable list.
579                  *
580                  * The other side is TestClearPageMlocked() or shmem_lock().
581                  */
582                 smp_mb();
583         }
584
585         /*
586          * page's status can change while we move it among lru. If an evictable
587          * page is on unevictable list, it never be freed. To avoid that,
588          * check after we added it to the list, again.
589          */
590         if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
591                 if (!isolate_lru_page(page)) {
592                         put_page(page);
593                         goto redo;
594                 }
595                 /* This means someone else dropped this page from LRU
596                  * So, it will be freed or putback to LRU again. There is
597                  * nothing to do here.
598                  */
599         }
600
601         if (was_unevictable && lru != LRU_UNEVICTABLE)
602                 count_vm_event(UNEVICTABLE_PGRESCUED);
603         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
604                 count_vm_event(UNEVICTABLE_PGCULLED);
605
606         put_page(page);         /* drop ref from isolate */
607 }
608
609 enum page_references {
610         PAGEREF_RECLAIM,
611         PAGEREF_RECLAIM_CLEAN,
612         PAGEREF_KEEP,
613         PAGEREF_ACTIVATE,
614 };
615
616 static enum page_references page_check_references(struct page *page,
617                                                   struct scan_control *sc)
618 {
619         int referenced_ptes, referenced_page;
620         unsigned long vm_flags;
621
622         referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
623                                           &vm_flags);
624         referenced_page = TestClearPageReferenced(page);
625
626         /*
627          * Mlock lost the isolation race with us.  Let try_to_unmap()
628          * move the page to the unevictable list.
629          */
630         if (vm_flags & VM_LOCKED)
631                 return PAGEREF_RECLAIM;
632
633         if (referenced_ptes) {
634                 if (PageSwapBacked(page))
635                         return PAGEREF_ACTIVATE;
636                 /*
637                  * All mapped pages start out with page table
638                  * references from the instantiating fault, so we need
639                  * to look twice if a mapped file page is used more
640                  * than once.
641                  *
642                  * Mark it and spare it for another trip around the
643                  * inactive list.  Another page table reference will
644                  * lead to its activation.
645                  *
646                  * Note: the mark is set for activated pages as well
647                  * so that recently deactivated but used pages are
648                  * quickly recovered.
649                  */
650                 SetPageReferenced(page);
651
652                 if (referenced_page || referenced_ptes > 1)
653                         return PAGEREF_ACTIVATE;
654
655                 /*
656                  * Activate file-backed executable pages after first usage.
657                  */
658                 if (vm_flags & VM_EXEC)
659                         return PAGEREF_ACTIVATE;
660
661                 return PAGEREF_KEEP;
662         }
663
664         /* Reclaim if clean, defer dirty pages to writeback */
665         if (referenced_page && !PageSwapBacked(page))
666                 return PAGEREF_RECLAIM_CLEAN;
667
668         return PAGEREF_RECLAIM;
669 }
670
671 /*
672  * shrink_page_list() returns the number of reclaimed pages
673  */
674 static unsigned long shrink_page_list(struct list_head *page_list,
675                                       struct zone *zone,
676                                       struct scan_control *sc,
677                                       unsigned long *ret_nr_dirty,
678                                       unsigned long *ret_nr_writeback)
679 {
680         LIST_HEAD(ret_pages);
681         LIST_HEAD(free_pages);
682         int pgactivate = 0;
683         unsigned long nr_dirty = 0;
684         unsigned long nr_congested = 0;
685         unsigned long nr_reclaimed = 0;
686         unsigned long nr_writeback = 0;
687
688         cond_resched();
689
690         while (!list_empty(page_list)) {
691                 enum page_references references;
692                 struct address_space *mapping;
693                 struct page *page;
694                 int may_enter_fs;
695
696                 cond_resched();
697
698                 page = lru_to_page(page_list);
699                 list_del(&page->lru);
700
701                 if (!trylock_page(page))
702                         goto keep;
703
704                 VM_BUG_ON(PageActive(page));
705                 VM_BUG_ON(page_zone(page) != zone);
706
707                 sc->nr_scanned++;
708
709                 if (unlikely(!page_evictable(page, NULL)))
710                         goto cull_mlocked;
711
712                 if (!sc->may_unmap && page_mapped(page))
713                         goto keep_locked;
714
715                 /* Double the slab pressure for mapped and swapcache pages */
716                 if (page_mapped(page) || PageSwapCache(page))
717                         sc->nr_scanned++;
718
719                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
720                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
721
722                 if (PageWriteback(page)) {
723                         nr_writeback++;
724                         unlock_page(page);
725                         goto keep;
726                 }
727
728                 references = page_check_references(page, sc);
729                 switch (references) {
730                 case PAGEREF_ACTIVATE:
731                         goto activate_locked;
732                 case PAGEREF_KEEP:
733                         goto keep_locked;
734                 case PAGEREF_RECLAIM:
735                 case PAGEREF_RECLAIM_CLEAN:
736                         ; /* try to reclaim the page below */
737                 }
738
739                 /*
740                  * Anonymous process memory has backing store?
741                  * Try to allocate it some swap space here.
742                  */
743                 if (PageAnon(page) && !PageSwapCache(page)) {
744                         if (!(sc->gfp_mask & __GFP_IO))
745                                 goto keep_locked;
746                         if (!add_to_swap(page))
747                                 goto activate_locked;
748                         may_enter_fs = 1;
749                 }
750
751                 mapping = page_mapping(page);
752
753                 /*
754                  * The page is mapped into the page tables of one or more
755                  * processes. Try to unmap it here.
756                  */
757                 if (page_mapped(page) && mapping) {
758                         switch (try_to_unmap(page, TTU_UNMAP)) {
759                         case SWAP_FAIL:
760                                 goto activate_locked;
761                         case SWAP_AGAIN:
762                                 goto keep_locked;
763                         case SWAP_MLOCK:
764                                 goto cull_mlocked;
765                         case SWAP_SUCCESS:
766                                 ; /* try to free the page below */
767                         }
768                 }
769
770                 if (PageDirty(page)) {
771                         nr_dirty++;
772
773                         /*
774                          * Only kswapd can writeback filesystem pages to
775                          * avoid risk of stack overflow but do not writeback
776                          * unless under significant pressure.
777                          */
778                         if (page_is_file_cache(page) &&
779                                         (!current_is_kswapd() ||
780                                          sc->priority >= DEF_PRIORITY - 2)) {
781                                 /*
782                                  * Immediately reclaim when written back.
783                                  * Similar in principal to deactivate_page()
784                                  * except we already have the page isolated
785                                  * and know it's dirty
786                                  */
787                                 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
788                                 SetPageReclaim(page);
789
790                                 goto keep_locked;
791                         }
792
793                         if (references == PAGEREF_RECLAIM_CLEAN)
794                                 goto keep_locked;
795                         if (!may_enter_fs)
796                                 goto keep_locked;
797                         if (!sc->may_writepage)
798                                 goto keep_locked;
799
800                         /* Page is dirty, try to write it out here */
801                         switch (pageout(page, mapping, sc)) {
802                         case PAGE_KEEP:
803                                 nr_congested++;
804                                 goto keep_locked;
805                         case PAGE_ACTIVATE:
806                                 goto activate_locked;
807                         case PAGE_SUCCESS:
808                                 if (PageWriteback(page))
809                                         goto keep;
810                                 if (PageDirty(page))
811                                         goto keep;
812
813                                 /*
814                                  * A synchronous write - probably a ramdisk.  Go
815                                  * ahead and try to reclaim the page.
816                                  */
817                                 if (!trylock_page(page))
818                                         goto keep;
819                                 if (PageDirty(page) || PageWriteback(page))
820                                         goto keep_locked;
821                                 mapping = page_mapping(page);
822                         case PAGE_CLEAN:
823                                 ; /* try to free the page below */
824                         }
825                 }
826
827                 /*
828                  * If the page has buffers, try to free the buffer mappings
829                  * associated with this page. If we succeed we try to free
830                  * the page as well.
831                  *
832                  * We do this even if the page is PageDirty().
833                  * try_to_release_page() does not perform I/O, but it is
834                  * possible for a page to have PageDirty set, but it is actually
835                  * clean (all its buffers are clean).  This happens if the
836                  * buffers were written out directly, with submit_bh(). ext3
837                  * will do this, as well as the blockdev mapping.
838                  * try_to_release_page() will discover that cleanness and will
839                  * drop the buffers and mark the page clean - it can be freed.
840                  *
841                  * Rarely, pages can have buffers and no ->mapping.  These are
842                  * the pages which were not successfully invalidated in
843                  * truncate_complete_page().  We try to drop those buffers here
844                  * and if that worked, and the page is no longer mapped into
845                  * process address space (page_count == 1) it can be freed.
846                  * Otherwise, leave the page on the LRU so it is swappable.
847                  */
848                 if (page_has_private(page)) {
849                         if (!try_to_release_page(page, sc->gfp_mask))
850                                 goto activate_locked;
851                         if (!mapping && page_count(page) == 1) {
852                                 unlock_page(page);
853                                 if (put_page_testzero(page))
854                                         goto free_it;
855                                 else {
856                                         /*
857                                          * rare race with speculative reference.
858                                          * the speculative reference will free
859                                          * this page shortly, so we may
860                                          * increment nr_reclaimed here (and
861                                          * leave it off the LRU).
862                                          */
863                                         nr_reclaimed++;
864                                         continue;
865                                 }
866                         }
867                 }
868
869                 if (!mapping || !__remove_mapping(mapping, page))
870                         goto keep_locked;
871
872                 /*
873                  * At this point, we have no other references and there is
874                  * no way to pick any more up (removed from LRU, removed
875                  * from pagecache). Can use non-atomic bitops now (and
876                  * we obviously don't have to worry about waking up a process
877                  * waiting on the page lock, because there are no references.
878                  */
879                 __clear_page_locked(page);
880 free_it:
881                 nr_reclaimed++;
882
883                 /*
884                  * Is there need to periodically free_page_list? It would
885                  * appear not as the counts should be low
886                  */
887                 list_add(&page->lru, &free_pages);
888                 continue;
889
890 cull_mlocked:
891                 if (PageSwapCache(page))
892                         try_to_free_swap(page);
893                 unlock_page(page);
894                 putback_lru_page(page);
895                 continue;
896
897 activate_locked:
898                 /* Not a candidate for swapping, so reclaim swap space. */
899                 if (PageSwapCache(page) && vm_swap_full())
900                         try_to_free_swap(page);
901                 VM_BUG_ON(PageActive(page));
902                 SetPageActive(page);
903                 pgactivate++;
904 keep_locked:
905                 unlock_page(page);
906 keep:
907                 list_add(&page->lru, &ret_pages);
908                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
909         }
910
911         /*
912          * Tag a zone as congested if all the dirty pages encountered were
913          * backed by a congested BDI. In this case, reclaimers should just
914          * back off and wait for congestion to clear because further reclaim
915          * will encounter the same problem
916          */
917         if (nr_dirty && nr_dirty == nr_congested && global_reclaim(sc))
918                 zone_set_flag(zone, ZONE_CONGESTED);
919
920         free_hot_cold_page_list(&free_pages, 1);
921
922         list_splice(&ret_pages, page_list);
923         count_vm_events(PGACTIVATE, pgactivate);
924         *ret_nr_dirty += nr_dirty;
925         *ret_nr_writeback += nr_writeback;
926         return nr_reclaimed;
927 }
928
929 /*
930  * Attempt to remove the specified page from its LRU.  Only take this page
931  * if it is of the appropriate PageActive status.  Pages which are being
932  * freed elsewhere are also ignored.
933  *
934  * page:        page to consider
935  * mode:        one of the LRU isolation modes defined above
936  *
937  * returns 0 on success, -ve errno on failure.
938  */
939 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
940 {
941         int ret = -EINVAL;
942
943         /* Only take pages on the LRU. */
944         if (!PageLRU(page))
945                 return ret;
946
947         /* Do not give back unevictable pages for compaction */
948         if (PageUnevictable(page))
949                 return ret;
950
951         ret = -EBUSY;
952
953         /*
954          * To minimise LRU disruption, the caller can indicate that it only
955          * wants to isolate pages it will be able to operate on without
956          * blocking - clean pages for the most part.
957          *
958          * ISOLATE_CLEAN means that only clean pages should be isolated. This
959          * is used by reclaim when it is cannot write to backing storage
960          *
961          * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
962          * that it is possible to migrate without blocking
963          */
964         if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
965                 /* All the caller can do on PageWriteback is block */
966                 if (PageWriteback(page))
967                         return ret;
968
969                 if (PageDirty(page)) {
970                         struct address_space *mapping;
971
972                         /* ISOLATE_CLEAN means only clean pages */
973                         if (mode & ISOLATE_CLEAN)
974                                 return ret;
975
976                         /*
977                          * Only pages without mappings or that have a
978                          * ->migratepage callback are possible to migrate
979                          * without blocking
980                          */
981                         mapping = page_mapping(page);
982                         if (mapping && !mapping->a_ops->migratepage)
983                                 return ret;
984                 }
985         }
986
987         if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
988                 return ret;
989
990         if (likely(get_page_unless_zero(page))) {
991                 /*
992                  * Be careful not to clear PageLRU until after we're
993                  * sure the page is not being freed elsewhere -- the
994                  * page release code relies on it.
995                  */
996                 ClearPageLRU(page);
997                 ret = 0;
998         }
999
1000         return ret;
1001 }
1002
1003 /*
1004  * zone->lru_lock is heavily contended.  Some of the functions that
1005  * shrink the lists perform better by taking out a batch of pages
1006  * and working on them outside the LRU lock.
1007  *
1008  * For pagecache intensive workloads, this function is the hottest
1009  * spot in the kernel (apart from copy_*_user functions).
1010  *
1011  * Appropriate locks must be held before calling this function.
1012  *
1013  * @nr_to_scan: The number of pages to look through on the list.
1014  * @lruvec:     The LRU vector to pull pages from.
1015  * @dst:        The temp list to put pages on to.
1016  * @nr_scanned: The number of pages that were scanned.
1017  * @sc:         The scan_control struct for this reclaim session
1018  * @mode:       One of the LRU isolation modes
1019  * @lru:        LRU list id for isolating
1020  *
1021  * returns how many pages were moved onto *@dst.
1022  */
1023 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1024                 struct lruvec *lruvec, struct list_head *dst,
1025                 unsigned long *nr_scanned, struct scan_control *sc,
1026                 isolate_mode_t mode, enum lru_list lru)
1027 {
1028         struct list_head *src = &lruvec->lists[lru];
1029         unsigned long nr_taken = 0;
1030         unsigned long scan;
1031
1032         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1033                 struct page *page;
1034                 int nr_pages;
1035
1036                 page = lru_to_page(src);
1037                 prefetchw_prev_lru_page(page, src, flags);
1038
1039                 VM_BUG_ON(!PageLRU(page));
1040
1041                 switch (__isolate_lru_page(page, mode)) {
1042                 case 0:
1043                         nr_pages = hpage_nr_pages(page);
1044                         mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
1045                         list_move(&page->lru, dst);
1046                         nr_taken += nr_pages;
1047                         break;
1048
1049                 case -EBUSY:
1050                         /* else it is being freed elsewhere */
1051                         list_move(&page->lru, src);
1052                         continue;
1053
1054                 default:
1055                         BUG();
1056                 }
1057         }
1058
1059         *nr_scanned = scan;
1060         trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1061                                     nr_taken, mode, is_file_lru(lru));
1062         return nr_taken;
1063 }
1064
1065 /**
1066  * isolate_lru_page - tries to isolate a page from its LRU list
1067  * @page: page to isolate from its LRU list
1068  *
1069  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1070  * vmstat statistic corresponding to whatever LRU list the page was on.
1071  *
1072  * Returns 0 if the page was removed from an LRU list.
1073  * Returns -EBUSY if the page was not on an LRU list.
1074  *
1075  * The returned page will have PageLRU() cleared.  If it was found on
1076  * the active list, it will have PageActive set.  If it was found on
1077  * the unevictable list, it will have the PageUnevictable bit set. That flag
1078  * may need to be cleared by the caller before letting the page go.
1079  *
1080  * The vmstat statistic corresponding to the list on which the page was
1081  * found will be decremented.
1082  *
1083  * Restrictions:
1084  * (1) Must be called with an elevated refcount on the page. This is a
1085  *     fundamentnal difference from isolate_lru_pages (which is called
1086  *     without a stable reference).
1087  * (2) the lru_lock must not be held.
1088  * (3) interrupts must be enabled.
1089  */
1090 int isolate_lru_page(struct page *page)
1091 {
1092         int ret = -EBUSY;
1093
1094         VM_BUG_ON(!page_count(page));
1095
1096         if (PageLRU(page)) {
1097                 struct zone *zone = page_zone(page);
1098                 struct lruvec *lruvec;
1099
1100                 spin_lock_irq(&zone->lru_lock);
1101                 lruvec = mem_cgroup_page_lruvec(page, zone);
1102                 if (PageLRU(page)) {
1103                         int lru = page_lru(page);
1104                         get_page(page);
1105                         ClearPageLRU(page);
1106                         del_page_from_lru_list(page, lruvec, lru);
1107                         ret = 0;
1108                 }
1109                 spin_unlock_irq(&zone->lru_lock);
1110         }
1111         return ret;
1112 }
1113
1114 /*
1115  * Are there way too many processes in the direct reclaim path already?
1116  */
1117 static int too_many_isolated(struct zone *zone, int file,
1118                 struct scan_control *sc)
1119 {
1120         unsigned long inactive, isolated;
1121
1122         if (current_is_kswapd())
1123                 return 0;
1124
1125         if (!global_reclaim(sc))
1126                 return 0;
1127
1128         if (file) {
1129                 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1130                 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1131         } else {
1132                 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1133                 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1134         }
1135
1136         return isolated > inactive;
1137 }
1138
1139 static noinline_for_stack void
1140 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1141 {
1142         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1143         struct zone *zone = lruvec_zone(lruvec);
1144         LIST_HEAD(pages_to_free);
1145
1146         /*
1147          * Put back any unfreeable pages.
1148          */
1149         while (!list_empty(page_list)) {
1150                 struct page *page = lru_to_page(page_list);
1151                 int lru;
1152
1153                 VM_BUG_ON(PageLRU(page));
1154                 list_del(&page->lru);
1155                 if (unlikely(!page_evictable(page, NULL))) {
1156                         spin_unlock_irq(&zone->lru_lock);
1157                         putback_lru_page(page);
1158                         spin_lock_irq(&zone->lru_lock);
1159                         continue;
1160                 }
1161
1162                 lruvec = mem_cgroup_page_lruvec(page, zone);
1163
1164                 SetPageLRU(page);
1165                 lru = page_lru(page);
1166                 add_page_to_lru_list(page, lruvec, lru);
1167
1168                 if (is_active_lru(lru)) {
1169                         int file = is_file_lru(lru);
1170                         int numpages = hpage_nr_pages(page);
1171                         reclaim_stat->recent_rotated[file] += numpages;
1172                 }
1173                 if (put_page_testzero(page)) {
1174                         __ClearPageLRU(page);
1175                         __ClearPageActive(page);
1176                         del_page_from_lru_list(page, lruvec, lru);
1177
1178                         if (unlikely(PageCompound(page))) {
1179                                 spin_unlock_irq(&zone->lru_lock);
1180                                 (*get_compound_page_dtor(page))(page);
1181                                 spin_lock_irq(&zone->lru_lock);
1182                         } else
1183                                 list_add(&page->lru, &pages_to_free);
1184                 }
1185         }
1186
1187         /*
1188          * To save our caller's stack, now use input list for pages to free.
1189          */
1190         list_splice(&pages_to_free, page_list);
1191 }
1192
1193 /*
1194  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1195  * of reclaimed pages
1196  */
1197 static noinline_for_stack unsigned long
1198 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1199                      struct scan_control *sc, enum lru_list lru)
1200 {
1201         LIST_HEAD(page_list);
1202         unsigned long nr_scanned;
1203         unsigned long nr_reclaimed = 0;
1204         unsigned long nr_taken;
1205         unsigned long nr_dirty = 0;
1206         unsigned long nr_writeback = 0;
1207         isolate_mode_t isolate_mode = 0;
1208         int file = is_file_lru(lru);
1209         struct zone *zone = lruvec_zone(lruvec);
1210         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1211
1212         while (unlikely(too_many_isolated(zone, file, sc))) {
1213                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1214
1215                 /* We are about to die and free our memory. Return now. */
1216                 if (fatal_signal_pending(current))
1217                         return SWAP_CLUSTER_MAX;
1218         }
1219
1220         lru_add_drain();
1221
1222         if (!sc->may_unmap)
1223                 isolate_mode |= ISOLATE_UNMAPPED;
1224         if (!sc->may_writepage)
1225                 isolate_mode |= ISOLATE_CLEAN;
1226
1227         spin_lock_irq(&zone->lru_lock);
1228
1229         nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1230                                      &nr_scanned, sc, isolate_mode, lru);
1231
1232         __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1233         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1234
1235         if (global_reclaim(sc)) {
1236                 zone->pages_scanned += nr_scanned;
1237                 if (current_is_kswapd())
1238                         __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
1239                 else
1240                         __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
1241         }
1242         spin_unlock_irq(&zone->lru_lock);
1243
1244         if (nr_taken == 0)
1245                 return 0;
1246
1247         nr_reclaimed = shrink_page_list(&page_list, zone, sc,
1248                                                 &nr_dirty, &nr_writeback);
1249
1250         spin_lock_irq(&zone->lru_lock);
1251
1252         reclaim_stat->recent_scanned[file] += nr_taken;
1253
1254         if (global_reclaim(sc)) {
1255                 if (current_is_kswapd())
1256                         __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1257                                                nr_reclaimed);
1258                 else
1259                         __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1260                                                nr_reclaimed);
1261         }
1262
1263         putback_inactive_pages(lruvec, &page_list);
1264
1265         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1266
1267         spin_unlock_irq(&zone->lru_lock);
1268
1269         free_hot_cold_page_list(&page_list, 1);
1270
1271         /*
1272          * If reclaim is isolating dirty pages under writeback, it implies
1273          * that the long-lived page allocation rate is exceeding the page
1274          * laundering rate. Either the global limits are not being effective
1275          * at throttling processes due to the page distribution throughout
1276          * zones or there is heavy usage of a slow backing device. The
1277          * only option is to throttle from reclaim context which is not ideal
1278          * as there is no guarantee the dirtying process is throttled in the
1279          * same way balance_dirty_pages() manages.
1280          *
1281          * This scales the number of dirty pages that must be under writeback
1282          * before throttling depending on priority. It is a simple backoff
1283          * function that has the most effect in the range DEF_PRIORITY to
1284          * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1285          * in trouble and reclaim is considered to be in trouble.
1286          *
1287          * DEF_PRIORITY   100% isolated pages must be PageWriteback to throttle
1288          * DEF_PRIORITY-1  50% must be PageWriteback
1289          * DEF_PRIORITY-2  25% must be PageWriteback, kswapd in trouble
1290          * ...
1291          * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1292          *                     isolated page is PageWriteback
1293          */
1294         if (nr_writeback && nr_writeback >=
1295                         (nr_taken >> (DEF_PRIORITY - sc->priority)))
1296                 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1297
1298         trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1299                 zone_idx(zone),
1300                 nr_scanned, nr_reclaimed,
1301                 sc->priority,
1302                 trace_shrink_flags(file));
1303         return nr_reclaimed;
1304 }
1305
1306 /*
1307  * This moves pages from the active list to the inactive list.
1308  *
1309  * We move them the other way if the page is referenced by one or more
1310  * processes, from rmap.
1311  *
1312  * If the pages are mostly unmapped, the processing is fast and it is
1313  * appropriate to hold zone->lru_lock across the whole operation.  But if
1314  * the pages are mapped, the processing is slow (page_referenced()) so we
1315  * should drop zone->lru_lock around each page.  It's impossible to balance
1316  * this, so instead we remove the pages from the LRU while processing them.
1317  * It is safe to rely on PG_active against the non-LRU pages in here because
1318  * nobody will play with that bit on a non-LRU page.
1319  *
1320  * The downside is that we have to touch page->_count against each page.
1321  * But we had to alter page->flags anyway.
1322  */
1323
1324 static void move_active_pages_to_lru(struct lruvec *lruvec,
1325                                      struct list_head *list,
1326                                      struct list_head *pages_to_free,
1327                                      enum lru_list lru)
1328 {
1329         struct zone *zone = lruvec_zone(lruvec);
1330         unsigned long pgmoved = 0;
1331         struct page *page;
1332         int nr_pages;
1333
1334         while (!list_empty(list)) {
1335                 page = lru_to_page(list);
1336                 lruvec = mem_cgroup_page_lruvec(page, zone);
1337
1338                 VM_BUG_ON(PageLRU(page));
1339                 SetPageLRU(page);
1340
1341                 nr_pages = hpage_nr_pages(page);
1342                 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
1343                 list_move(&page->lru, &lruvec->lists[lru]);
1344                 pgmoved += nr_pages;
1345
1346                 if (put_page_testzero(page)) {
1347                         __ClearPageLRU(page);
1348                         __ClearPageActive(page);
1349                         del_page_from_lru_list(page, lruvec, lru);
1350
1351                         if (unlikely(PageCompound(page))) {
1352                                 spin_unlock_irq(&zone->lru_lock);
1353                                 (*get_compound_page_dtor(page))(page);
1354                                 spin_lock_irq(&zone->lru_lock);
1355                         } else
1356                                 list_add(&page->lru, pages_to_free);
1357                 }
1358         }
1359         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1360         if (!is_active_lru(lru))
1361                 __count_vm_events(PGDEACTIVATE, pgmoved);
1362 }
1363
1364 static void shrink_active_list(unsigned long nr_to_scan,
1365                                struct lruvec *lruvec,
1366                                struct scan_control *sc,
1367                                enum lru_list lru)
1368 {
1369         unsigned long nr_taken;
1370         unsigned long nr_scanned;
1371         unsigned long vm_flags;
1372         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1373         LIST_HEAD(l_active);
1374         LIST_HEAD(l_inactive);
1375         struct page *page;
1376         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1377         unsigned long nr_rotated = 0;
1378         isolate_mode_t isolate_mode = 0;
1379         int file = is_file_lru(lru);
1380         struct zone *zone = lruvec_zone(lruvec);
1381
1382         lru_add_drain();
1383
1384         if (!sc->may_unmap)
1385                 isolate_mode |= ISOLATE_UNMAPPED;
1386         if (!sc->may_writepage)
1387                 isolate_mode |= ISOLATE_CLEAN;
1388
1389         spin_lock_irq(&zone->lru_lock);
1390
1391         nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1392                                      &nr_scanned, sc, isolate_mode, lru);
1393         if (global_reclaim(sc))
1394                 zone->pages_scanned += nr_scanned;
1395
1396         reclaim_stat->recent_scanned[file] += nr_taken;
1397
1398         __count_zone_vm_events(PGREFILL, zone, nr_scanned);
1399         __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1400         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1401         spin_unlock_irq(&zone->lru_lock);
1402
1403         while (!list_empty(&l_hold)) {
1404                 cond_resched();
1405                 page = lru_to_page(&l_hold);
1406                 list_del(&page->lru);
1407
1408                 if (unlikely(!page_evictable(page, NULL))) {
1409                         putback_lru_page(page);
1410                         continue;
1411                 }
1412
1413                 if (unlikely(buffer_heads_over_limit)) {
1414                         if (page_has_private(page) && trylock_page(page)) {
1415                                 if (page_has_private(page))
1416                                         try_to_release_page(page, 0);
1417                                 unlock_page(page);
1418                         }
1419                 }
1420
1421                 if (page_referenced(page, 0, sc->target_mem_cgroup,
1422                                     &vm_flags)) {
1423                         nr_rotated += hpage_nr_pages(page);
1424                         /*
1425                          * Identify referenced, file-backed active pages and
1426                          * give them one more trip around the active list. So
1427                          * that executable code get better chances to stay in
1428                          * memory under moderate memory pressure.  Anon pages
1429                          * are not likely to be evicted by use-once streaming
1430                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1431                          * so we ignore them here.
1432                          */
1433                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1434                                 list_add(&page->lru, &l_active);
1435                                 continue;
1436                         }
1437                 }
1438
1439                 ClearPageActive(page);  /* we are de-activating */
1440                 list_add(&page->lru, &l_inactive);
1441         }
1442
1443         /*
1444          * Move pages back to the lru list.
1445          */
1446         spin_lock_irq(&zone->lru_lock);
1447         /*
1448          * Count referenced pages from currently used mappings as rotated,
1449          * even though only some of them are actually re-activated.  This
1450          * helps balance scan pressure between file and anonymous pages in
1451          * get_scan_ratio.
1452          */
1453         reclaim_stat->recent_rotated[file] += nr_rotated;
1454
1455         move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1456         move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
1457         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1458         spin_unlock_irq(&zone->lru_lock);
1459
1460         free_hot_cold_page_list(&l_hold, 1);
1461 }
1462
1463 #ifdef CONFIG_SWAP
1464 static int inactive_anon_is_low_global(struct zone *zone)
1465 {
1466         unsigned long active, inactive;
1467
1468         active = zone_page_state(zone, NR_ACTIVE_ANON);
1469         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1470
1471         if (inactive * zone->inactive_ratio < active)
1472                 return 1;
1473
1474         return 0;
1475 }
1476
1477 /**
1478  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1479  * @lruvec: LRU vector to check
1480  *
1481  * Returns true if the zone does not have enough inactive anon pages,
1482  * meaning some active anon pages need to be deactivated.
1483  */
1484 static int inactive_anon_is_low(struct lruvec *lruvec)
1485 {
1486         /*
1487          * If we don't have swap space, anonymous page deactivation
1488          * is pointless.
1489          */
1490         if (!total_swap_pages)
1491                 return 0;
1492
1493         if (!mem_cgroup_disabled())
1494                 return mem_cgroup_inactive_anon_is_low(lruvec);
1495
1496         return inactive_anon_is_low_global(lruvec_zone(lruvec));
1497 }
1498 #else
1499 static inline int inactive_anon_is_low(struct lruvec *lruvec)
1500 {
1501         return 0;
1502 }
1503 #endif
1504
1505 static int inactive_file_is_low_global(struct zone *zone)
1506 {
1507         unsigned long active, inactive;
1508
1509         active = zone_page_state(zone, NR_ACTIVE_FILE);
1510         inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1511
1512         return (active > inactive);
1513 }
1514
1515 /**
1516  * inactive_file_is_low - check if file pages need to be deactivated
1517  * @lruvec: LRU vector to check
1518  *
1519  * When the system is doing streaming IO, memory pressure here
1520  * ensures that active file pages get deactivated, until more
1521  * than half of the file pages are on the inactive list.
1522  *
1523  * Once we get to that situation, protect the system's working
1524  * set from being evicted by disabling active file page aging.
1525  *
1526  * This uses a different ratio than the anonymous pages, because
1527  * the page cache uses a use-once replacement algorithm.
1528  */
1529 static int inactive_file_is_low(struct lruvec *lruvec)
1530 {
1531         if (!mem_cgroup_disabled())
1532                 return mem_cgroup_inactive_file_is_low(lruvec);
1533
1534         return inactive_file_is_low_global(lruvec_zone(lruvec));
1535 }
1536
1537 static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
1538 {
1539         if (is_file_lru(lru))
1540                 return inactive_file_is_low(lruvec);
1541         else
1542                 return inactive_anon_is_low(lruvec);
1543 }
1544
1545 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1546                                  struct lruvec *lruvec, struct scan_control *sc)
1547 {
1548         if (is_active_lru(lru)) {
1549                 if (inactive_list_is_low(lruvec, lru))
1550                         shrink_active_list(nr_to_scan, lruvec, sc, lru);
1551                 return 0;
1552         }
1553
1554         return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
1555 }
1556
1557 static int vmscan_swappiness(struct scan_control *sc)
1558 {
1559         if (global_reclaim(sc))
1560                 return vm_swappiness;
1561         return mem_cgroup_swappiness(sc->target_mem_cgroup);
1562 }
1563
1564 /*
1565  * Determine how aggressively the anon and file LRU lists should be
1566  * scanned.  The relative value of each set of LRU lists is determined
1567  * by looking at the fraction of the pages scanned we did rotate back
1568  * onto the active list instead of evict.
1569  *
1570  * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1571  * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
1572  */
1573 static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
1574                            unsigned long *nr)
1575 {
1576         unsigned long anon, file, free;
1577         unsigned long anon_prio, file_prio;
1578         unsigned long ap, fp;
1579         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1580         u64 fraction[2], denominator;
1581         enum lru_list lru;
1582         int noswap = 0;
1583         bool force_scan = false;
1584         struct zone *zone = lruvec_zone(lruvec);
1585
1586         /*
1587          * If the zone or memcg is small, nr[l] can be 0.  This
1588          * results in no scanning on this priority and a potential
1589          * priority drop.  Global direct reclaim can go to the next
1590          * zone and tends to have no problems. Global kswapd is for
1591          * zone balancing and it needs to scan a minimum amount. When
1592          * reclaiming for a memcg, a priority drop can cause high
1593          * latencies, so it's better to scan a minimum amount there as
1594          * well.
1595          */
1596         if (current_is_kswapd() && zone->all_unreclaimable)
1597                 force_scan = true;
1598         if (!global_reclaim(sc))
1599                 force_scan = true;
1600
1601         /* If we have no swap space, do not bother scanning anon pages. */
1602         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1603                 noswap = 1;
1604                 fraction[0] = 0;
1605                 fraction[1] = 1;
1606                 denominator = 1;
1607                 goto out;
1608         }
1609
1610         anon  = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
1611                 get_lru_size(lruvec, LRU_INACTIVE_ANON);
1612         file  = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
1613                 get_lru_size(lruvec, LRU_INACTIVE_FILE);
1614
1615         if (global_reclaim(sc)) {
1616                 free  = zone_page_state(zone, NR_FREE_PAGES);
1617                 /* If we have very few page cache pages,
1618                    force-scan anon pages. */
1619                 if (unlikely(file + free <= high_wmark_pages(zone))) {
1620                         fraction[0] = 1;
1621                         fraction[1] = 0;
1622                         denominator = 1;
1623                         goto out;
1624                 }
1625         }
1626
1627         /*
1628          * With swappiness at 100, anonymous and file have the same priority.
1629          * This scanning priority is essentially the inverse of IO cost.
1630          */
1631         anon_prio = vmscan_swappiness(sc);
1632         file_prio = 200 - anon_prio;
1633
1634         /*
1635          * OK, so we have swap space and a fair amount of page cache
1636          * pages.  We use the recently rotated / recently scanned
1637          * ratios to determine how valuable each cache is.
1638          *
1639          * Because workloads change over time (and to avoid overflow)
1640          * we keep these statistics as a floating average, which ends
1641          * up weighing recent references more than old ones.
1642          *
1643          * anon in [0], file in [1]
1644          */
1645         spin_lock_irq(&zone->lru_lock);
1646         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1647                 reclaim_stat->recent_scanned[0] /= 2;
1648                 reclaim_stat->recent_rotated[0] /= 2;
1649         }
1650
1651         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1652                 reclaim_stat->recent_scanned[1] /= 2;
1653                 reclaim_stat->recent_rotated[1] /= 2;
1654         }
1655
1656         /*
1657          * The amount of pressure on anon vs file pages is inversely
1658          * proportional to the fraction of recently scanned pages on
1659          * each list that were recently referenced and in active use.
1660          */
1661         ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
1662         ap /= reclaim_stat->recent_rotated[0] + 1;
1663
1664         fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
1665         fp /= reclaim_stat->recent_rotated[1] + 1;
1666         spin_unlock_irq(&zone->lru_lock);
1667
1668         fraction[0] = ap;
1669         fraction[1] = fp;
1670         denominator = ap + fp + 1;
1671 out:
1672         for_each_evictable_lru(lru) {
1673                 int file = is_file_lru(lru);
1674                 unsigned long scan;
1675
1676                 scan = get_lru_size(lruvec, lru);
1677                 if (sc->priority || noswap || !vmscan_swappiness(sc)) {
1678                         scan >>= sc->priority;
1679                         if (!scan && force_scan)
1680                                 scan = SWAP_CLUSTER_MAX;
1681                         scan = div64_u64(scan * fraction[file], denominator);
1682                 }
1683                 nr[lru] = scan;
1684         }
1685 }
1686
1687 /* Use reclaim/compaction for costly allocs or under memory pressure */
1688 static bool in_reclaim_compaction(struct scan_control *sc)
1689 {
1690         if (COMPACTION_BUILD && sc->order &&
1691                         (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
1692                          sc->priority < DEF_PRIORITY - 2))
1693                 return true;
1694
1695         return false;
1696 }
1697
1698 /*
1699  * Reclaim/compaction is used for high-order allocation requests. It reclaims
1700  * order-0 pages before compacting the zone. should_continue_reclaim() returns
1701  * true if more pages should be reclaimed such that when the page allocator
1702  * calls try_to_compact_zone() that it will have enough free pages to succeed.
1703  * It will give up earlier than that if there is difficulty reclaiming pages.
1704  */
1705 static inline bool should_continue_reclaim(struct lruvec *lruvec,
1706                                         unsigned long nr_reclaimed,
1707                                         unsigned long nr_scanned,
1708                                         struct scan_control *sc)
1709 {
1710         unsigned long pages_for_compaction;
1711         unsigned long inactive_lru_pages;
1712
1713         /* If not in reclaim/compaction mode, stop */
1714         if (!in_reclaim_compaction(sc))
1715                 return false;
1716
1717         /* Consider stopping depending on scan and reclaim activity */
1718         if (sc->gfp_mask & __GFP_REPEAT) {
1719                 /*
1720                  * For __GFP_REPEAT allocations, stop reclaiming if the
1721                  * full LRU list has been scanned and we are still failing
1722                  * to reclaim pages. This full LRU scan is potentially
1723                  * expensive but a __GFP_REPEAT caller really wants to succeed
1724                  */
1725                 if (!nr_reclaimed && !nr_scanned)
1726                         return false;
1727         } else {
1728                 /*
1729                  * For non-__GFP_REPEAT allocations which can presumably
1730                  * fail without consequence, stop if we failed to reclaim
1731                  * any pages from the last SWAP_CLUSTER_MAX number of
1732                  * pages that were scanned. This will return to the
1733                  * caller faster at the risk reclaim/compaction and
1734                  * the resulting allocation attempt fails
1735                  */
1736                 if (!nr_reclaimed)
1737                         return false;
1738         }
1739
1740         /*
1741          * If we have not reclaimed enough pages for compaction and the
1742          * inactive lists are large enough, continue reclaiming
1743          */
1744         pages_for_compaction = (2UL << sc->order);
1745         inactive_lru_pages = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1746         if (nr_swap_pages > 0)
1747                 inactive_lru_pages += get_lru_size(lruvec, LRU_INACTIVE_ANON);
1748         if (sc->nr_reclaimed < pages_for_compaction &&
1749                         inactive_lru_pages > pages_for_compaction)
1750                 return true;
1751
1752         /* If compaction would go ahead or the allocation would succeed, stop */
1753         switch (compaction_suitable(lruvec_zone(lruvec), sc->order)) {
1754         case COMPACT_PARTIAL:
1755         case COMPACT_CONTINUE:
1756                 return false;
1757         default:
1758                 return true;
1759         }
1760 }
1761
1762 /*
1763  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1764  */
1765 static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
1766 {
1767         unsigned long nr[NR_LRU_LISTS];
1768         unsigned long nr_to_scan;
1769         enum lru_list lru;
1770         unsigned long nr_reclaimed, nr_scanned;
1771         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1772         struct blk_plug plug;
1773
1774 restart:
1775         nr_reclaimed = 0;
1776         nr_scanned = sc->nr_scanned;
1777         get_scan_count(lruvec, sc, nr);
1778
1779         blk_start_plug(&plug);
1780         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1781                                         nr[LRU_INACTIVE_FILE]) {
1782                 for_each_evictable_lru(lru) {
1783                         if (nr[lru]) {
1784                                 nr_to_scan = min_t(unsigned long,
1785                                                    nr[lru], SWAP_CLUSTER_MAX);
1786                                 nr[lru] -= nr_to_scan;
1787
1788                                 nr_reclaimed += shrink_list(lru, nr_to_scan,
1789                                                             lruvec, sc);
1790                         }
1791                 }
1792                 /*
1793                  * On large memory systems, scan >> priority can become
1794                  * really large. This is fine for the starting priority;
1795                  * we want to put equal scanning pressure on each zone.
1796                  * However, if the VM has a harder time of freeing pages,
1797                  * with multiple processes reclaiming pages, the total
1798                  * freeing target can get unreasonably large.
1799                  */
1800                 if (nr_reclaimed >= nr_to_reclaim &&
1801                     sc->priority < DEF_PRIORITY)
1802                         break;
1803         }
1804         blk_finish_plug(&plug);
1805         sc->nr_reclaimed += nr_reclaimed;
1806
1807         /*
1808          * Even if we did not try to evict anon pages at all, we want to
1809          * rebalance the anon lru active/inactive ratio.
1810          */
1811         if (inactive_anon_is_low(lruvec))
1812                 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
1813                                    sc, LRU_ACTIVE_ANON);
1814
1815         /* reclaim/compaction might need reclaim to continue */
1816         if (should_continue_reclaim(lruvec, nr_reclaimed,
1817                                     sc->nr_scanned - nr_scanned, sc))
1818                 goto restart;
1819
1820         throttle_vm_writeout(sc->gfp_mask);
1821 }
1822
1823 static void shrink_zone(struct zone *zone, struct scan_control *sc)
1824 {
1825         struct mem_cgroup *root = sc->target_mem_cgroup;
1826         struct mem_cgroup_reclaim_cookie reclaim = {
1827                 .zone = zone,
1828                 .priority = sc->priority,
1829         };
1830         struct mem_cgroup *memcg;
1831
1832         memcg = mem_cgroup_iter(root, NULL, &reclaim);
1833         do {
1834                 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
1835
1836                 shrink_lruvec(lruvec, sc);
1837
1838                 /*
1839                  * Limit reclaim has historically picked one memcg and
1840                  * scanned it with decreasing priority levels until
1841                  * nr_to_reclaim had been reclaimed.  This priority
1842                  * cycle is thus over after a single memcg.
1843                  *
1844                  * Direct reclaim and kswapd, on the other hand, have
1845                  * to scan all memory cgroups to fulfill the overall
1846                  * scan target for the zone.
1847                  */
1848                 if (!global_reclaim(sc)) {
1849                         mem_cgroup_iter_break(root, memcg);
1850                         break;
1851                 }
1852                 memcg = mem_cgroup_iter(root, memcg, &reclaim);
1853         } while (memcg);
1854 }
1855
1856 /* Returns true if compaction should go ahead for a high-order request */
1857 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
1858 {
1859         unsigned long balance_gap, watermark;
1860         bool watermark_ok;
1861
1862         /* Do not consider compaction for orders reclaim is meant to satisfy */
1863         if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
1864                 return false;
1865
1866         /*
1867          * Compaction takes time to run and there are potentially other
1868          * callers using the pages just freed. Continue reclaiming until
1869          * there is a buffer of free pages available to give compaction
1870          * a reasonable chance of completing and allocating the page
1871          */
1872         balance_gap = min(low_wmark_pages(zone),
1873                 (zone->present_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
1874                         KSWAPD_ZONE_BALANCE_GAP_RATIO);
1875         watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
1876         watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
1877
1878         /*
1879          * If compaction is deferred, reclaim up to a point where
1880          * compaction will have a chance of success when re-enabled
1881          */
1882         if (compaction_deferred(zone, sc->order))
1883                 return watermark_ok;
1884
1885         /* If compaction is not ready to start, keep reclaiming */
1886         if (!compaction_suitable(zone, sc->order))
1887                 return false;
1888
1889         return watermark_ok;
1890 }
1891
1892 /*
1893  * This is the direct reclaim path, for page-allocating processes.  We only
1894  * try to reclaim pages from zones which will satisfy the caller's allocation
1895  * request.
1896  *
1897  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1898  * Because:
1899  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1900  *    allocation or
1901  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1902  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1903  *    zone defense algorithm.
1904  *
1905  * If a zone is deemed to be full of pinned pages then just give it a light
1906  * scan then give up on it.
1907  *
1908  * This function returns true if a zone is being reclaimed for a costly
1909  * high-order allocation and compaction is ready to begin. This indicates to
1910  * the caller that it should consider retrying the allocation instead of
1911  * further reclaim.
1912  */
1913 static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
1914 {
1915         struct zoneref *z;
1916         struct zone *zone;
1917         unsigned long nr_soft_reclaimed;
1918         unsigned long nr_soft_scanned;
1919         bool aborted_reclaim = false;
1920
1921         /*
1922          * If the number of buffer_heads in the machine exceeds the maximum
1923          * allowed level, force direct reclaim to scan the highmem zone as
1924          * highmem pages could be pinning lowmem pages storing buffer_heads
1925          */
1926         if (buffer_heads_over_limit)
1927                 sc->gfp_mask |= __GFP_HIGHMEM;
1928
1929         for_each_zone_zonelist_nodemask(zone, z, zonelist,
1930                                         gfp_zone(sc->gfp_mask), sc->nodemask) {
1931                 if (!populated_zone(zone))
1932                         continue;
1933                 /*
1934                  * Take care memory controller reclaiming has small influence
1935                  * to global LRU.
1936                  */
1937                 if (global_reclaim(sc)) {
1938                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1939                                 continue;
1940                         if (zone->all_unreclaimable &&
1941                                         sc->priority != DEF_PRIORITY)
1942                                 continue;       /* Let kswapd poll it */
1943                         if (COMPACTION_BUILD) {
1944                                 /*
1945                                  * If we already have plenty of memory free for
1946                                  * compaction in this zone, don't free any more.
1947                                  * Even though compaction is invoked for any
1948                                  * non-zero order, only frequent costly order
1949                                  * reclamation is disruptive enough to become a
1950                                  * noticeable problem, like transparent huge
1951                                  * page allocations.
1952                                  */
1953                                 if (compaction_ready(zone, sc)) {
1954                                         aborted_reclaim = true;
1955                                         continue;
1956                                 }
1957                         }
1958                         /*
1959                          * This steals pages from memory cgroups over softlimit
1960                          * and returns the number of reclaimed pages and
1961                          * scanned pages. This works for global memory pressure
1962                          * and balancing, not for a memcg's limit.
1963                          */
1964                         nr_soft_scanned = 0;
1965                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
1966                                                 sc->order, sc->gfp_mask,
1967                                                 &nr_soft_scanned);
1968                         sc->nr_reclaimed += nr_soft_reclaimed;
1969                         sc->nr_scanned += nr_soft_scanned;
1970                         /* need some check for avoid more shrink_zone() */
1971                 }
1972
1973                 shrink_zone(zone, sc);
1974         }
1975
1976         return aborted_reclaim;
1977 }
1978
1979 static bool zone_reclaimable(struct zone *zone)
1980 {
1981         return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
1982 }
1983
1984 /* All zones in zonelist are unreclaimable? */
1985 static bool all_unreclaimable(struct zonelist *zonelist,
1986                 struct scan_control *sc)
1987 {
1988         struct zoneref *z;
1989         struct zone *zone;
1990
1991         for_each_zone_zonelist_nodemask(zone, z, zonelist,
1992                         gfp_zone(sc->gfp_mask), sc->nodemask) {
1993                 if (!populated_zone(zone))
1994                         continue;
1995                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1996                         continue;
1997                 if (!zone->all_unreclaimable)
1998                         return false;
1999         }
2000
2001         return true;
2002 }
2003
2004 /*
2005  * This is the main entry point to direct page reclaim.
2006  *
2007  * If a full scan of the inactive list fails to free enough memory then we
2008  * are "out of memory" and something needs to be killed.
2009  *
2010  * If the caller is !__GFP_FS then the probability of a failure is reasonably
2011  * high - the zone may be full of dirty or under-writeback pages, which this
2012  * caller can't do much about.  We kick the writeback threads and take explicit
2013  * naps in the hope that some of these pages can be written.  But if the
2014  * allocating task holds filesystem locks which prevent writeout this might not
2015  * work, and the allocation attempt will fail.
2016  *
2017  * returns:     0, if no pages reclaimed
2018  *              else, the number of pages reclaimed
2019  */
2020 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2021                                         struct scan_control *sc,
2022                                         struct shrink_control *shrink)
2023 {
2024         unsigned long total_scanned = 0;
2025         struct reclaim_state *reclaim_state = current->reclaim_state;
2026         struct zoneref *z;
2027         struct zone *zone;
2028         unsigned long writeback_threshold;
2029         bool aborted_reclaim;
2030
2031         delayacct_freepages_start();
2032
2033         if (global_reclaim(sc))
2034                 count_vm_event(ALLOCSTALL);
2035
2036         do {
2037                 sc->nr_scanned = 0;
2038                 aborted_reclaim = shrink_zones(zonelist, sc);
2039
2040                 /*
2041                  * Don't shrink slabs when reclaiming memory from
2042                  * over limit cgroups
2043                  */
2044                 if (global_reclaim(sc)) {
2045                         unsigned long lru_pages = 0;
2046                         for_each_zone_zonelist(zone, z, zonelist,
2047                                         gfp_zone(sc->gfp_mask)) {
2048                                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2049                                         continue;
2050
2051                                 lru_pages += zone_reclaimable_pages(zone);
2052                         }
2053
2054                         shrink_slab(shrink, sc->nr_scanned, lru_pages);
2055                         if (reclaim_state) {
2056                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2057                                 reclaim_state->reclaimed_slab = 0;
2058                         }
2059                 }
2060                 total_scanned += sc->nr_scanned;
2061                 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2062                         goto out;
2063
2064                 /*
2065                  * Try to write back as many pages as we just scanned.  This
2066                  * tends to cause slow streaming writers to write data to the
2067                  * disk smoothly, at the dirtying rate, which is nice.   But
2068                  * that's undesirable in laptop mode, where we *want* lumpy
2069                  * writeout.  So in laptop mode, write out the whole world.
2070                  */
2071                 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2072                 if (total_scanned > writeback_threshold) {
2073                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2074                                                 WB_REASON_TRY_TO_FREE_PAGES);
2075                         sc->may_writepage = 1;
2076                 }
2077
2078                 /* Take a nap, wait for some writeback to complete */
2079                 if (!sc->hibernation_mode && sc->nr_scanned &&
2080                     sc->priority < DEF_PRIORITY - 2) {
2081                         struct zone *preferred_zone;
2082
2083                         first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2084                                                 &cpuset_current_mems_allowed,
2085                                                 &preferred_zone);
2086                         wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2087                 }
2088         } while (--sc->priority >= 0);
2089
2090 out:
2091         delayacct_freepages_end();
2092
2093         if (sc->nr_reclaimed)
2094                 return sc->nr_reclaimed;
2095
2096         /*
2097          * As hibernation is going on, kswapd is freezed so that it can't mark
2098          * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2099          * check.
2100          */
2101         if (oom_killer_disabled)
2102                 return 0;
2103
2104         /* Aborted reclaim to try compaction? don't OOM, then */
2105         if (aborted_reclaim)
2106                 return 1;
2107
2108         /* top priority shrink_zones still had more to do? don't OOM, then */
2109         if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
2110                 return 1;
2111
2112         return 0;
2113 }
2114
2115 static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2116 {
2117         struct zone *zone;
2118         unsigned long pfmemalloc_reserve = 0;
2119         unsigned long free_pages = 0;
2120         int i;
2121         bool wmark_ok;
2122
2123         for (i = 0; i <= ZONE_NORMAL; i++) {
2124                 zone = &pgdat->node_zones[i];
2125                 pfmemalloc_reserve += min_wmark_pages(zone);
2126                 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2127         }
2128
2129         wmark_ok = free_pages > pfmemalloc_reserve / 2;
2130
2131         /* kswapd must be awake if processes are being throttled */
2132         if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2133                 pgdat->classzone_idx = min(pgdat->classzone_idx,
2134                                                 (enum zone_type)ZONE_NORMAL);
2135                 wake_up_interruptible(&pgdat->kswapd_wait);
2136         }
2137
2138         return wmark_ok;
2139 }
2140
2141 /*
2142  * Throttle direct reclaimers if backing storage is backed by the network
2143  * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2144  * depleted. kswapd will continue to make progress and wake the processes
2145  * when the low watermark is reached
2146  */
2147 static void throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2148                                         nodemask_t *nodemask)
2149 {
2150         struct zone *zone;
2151         int high_zoneidx = gfp_zone(gfp_mask);
2152         pg_data_t *pgdat;
2153
2154         /*
2155          * Kernel threads should not be throttled as they may be indirectly
2156          * responsible for cleaning pages necessary for reclaim to make forward
2157          * progress. kjournald for example may enter direct reclaim while
2158          * committing a transaction where throttling it could forcing other
2159          * processes to block on log_wait_commit().
2160          */
2161         if (current->flags & PF_KTHREAD)
2162                 return;
2163
2164         /* Check if the pfmemalloc reserves are ok */
2165         first_zones_zonelist(zonelist, high_zoneidx, NULL, &zone);
2166         pgdat = zone->zone_pgdat;
2167         if (pfmemalloc_watermark_ok(pgdat))
2168                 return;
2169
2170         /* Account for the throttling */
2171         count_vm_event(PGSCAN_DIRECT_THROTTLE);
2172
2173         /*
2174          * If the caller cannot enter the filesystem, it's possible that it
2175          * is due to the caller holding an FS lock or performing a journal
2176          * transaction in the case of a filesystem like ext[3|4]. In this case,
2177          * it is not safe to block on pfmemalloc_wait as kswapd could be
2178          * blocked waiting on the same lock. Instead, throttle for up to a
2179          * second before continuing.
2180          */
2181         if (!(gfp_mask & __GFP_FS)) {
2182                 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2183                         pfmemalloc_watermark_ok(pgdat), HZ);
2184                 return;
2185         }
2186
2187         /* Throttle until kswapd wakes the process */
2188         wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2189                 pfmemalloc_watermark_ok(pgdat));
2190 }
2191
2192 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2193                                 gfp_t gfp_mask, nodemask_t *nodemask)
2194 {
2195         unsigned long nr_reclaimed;
2196         struct scan_control sc = {
2197                 .gfp_mask = gfp_mask,
2198                 .may_writepage = !laptop_mode,
2199                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2200                 .may_unmap = 1,
2201                 .may_swap = 1,
2202                 .order = order,
2203                 .priority = DEF_PRIORITY,
2204                 .target_mem_cgroup = NULL,
2205                 .nodemask = nodemask,
2206         };
2207         struct shrink_control shrink = {
2208                 .gfp_mask = sc.gfp_mask,
2209         };
2210
2211         throttle_direct_reclaim(gfp_mask, zonelist, nodemask);
2212
2213         /*
2214          * Do not enter reclaim if fatal signal is pending. 1 is returned so
2215          * that the page allocator does not consider triggering OOM
2216          */
2217         if (fatal_signal_pending(current))
2218                 return 1;
2219
2220         trace_mm_vmscan_direct_reclaim_begin(order,
2221                                 sc.may_writepage,
2222                                 gfp_mask);
2223
2224         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2225
2226         trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2227
2228         return nr_reclaimed;
2229 }
2230
2231 #ifdef CONFIG_MEMCG
2232
2233 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
2234                                                 gfp_t gfp_mask, bool noswap,
2235                                                 struct zone *zone,
2236                                                 unsigned long *nr_scanned)
2237 {
2238         struct scan_control sc = {
2239                 .nr_scanned = 0,
2240                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2241                 .may_writepage = !laptop_mode,
2242                 .may_unmap = 1,
2243                 .may_swap = !noswap,
2244                 .order = 0,
2245                 .priority = 0,
2246                 .target_mem_cgroup = memcg,
2247         };
2248         struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2249
2250         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2251                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2252
2253         trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
2254                                                       sc.may_writepage,
2255                                                       sc.gfp_mask);
2256
2257         /*
2258          * NOTE: Although we can get the priority field, using it
2259          * here is not a good idea, since it limits the pages we can scan.
2260          * if we don't reclaim here, the shrink_zone from balance_pgdat
2261          * will pick up pages from other mem cgroup's as well. We hack
2262          * the priority and make it zero.
2263          */
2264         shrink_lruvec(lruvec, &sc);
2265
2266         trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2267
2268         *nr_scanned = sc.nr_scanned;
2269         return sc.nr_reclaimed;
2270 }
2271
2272 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2273                                            gfp_t gfp_mask,
2274                                            bool noswap)
2275 {
2276         struct zonelist *zonelist;
2277         unsigned long nr_reclaimed;
2278         int nid;
2279         struct scan_control sc = {
2280                 .may_writepage = !laptop_mode,
2281                 .may_unmap = 1,
2282                 .may_swap = !noswap,
2283                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2284                 .order = 0,
2285                 .priority = DEF_PRIORITY,
2286                 .target_mem_cgroup = memcg,
2287                 .nodemask = NULL, /* we don't care the placement */
2288                 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2289                                 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2290         };
2291         struct shrink_control shrink = {
2292                 .gfp_mask = sc.gfp_mask,
2293         };
2294
2295         /*
2296          * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2297          * take care of from where we get pages. So the node where we start the
2298          * scan does not need to be the current node.
2299          */
2300         nid = mem_cgroup_select_victim_node(memcg);
2301
2302         zonelist = NODE_DATA(nid)->node_zonelists;
2303
2304         trace_mm_vmscan_memcg_reclaim_begin(0,
2305                                             sc.may_writepage,
2306                                             sc.gfp_mask);
2307
2308         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2309
2310         trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2311
2312         return nr_reclaimed;
2313 }
2314 #endif
2315
2316 static void age_active_anon(struct zone *zone, struct scan_control *sc)
2317 {
2318         struct mem_cgroup *memcg;
2319
2320         if (!total_swap_pages)
2321                 return;
2322
2323         memcg = mem_cgroup_iter(NULL, NULL, NULL);
2324         do {
2325                 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2326
2327                 if (inactive_anon_is_low(lruvec))
2328                         shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2329                                            sc, LRU_ACTIVE_ANON);
2330
2331                 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2332         } while (memcg);
2333 }
2334
2335 /*
2336  * pgdat_balanced is used when checking if a node is balanced for high-order
2337  * allocations. Only zones that meet watermarks and are in a zone allowed
2338  * by the callers classzone_idx are added to balanced_pages. The total of
2339  * balanced pages must be at least 25% of the zones allowed by classzone_idx
2340  * for the node to be considered balanced. Forcing all zones to be balanced
2341  * for high orders can cause excessive reclaim when there are imbalanced zones.
2342  * The choice of 25% is due to
2343  *   o a 16M DMA zone that is balanced will not balance a zone on any
2344  *     reasonable sized machine
2345  *   o On all other machines, the top zone must be at least a reasonable
2346  *     percentage of the middle zones. For example, on 32-bit x86, highmem
2347  *     would need to be at least 256M for it to be balance a whole node.
2348  *     Similarly, on x86-64 the Normal zone would need to be at least 1G
2349  *     to balance a node on its own. These seemed like reasonable ratios.
2350  */
2351 static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2352                                                 int classzone_idx)
2353 {
2354         unsigned long present_pages = 0;
2355         int i;
2356
2357         for (i = 0; i <= classzone_idx; i++)
2358                 present_pages += pgdat->node_zones[i].present_pages;
2359
2360         /* A special case here: if zone has no page, we think it's balanced */
2361         return balanced_pages >= (present_pages >> 2);
2362 }
2363
2364 /*
2365  * Prepare kswapd for sleeping. This verifies that there are no processes
2366  * waiting in throttle_direct_reclaim() and that watermarks have been met.
2367  *
2368  * Returns true if kswapd is ready to sleep
2369  */
2370 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
2371                                         int classzone_idx)
2372 {
2373         int i;
2374         unsigned long balanced = 0;
2375         bool all_zones_ok = true;
2376
2377         /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2378         if (remaining)
2379                 return false;
2380
2381         /*
2382          * There is a potential race between when kswapd checks its watermarks
2383          * and a process gets throttled. There is also a potential race if
2384          * processes get throttled, kswapd wakes, a large process exits therby
2385          * balancing the zones that causes kswapd to miss a wakeup. If kswapd
2386          * is going to sleep, no process should be sleeping on pfmemalloc_wait
2387          * so wake them now if necessary. If necessary, processes will wake
2388          * kswapd and get throttled again
2389          */
2390         if (waitqueue_active(&pgdat->pfmemalloc_wait)) {
2391                 wake_up(&pgdat->pfmemalloc_wait);
2392                 return false;
2393         }
2394
2395         /* Check the watermark levels */
2396         for (i = 0; i <= classzone_idx; i++) {
2397                 struct zone *zone = pgdat->node_zones + i;
2398
2399                 if (!populated_zone(zone))
2400                         continue;
2401
2402                 /*
2403                  * balance_pgdat() skips over all_unreclaimable after
2404                  * DEF_PRIORITY. Effectively, it considers them balanced so
2405                  * they must be considered balanced here as well if kswapd
2406                  * is to sleep
2407                  */
2408                 if (zone->all_unreclaimable) {
2409                         balanced += zone->present_pages;
2410                         continue;
2411                 }
2412
2413                 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2414                                                         i, 0))
2415                         all_zones_ok = false;
2416                 else
2417                         balanced += zone->present_pages;
2418         }
2419
2420         /*
2421          * For high-order requests, the balanced zones must contain at least
2422          * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2423          * must be balanced
2424          */
2425         if (order)
2426                 return pgdat_balanced(pgdat, balanced, classzone_idx);
2427         else
2428                 return all_zones_ok;
2429 }
2430
2431 /*
2432  * For kswapd, balance_pgdat() will work across all this node's zones until
2433  * they are all at high_wmark_pages(zone).
2434  *
2435  * Returns the final order kswapd was reclaiming at
2436  *
2437  * There is special handling here for zones which are full of pinned pages.
2438  * This can happen if the pages are all mlocked, or if they are all used by
2439  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
2440  * What we do is to detect the case where all pages in the zone have been
2441  * scanned twice and there has been zero successful reclaim.  Mark the zone as
2442  * dead and from now on, only perform a short scan.  Basically we're polling
2443  * the zone for when the problem goes away.
2444  *
2445  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
2446  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2447  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2448  * lower zones regardless of the number of free pages in the lower zones. This
2449  * interoperates with the page allocator fallback scheme to ensure that aging
2450  * of pages is balanced across the zones.
2451  */
2452 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2453                                                         int *classzone_idx)
2454 {
2455         int all_zones_ok;
2456         unsigned long balanced;
2457         int i;
2458         int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
2459         unsigned long total_scanned;
2460         struct reclaim_state *reclaim_state = current->reclaim_state;
2461         unsigned long nr_soft_reclaimed;
2462         unsigned long nr_soft_scanned;
2463         struct scan_control sc = {
2464                 .gfp_mask = GFP_KERNEL,
2465                 .may_unmap = 1,
2466                 .may_swap = 1,
2467                 /*
2468                  * kswapd doesn't want to be bailed out while reclaim. because
2469                  * we want to put equal scanning pressure on each zone.
2470                  */
2471                 .nr_to_reclaim = ULONG_MAX,
2472                 .order = order,
2473                 .target_mem_cgroup = NULL,
2474         };
2475         struct shrink_control shrink = {
2476                 .gfp_mask = sc.gfp_mask,
2477         };
2478 loop_again:
2479         total_scanned = 0;
2480         sc.priority = DEF_PRIORITY;
2481         sc.nr_reclaimed = 0;
2482         sc.may_writepage = !laptop_mode;
2483         count_vm_event(PAGEOUTRUN);
2484
2485         do {
2486                 unsigned long lru_pages = 0;
2487                 int has_under_min_watermark_zone = 0;
2488
2489                 all_zones_ok = 1;
2490                 balanced = 0;
2491
2492                 /*
2493                  * Scan in the highmem->dma direction for the highest
2494                  * zone which needs scanning
2495                  */
2496                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2497                         struct zone *zone = pgdat->node_zones + i;
2498
2499                         if (!populated_zone(zone))
2500                                 continue;
2501
2502                         if (zone->all_unreclaimable &&
2503                             sc.priority != DEF_PRIORITY)
2504                                 continue;
2505
2506                         /*
2507                          * Do some background aging of the anon list, to give
2508                          * pages a chance to be referenced before reclaiming.
2509                          */
2510                         age_active_anon(zone, &sc);
2511
2512                         /*
2513                          * If the number of buffer_heads in the machine
2514                          * exceeds the maximum allowed level and this node
2515                          * has a highmem zone, force kswapd to reclaim from
2516                          * it to relieve lowmem pressure.
2517                          */
2518                         if (buffer_heads_over_limit && is_highmem_idx(i)) {
2519                                 end_zone = i;
2520                                 break;
2521                         }
2522
2523                         if (!zone_watermark_ok_safe(zone, order,
2524                                         high_wmark_pages(zone), 0, 0)) {
2525                                 end_zone = i;
2526                                 break;
2527                         } else {
2528                                 /* If balanced, clear the congested flag */
2529                                 zone_clear_flag(zone, ZONE_CONGESTED);
2530                         }
2531                 }
2532                 if (i < 0)
2533                         goto out;
2534
2535                 for (i = 0; i <= end_zone; i++) {
2536                         struct zone *zone = pgdat->node_zones + i;
2537
2538                         lru_pages += zone_reclaimable_pages(zone);
2539                 }
2540
2541                 /*
2542                  * Now scan the zone in the dma->highmem direction, stopping
2543                  * at the last zone which needs scanning.
2544                  *
2545                  * We do this because the page allocator works in the opposite
2546                  * direction.  This prevents the page allocator from allocating
2547                  * pages behind kswapd's direction of progress, which would
2548                  * cause too much scanning of the lower zones.
2549                  */
2550                 for (i = 0; i <= end_zone; i++) {
2551                         struct zone *zone = pgdat->node_zones + i;
2552                         int nr_slab, testorder;
2553                         unsigned long balance_gap;
2554
2555                         if (!populated_zone(zone))
2556                                 continue;
2557
2558                         if (zone->all_unreclaimable &&
2559                             sc.priority != DEF_PRIORITY)
2560                                 continue;
2561
2562                         sc.nr_scanned = 0;
2563
2564                         nr_soft_scanned = 0;
2565                         /*
2566                          * Call soft limit reclaim before calling shrink_zone.
2567                          */
2568                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2569                                                         order, sc.gfp_mask,
2570                                                         &nr_soft_scanned);
2571                         sc.nr_reclaimed += nr_soft_reclaimed;
2572                         total_scanned += nr_soft_scanned;
2573
2574                         /*
2575                          * We put equal pressure on every zone, unless
2576                          * one zone has way too many pages free
2577                          * already. The "too many pages" is defined
2578                          * as the high wmark plus a "gap" where the
2579                          * gap is either the low watermark or 1%
2580                          * of the zone, whichever is smaller.
2581                          */
2582                         balance_gap = min(low_wmark_pages(zone),
2583                                 (zone->present_pages +
2584                                         KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2585                                 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2586                         /*
2587                          * Kswapd reclaims only single pages with compaction
2588                          * enabled. Trying too hard to reclaim until contiguous
2589                          * free pages have become available can hurt performance
2590                          * by evicting too much useful data from memory.
2591                          * Do not reclaim more than needed for compaction.
2592                          */
2593                         testorder = order;
2594                         if (COMPACTION_BUILD && order &&
2595                                         compaction_suitable(zone, order) !=
2596                                                 COMPACT_SKIPPED)
2597                                 testorder = 0;
2598
2599                         if ((buffer_heads_over_limit && is_highmem_idx(i)) ||
2600                                     !zone_watermark_ok_safe(zone, testorder,
2601                                         high_wmark_pages(zone) + balance_gap,
2602                                         end_zone, 0)) {
2603                                 shrink_zone(zone, &sc);
2604
2605                                 reclaim_state->reclaimed_slab = 0;
2606                                 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2607                                 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2608                                 total_scanned += sc.nr_scanned;
2609
2610                                 if (nr_slab == 0 && !zone_reclaimable(zone))
2611                                         zone->all_unreclaimable = 1;
2612                         }
2613
2614                         /*
2615                          * If we've done a decent amount of scanning and
2616                          * the reclaim ratio is low, start doing writepage
2617                          * even in laptop mode
2618                          */
2619                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2620                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2621                                 sc.may_writepage = 1;
2622
2623                         if (zone->all_unreclaimable) {
2624                                 if (end_zone && end_zone == i)
2625                                         end_zone--;
2626                                 continue;
2627                         }
2628
2629                         if (!zone_watermark_ok_safe(zone, testorder,
2630                                         high_wmark_pages(zone), end_zone, 0)) {
2631                                 all_zones_ok = 0;
2632                                 /*
2633                                  * We are still under min water mark.  This
2634                                  * means that we have a GFP_ATOMIC allocation
2635                                  * failure risk. Hurry up!
2636                                  */
2637                                 if (!zone_watermark_ok_safe(zone, order,
2638                                             min_wmark_pages(zone), end_zone, 0))
2639                                         has_under_min_watermark_zone = 1;
2640                         } else {
2641                                 /*
2642                                  * If a zone reaches its high watermark,
2643                                  * consider it to be no longer congested. It's
2644                                  * possible there are dirty pages backed by
2645                                  * congested BDIs but as pressure is relieved,
2646                                  * speculatively avoid congestion waits
2647                                  */
2648                                 zone_clear_flag(zone, ZONE_CONGESTED);
2649                                 if (i <= *classzone_idx)
2650                                         balanced += zone->present_pages;
2651                         }
2652
2653                 }
2654
2655                 /*
2656                  * If the low watermark is met there is no need for processes
2657                  * to be throttled on pfmemalloc_wait as they should not be
2658                  * able to safely make forward progress. Wake them
2659                  */
2660                 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
2661                                 pfmemalloc_watermark_ok(pgdat))
2662                         wake_up(&pgdat->pfmemalloc_wait);
2663
2664                 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2665                         break;          /* kswapd: all done */
2666                 /*
2667                  * OK, kswapd is getting into trouble.  Take a nap, then take
2668                  * another pass across the zones.
2669                  */
2670                 if (total_scanned && (sc.priority < DEF_PRIORITY - 2)) {
2671                         if (has_under_min_watermark_zone)
2672                                 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2673                         else
2674                                 congestion_wait(BLK_RW_ASYNC, HZ/10);
2675                 }
2676
2677                 /*
2678                  * We do this so kswapd doesn't build up large priorities for
2679                  * example when it is freeing in parallel with allocators. It
2680                  * matches the direct reclaim path behaviour in terms of impact
2681                  * on zone->*_priority.
2682                  */
2683                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2684                         break;
2685         } while (--sc.priority >= 0);
2686 out:
2687
2688         /*
2689          * order-0: All zones must meet high watermark for a balanced node
2690          * high-order: Balanced zones must make up at least 25% of the node
2691          *             for the node to be balanced
2692          */
2693         if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2694                 cond_resched();
2695
2696                 try_to_freeze();
2697
2698                 /*
2699                  * Fragmentation may mean that the system cannot be
2700                  * rebalanced for high-order allocations in all zones.
2701                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2702                  * it means the zones have been fully scanned and are still
2703                  * not balanced. For high-order allocations, there is
2704                  * little point trying all over again as kswapd may
2705                  * infinite loop.
2706                  *
2707                  * Instead, recheck all watermarks at order-0 as they
2708                  * are the most important. If watermarks are ok, kswapd will go
2709                  * back to sleep. High-order users can still perform direct
2710                  * reclaim if they wish.
2711                  */
2712                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2713                         order = sc.order = 0;
2714
2715                 goto loop_again;
2716         }
2717
2718         /*
2719          * If kswapd was reclaiming at a higher order, it has the option of
2720          * sleeping without all zones being balanced. Before it does, it must
2721          * ensure that the watermarks for order-0 on *all* zones are met and
2722          * that the congestion flags are cleared. The congestion flag must
2723          * be cleared as kswapd is the only mechanism that clears the flag
2724          * and it is potentially going to sleep here.
2725          */
2726         if (order) {
2727                 int zones_need_compaction = 1;
2728
2729                 for (i = 0; i <= end_zone; i++) {
2730                         struct zone *zone = pgdat->node_zones + i;
2731
2732                         if (!populated_zone(zone))
2733                                 continue;
2734
2735                         if (zone->all_unreclaimable &&
2736                             sc.priority != DEF_PRIORITY)
2737                                 continue;
2738
2739                         /* Would compaction fail due to lack of free memory? */
2740                         if (COMPACTION_BUILD &&
2741                             compaction_suitable(zone, order) == COMPACT_SKIPPED)
2742                                 goto loop_again;
2743
2744                         /* Confirm the zone is balanced for order-0 */
2745                         if (!zone_watermark_ok(zone, 0,
2746                                         high_wmark_pages(zone), 0, 0)) {
2747                                 order = sc.order = 0;
2748                                 goto loop_again;
2749                         }
2750
2751                         /* Check if the memory needs to be defragmented. */
2752                         if (zone_watermark_ok(zone, order,
2753                                     low_wmark_pages(zone), *classzone_idx, 0))
2754                                 zones_need_compaction = 0;
2755
2756                         /* If balanced, clear the congested flag */
2757                         zone_clear_flag(zone, ZONE_CONGESTED);
2758                 }
2759
2760                 if (zones_need_compaction)
2761                         compact_pgdat(pgdat, order);
2762         }
2763
2764         /*
2765          * Return the order we were reclaiming at so prepare_kswapd_sleep()
2766          * makes a decision on the order we were last reclaiming at. However,
2767          * if another caller entered the allocator slow path while kswapd
2768          * was awake, order will remain at the higher level
2769          */
2770         *classzone_idx = end_zone;
2771         return order;
2772 }
2773
2774 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2775 {
2776         long remaining = 0;
2777         DEFINE_WAIT(wait);
2778
2779         if (freezing(current) || kthread_should_stop())
2780                 return;
2781
2782         prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2783
2784         /* Try to sleep for a short interval */
2785         if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
2786                 remaining = schedule_timeout(HZ/10);
2787                 finish_wait(&pgdat->kswapd_wait, &wait);
2788                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2789         }
2790
2791         /*
2792          * After a short sleep, check if it was a premature sleep. If not, then
2793          * go fully to sleep until explicitly woken up.
2794          */
2795         if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
2796                 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2797
2798                 /*
2799                  * vmstat counters are not perfectly accurate and the estimated
2800                  * value for counters such as NR_FREE_PAGES can deviate from the
2801                  * true value by nr_online_cpus * threshold. To avoid the zone
2802                  * watermarks being breached while under pressure, we reduce the
2803                  * per-cpu vmstat threshold while kswapd is awake and restore
2804                  * them before going back to sleep.
2805                  */
2806                 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2807
2808                 if (!kthread_should_stop())
2809                         schedule();
2810
2811                 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2812         } else {
2813                 if (remaining)
2814                         count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2815                 else
2816                         count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2817         }
2818         finish_wait(&pgdat->kswapd_wait, &wait);
2819 }
2820
2821 /*
2822  * The background pageout daemon, started as a kernel thread
2823  * from the init process.
2824  *
2825  * This basically trickles out pages so that we have _some_
2826  * free memory available even if there is no other activity
2827  * that frees anything up. This is needed for things like routing
2828  * etc, where we otherwise might have all activity going on in
2829  * asynchronous contexts that cannot page things out.
2830  *
2831  * If there are applications that are active memory-allocators
2832  * (most normal use), this basically shouldn't matter.
2833  */
2834 static int kswapd(void *p)
2835 {
2836         unsigned long order, new_order;
2837         unsigned balanced_order;
2838         int classzone_idx, new_classzone_idx;
2839         int balanced_classzone_idx;
2840         pg_data_t *pgdat = (pg_data_t*)p;
2841         struct task_struct *tsk = current;
2842
2843         struct reclaim_state reclaim_state = {
2844                 .reclaimed_slab = 0,
2845         };
2846         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2847
2848         lockdep_set_current_reclaim_state(GFP_KERNEL);
2849
2850         if (!cpumask_empty(cpumask))
2851                 set_cpus_allowed_ptr(tsk, cpumask);
2852         current->reclaim_state = &reclaim_state;
2853
2854         /*
2855          * Tell the memory management that we're a "memory allocator",
2856          * and that if we need more memory we should get access to it
2857          * regardless (see "__alloc_pages()"). "kswapd" should
2858          * never get caught in the normal page freeing logic.
2859          *
2860          * (Kswapd normally doesn't need memory anyway, but sometimes
2861          * you need a small amount of memory in order to be able to
2862          * page out something else, and this flag essentially protects
2863          * us from recursively trying to free more memory as we're
2864          * trying to free the first piece of memory in the first place).
2865          */
2866         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2867         set_freezable();
2868
2869         order = new_order = 0;
2870         balanced_order = 0;
2871         classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
2872         balanced_classzone_idx = classzone_idx;
2873         for ( ; ; ) {
2874                 int ret;
2875
2876                 /*
2877                  * If the last balance_pgdat was unsuccessful it's unlikely a
2878                  * new request of a similar or harder type will succeed soon
2879                  * so consider going to sleep on the basis we reclaimed at
2880                  */
2881                 if (balanced_classzone_idx >= new_classzone_idx &&
2882                                         balanced_order == new_order) {
2883                         new_order = pgdat->kswapd_max_order;
2884                         new_classzone_idx = pgdat->classzone_idx;
2885                         pgdat->kswapd_max_order =  0;
2886                         pgdat->classzone_idx = pgdat->nr_zones - 1;
2887                 }
2888
2889                 if (order < new_order || classzone_idx > new_classzone_idx) {
2890                         /*
2891                          * Don't sleep if someone wants a larger 'order'
2892                          * allocation or has tigher zone constraints
2893                          */
2894                         order = new_order;
2895                         classzone_idx = new_classzone_idx;
2896                 } else {
2897                         kswapd_try_to_sleep(pgdat, balanced_order,
2898                                                 balanced_classzone_idx);
2899                         order = pgdat->kswapd_max_order;
2900                         classzone_idx = pgdat->classzone_idx;
2901                         new_order = order;
2902                         new_classzone_idx = classzone_idx;
2903                         pgdat->kswapd_max_order = 0;
2904                         pgdat->classzone_idx = pgdat->nr_zones - 1;
2905                 }
2906
2907                 ret = try_to_freeze();
2908                 if (kthread_should_stop())
2909                         break;
2910
2911                 /*
2912                  * We can speed up thawing tasks if we don't call balance_pgdat
2913                  * after returning from the refrigerator
2914                  */
2915                 if (!ret) {
2916                         trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2917                         balanced_classzone_idx = classzone_idx;
2918                         balanced_order = balance_pgdat(pgdat, order,
2919                                                 &balanced_classzone_idx);
2920                 }
2921         }
2922         return 0;
2923 }
2924
2925 /*
2926  * A zone is low on free memory, so wake its kswapd task to service it.
2927  */
2928 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
2929 {
2930         pg_data_t *pgdat;
2931
2932         if (!populated_zone(zone))
2933                 return;
2934
2935         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2936                 return;
2937         pgdat = zone->zone_pgdat;
2938         if (pgdat->kswapd_max_order < order) {
2939                 pgdat->kswapd_max_order = order;
2940                 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
2941         }
2942         if (!waitqueue_active(&pgdat->kswapd_wait))
2943                 return;
2944         if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
2945                 return;
2946
2947         trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
2948         wake_up_interruptible(&pgdat->kswapd_wait);
2949 }
2950
2951 /*
2952  * The reclaimable count would be mostly accurate.
2953  * The less reclaimable pages may be
2954  * - mlocked pages, which will be moved to unevictable list when encountered
2955  * - mapped pages, which may require several travels to be reclaimed
2956  * - dirty pages, which is not "instantly" reclaimable
2957  */
2958 unsigned long global_reclaimable_pages(void)
2959 {
2960         int nr;
2961
2962         nr = global_page_state(NR_ACTIVE_FILE) +
2963              global_page_state(NR_INACTIVE_FILE);
2964
2965         if (nr_swap_pages > 0)
2966                 nr += global_page_state(NR_ACTIVE_ANON) +
2967                       global_page_state(NR_INACTIVE_ANON);
2968
2969         return nr;
2970 }
2971
2972 unsigned long zone_reclaimable_pages(struct zone *zone)
2973 {
2974         int nr;
2975
2976         nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2977              zone_page_state(zone, NR_INACTIVE_FILE);
2978
2979         if (nr_swap_pages > 0)
2980                 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2981                       zone_page_state(zone, NR_INACTIVE_ANON);
2982
2983         return nr;
2984 }
2985
2986 #ifdef CONFIG_HIBERNATION
2987 /*
2988  * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2989  * freed pages.
2990  *
2991  * Rather than trying to age LRUs the aim is to preserve the overall
2992  * LRU order by reclaiming preferentially
2993  * inactive > active > active referenced > active mapped
2994  */
2995 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
2996 {
2997         struct reclaim_state reclaim_state;
2998         struct scan_control sc = {
2999                 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3000                 .may_swap = 1,
3001                 .may_unmap = 1,
3002                 .may_writepage = 1,
3003                 .nr_to_reclaim = nr_to_reclaim,
3004                 .hibernation_mode = 1,
3005                 .order = 0,
3006                 .priority = DEF_PRIORITY,
3007         };
3008         struct shrink_control shrink = {
3009                 .gfp_mask = sc.gfp_mask,
3010         };
3011         struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3012         struct task_struct *p = current;
3013         unsigned long nr_reclaimed;
3014
3015         p->flags |= PF_MEMALLOC;
3016         lockdep_set_current_reclaim_state(sc.gfp_mask);
3017         reclaim_state.reclaimed_slab = 0;
3018         p->reclaim_state = &reclaim_state;
3019
3020         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
3021
3022         p->reclaim_state = NULL;
3023         lockdep_clear_current_reclaim_state();
3024         p->flags &= ~PF_MEMALLOC;
3025
3026         return nr_reclaimed;
3027 }
3028 #endif /* CONFIG_HIBERNATION */
3029
3030 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3031    not required for correctness.  So if the last cpu in a node goes
3032    away, we get changed to run anywhere: as the first one comes back,
3033    restore their cpu bindings. */
3034 static int __devinit cpu_callback(struct notifier_block *nfb,
3035                                   unsigned long action, void *hcpu)
3036 {
3037         int nid;
3038
3039         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3040                 for_each_node_state(nid, N_HIGH_MEMORY) {
3041                         pg_data_t *pgdat = NODE_DATA(nid);
3042                         const struct cpumask *mask;
3043
3044                         mask = cpumask_of_node(pgdat->node_id);
3045
3046                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3047                                 /* One of our CPUs online: restore mask */
3048                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3049                 }
3050         }
3051         return NOTIFY_OK;
3052 }
3053
3054 /*
3055  * This kswapd start function will be called by init and node-hot-add.
3056  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3057  */
3058 int kswapd_run(int nid)
3059 {
3060         pg_data_t *pgdat = NODE_DATA(nid);
3061         int ret = 0;
3062
3063         if (pgdat->kswapd)
3064                 return 0;
3065
3066         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3067         if (IS_ERR(pgdat->kswapd)) {
3068                 /* failure at boot is fatal */
3069                 BUG_ON(system_state == SYSTEM_BOOTING);
3070                 printk("Failed to start kswapd on node %d\n",nid);
3071                 ret = -1;
3072         }
3073         return ret;
3074 }
3075
3076 /*
3077  * Called by memory hotplug when all memory in a node is offlined.  Caller must
3078  * hold lock_memory_hotplug().
3079  */
3080 void kswapd_stop(int nid)
3081 {
3082         struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3083
3084         if (kswapd) {
3085                 kthread_stop(kswapd);
3086                 NODE_DATA(nid)->kswapd = NULL;
3087         }
3088 }
3089
3090 static int __init kswapd_init(void)
3091 {
3092         int nid;
3093
3094         swap_setup();
3095         for_each_node_state(nid, N_HIGH_MEMORY)
3096                 kswapd_run(nid);
3097         hotcpu_notifier(cpu_callback, 0);
3098         return 0;
3099 }
3100
3101 module_init(kswapd_init)
3102
3103 #ifdef CONFIG_NUMA
3104 /*
3105  * Zone reclaim mode
3106  *
3107  * If non-zero call zone_reclaim when the number of free pages falls below
3108  * the watermarks.
3109  */
3110 int zone_reclaim_mode __read_mostly;
3111
3112 #define RECLAIM_OFF 0
3113 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
3114 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
3115 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
3116
3117 /*
3118  * Priority for ZONE_RECLAIM. This determines the fraction of pages
3119  * of a node considered for each zone_reclaim. 4 scans 1/16th of
3120  * a zone.
3121  */
3122 #define ZONE_RECLAIM_PRIORITY 4
3123
3124 /*
3125  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3126  * occur.
3127  */
3128 int sysctl_min_unmapped_ratio = 1;
3129
3130 /*
3131  * If the number of slab pages in a zone grows beyond this percentage then
3132  * slab reclaim needs to occur.
3133  */
3134 int sysctl_min_slab_ratio = 5;
3135
3136 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3137 {
3138         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3139         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3140                 zone_page_state(zone, NR_ACTIVE_FILE);
3141
3142         /*
3143          * It's possible for there to be more file mapped pages than
3144          * accounted for by the pages on the file LRU lists because
3145          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3146          */
3147         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3148 }
3149
3150 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3151 static long zone_pagecache_reclaimable(struct zone *zone)
3152 {
3153         long nr_pagecache_reclaimable;
3154         long delta = 0;
3155
3156         /*
3157          * If RECLAIM_SWAP is set, then all file pages are considered
3158          * potentially reclaimable. Otherwise, we have to worry about
3159          * pages like swapcache and zone_unmapped_file_pages() provides
3160          * a better estimate
3161          */
3162         if (zone_reclaim_mode & RECLAIM_SWAP)
3163                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3164         else
3165                 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3166
3167         /* If we can't clean pages, remove dirty pages from consideration */
3168         if (!(zone_reclaim_mode & RECLAIM_WRITE))
3169                 delta += zone_page_state(zone, NR_FILE_DIRTY);
3170
3171         /* Watch for any possible underflows due to delta */
3172         if (unlikely(delta > nr_pagecache_reclaimable))
3173                 delta = nr_pagecache_reclaimable;
3174
3175         return nr_pagecache_reclaimable - delta;
3176 }
3177
3178 /*
3179  * Try to free up some pages from this zone through reclaim.
3180  */
3181 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3182 {
3183         /* Minimum pages needed in order to stay on node */
3184         const unsigned long nr_pages = 1 << order;
3185         struct task_struct *p = current;
3186         struct reclaim_state reclaim_state;
3187         struct scan_control sc = {
3188                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3189                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3190                 .may_swap = 1,
3191                 .nr_to_reclaim = max_t(unsigned long, nr_pages,
3192                                        SWAP_CLUSTER_MAX),
3193                 .gfp_mask = gfp_mask,
3194                 .order = order,
3195                 .priority = ZONE_RECLAIM_PRIORITY,
3196         };
3197         struct shrink_control shrink = {
3198                 .gfp_mask = sc.gfp_mask,
3199         };
3200         unsigned long nr_slab_pages0, nr_slab_pages1;
3201
3202         cond_resched();
3203         /*
3204          * We need to be able to allocate from the reserves for RECLAIM_SWAP
3205          * and we also need to be able to write out pages for RECLAIM_WRITE
3206          * and RECLAIM_SWAP.
3207          */
3208         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3209         lockdep_set_current_reclaim_state(gfp_mask);
3210         reclaim_state.reclaimed_slab = 0;
3211         p->reclaim_state = &reclaim_state;
3212
3213         if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3214                 /*
3215                  * Free memory by calling shrink zone with increasing
3216                  * priorities until we have enough memory freed.
3217                  */
3218                 do {
3219                         shrink_zone(zone, &sc);
3220                 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3221         }
3222
3223         nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3224         if (nr_slab_pages0 > zone->min_slab_pages) {
3225                 /*
3226                  * shrink_slab() does not currently allow us to determine how
3227                  * many pages were freed in this zone. So we take the current
3228                  * number of slab pages and shake the slab until it is reduced
3229                  * by the same nr_pages that we used for reclaiming unmapped
3230                  * pages.
3231                  *
3232                  * Note that shrink_slab will free memory on all zones and may
3233                  * take a long time.
3234                  */
3235                 for (;;) {
3236                         unsigned long lru_pages = zone_reclaimable_pages(zone);
3237
3238                         /* No reclaimable slab or very low memory pressure */
3239                         if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3240                                 break;
3241
3242                         /* Freed enough memory */
3243                         nr_slab_pages1 = zone_page_state(zone,
3244                                                         NR_SLAB_RECLAIMABLE);
3245                         if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3246                                 break;
3247                 }
3248
3249                 /*
3250                  * Update nr_reclaimed by the number of slab pages we
3251                  * reclaimed from this zone.
3252                  */
3253                 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3254                 if (nr_slab_pages1 < nr_slab_pages0)
3255                         sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3256         }
3257
3258         p->reclaim_state = NULL;
3259         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3260         lockdep_clear_current_reclaim_state();
3261         return sc.nr_reclaimed >= nr_pages;
3262 }
3263
3264 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3265 {
3266         int node_id;
3267         int ret;
3268
3269         /*
3270          * Zone reclaim reclaims unmapped file backed pages and
3271          * slab pages if we are over the defined limits.
3272          *
3273          * A small portion of unmapped file backed pages is needed for
3274          * file I/O otherwise pages read by file I/O will be immediately
3275          * thrown out if the zone is overallocated. So we do not reclaim
3276          * if less than a specified percentage of the zone is used by
3277          * unmapped file backed pages.
3278          */
3279         if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3280             zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3281                 return ZONE_RECLAIM_FULL;
3282
3283         if (zone->all_unreclaimable)
3284                 return ZONE_RECLAIM_FULL;
3285
3286         /*
3287          * Do not scan if the allocation should not be delayed.
3288          */
3289         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3290                 return ZONE_RECLAIM_NOSCAN;
3291
3292         /*
3293          * Only run zone reclaim on the local zone or on zones that do not
3294          * have associated processors. This will favor the local processor
3295          * over remote processors and spread off node memory allocations
3296          * as wide as possible.
3297          */
3298         node_id = zone_to_nid(zone);
3299         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3300                 return ZONE_RECLAIM_NOSCAN;
3301
3302         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3303                 return ZONE_RECLAIM_NOSCAN;
3304
3305         ret = __zone_reclaim(zone, gfp_mask, order);
3306         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3307
3308         if (!ret)
3309                 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3310
3311         return ret;
3312 }
3313 #endif
3314
3315 /*
3316  * page_evictable - test whether a page is evictable
3317  * @page: the page to test
3318  * @vma: the VMA in which the page is or will be mapped, may be NULL
3319  *
3320  * Test whether page is evictable--i.e., should be placed on active/inactive
3321  * lists vs unevictable list.  The vma argument is !NULL when called from the
3322  * fault path to determine how to instantate a new page.
3323  *
3324  * Reasons page might not be evictable:
3325  * (1) page's mapping marked unevictable
3326  * (2) page is part of an mlocked VMA
3327  *
3328  */
3329 int page_evictable(struct page *page, struct vm_area_struct *vma)
3330 {
3331
3332         if (mapping_unevictable(page_mapping(page)))
3333                 return 0;
3334
3335         if (PageMlocked(page) || (vma && mlocked_vma_newpage(vma, page)))
3336                 return 0;
3337
3338         return 1;
3339 }
3340
3341 #ifdef CONFIG_SHMEM
3342 /**
3343  * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3344  * @pages:      array of pages to check
3345  * @nr_pages:   number of pages to check
3346  *
3347  * Checks pages for evictability and moves them to the appropriate lru list.
3348  *
3349  * This function is only used for SysV IPC SHM_UNLOCK.
3350  */
3351 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3352 {
3353         struct lruvec *lruvec;
3354         struct zone *zone = NULL;
3355         int pgscanned = 0;
3356         int pgrescued = 0;
3357         int i;
3358
3359         for (i = 0; i < nr_pages; i++) {
3360                 struct page *page = pages[i];
3361                 struct zone *pagezone;
3362
3363                 pgscanned++;
3364                 pagezone = page_zone(page);
3365                 if (pagezone != zone) {
3366                         if (zone)
3367                                 spin_unlock_irq(&zone->lru_lock);
3368                         zone = pagezone;
3369                         spin_lock_irq(&zone->lru_lock);
3370                 }
3371                 lruvec = mem_cgroup_page_lruvec(page, zone);
3372
3373                 if (!PageLRU(page) || !PageUnevictable(page))
3374                         continue;
3375
3376                 if (page_evictable(page, NULL)) {
3377                         enum lru_list lru = page_lru_base_type(page);
3378
3379                         VM_BUG_ON(PageActive(page));
3380                         ClearPageUnevictable(page);
3381                         del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3382                         add_page_to_lru_list(page, lruvec, lru);
3383                         pgrescued++;
3384                 }
3385         }
3386
3387         if (zone) {
3388                 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3389                 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3390                 spin_unlock_irq(&zone->lru_lock);
3391         }
3392 }
3393 #endif /* CONFIG_SHMEM */
3394
3395 static void warn_scan_unevictable_pages(void)
3396 {
3397         printk_once(KERN_WARNING
3398                     "%s: The scan_unevictable_pages sysctl/node-interface has been "
3399                     "disabled for lack of a legitimate use case.  If you have "
3400                     "one, please send an email to linux-mm@kvack.org.\n",
3401                     current->comm);
3402 }
3403
3404 /*
3405  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
3406  * all nodes' unevictable lists for evictable pages
3407  */
3408 unsigned long scan_unevictable_pages;
3409
3410 int scan_unevictable_handler(struct ctl_table *table, int write,
3411                            void __user *buffer,
3412                            size_t *length, loff_t *ppos)
3413 {
3414         warn_scan_unevictable_pages();
3415         proc_doulongvec_minmax(table, write, buffer, length, ppos);
3416         scan_unevictable_pages = 0;
3417         return 0;
3418 }
3419
3420 #ifdef CONFIG_NUMA
3421 /*
3422  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
3423  * a specified node's per zone unevictable lists for evictable pages.
3424  */
3425
3426 static ssize_t read_scan_unevictable_node(struct device *dev,
3427                                           struct device_attribute *attr,
3428                                           char *buf)
3429 {
3430         warn_scan_unevictable_pages();
3431         return sprintf(buf, "0\n");     /* always zero; should fit... */
3432 }
3433
3434 static ssize_t write_scan_unevictable_node(struct device *dev,
3435                                            struct device_attribute *attr,
3436                                         const char *buf, size_t count)
3437 {
3438         warn_scan_unevictable_pages();
3439         return 1;
3440 }
3441
3442
3443 static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3444                         read_scan_unevictable_node,
3445                         write_scan_unevictable_node);
3446
3447 int scan_unevictable_register_node(struct node *node)
3448 {
3449         return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
3450 }
3451
3452 void scan_unevictable_unregister_node(struct node *node)
3453 {
3454         device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);
3455 }
3456 #endif