drivers: dma-contiguous: refactor dma_alloc_from_contiguous()
[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         mem_cgroup_uncharge_start();
691         while (!list_empty(page_list)) {
692                 enum page_references references;
693                 struct address_space *mapping;
694                 struct page *page;
695                 int may_enter_fs;
696
697                 cond_resched();
698
699                 page = lru_to_page(page_list);
700                 list_del(&page->lru);
701
702                 if (!trylock_page(page))
703                         goto keep;
704
705                 VM_BUG_ON(PageActive(page));
706                 VM_BUG_ON(page_zone(page) != zone);
707
708                 sc->nr_scanned++;
709
710                 if (unlikely(!page_evictable(page, NULL)))
711                         goto cull_mlocked;
712
713                 if (!sc->may_unmap && page_mapped(page))
714                         goto keep_locked;
715
716                 /* Double the slab pressure for mapped and swapcache pages */
717                 if (page_mapped(page) || PageSwapCache(page))
718                         sc->nr_scanned++;
719
720                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
721                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
722
723                 if (PageWriteback(page)) {
724                         /*
725                          * memcg doesn't have any dirty pages throttling so we
726                          * could easily OOM just because too many pages are in
727                          * writeback and there is nothing else to reclaim.
728                          *
729                          * Check __GFP_IO, certainly because a loop driver
730                          * thread might enter reclaim, and deadlock if it waits
731                          * on a page for which it is needed to do the write
732                          * (loop masks off __GFP_IO|__GFP_FS for this reason);
733                          * but more thought would probably show more reasons.
734                          *
735                          * Don't require __GFP_FS, since we're not going into
736                          * the FS, just waiting on its writeback completion.
737                          * Worryingly, ext4 gfs2 and xfs allocate pages with
738                          * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so
739                          * testing may_enter_fs here is liable to OOM on them.
740                          */
741                         if (global_reclaim(sc) ||
742                             !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) {
743                                 /*
744                                  * This is slightly racy - end_page_writeback()
745                                  * might have just cleared PageReclaim, then
746                                  * setting PageReclaim here end up interpreted
747                                  * as PageReadahead - but that does not matter
748                                  * enough to care.  What we do want is for this
749                                  * page to have PageReclaim set next time memcg
750                                  * reclaim reaches the tests above, so it will
751                                  * then wait_on_page_writeback() to avoid OOM;
752                                  * and it's also appropriate in global reclaim.
753                                  */
754                                 SetPageReclaim(page);
755                                 nr_writeback++;
756                                 goto keep_locked;
757                         }
758                         wait_on_page_writeback(page);
759                 }
760
761                 references = page_check_references(page, sc);
762                 switch (references) {
763                 case PAGEREF_ACTIVATE:
764                         goto activate_locked;
765                 case PAGEREF_KEEP:
766                         goto keep_locked;
767                 case PAGEREF_RECLAIM:
768                 case PAGEREF_RECLAIM_CLEAN:
769                         ; /* try to reclaim the page below */
770                 }
771
772                 /*
773                  * Anonymous process memory has backing store?
774                  * Try to allocate it some swap space here.
775                  */
776                 if (PageAnon(page) && !PageSwapCache(page)) {
777                         if (!(sc->gfp_mask & __GFP_IO))
778                                 goto keep_locked;
779                         if (!add_to_swap(page))
780                                 goto activate_locked;
781                         may_enter_fs = 1;
782                 }
783
784                 mapping = page_mapping(page);
785
786                 /*
787                  * The page is mapped into the page tables of one or more
788                  * processes. Try to unmap it here.
789                  */
790                 if (page_mapped(page) && mapping) {
791                         switch (try_to_unmap(page, TTU_UNMAP)) {
792                         case SWAP_FAIL:
793                                 goto activate_locked;
794                         case SWAP_AGAIN:
795                                 goto keep_locked;
796                         case SWAP_MLOCK:
797                                 goto cull_mlocked;
798                         case SWAP_SUCCESS:
799                                 ; /* try to free the page below */
800                         }
801                 }
802
803                 if (PageDirty(page)) {
804                         nr_dirty++;
805
806                         /*
807                          * Only kswapd can writeback filesystem pages to
808                          * avoid risk of stack overflow but do not writeback
809                          * unless under significant pressure.
810                          */
811                         if (page_is_file_cache(page) &&
812                                         (!current_is_kswapd() ||
813                                          sc->priority >= DEF_PRIORITY - 2)) {
814                                 /*
815                                  * Immediately reclaim when written back.
816                                  * Similar in principal to deactivate_page()
817                                  * except we already have the page isolated
818                                  * and know it's dirty
819                                  */
820                                 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
821                                 SetPageReclaim(page);
822
823                                 goto keep_locked;
824                         }
825
826                         if (references == PAGEREF_RECLAIM_CLEAN)
827                                 goto keep_locked;
828                         if (!may_enter_fs)
829                                 goto keep_locked;
830                         if (!sc->may_writepage)
831                                 goto keep_locked;
832
833                         /* Page is dirty, try to write it out here */
834                         switch (pageout(page, mapping, sc)) {
835                         case PAGE_KEEP:
836                                 nr_congested++;
837                                 goto keep_locked;
838                         case PAGE_ACTIVATE:
839                                 goto activate_locked;
840                         case PAGE_SUCCESS:
841                                 if (PageWriteback(page))
842                                         goto keep;
843                                 if (PageDirty(page))
844                                         goto keep;
845
846                                 /*
847                                  * A synchronous write - probably a ramdisk.  Go
848                                  * ahead and try to reclaim the page.
849                                  */
850                                 if (!trylock_page(page))
851                                         goto keep;
852                                 if (PageDirty(page) || PageWriteback(page))
853                                         goto keep_locked;
854                                 mapping = page_mapping(page);
855                         case PAGE_CLEAN:
856                                 ; /* try to free the page below */
857                         }
858                 }
859
860                 /*
861                  * If the page has buffers, try to free the buffer mappings
862                  * associated with this page. If we succeed we try to free
863                  * the page as well.
864                  *
865                  * We do this even if the page is PageDirty().
866                  * try_to_release_page() does not perform I/O, but it is
867                  * possible for a page to have PageDirty set, but it is actually
868                  * clean (all its buffers are clean).  This happens if the
869                  * buffers were written out directly, with submit_bh(). ext3
870                  * will do this, as well as the blockdev mapping.
871                  * try_to_release_page() will discover that cleanness and will
872                  * drop the buffers and mark the page clean - it can be freed.
873                  *
874                  * Rarely, pages can have buffers and no ->mapping.  These are
875                  * the pages which were not successfully invalidated in
876                  * truncate_complete_page().  We try to drop those buffers here
877                  * and if that worked, and the page is no longer mapped into
878                  * process address space (page_count == 1) it can be freed.
879                  * Otherwise, leave the page on the LRU so it is swappable.
880                  */
881                 if (page_has_private(page)) {
882                         if (!try_to_release_page(page, sc->gfp_mask))
883                                 goto activate_locked;
884                         if (!mapping && page_count(page) == 1) {
885                                 unlock_page(page);
886                                 if (put_page_testzero(page))
887                                         goto free_it;
888                                 else {
889                                         /*
890                                          * rare race with speculative reference.
891                                          * the speculative reference will free
892                                          * this page shortly, so we may
893                                          * increment nr_reclaimed here (and
894                                          * leave it off the LRU).
895                                          */
896                                         nr_reclaimed++;
897                                         continue;
898                                 }
899                         }
900                 }
901
902                 if (!mapping || !__remove_mapping(mapping, page))
903                         goto keep_locked;
904
905                 /*
906                  * At this point, we have no other references and there is
907                  * no way to pick any more up (removed from LRU, removed
908                  * from pagecache). Can use non-atomic bitops now (and
909                  * we obviously don't have to worry about waking up a process
910                  * waiting on the page lock, because there are no references.
911                  */
912                 __clear_page_locked(page);
913 free_it:
914                 nr_reclaimed++;
915
916                 /*
917                  * Is there need to periodically free_page_list? It would
918                  * appear not as the counts should be low
919                  */
920                 list_add(&page->lru, &free_pages);
921                 continue;
922
923 cull_mlocked:
924                 if (PageSwapCache(page))
925                         try_to_free_swap(page);
926                 unlock_page(page);
927                 putback_lru_page(page);
928                 continue;
929
930 activate_locked:
931                 /* Not a candidate for swapping, so reclaim swap space. */
932                 if (PageSwapCache(page) && vm_swap_full())
933                         try_to_free_swap(page);
934                 VM_BUG_ON(PageActive(page));
935                 SetPageActive(page);
936                 pgactivate++;
937 keep_locked:
938                 unlock_page(page);
939 keep:
940                 list_add(&page->lru, &ret_pages);
941                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
942         }
943
944         /*
945          * Tag a zone as congested if all the dirty pages encountered were
946          * backed by a congested BDI. In this case, reclaimers should just
947          * back off and wait for congestion to clear because further reclaim
948          * will encounter the same problem
949          */
950         if (nr_dirty && nr_dirty == nr_congested && global_reclaim(sc))
951                 zone_set_flag(zone, ZONE_CONGESTED);
952
953         free_hot_cold_page_list(&free_pages, 1);
954
955         list_splice(&ret_pages, page_list);
956         count_vm_events(PGACTIVATE, pgactivate);
957         mem_cgroup_uncharge_end();
958         *ret_nr_dirty += nr_dirty;
959         *ret_nr_writeback += nr_writeback;
960         return nr_reclaimed;
961 }
962
963 /*
964  * Attempt to remove the specified page from its LRU.  Only take this page
965  * if it is of the appropriate PageActive status.  Pages which are being
966  * freed elsewhere are also ignored.
967  *
968  * page:        page to consider
969  * mode:        one of the LRU isolation modes defined above
970  *
971  * returns 0 on success, -ve errno on failure.
972  */
973 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
974 {
975         int ret = -EINVAL;
976
977         /* Only take pages on the LRU. */
978         if (!PageLRU(page))
979                 return ret;
980
981         /* Do not give back unevictable pages for compaction */
982         if (PageUnevictable(page))
983                 return ret;
984
985         ret = -EBUSY;
986
987         /*
988          * To minimise LRU disruption, the caller can indicate that it only
989          * wants to isolate pages it will be able to operate on without
990          * blocking - clean pages for the most part.
991          *
992          * ISOLATE_CLEAN means that only clean pages should be isolated. This
993          * is used by reclaim when it is cannot write to backing storage
994          *
995          * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
996          * that it is possible to migrate without blocking
997          */
998         if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
999                 /* All the caller can do on PageWriteback is block */
1000                 if (PageWriteback(page))
1001                         return ret;
1002
1003                 if (PageDirty(page)) {
1004                         struct address_space *mapping;
1005
1006                         /* ISOLATE_CLEAN means only clean pages */
1007                         if (mode & ISOLATE_CLEAN)
1008                                 return ret;
1009
1010                         /*
1011                          * Only pages without mappings or that have a
1012                          * ->migratepage callback are possible to migrate
1013                          * without blocking
1014                          */
1015                         mapping = page_mapping(page);
1016                         if (mapping && !mapping->a_ops->migratepage)
1017                                 return ret;
1018                 }
1019         }
1020
1021         if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1022                 return ret;
1023
1024         if (likely(get_page_unless_zero(page))) {
1025                 /*
1026                  * Be careful not to clear PageLRU until after we're
1027                  * sure the page is not being freed elsewhere -- the
1028                  * page release code relies on it.
1029                  */
1030                 ClearPageLRU(page);
1031                 ret = 0;
1032         }
1033
1034         return ret;
1035 }
1036
1037 /*
1038  * zone->lru_lock is heavily contended.  Some of the functions that
1039  * shrink the lists perform better by taking out a batch of pages
1040  * and working on them outside the LRU lock.
1041  *
1042  * For pagecache intensive workloads, this function is the hottest
1043  * spot in the kernel (apart from copy_*_user functions).
1044  *
1045  * Appropriate locks must be held before calling this function.
1046  *
1047  * @nr_to_scan: The number of pages to look through on the list.
1048  * @lruvec:     The LRU vector to pull pages from.
1049  * @dst:        The temp list to put pages on to.
1050  * @nr_scanned: The number of pages that were scanned.
1051  * @sc:         The scan_control struct for this reclaim session
1052  * @mode:       One of the LRU isolation modes
1053  * @lru:        LRU list id for isolating
1054  *
1055  * returns how many pages were moved onto *@dst.
1056  */
1057 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1058                 struct lruvec *lruvec, struct list_head *dst,
1059                 unsigned long *nr_scanned, struct scan_control *sc,
1060                 isolate_mode_t mode, enum lru_list lru)
1061 {
1062         struct list_head *src = &lruvec->lists[lru];
1063         unsigned long nr_taken = 0;
1064         unsigned long scan;
1065
1066         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1067                 struct page *page;
1068                 int nr_pages;
1069
1070                 page = lru_to_page(src);
1071                 prefetchw_prev_lru_page(page, src, flags);
1072
1073                 VM_BUG_ON(!PageLRU(page));
1074
1075                 switch (__isolate_lru_page(page, mode)) {
1076                 case 0:
1077                         nr_pages = hpage_nr_pages(page);
1078                         mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
1079                         list_move(&page->lru, dst);
1080                         nr_taken += nr_pages;
1081                         break;
1082
1083                 case -EBUSY:
1084                         /* else it is being freed elsewhere */
1085                         list_move(&page->lru, src);
1086                         continue;
1087
1088                 default:
1089                         BUG();
1090                 }
1091         }
1092
1093         *nr_scanned = scan;
1094         trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1095                                     nr_taken, mode, is_file_lru(lru));
1096         return nr_taken;
1097 }
1098
1099 /**
1100  * isolate_lru_page - tries to isolate a page from its LRU list
1101  * @page: page to isolate from its LRU list
1102  *
1103  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1104  * vmstat statistic corresponding to whatever LRU list the page was on.
1105  *
1106  * Returns 0 if the page was removed from an LRU list.
1107  * Returns -EBUSY if the page was not on an LRU list.
1108  *
1109  * The returned page will have PageLRU() cleared.  If it was found on
1110  * the active list, it will have PageActive set.  If it was found on
1111  * the unevictable list, it will have the PageUnevictable bit set. That flag
1112  * may need to be cleared by the caller before letting the page go.
1113  *
1114  * The vmstat statistic corresponding to the list on which the page was
1115  * found will be decremented.
1116  *
1117  * Restrictions:
1118  * (1) Must be called with an elevated refcount on the page. This is a
1119  *     fundamentnal difference from isolate_lru_pages (which is called
1120  *     without a stable reference).
1121  * (2) the lru_lock must not be held.
1122  * (3) interrupts must be enabled.
1123  */
1124 int isolate_lru_page(struct page *page)
1125 {
1126         int ret = -EBUSY;
1127
1128         VM_BUG_ON(!page_count(page));
1129
1130         if (PageLRU(page)) {
1131                 struct zone *zone = page_zone(page);
1132                 struct lruvec *lruvec;
1133
1134                 spin_lock_irq(&zone->lru_lock);
1135                 lruvec = mem_cgroup_page_lruvec(page, zone);
1136                 if (PageLRU(page)) {
1137                         int lru = page_lru(page);
1138                         get_page(page);
1139                         ClearPageLRU(page);
1140                         del_page_from_lru_list(page, lruvec, lru);
1141                         ret = 0;
1142                 }
1143                 spin_unlock_irq(&zone->lru_lock);
1144         }
1145         return ret;
1146 }
1147
1148 /*
1149  * Are there way too many processes in the direct reclaim path already?
1150  */
1151 static int too_many_isolated(struct zone *zone, int file,
1152                 struct scan_control *sc)
1153 {
1154         unsigned long inactive, isolated;
1155
1156         if (current_is_kswapd())
1157                 return 0;
1158
1159         if (!global_reclaim(sc))
1160                 return 0;
1161
1162         if (file) {
1163                 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1164                 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1165         } else {
1166                 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1167                 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1168         }
1169
1170         return isolated > inactive;
1171 }
1172
1173 static noinline_for_stack void
1174 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1175 {
1176         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1177         struct zone *zone = lruvec_zone(lruvec);
1178         LIST_HEAD(pages_to_free);
1179
1180         /*
1181          * Put back any unfreeable pages.
1182          */
1183         while (!list_empty(page_list)) {
1184                 struct page *page = lru_to_page(page_list);
1185                 int lru;
1186
1187                 VM_BUG_ON(PageLRU(page));
1188                 list_del(&page->lru);
1189                 if (unlikely(!page_evictable(page, NULL))) {
1190                         spin_unlock_irq(&zone->lru_lock);
1191                         putback_lru_page(page);
1192                         spin_lock_irq(&zone->lru_lock);
1193                         continue;
1194                 }
1195
1196                 lruvec = mem_cgroup_page_lruvec(page, zone);
1197
1198                 SetPageLRU(page);
1199                 lru = page_lru(page);
1200                 add_page_to_lru_list(page, lruvec, lru);
1201
1202                 if (is_active_lru(lru)) {
1203                         int file = is_file_lru(lru);
1204                         int numpages = hpage_nr_pages(page);
1205                         reclaim_stat->recent_rotated[file] += numpages;
1206                 }
1207                 if (put_page_testzero(page)) {
1208                         __ClearPageLRU(page);
1209                         __ClearPageActive(page);
1210                         del_page_from_lru_list(page, lruvec, lru);
1211
1212                         if (unlikely(PageCompound(page))) {
1213                                 spin_unlock_irq(&zone->lru_lock);
1214                                 (*get_compound_page_dtor(page))(page);
1215                                 spin_lock_irq(&zone->lru_lock);
1216                         } else
1217                                 list_add(&page->lru, &pages_to_free);
1218                 }
1219         }
1220
1221         /*
1222          * To save our caller's stack, now use input list for pages to free.
1223          */
1224         list_splice(&pages_to_free, page_list);
1225 }
1226
1227 /*
1228  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1229  * of reclaimed pages
1230  */
1231 static noinline_for_stack unsigned long
1232 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1233                      struct scan_control *sc, enum lru_list lru)
1234 {
1235         LIST_HEAD(page_list);
1236         unsigned long nr_scanned;
1237         unsigned long nr_reclaimed = 0;
1238         unsigned long nr_taken;
1239         unsigned long nr_dirty = 0;
1240         unsigned long nr_writeback = 0;
1241         isolate_mode_t isolate_mode = 0;
1242         int file = is_file_lru(lru);
1243         struct zone *zone = lruvec_zone(lruvec);
1244         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1245
1246         while (unlikely(too_many_isolated(zone, file, sc))) {
1247                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1248
1249                 /* We are about to die and free our memory. Return now. */
1250                 if (fatal_signal_pending(current))
1251                         return SWAP_CLUSTER_MAX;
1252         }
1253
1254         lru_add_drain();
1255
1256         if (!sc->may_unmap)
1257                 isolate_mode |= ISOLATE_UNMAPPED;
1258         if (!sc->may_writepage)
1259                 isolate_mode |= ISOLATE_CLEAN;
1260
1261         spin_lock_irq(&zone->lru_lock);
1262
1263         nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1264                                      &nr_scanned, sc, isolate_mode, lru);
1265
1266         __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1267         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1268
1269         if (global_reclaim(sc)) {
1270                 zone->pages_scanned += nr_scanned;
1271                 if (current_is_kswapd())
1272                         __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
1273                 else
1274                         __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
1275         }
1276         spin_unlock_irq(&zone->lru_lock);
1277
1278         if (nr_taken == 0)
1279                 return 0;
1280
1281         nr_reclaimed = shrink_page_list(&page_list, zone, sc,
1282                                                 &nr_dirty, &nr_writeback);
1283
1284         spin_lock_irq(&zone->lru_lock);
1285
1286         reclaim_stat->recent_scanned[file] += nr_taken;
1287
1288         if (global_reclaim(sc)) {
1289                 if (current_is_kswapd())
1290                         __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1291                                                nr_reclaimed);
1292                 else
1293                         __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1294                                                nr_reclaimed);
1295         }
1296
1297         putback_inactive_pages(lruvec, &page_list);
1298
1299         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1300
1301         spin_unlock_irq(&zone->lru_lock);
1302
1303         free_hot_cold_page_list(&page_list, 1);
1304
1305         /*
1306          * If reclaim is isolating dirty pages under writeback, it implies
1307          * that the long-lived page allocation rate is exceeding the page
1308          * laundering rate. Either the global limits are not being effective
1309          * at throttling processes due to the page distribution throughout
1310          * zones or there is heavy usage of a slow backing device. The
1311          * only option is to throttle from reclaim context which is not ideal
1312          * as there is no guarantee the dirtying process is throttled in the
1313          * same way balance_dirty_pages() manages.
1314          *
1315          * This scales the number of dirty pages that must be under writeback
1316          * before throttling depending on priority. It is a simple backoff
1317          * function that has the most effect in the range DEF_PRIORITY to
1318          * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1319          * in trouble and reclaim is considered to be in trouble.
1320          *
1321          * DEF_PRIORITY   100% isolated pages must be PageWriteback to throttle
1322          * DEF_PRIORITY-1  50% must be PageWriteback
1323          * DEF_PRIORITY-2  25% must be PageWriteback, kswapd in trouble
1324          * ...
1325          * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1326          *                     isolated page is PageWriteback
1327          */
1328         if (nr_writeback && nr_writeback >=
1329                         (nr_taken >> (DEF_PRIORITY - sc->priority)))
1330                 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1331
1332         trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1333                 zone_idx(zone),
1334                 nr_scanned, nr_reclaimed,
1335                 sc->priority,
1336                 trace_shrink_flags(file));
1337         return nr_reclaimed;
1338 }
1339
1340 /*
1341  * This moves pages from the active list to the inactive list.
1342  *
1343  * We move them the other way if the page is referenced by one or more
1344  * processes, from rmap.
1345  *
1346  * If the pages are mostly unmapped, the processing is fast and it is
1347  * appropriate to hold zone->lru_lock across the whole operation.  But if
1348  * the pages are mapped, the processing is slow (page_referenced()) so we
1349  * should drop zone->lru_lock around each page.  It's impossible to balance
1350  * this, so instead we remove the pages from the LRU while processing them.
1351  * It is safe to rely on PG_active against the non-LRU pages in here because
1352  * nobody will play with that bit on a non-LRU page.
1353  *
1354  * The downside is that we have to touch page->_count against each page.
1355  * But we had to alter page->flags anyway.
1356  */
1357
1358 static void move_active_pages_to_lru(struct lruvec *lruvec,
1359                                      struct list_head *list,
1360                                      struct list_head *pages_to_free,
1361                                      enum lru_list lru)
1362 {
1363         struct zone *zone = lruvec_zone(lruvec);
1364         unsigned long pgmoved = 0;
1365         struct page *page;
1366         int nr_pages;
1367
1368         while (!list_empty(list)) {
1369                 page = lru_to_page(list);
1370                 lruvec = mem_cgroup_page_lruvec(page, zone);
1371
1372                 VM_BUG_ON(PageLRU(page));
1373                 SetPageLRU(page);
1374
1375                 nr_pages = hpage_nr_pages(page);
1376                 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
1377                 list_move(&page->lru, &lruvec->lists[lru]);
1378                 pgmoved += nr_pages;
1379
1380                 if (put_page_testzero(page)) {
1381                         __ClearPageLRU(page);
1382                         __ClearPageActive(page);
1383                         del_page_from_lru_list(page, lruvec, lru);
1384
1385                         if (unlikely(PageCompound(page))) {
1386                                 spin_unlock_irq(&zone->lru_lock);
1387                                 (*get_compound_page_dtor(page))(page);
1388                                 spin_lock_irq(&zone->lru_lock);
1389                         } else
1390                                 list_add(&page->lru, pages_to_free);
1391                 }
1392         }
1393         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1394         if (!is_active_lru(lru))
1395                 __count_vm_events(PGDEACTIVATE, pgmoved);
1396 }
1397
1398 static void shrink_active_list(unsigned long nr_to_scan,
1399                                struct lruvec *lruvec,
1400                                struct scan_control *sc,
1401                                enum lru_list lru)
1402 {
1403         unsigned long nr_taken;
1404         unsigned long nr_scanned;
1405         unsigned long vm_flags;
1406         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1407         LIST_HEAD(l_active);
1408         LIST_HEAD(l_inactive);
1409         struct page *page;
1410         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1411         unsigned long nr_rotated = 0;
1412         isolate_mode_t isolate_mode = 0;
1413         int file = is_file_lru(lru);
1414         struct zone *zone = lruvec_zone(lruvec);
1415
1416         lru_add_drain();
1417
1418         if (!sc->may_unmap)
1419                 isolate_mode |= ISOLATE_UNMAPPED;
1420         if (!sc->may_writepage)
1421                 isolate_mode |= ISOLATE_CLEAN;
1422
1423         spin_lock_irq(&zone->lru_lock);
1424
1425         nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1426                                      &nr_scanned, sc, isolate_mode, lru);
1427         if (global_reclaim(sc))
1428                 zone->pages_scanned += nr_scanned;
1429
1430         reclaim_stat->recent_scanned[file] += nr_taken;
1431
1432         __count_zone_vm_events(PGREFILL, zone, nr_scanned);
1433         __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1434         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1435         spin_unlock_irq(&zone->lru_lock);
1436
1437         while (!list_empty(&l_hold)) {
1438                 cond_resched();
1439                 page = lru_to_page(&l_hold);
1440                 list_del(&page->lru);
1441
1442                 if (unlikely(!page_evictable(page, NULL))) {
1443                         putback_lru_page(page);
1444                         continue;
1445                 }
1446
1447                 if (unlikely(buffer_heads_over_limit)) {
1448                         if (page_has_private(page) && trylock_page(page)) {
1449                                 if (page_has_private(page))
1450                                         try_to_release_page(page, 0);
1451                                 unlock_page(page);
1452                         }
1453                 }
1454
1455                 if (page_referenced(page, 0, sc->target_mem_cgroup,
1456                                     &vm_flags)) {
1457                         nr_rotated += hpage_nr_pages(page);
1458                         /*
1459                          * Identify referenced, file-backed active pages and
1460                          * give them one more trip around the active list. So
1461                          * that executable code get better chances to stay in
1462                          * memory under moderate memory pressure.  Anon pages
1463                          * are not likely to be evicted by use-once streaming
1464                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1465                          * so we ignore them here.
1466                          */
1467                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1468                                 list_add(&page->lru, &l_active);
1469                                 continue;
1470                         }
1471                 }
1472
1473                 ClearPageActive(page);  /* we are de-activating */
1474                 list_add(&page->lru, &l_inactive);
1475         }
1476
1477         /*
1478          * Move pages back to the lru list.
1479          */
1480         spin_lock_irq(&zone->lru_lock);
1481         /*
1482          * Count referenced pages from currently used mappings as rotated,
1483          * even though only some of them are actually re-activated.  This
1484          * helps balance scan pressure between file and anonymous pages in
1485          * get_scan_ratio.
1486          */
1487         reclaim_stat->recent_rotated[file] += nr_rotated;
1488
1489         move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1490         move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
1491         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1492         spin_unlock_irq(&zone->lru_lock);
1493
1494         free_hot_cold_page_list(&l_hold, 1);
1495 }
1496
1497 #ifdef CONFIG_SWAP
1498 static int inactive_anon_is_low_global(struct zone *zone)
1499 {
1500         unsigned long active, inactive;
1501
1502         active = zone_page_state(zone, NR_ACTIVE_ANON);
1503         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1504
1505         if (inactive * zone->inactive_ratio < active)
1506                 return 1;
1507
1508         return 0;
1509 }
1510
1511 /**
1512  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1513  * @lruvec: LRU vector to check
1514  *
1515  * Returns true if the zone does not have enough inactive anon pages,
1516  * meaning some active anon pages need to be deactivated.
1517  */
1518 static int inactive_anon_is_low(struct lruvec *lruvec)
1519 {
1520         /*
1521          * If we don't have swap space, anonymous page deactivation
1522          * is pointless.
1523          */
1524         if (!total_swap_pages)
1525                 return 0;
1526
1527         if (!mem_cgroup_disabled())
1528                 return mem_cgroup_inactive_anon_is_low(lruvec);
1529
1530         return inactive_anon_is_low_global(lruvec_zone(lruvec));
1531 }
1532 #else
1533 static inline int inactive_anon_is_low(struct lruvec *lruvec)
1534 {
1535         return 0;
1536 }
1537 #endif
1538
1539 static int inactive_file_is_low_global(struct zone *zone)
1540 {
1541         unsigned long active, inactive;
1542
1543         active = zone_page_state(zone, NR_ACTIVE_FILE);
1544         inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1545
1546         return (active > inactive);
1547 }
1548
1549 /**
1550  * inactive_file_is_low - check if file pages need to be deactivated
1551  * @lruvec: LRU vector to check
1552  *
1553  * When the system is doing streaming IO, memory pressure here
1554  * ensures that active file pages get deactivated, until more
1555  * than half of the file pages are on the inactive list.
1556  *
1557  * Once we get to that situation, protect the system's working
1558  * set from being evicted by disabling active file page aging.
1559  *
1560  * This uses a different ratio than the anonymous pages, because
1561  * the page cache uses a use-once replacement algorithm.
1562  */
1563 static int inactive_file_is_low(struct lruvec *lruvec)
1564 {
1565         if (!mem_cgroup_disabled())
1566                 return mem_cgroup_inactive_file_is_low(lruvec);
1567
1568         return inactive_file_is_low_global(lruvec_zone(lruvec));
1569 }
1570
1571 static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
1572 {
1573         if (is_file_lru(lru))
1574                 return inactive_file_is_low(lruvec);
1575         else
1576                 return inactive_anon_is_low(lruvec);
1577 }
1578
1579 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1580                                  struct lruvec *lruvec, struct scan_control *sc)
1581 {
1582         if (is_active_lru(lru)) {
1583                 if (inactive_list_is_low(lruvec, lru))
1584                         shrink_active_list(nr_to_scan, lruvec, sc, lru);
1585                 return 0;
1586         }
1587
1588         return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
1589 }
1590
1591 static int vmscan_swappiness(struct scan_control *sc)
1592 {
1593         if (global_reclaim(sc))
1594                 return vm_swappiness;
1595         return mem_cgroup_swappiness(sc->target_mem_cgroup);
1596 }
1597
1598 /*
1599  * Determine how aggressively the anon and file LRU lists should be
1600  * scanned.  The relative value of each set of LRU lists is determined
1601  * by looking at the fraction of the pages scanned we did rotate back
1602  * onto the active list instead of evict.
1603  *
1604  * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1605  * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
1606  */
1607 static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
1608                            unsigned long *nr)
1609 {
1610         unsigned long anon, file, free;
1611         unsigned long anon_prio, file_prio;
1612         unsigned long ap, fp;
1613         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1614         u64 fraction[2], denominator;
1615         enum lru_list lru;
1616         int noswap = 0;
1617         bool force_scan = false;
1618         struct zone *zone = lruvec_zone(lruvec);
1619
1620         /*
1621          * If the zone or memcg is small, nr[l] can be 0.  This
1622          * results in no scanning on this priority and a potential
1623          * priority drop.  Global direct reclaim can go to the next
1624          * zone and tends to have no problems. Global kswapd is for
1625          * zone balancing and it needs to scan a minimum amount. When
1626          * reclaiming for a memcg, a priority drop can cause high
1627          * latencies, so it's better to scan a minimum amount there as
1628          * well.
1629          */
1630         if (current_is_kswapd() && zone->all_unreclaimable)
1631                 force_scan = true;
1632         if (!global_reclaim(sc))
1633                 force_scan = true;
1634
1635         /* If we have no swap space, do not bother scanning anon pages. */
1636         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1637                 noswap = 1;
1638                 fraction[0] = 0;
1639                 fraction[1] = 1;
1640                 denominator = 1;
1641                 goto out;
1642         }
1643
1644         anon  = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
1645                 get_lru_size(lruvec, LRU_INACTIVE_ANON);
1646         file  = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
1647                 get_lru_size(lruvec, LRU_INACTIVE_FILE);
1648
1649         if (global_reclaim(sc)) {
1650                 free  = zone_page_state(zone, NR_FREE_PAGES);
1651                 /* If we have very few page cache pages,
1652                    force-scan anon pages. */
1653                 if (unlikely(file + free <= high_wmark_pages(zone))) {
1654                         fraction[0] = 1;
1655                         fraction[1] = 0;
1656                         denominator = 1;
1657                         goto out;
1658                 }
1659         }
1660
1661         /*
1662          * With swappiness at 100, anonymous and file have the same priority.
1663          * This scanning priority is essentially the inverse of IO cost.
1664          */
1665         anon_prio = vmscan_swappiness(sc);
1666         file_prio = 200 - anon_prio;
1667
1668         /*
1669          * OK, so we have swap space and a fair amount of page cache
1670          * pages.  We use the recently rotated / recently scanned
1671          * ratios to determine how valuable each cache is.
1672          *
1673          * Because workloads change over time (and to avoid overflow)
1674          * we keep these statistics as a floating average, which ends
1675          * up weighing recent references more than old ones.
1676          *
1677          * anon in [0], file in [1]
1678          */
1679         spin_lock_irq(&zone->lru_lock);
1680         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1681                 reclaim_stat->recent_scanned[0] /= 2;
1682                 reclaim_stat->recent_rotated[0] /= 2;
1683         }
1684
1685         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1686                 reclaim_stat->recent_scanned[1] /= 2;
1687                 reclaim_stat->recent_rotated[1] /= 2;
1688         }
1689
1690         /*
1691          * The amount of pressure on anon vs file pages is inversely
1692          * proportional to the fraction of recently scanned pages on
1693          * each list that were recently referenced and in active use.
1694          */
1695         ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
1696         ap /= reclaim_stat->recent_rotated[0] + 1;
1697
1698         fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
1699         fp /= reclaim_stat->recent_rotated[1] + 1;
1700         spin_unlock_irq(&zone->lru_lock);
1701
1702         fraction[0] = ap;
1703         fraction[1] = fp;
1704         denominator = ap + fp + 1;
1705 out:
1706         for_each_evictable_lru(lru) {
1707                 int file = is_file_lru(lru);
1708                 unsigned long scan;
1709
1710                 scan = get_lru_size(lruvec, lru);
1711                 if (sc->priority || noswap || !vmscan_swappiness(sc)) {
1712                         scan >>= sc->priority;
1713                         if (!scan && force_scan)
1714                                 scan = SWAP_CLUSTER_MAX;
1715                         scan = div64_u64(scan * fraction[file], denominator);
1716                 }
1717                 nr[lru] = scan;
1718         }
1719 }
1720
1721 /* Use reclaim/compaction for costly allocs or under memory pressure */
1722 static bool in_reclaim_compaction(struct scan_control *sc)
1723 {
1724         if (COMPACTION_BUILD && sc->order &&
1725                         (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
1726                          sc->priority < DEF_PRIORITY - 2))
1727                 return true;
1728
1729         return false;
1730 }
1731
1732 /*
1733  * Reclaim/compaction is used for high-order allocation requests. It reclaims
1734  * order-0 pages before compacting the zone. should_continue_reclaim() returns
1735  * true if more pages should be reclaimed such that when the page allocator
1736  * calls try_to_compact_zone() that it will have enough free pages to succeed.
1737  * It will give up earlier than that if there is difficulty reclaiming pages.
1738  */
1739 static inline bool should_continue_reclaim(struct lruvec *lruvec,
1740                                         unsigned long nr_reclaimed,
1741                                         unsigned long nr_scanned,
1742                                         struct scan_control *sc)
1743 {
1744         unsigned long pages_for_compaction;
1745         unsigned long inactive_lru_pages;
1746
1747         /* If not in reclaim/compaction mode, stop */
1748         if (!in_reclaim_compaction(sc))
1749                 return false;
1750
1751         /* Consider stopping depending on scan and reclaim activity */
1752         if (sc->gfp_mask & __GFP_REPEAT) {
1753                 /*
1754                  * For __GFP_REPEAT allocations, stop reclaiming if the
1755                  * full LRU list has been scanned and we are still failing
1756                  * to reclaim pages. This full LRU scan is potentially
1757                  * expensive but a __GFP_REPEAT caller really wants to succeed
1758                  */
1759                 if (!nr_reclaimed && !nr_scanned)
1760                         return false;
1761         } else {
1762                 /*
1763                  * For non-__GFP_REPEAT allocations which can presumably
1764                  * fail without consequence, stop if we failed to reclaim
1765                  * any pages from the last SWAP_CLUSTER_MAX number of
1766                  * pages that were scanned. This will return to the
1767                  * caller faster at the risk reclaim/compaction and
1768                  * the resulting allocation attempt fails
1769                  */
1770                 if (!nr_reclaimed)
1771                         return false;
1772         }
1773
1774         /*
1775          * If we have not reclaimed enough pages for compaction and the
1776          * inactive lists are large enough, continue reclaiming
1777          */
1778         pages_for_compaction = (2UL << sc->order);
1779         inactive_lru_pages = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1780         if (nr_swap_pages > 0)
1781                 inactive_lru_pages += get_lru_size(lruvec, LRU_INACTIVE_ANON);
1782         if (sc->nr_reclaimed < pages_for_compaction &&
1783                         inactive_lru_pages > pages_for_compaction)
1784                 return true;
1785
1786         /* If compaction would go ahead or the allocation would succeed, stop */
1787         switch (compaction_suitable(lruvec_zone(lruvec), sc->order)) {
1788         case COMPACT_PARTIAL:
1789         case COMPACT_CONTINUE:
1790                 return false;
1791         default:
1792                 return true;
1793         }
1794 }
1795
1796 /*
1797  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1798  */
1799 static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
1800 {
1801         unsigned long nr[NR_LRU_LISTS];
1802         unsigned long nr_to_scan;
1803         enum lru_list lru;
1804         unsigned long nr_reclaimed, nr_scanned;
1805         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1806         struct blk_plug plug;
1807
1808 restart:
1809         nr_reclaimed = 0;
1810         nr_scanned = sc->nr_scanned;
1811         get_scan_count(lruvec, sc, nr);
1812
1813         blk_start_plug(&plug);
1814         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1815                                         nr[LRU_INACTIVE_FILE]) {
1816                 for_each_evictable_lru(lru) {
1817                         if (nr[lru]) {
1818                                 nr_to_scan = min_t(unsigned long,
1819                                                    nr[lru], SWAP_CLUSTER_MAX);
1820                                 nr[lru] -= nr_to_scan;
1821
1822                                 nr_reclaimed += shrink_list(lru, nr_to_scan,
1823                                                             lruvec, sc);
1824                         }
1825                 }
1826                 /*
1827                  * On large memory systems, scan >> priority can become
1828                  * really large. This is fine for the starting priority;
1829                  * we want to put equal scanning pressure on each zone.
1830                  * However, if the VM has a harder time of freeing pages,
1831                  * with multiple processes reclaiming pages, the total
1832                  * freeing target can get unreasonably large.
1833                  */
1834                 if (nr_reclaimed >= nr_to_reclaim &&
1835                     sc->priority < DEF_PRIORITY)
1836                         break;
1837         }
1838         blk_finish_plug(&plug);
1839         sc->nr_reclaimed += nr_reclaimed;
1840
1841         /*
1842          * Even if we did not try to evict anon pages at all, we want to
1843          * rebalance the anon lru active/inactive ratio.
1844          */
1845         if (inactive_anon_is_low(lruvec))
1846                 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
1847                                    sc, LRU_ACTIVE_ANON);
1848
1849         /* reclaim/compaction might need reclaim to continue */
1850         if (should_continue_reclaim(lruvec, nr_reclaimed,
1851                                     sc->nr_scanned - nr_scanned, sc))
1852                 goto restart;
1853
1854         throttle_vm_writeout(sc->gfp_mask);
1855 }
1856
1857 static void shrink_zone(struct zone *zone, struct scan_control *sc)
1858 {
1859         struct mem_cgroup *root = sc->target_mem_cgroup;
1860         struct mem_cgroup_reclaim_cookie reclaim = {
1861                 .zone = zone,
1862                 .priority = sc->priority,
1863         };
1864         struct mem_cgroup *memcg;
1865
1866         memcg = mem_cgroup_iter(root, NULL, &reclaim);
1867         do {
1868                 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
1869
1870                 shrink_lruvec(lruvec, sc);
1871
1872                 /*
1873                  * Limit reclaim has historically picked one memcg and
1874                  * scanned it with decreasing priority levels until
1875                  * nr_to_reclaim had been reclaimed.  This priority
1876                  * cycle is thus over after a single memcg.
1877                  *
1878                  * Direct reclaim and kswapd, on the other hand, have
1879                  * to scan all memory cgroups to fulfill the overall
1880                  * scan target for the zone.
1881                  */
1882                 if (!global_reclaim(sc)) {
1883                         mem_cgroup_iter_break(root, memcg);
1884                         break;
1885                 }
1886                 memcg = mem_cgroup_iter(root, memcg, &reclaim);
1887         } while (memcg);
1888 }
1889
1890 /* Returns true if compaction should go ahead for a high-order request */
1891 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
1892 {
1893         unsigned long balance_gap, watermark;
1894         bool watermark_ok;
1895
1896         /* Do not consider compaction for orders reclaim is meant to satisfy */
1897         if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
1898                 return false;
1899
1900         /*
1901          * Compaction takes time to run and there are potentially other
1902          * callers using the pages just freed. Continue reclaiming until
1903          * there is a buffer of free pages available to give compaction
1904          * a reasonable chance of completing and allocating the page
1905          */
1906         balance_gap = min(low_wmark_pages(zone),
1907                 (zone->present_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
1908                         KSWAPD_ZONE_BALANCE_GAP_RATIO);
1909         watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
1910         watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
1911
1912         /*
1913          * If compaction is deferred, reclaim up to a point where
1914          * compaction will have a chance of success when re-enabled
1915          */
1916         if (compaction_deferred(zone, sc->order))
1917                 return watermark_ok;
1918
1919         /* If compaction is not ready to start, keep reclaiming */
1920         if (!compaction_suitable(zone, sc->order))
1921                 return false;
1922
1923         return watermark_ok;
1924 }
1925
1926 /*
1927  * This is the direct reclaim path, for page-allocating processes.  We only
1928  * try to reclaim pages from zones which will satisfy the caller's allocation
1929  * request.
1930  *
1931  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1932  * Because:
1933  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1934  *    allocation or
1935  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1936  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1937  *    zone defense algorithm.
1938  *
1939  * If a zone is deemed to be full of pinned pages then just give it a light
1940  * scan then give up on it.
1941  *
1942  * This function returns true if a zone is being reclaimed for a costly
1943  * high-order allocation and compaction is ready to begin. This indicates to
1944  * the caller that it should consider retrying the allocation instead of
1945  * further reclaim.
1946  */
1947 static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
1948 {
1949         struct zoneref *z;
1950         struct zone *zone;
1951         unsigned long nr_soft_reclaimed;
1952         unsigned long nr_soft_scanned;
1953         bool aborted_reclaim = false;
1954
1955         /*
1956          * If the number of buffer_heads in the machine exceeds the maximum
1957          * allowed level, force direct reclaim to scan the highmem zone as
1958          * highmem pages could be pinning lowmem pages storing buffer_heads
1959          */
1960         if (buffer_heads_over_limit)
1961                 sc->gfp_mask |= __GFP_HIGHMEM;
1962
1963         for_each_zone_zonelist_nodemask(zone, z, zonelist,
1964                                         gfp_zone(sc->gfp_mask), sc->nodemask) {
1965                 if (!populated_zone(zone))
1966                         continue;
1967                 /*
1968                  * Take care memory controller reclaiming has small influence
1969                  * to global LRU.
1970                  */
1971                 if (global_reclaim(sc)) {
1972                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1973                                 continue;
1974                         if (zone->all_unreclaimable &&
1975                                         sc->priority != DEF_PRIORITY)
1976                                 continue;       /* Let kswapd poll it */
1977                         if (COMPACTION_BUILD) {
1978                                 /*
1979                                  * If we already have plenty of memory free for
1980                                  * compaction in this zone, don't free any more.
1981                                  * Even though compaction is invoked for any
1982                                  * non-zero order, only frequent costly order
1983                                  * reclamation is disruptive enough to become a
1984                                  * noticeable problem, like transparent huge
1985                                  * page allocations.
1986                                  */
1987                                 if (compaction_ready(zone, sc)) {
1988                                         aborted_reclaim = true;
1989                                         continue;
1990                                 }
1991                         }
1992                         /*
1993                          * This steals pages from memory cgroups over softlimit
1994                          * and returns the number of reclaimed pages and
1995                          * scanned pages. This works for global memory pressure
1996                          * and balancing, not for a memcg's limit.
1997                          */
1998                         nr_soft_scanned = 0;
1999                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2000                                                 sc->order, sc->gfp_mask,
2001                                                 &nr_soft_scanned);
2002                         sc->nr_reclaimed += nr_soft_reclaimed;
2003                         sc->nr_scanned += nr_soft_scanned;
2004                         /* need some check for avoid more shrink_zone() */
2005                 }
2006
2007                 shrink_zone(zone, sc);
2008         }
2009
2010         return aborted_reclaim;
2011 }
2012
2013 static bool zone_reclaimable(struct zone *zone)
2014 {
2015         return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2016 }
2017
2018 /* All zones in zonelist are unreclaimable? */
2019 static bool all_unreclaimable(struct zonelist *zonelist,
2020                 struct scan_control *sc)
2021 {
2022         struct zoneref *z;
2023         struct zone *zone;
2024
2025         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2026                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2027                 if (!populated_zone(zone))
2028                         continue;
2029                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2030                         continue;
2031                 if (!zone->all_unreclaimable)
2032                         return false;
2033         }
2034
2035         return true;
2036 }
2037
2038 /*
2039  * This is the main entry point to direct page reclaim.
2040  *
2041  * If a full scan of the inactive list fails to free enough memory then we
2042  * are "out of memory" and something needs to be killed.
2043  *
2044  * If the caller is !__GFP_FS then the probability of a failure is reasonably
2045  * high - the zone may be full of dirty or under-writeback pages, which this
2046  * caller can't do much about.  We kick the writeback threads and take explicit
2047  * naps in the hope that some of these pages can be written.  But if the
2048  * allocating task holds filesystem locks which prevent writeout this might not
2049  * work, and the allocation attempt will fail.
2050  *
2051  * returns:     0, if no pages reclaimed
2052  *              else, the number of pages reclaimed
2053  */
2054 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2055                                         struct scan_control *sc,
2056                                         struct shrink_control *shrink)
2057 {
2058         unsigned long total_scanned = 0;
2059         struct reclaim_state *reclaim_state = current->reclaim_state;
2060         struct zoneref *z;
2061         struct zone *zone;
2062         unsigned long writeback_threshold;
2063         bool aborted_reclaim;
2064
2065         delayacct_freepages_start();
2066
2067         if (global_reclaim(sc))
2068                 count_vm_event(ALLOCSTALL);
2069
2070         do {
2071                 sc->nr_scanned = 0;
2072                 aborted_reclaim = shrink_zones(zonelist, sc);
2073
2074                 /*
2075                  * Don't shrink slabs when reclaiming memory from
2076                  * over limit cgroups
2077                  */
2078                 if (global_reclaim(sc)) {
2079                         unsigned long lru_pages = 0;
2080                         for_each_zone_zonelist(zone, z, zonelist,
2081                                         gfp_zone(sc->gfp_mask)) {
2082                                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2083                                         continue;
2084
2085                                 lru_pages += zone_reclaimable_pages(zone);
2086                         }
2087
2088                         shrink_slab(shrink, sc->nr_scanned, lru_pages);
2089                         if (reclaim_state) {
2090                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2091                                 reclaim_state->reclaimed_slab = 0;
2092                         }
2093                 }
2094                 total_scanned += sc->nr_scanned;
2095                 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2096                         goto out;
2097
2098                 /*
2099                  * Try to write back as many pages as we just scanned.  This
2100                  * tends to cause slow streaming writers to write data to the
2101                  * disk smoothly, at the dirtying rate, which is nice.   But
2102                  * that's undesirable in laptop mode, where we *want* lumpy
2103                  * writeout.  So in laptop mode, write out the whole world.
2104                  */
2105                 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2106                 if (total_scanned > writeback_threshold) {
2107                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2108                                                 WB_REASON_TRY_TO_FREE_PAGES);
2109                         sc->may_writepage = 1;
2110                 }
2111
2112                 /* Take a nap, wait for some writeback to complete */
2113                 if (!sc->hibernation_mode && sc->nr_scanned &&
2114                     sc->priority < DEF_PRIORITY - 2) {
2115                         struct zone *preferred_zone;
2116
2117                         first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2118                                                 &cpuset_current_mems_allowed,
2119                                                 &preferred_zone);
2120                         wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2121                 }
2122         } while (--sc->priority >= 0);
2123
2124 out:
2125         delayacct_freepages_end();
2126
2127         if (sc->nr_reclaimed)
2128                 return sc->nr_reclaimed;
2129
2130         /*
2131          * As hibernation is going on, kswapd is freezed so that it can't mark
2132          * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2133          * check.
2134          */
2135         if (oom_killer_disabled)
2136                 return 0;
2137
2138         /* Aborted reclaim to try compaction? don't OOM, then */
2139         if (aborted_reclaim)
2140                 return 1;
2141
2142         /* top priority shrink_zones still had more to do? don't OOM, then */
2143         if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
2144                 return 1;
2145
2146         return 0;
2147 }
2148
2149 static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2150 {
2151         struct zone *zone;
2152         unsigned long pfmemalloc_reserve = 0;
2153         unsigned long free_pages = 0;
2154         int i;
2155         bool wmark_ok;
2156
2157         for (i = 0; i <= ZONE_NORMAL; i++) {
2158                 zone = &pgdat->node_zones[i];
2159                 pfmemalloc_reserve += min_wmark_pages(zone);
2160                 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2161         }
2162
2163         wmark_ok = free_pages > pfmemalloc_reserve / 2;
2164
2165         /* kswapd must be awake if processes are being throttled */
2166         if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2167                 pgdat->classzone_idx = min(pgdat->classzone_idx,
2168                                                 (enum zone_type)ZONE_NORMAL);
2169                 wake_up_interruptible(&pgdat->kswapd_wait);
2170         }
2171
2172         return wmark_ok;
2173 }
2174
2175 /*
2176  * Throttle direct reclaimers if backing storage is backed by the network
2177  * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2178  * depleted. kswapd will continue to make progress and wake the processes
2179  * when the low watermark is reached
2180  */
2181 static void throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2182                                         nodemask_t *nodemask)
2183 {
2184         struct zone *zone;
2185         int high_zoneidx = gfp_zone(gfp_mask);
2186         pg_data_t *pgdat;
2187
2188         /*
2189          * Kernel threads should not be throttled as they may be indirectly
2190          * responsible for cleaning pages necessary for reclaim to make forward
2191          * progress. kjournald for example may enter direct reclaim while
2192          * committing a transaction where throttling it could forcing other
2193          * processes to block on log_wait_commit().
2194          */
2195         if (current->flags & PF_KTHREAD)
2196                 return;
2197
2198         /* Check if the pfmemalloc reserves are ok */
2199         first_zones_zonelist(zonelist, high_zoneidx, NULL, &zone);
2200         pgdat = zone->zone_pgdat;
2201         if (pfmemalloc_watermark_ok(pgdat))
2202                 return;
2203
2204         /* Account for the throttling */
2205         count_vm_event(PGSCAN_DIRECT_THROTTLE);
2206
2207         /*
2208          * If the caller cannot enter the filesystem, it's possible that it
2209          * is due to the caller holding an FS lock or performing a journal
2210          * transaction in the case of a filesystem like ext[3|4]. In this case,
2211          * it is not safe to block on pfmemalloc_wait as kswapd could be
2212          * blocked waiting on the same lock. Instead, throttle for up to a
2213          * second before continuing.
2214          */
2215         if (!(gfp_mask & __GFP_FS)) {
2216                 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2217                         pfmemalloc_watermark_ok(pgdat), HZ);
2218                 return;
2219         }
2220
2221         /* Throttle until kswapd wakes the process */
2222         wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2223                 pfmemalloc_watermark_ok(pgdat));
2224 }
2225
2226 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2227                                 gfp_t gfp_mask, nodemask_t *nodemask)
2228 {
2229         unsigned long nr_reclaimed;
2230         struct scan_control sc = {
2231                 .gfp_mask = gfp_mask,
2232                 .may_writepage = !laptop_mode,
2233                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2234                 .may_unmap = 1,
2235                 .may_swap = 1,
2236                 .order = order,
2237                 .priority = DEF_PRIORITY,
2238                 .target_mem_cgroup = NULL,
2239                 .nodemask = nodemask,
2240         };
2241         struct shrink_control shrink = {
2242                 .gfp_mask = sc.gfp_mask,
2243         };
2244
2245         throttle_direct_reclaim(gfp_mask, zonelist, nodemask);
2246
2247         /*
2248          * Do not enter reclaim if fatal signal is pending. 1 is returned so
2249          * that the page allocator does not consider triggering OOM
2250          */
2251         if (fatal_signal_pending(current))
2252                 return 1;
2253
2254         trace_mm_vmscan_direct_reclaim_begin(order,
2255                                 sc.may_writepage,
2256                                 gfp_mask);
2257
2258         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2259
2260         trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2261
2262         return nr_reclaimed;
2263 }
2264
2265 #ifdef CONFIG_MEMCG
2266
2267 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
2268                                                 gfp_t gfp_mask, bool noswap,
2269                                                 struct zone *zone,
2270                                                 unsigned long *nr_scanned)
2271 {
2272         struct scan_control sc = {
2273                 .nr_scanned = 0,
2274                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2275                 .may_writepage = !laptop_mode,
2276                 .may_unmap = 1,
2277                 .may_swap = !noswap,
2278                 .order = 0,
2279                 .priority = 0,
2280                 .target_mem_cgroup = memcg,
2281         };
2282         struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2283
2284         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2285                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2286
2287         trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
2288                                                       sc.may_writepage,
2289                                                       sc.gfp_mask);
2290
2291         /*
2292          * NOTE: Although we can get the priority field, using it
2293          * here is not a good idea, since it limits the pages we can scan.
2294          * if we don't reclaim here, the shrink_zone from balance_pgdat
2295          * will pick up pages from other mem cgroup's as well. We hack
2296          * the priority and make it zero.
2297          */
2298         shrink_lruvec(lruvec, &sc);
2299
2300         trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2301
2302         *nr_scanned = sc.nr_scanned;
2303         return sc.nr_reclaimed;
2304 }
2305
2306 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2307                                            gfp_t gfp_mask,
2308                                            bool noswap)
2309 {
2310         struct zonelist *zonelist;
2311         unsigned long nr_reclaimed;
2312         int nid;
2313         struct scan_control sc = {
2314                 .may_writepage = !laptop_mode,
2315                 .may_unmap = 1,
2316                 .may_swap = !noswap,
2317                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2318                 .order = 0,
2319                 .priority = DEF_PRIORITY,
2320                 .target_mem_cgroup = memcg,
2321                 .nodemask = NULL, /* we don't care the placement */
2322                 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2323                                 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2324         };
2325         struct shrink_control shrink = {
2326                 .gfp_mask = sc.gfp_mask,
2327         };
2328
2329         /*
2330          * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2331          * take care of from where we get pages. So the node where we start the
2332          * scan does not need to be the current node.
2333          */
2334         nid = mem_cgroup_select_victim_node(memcg);
2335
2336         zonelist = NODE_DATA(nid)->node_zonelists;
2337
2338         trace_mm_vmscan_memcg_reclaim_begin(0,
2339                                             sc.may_writepage,
2340                                             sc.gfp_mask);
2341
2342         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2343
2344         trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2345
2346         return nr_reclaimed;
2347 }
2348 #endif
2349
2350 static void age_active_anon(struct zone *zone, struct scan_control *sc)
2351 {
2352         struct mem_cgroup *memcg;
2353
2354         if (!total_swap_pages)
2355                 return;
2356
2357         memcg = mem_cgroup_iter(NULL, NULL, NULL);
2358         do {
2359                 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2360
2361                 if (inactive_anon_is_low(lruvec))
2362                         shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2363                                            sc, LRU_ACTIVE_ANON);
2364
2365                 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2366         } while (memcg);
2367 }
2368
2369 /*
2370  * pgdat_balanced is used when checking if a node is balanced for high-order
2371  * allocations. Only zones that meet watermarks and are in a zone allowed
2372  * by the callers classzone_idx are added to balanced_pages. The total of
2373  * balanced pages must be at least 25% of the zones allowed by classzone_idx
2374  * for the node to be considered balanced. Forcing all zones to be balanced
2375  * for high orders can cause excessive reclaim when there are imbalanced zones.
2376  * The choice of 25% is due to
2377  *   o a 16M DMA zone that is balanced will not balance a zone on any
2378  *     reasonable sized machine
2379  *   o On all other machines, the top zone must be at least a reasonable
2380  *     percentage of the middle zones. For example, on 32-bit x86, highmem
2381  *     would need to be at least 256M for it to be balance a whole node.
2382  *     Similarly, on x86-64 the Normal zone would need to be at least 1G
2383  *     to balance a node on its own. These seemed like reasonable ratios.
2384  */
2385 static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2386                                                 int classzone_idx)
2387 {
2388         unsigned long present_pages = 0;
2389         int i;
2390
2391         for (i = 0; i <= classzone_idx; i++)
2392                 present_pages += pgdat->node_zones[i].present_pages;
2393
2394         /* A special case here: if zone has no page, we think it's balanced */
2395         return balanced_pages >= (present_pages >> 2);
2396 }
2397
2398 /*
2399  * Prepare kswapd for sleeping. This verifies that there are no processes
2400  * waiting in throttle_direct_reclaim() and that watermarks have been met.
2401  *
2402  * Returns true if kswapd is ready to sleep
2403  */
2404 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
2405                                         int classzone_idx)
2406 {
2407         int i;
2408         unsigned long balanced = 0;
2409         bool all_zones_ok = true;
2410
2411         /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2412         if (remaining)
2413                 return false;
2414
2415         /*
2416          * There is a potential race between when kswapd checks its watermarks
2417          * and a process gets throttled. There is also a potential race if
2418          * processes get throttled, kswapd wakes, a large process exits therby
2419          * balancing the zones that causes kswapd to miss a wakeup. If kswapd
2420          * is going to sleep, no process should be sleeping on pfmemalloc_wait
2421          * so wake them now if necessary. If necessary, processes will wake
2422          * kswapd and get throttled again
2423          */
2424         if (waitqueue_active(&pgdat->pfmemalloc_wait)) {
2425                 wake_up(&pgdat->pfmemalloc_wait);
2426                 return false;
2427         }
2428
2429         /* Check the watermark levels */
2430         for (i = 0; i <= classzone_idx; i++) {
2431                 struct zone *zone = pgdat->node_zones + i;
2432
2433                 if (!populated_zone(zone))
2434                         continue;
2435
2436                 /*
2437                  * balance_pgdat() skips over all_unreclaimable after
2438                  * DEF_PRIORITY. Effectively, it considers them balanced so
2439                  * they must be considered balanced here as well if kswapd
2440                  * is to sleep
2441                  */
2442                 if (zone->all_unreclaimable) {
2443                         balanced += zone->present_pages;
2444                         continue;
2445                 }
2446
2447                 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2448                                                         i, 0))
2449                         all_zones_ok = false;
2450                 else
2451                         balanced += zone->present_pages;
2452         }
2453
2454         /*
2455          * For high-order requests, the balanced zones must contain at least
2456          * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2457          * must be balanced
2458          */
2459         if (order)
2460                 return pgdat_balanced(pgdat, balanced, classzone_idx);
2461         else
2462                 return all_zones_ok;
2463 }
2464
2465 /*
2466  * For kswapd, balance_pgdat() will work across all this node's zones until
2467  * they are all at high_wmark_pages(zone).
2468  *
2469  * Returns the final order kswapd was reclaiming at
2470  *
2471  * There is special handling here for zones which are full of pinned pages.
2472  * This can happen if the pages are all mlocked, or if they are all used by
2473  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
2474  * What we do is to detect the case where all pages in the zone have been
2475  * scanned twice and there has been zero successful reclaim.  Mark the zone as
2476  * dead and from now on, only perform a short scan.  Basically we're polling
2477  * the zone for when the problem goes away.
2478  *
2479  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
2480  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2481  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2482  * lower zones regardless of the number of free pages in the lower zones. This
2483  * interoperates with the page allocator fallback scheme to ensure that aging
2484  * of pages is balanced across the zones.
2485  */
2486 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2487                                                         int *classzone_idx)
2488 {
2489         int all_zones_ok;
2490         unsigned long balanced;
2491         int i;
2492         int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
2493         unsigned long total_scanned;
2494         struct reclaim_state *reclaim_state = current->reclaim_state;
2495         unsigned long nr_soft_reclaimed;
2496         unsigned long nr_soft_scanned;
2497         struct scan_control sc = {
2498                 .gfp_mask = GFP_KERNEL,
2499                 .may_unmap = 1,
2500                 .may_swap = 1,
2501                 /*
2502                  * kswapd doesn't want to be bailed out while reclaim. because
2503                  * we want to put equal scanning pressure on each zone.
2504                  */
2505                 .nr_to_reclaim = ULONG_MAX,
2506                 .order = order,
2507                 .target_mem_cgroup = NULL,
2508         };
2509         struct shrink_control shrink = {
2510                 .gfp_mask = sc.gfp_mask,
2511         };
2512 loop_again:
2513         total_scanned = 0;
2514         sc.priority = DEF_PRIORITY;
2515         sc.nr_reclaimed = 0;
2516         sc.may_writepage = !laptop_mode;
2517         count_vm_event(PAGEOUTRUN);
2518
2519         do {
2520                 unsigned long lru_pages = 0;
2521                 int has_under_min_watermark_zone = 0;
2522
2523                 all_zones_ok = 1;
2524                 balanced = 0;
2525
2526                 /*
2527                  * Scan in the highmem->dma direction for the highest
2528                  * zone which needs scanning
2529                  */
2530                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2531                         struct zone *zone = pgdat->node_zones + i;
2532
2533                         if (!populated_zone(zone))
2534                                 continue;
2535
2536                         if (zone->all_unreclaimable &&
2537                             sc.priority != DEF_PRIORITY)
2538                                 continue;
2539
2540                         /*
2541                          * Do some background aging of the anon list, to give
2542                          * pages a chance to be referenced before reclaiming.
2543                          */
2544                         age_active_anon(zone, &sc);
2545
2546                         /*
2547                          * If the number of buffer_heads in the machine
2548                          * exceeds the maximum allowed level and this node
2549                          * has a highmem zone, force kswapd to reclaim from
2550                          * it to relieve lowmem pressure.
2551                          */
2552                         if (buffer_heads_over_limit && is_highmem_idx(i)) {
2553                                 end_zone = i;
2554                                 break;
2555                         }
2556
2557                         if (!zone_watermark_ok_safe(zone, order,
2558                                         high_wmark_pages(zone), 0, 0)) {
2559                                 end_zone = i;
2560                                 break;
2561                         } else {
2562                                 /* If balanced, clear the congested flag */
2563                                 zone_clear_flag(zone, ZONE_CONGESTED);
2564                         }
2565                 }
2566                 if (i < 0)
2567                         goto out;
2568
2569                 for (i = 0; i <= end_zone; i++) {
2570                         struct zone *zone = pgdat->node_zones + i;
2571
2572                         lru_pages += zone_reclaimable_pages(zone);
2573                 }
2574
2575                 /*
2576                  * Now scan the zone in the dma->highmem direction, stopping
2577                  * at the last zone which needs scanning.
2578                  *
2579                  * We do this because the page allocator works in the opposite
2580                  * direction.  This prevents the page allocator from allocating
2581                  * pages behind kswapd's direction of progress, which would
2582                  * cause too much scanning of the lower zones.
2583                  */
2584                 for (i = 0; i <= end_zone; i++) {
2585                         struct zone *zone = pgdat->node_zones + i;
2586                         int nr_slab, testorder;
2587                         unsigned long balance_gap;
2588
2589                         if (!populated_zone(zone))
2590                                 continue;
2591
2592                         if (zone->all_unreclaimable &&
2593                             sc.priority != DEF_PRIORITY)
2594                                 continue;
2595
2596                         sc.nr_scanned = 0;
2597
2598                         nr_soft_scanned = 0;
2599                         /*
2600                          * Call soft limit reclaim before calling shrink_zone.
2601                          */
2602                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2603                                                         order, sc.gfp_mask,
2604                                                         &nr_soft_scanned);
2605                         sc.nr_reclaimed += nr_soft_reclaimed;
2606                         total_scanned += nr_soft_scanned;
2607
2608                         /*
2609                          * We put equal pressure on every zone, unless
2610                          * one zone has way too many pages free
2611                          * already. The "too many pages" is defined
2612                          * as the high wmark plus a "gap" where the
2613                          * gap is either the low watermark or 1%
2614                          * of the zone, whichever is smaller.
2615                          */
2616                         balance_gap = min(low_wmark_pages(zone),
2617                                 (zone->present_pages +
2618                                         KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2619                                 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2620                         /*
2621                          * Kswapd reclaims only single pages with compaction
2622                          * enabled. Trying too hard to reclaim until contiguous
2623                          * free pages have become available can hurt performance
2624                          * by evicting too much useful data from memory.
2625                          * Do not reclaim more than needed for compaction.
2626                          */
2627                         testorder = order;
2628                         if (COMPACTION_BUILD && order &&
2629                                         compaction_suitable(zone, order) !=
2630                                                 COMPACT_SKIPPED)
2631                                 testorder = 0;
2632
2633                         if ((buffer_heads_over_limit && is_highmem_idx(i)) ||
2634                                     !zone_watermark_ok_safe(zone, testorder,
2635                                         high_wmark_pages(zone) + balance_gap,
2636                                         end_zone, 0)) {
2637                                 shrink_zone(zone, &sc);
2638
2639                                 reclaim_state->reclaimed_slab = 0;
2640                                 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2641                                 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2642                                 total_scanned += sc.nr_scanned;
2643
2644                                 if (nr_slab == 0 && !zone_reclaimable(zone))
2645                                         zone->all_unreclaimable = 1;
2646                         }
2647
2648                         /*
2649                          * If we've done a decent amount of scanning and
2650                          * the reclaim ratio is low, start doing writepage
2651                          * even in laptop mode
2652                          */
2653                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2654                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2655                                 sc.may_writepage = 1;
2656
2657                         if (zone->all_unreclaimable) {
2658                                 if (end_zone && end_zone == i)
2659                                         end_zone--;
2660                                 continue;
2661                         }
2662
2663                         if (!zone_watermark_ok_safe(zone, testorder,
2664                                         high_wmark_pages(zone), end_zone, 0)) {
2665                                 all_zones_ok = 0;
2666                                 /*
2667                                  * We are still under min water mark.  This
2668                                  * means that we have a GFP_ATOMIC allocation
2669                                  * failure risk. Hurry up!
2670                                  */
2671                                 if (!zone_watermark_ok_safe(zone, order,
2672                                             min_wmark_pages(zone), end_zone, 0))
2673                                         has_under_min_watermark_zone = 1;
2674                         } else {
2675                                 /*
2676                                  * If a zone reaches its high watermark,
2677                                  * consider it to be no longer congested. It's
2678                                  * possible there are dirty pages backed by
2679                                  * congested BDIs but as pressure is relieved,
2680                                  * speculatively avoid congestion waits
2681                                  */
2682                                 zone_clear_flag(zone, ZONE_CONGESTED);
2683                                 if (i <= *classzone_idx)
2684                                         balanced += zone->present_pages;
2685                         }
2686
2687                 }
2688
2689                 /*
2690                  * If the low watermark is met there is no need for processes
2691                  * to be throttled on pfmemalloc_wait as they should not be
2692                  * able to safely make forward progress. Wake them
2693                  */
2694                 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
2695                                 pfmemalloc_watermark_ok(pgdat))
2696                         wake_up(&pgdat->pfmemalloc_wait);
2697
2698                 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2699                         break;          /* kswapd: all done */
2700                 /*
2701                  * OK, kswapd is getting into trouble.  Take a nap, then take
2702                  * another pass across the zones.
2703                  */
2704                 if (total_scanned && (sc.priority < DEF_PRIORITY - 2)) {
2705                         if (has_under_min_watermark_zone)
2706                                 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2707                         else
2708                                 congestion_wait(BLK_RW_ASYNC, HZ/10);
2709                 }
2710
2711                 /*
2712                  * We do this so kswapd doesn't build up large priorities for
2713                  * example when it is freeing in parallel with allocators. It
2714                  * matches the direct reclaim path behaviour in terms of impact
2715                  * on zone->*_priority.
2716                  */
2717                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2718                         break;
2719         } while (--sc.priority >= 0);
2720 out:
2721
2722         /*
2723          * order-0: All zones must meet high watermark for a balanced node
2724          * high-order: Balanced zones must make up at least 25% of the node
2725          *             for the node to be balanced
2726          */
2727         if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2728                 cond_resched();
2729
2730                 try_to_freeze();
2731
2732                 /*
2733                  * Fragmentation may mean that the system cannot be
2734                  * rebalanced for high-order allocations in all zones.
2735                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2736                  * it means the zones have been fully scanned and are still
2737                  * not balanced. For high-order allocations, there is
2738                  * little point trying all over again as kswapd may
2739                  * infinite loop.
2740                  *
2741                  * Instead, recheck all watermarks at order-0 as they
2742                  * are the most important. If watermarks are ok, kswapd will go
2743                  * back to sleep. High-order users can still perform direct
2744                  * reclaim if they wish.
2745                  */
2746                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2747                         order = sc.order = 0;
2748
2749                 goto loop_again;
2750         }
2751
2752         /*
2753          * If kswapd was reclaiming at a higher order, it has the option of
2754          * sleeping without all zones being balanced. Before it does, it must
2755          * ensure that the watermarks for order-0 on *all* zones are met and
2756          * that the congestion flags are cleared. The congestion flag must
2757          * be cleared as kswapd is the only mechanism that clears the flag
2758          * and it is potentially going to sleep here.
2759          */
2760         if (order) {
2761                 int zones_need_compaction = 1;
2762
2763                 for (i = 0; i <= end_zone; i++) {
2764                         struct zone *zone = pgdat->node_zones + i;
2765
2766                         if (!populated_zone(zone))
2767                                 continue;
2768
2769                         if (zone->all_unreclaimable &&
2770                             sc.priority != DEF_PRIORITY)
2771                                 continue;
2772
2773                         /* Would compaction fail due to lack of free memory? */
2774                         if (COMPACTION_BUILD &&
2775                             compaction_suitable(zone, order) == COMPACT_SKIPPED)
2776                                 goto loop_again;
2777
2778                         /* Confirm the zone is balanced for order-0 */
2779                         if (!zone_watermark_ok(zone, 0,
2780                                         high_wmark_pages(zone), 0, 0)) {
2781                                 order = sc.order = 0;
2782                                 goto loop_again;
2783                         }
2784
2785                         /* Check if the memory needs to be defragmented. */
2786                         if (zone_watermark_ok(zone, order,
2787                                     low_wmark_pages(zone), *classzone_idx, 0))
2788                                 zones_need_compaction = 0;
2789
2790                         /* If balanced, clear the congested flag */
2791                         zone_clear_flag(zone, ZONE_CONGESTED);
2792                 }
2793
2794                 if (zones_need_compaction)
2795                         compact_pgdat(pgdat, order);
2796         }
2797
2798         /*
2799          * Return the order we were reclaiming at so prepare_kswapd_sleep()
2800          * makes a decision on the order we were last reclaiming at. However,
2801          * if another caller entered the allocator slow path while kswapd
2802          * was awake, order will remain at the higher level
2803          */
2804         *classzone_idx = end_zone;
2805         return order;
2806 }
2807
2808 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2809 {
2810         long remaining = 0;
2811         DEFINE_WAIT(wait);
2812
2813         if (freezing(current) || kthread_should_stop())
2814                 return;
2815
2816         prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2817
2818         /* Try to sleep for a short interval */
2819         if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
2820                 remaining = schedule_timeout(HZ/10);
2821                 finish_wait(&pgdat->kswapd_wait, &wait);
2822                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2823         }
2824
2825         /*
2826          * After a short sleep, check if it was a premature sleep. If not, then
2827          * go fully to sleep until explicitly woken up.
2828          */
2829         if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
2830                 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2831
2832                 /*
2833                  * vmstat counters are not perfectly accurate and the estimated
2834                  * value for counters such as NR_FREE_PAGES can deviate from the
2835                  * true value by nr_online_cpus * threshold. To avoid the zone
2836                  * watermarks being breached while under pressure, we reduce the
2837                  * per-cpu vmstat threshold while kswapd is awake and restore
2838                  * them before going back to sleep.
2839                  */
2840                 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2841
2842                 if (!kthread_should_stop())
2843                         schedule();
2844
2845                 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2846         } else {
2847                 if (remaining)
2848                         count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2849                 else
2850                         count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2851         }
2852         finish_wait(&pgdat->kswapd_wait, &wait);
2853 }
2854
2855 /*
2856  * The background pageout daemon, started as a kernel thread
2857  * from the init process.
2858  *
2859  * This basically trickles out pages so that we have _some_
2860  * free memory available even if there is no other activity
2861  * that frees anything up. This is needed for things like routing
2862  * etc, where we otherwise might have all activity going on in
2863  * asynchronous contexts that cannot page things out.
2864  *
2865  * If there are applications that are active memory-allocators
2866  * (most normal use), this basically shouldn't matter.
2867  */
2868 static int kswapd(void *p)
2869 {
2870         unsigned long order, new_order;
2871         unsigned balanced_order;
2872         int classzone_idx, new_classzone_idx;
2873         int balanced_classzone_idx;
2874         pg_data_t *pgdat = (pg_data_t*)p;
2875         struct task_struct *tsk = current;
2876
2877         struct reclaim_state reclaim_state = {
2878                 .reclaimed_slab = 0,
2879         };
2880         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2881
2882         lockdep_set_current_reclaim_state(GFP_KERNEL);
2883
2884         if (!cpumask_empty(cpumask))
2885                 set_cpus_allowed_ptr(tsk, cpumask);
2886         current->reclaim_state = &reclaim_state;
2887
2888         /*
2889          * Tell the memory management that we're a "memory allocator",
2890          * and that if we need more memory we should get access to it
2891          * regardless (see "__alloc_pages()"). "kswapd" should
2892          * never get caught in the normal page freeing logic.
2893          *
2894          * (Kswapd normally doesn't need memory anyway, but sometimes
2895          * you need a small amount of memory in order to be able to
2896          * page out something else, and this flag essentially protects
2897          * us from recursively trying to free more memory as we're
2898          * trying to free the first piece of memory in the first place).
2899          */
2900         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2901         set_freezable();
2902
2903         order = new_order = 0;
2904         balanced_order = 0;
2905         classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
2906         balanced_classzone_idx = classzone_idx;
2907         for ( ; ; ) {
2908                 int ret;
2909
2910                 /*
2911                  * If the last balance_pgdat was unsuccessful it's unlikely a
2912                  * new request of a similar or harder type will succeed soon
2913                  * so consider going to sleep on the basis we reclaimed at
2914                  */
2915                 if (balanced_classzone_idx >= new_classzone_idx &&
2916                                         balanced_order == new_order) {
2917                         new_order = pgdat->kswapd_max_order;
2918                         new_classzone_idx = pgdat->classzone_idx;
2919                         pgdat->kswapd_max_order =  0;
2920                         pgdat->classzone_idx = pgdat->nr_zones - 1;
2921                 }
2922
2923                 if (order < new_order || classzone_idx > new_classzone_idx) {
2924                         /*
2925                          * Don't sleep if someone wants a larger 'order'
2926                          * allocation or has tigher zone constraints
2927                          */
2928                         order = new_order;
2929                         classzone_idx = new_classzone_idx;
2930                 } else {
2931                         kswapd_try_to_sleep(pgdat, balanced_order,
2932                                                 balanced_classzone_idx);
2933                         order = pgdat->kswapd_max_order;
2934                         classzone_idx = pgdat->classzone_idx;
2935                         new_order = order;
2936                         new_classzone_idx = classzone_idx;
2937                         pgdat->kswapd_max_order = 0;
2938                         pgdat->classzone_idx = pgdat->nr_zones - 1;
2939                 }
2940
2941                 ret = try_to_freeze();
2942                 if (kthread_should_stop())
2943                         break;
2944
2945                 /*
2946                  * We can speed up thawing tasks if we don't call balance_pgdat
2947                  * after returning from the refrigerator
2948                  */
2949                 if (!ret) {
2950                         trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2951                         balanced_classzone_idx = classzone_idx;
2952                         balanced_order = balance_pgdat(pgdat, order,
2953                                                 &balanced_classzone_idx);
2954                 }
2955         }
2956         return 0;
2957 }
2958
2959 /*
2960  * A zone is low on free memory, so wake its kswapd task to service it.
2961  */
2962 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
2963 {
2964         pg_data_t *pgdat;
2965
2966         if (!populated_zone(zone))
2967                 return;
2968
2969         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2970                 return;
2971         pgdat = zone->zone_pgdat;
2972         if (pgdat->kswapd_max_order < order) {
2973                 pgdat->kswapd_max_order = order;
2974                 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
2975         }
2976         if (!waitqueue_active(&pgdat->kswapd_wait))
2977                 return;
2978         if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
2979                 return;
2980
2981         trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
2982         wake_up_interruptible(&pgdat->kswapd_wait);
2983 }
2984
2985 /*
2986  * The reclaimable count would be mostly accurate.
2987  * The less reclaimable pages may be
2988  * - mlocked pages, which will be moved to unevictable list when encountered
2989  * - mapped pages, which may require several travels to be reclaimed
2990  * - dirty pages, which is not "instantly" reclaimable
2991  */
2992 unsigned long global_reclaimable_pages(void)
2993 {
2994         int nr;
2995
2996         nr = global_page_state(NR_ACTIVE_FILE) +
2997              global_page_state(NR_INACTIVE_FILE);
2998
2999         if (nr_swap_pages > 0)
3000                 nr += global_page_state(NR_ACTIVE_ANON) +
3001                       global_page_state(NR_INACTIVE_ANON);
3002
3003         return nr;
3004 }
3005
3006 unsigned long zone_reclaimable_pages(struct zone *zone)
3007 {
3008         int nr;
3009
3010         nr = zone_page_state(zone, NR_ACTIVE_FILE) +
3011              zone_page_state(zone, NR_INACTIVE_FILE);
3012
3013         if (nr_swap_pages > 0)
3014                 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
3015                       zone_page_state(zone, NR_INACTIVE_ANON);
3016
3017         return nr;
3018 }
3019
3020 #ifdef CONFIG_HIBERNATION
3021 /*
3022  * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3023  * freed pages.
3024  *
3025  * Rather than trying to age LRUs the aim is to preserve the overall
3026  * LRU order by reclaiming preferentially
3027  * inactive > active > active referenced > active mapped
3028  */
3029 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3030 {
3031         struct reclaim_state reclaim_state;
3032         struct scan_control sc = {
3033                 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3034                 .may_swap = 1,
3035                 .may_unmap = 1,
3036                 .may_writepage = 1,
3037                 .nr_to_reclaim = nr_to_reclaim,
3038                 .hibernation_mode = 1,
3039                 .order = 0,
3040                 .priority = DEF_PRIORITY,
3041         };
3042         struct shrink_control shrink = {
3043                 .gfp_mask = sc.gfp_mask,
3044         };
3045         struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3046         struct task_struct *p = current;
3047         unsigned long nr_reclaimed;
3048
3049         p->flags |= PF_MEMALLOC;
3050         lockdep_set_current_reclaim_state(sc.gfp_mask);
3051         reclaim_state.reclaimed_slab = 0;
3052         p->reclaim_state = &reclaim_state;
3053
3054         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
3055
3056         p->reclaim_state = NULL;
3057         lockdep_clear_current_reclaim_state();
3058         p->flags &= ~PF_MEMALLOC;
3059
3060         return nr_reclaimed;
3061 }
3062 #endif /* CONFIG_HIBERNATION */
3063
3064 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3065    not required for correctness.  So if the last cpu in a node goes
3066    away, we get changed to run anywhere: as the first one comes back,
3067    restore their cpu bindings. */
3068 static int __devinit cpu_callback(struct notifier_block *nfb,
3069                                   unsigned long action, void *hcpu)
3070 {
3071         int nid;
3072
3073         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3074                 for_each_node_state(nid, N_HIGH_MEMORY) {
3075                         pg_data_t *pgdat = NODE_DATA(nid);
3076                         const struct cpumask *mask;
3077
3078                         mask = cpumask_of_node(pgdat->node_id);
3079
3080                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3081                                 /* One of our CPUs online: restore mask */
3082                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3083                 }
3084         }
3085         return NOTIFY_OK;
3086 }
3087
3088 /*
3089  * This kswapd start function will be called by init and node-hot-add.
3090  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3091  */
3092 int kswapd_run(int nid)
3093 {
3094         pg_data_t *pgdat = NODE_DATA(nid);
3095         int ret = 0;
3096
3097         if (pgdat->kswapd)
3098                 return 0;
3099
3100         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3101         if (IS_ERR(pgdat->kswapd)) {
3102                 /* failure at boot is fatal */
3103                 BUG_ON(system_state == SYSTEM_BOOTING);
3104                 printk("Failed to start kswapd on node %d\n",nid);
3105                 pgdat->kswapd = NULL;
3106                 ret = -1;
3107         }
3108         return ret;
3109 }
3110
3111 /*
3112  * Called by memory hotplug when all memory in a node is offlined.  Caller must
3113  * hold lock_memory_hotplug().
3114  */
3115 void kswapd_stop(int nid)
3116 {
3117         struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3118
3119         if (kswapd) {
3120                 kthread_stop(kswapd);
3121                 NODE_DATA(nid)->kswapd = NULL;
3122         }
3123 }
3124
3125 static int __init kswapd_init(void)
3126 {
3127         int nid;
3128
3129         swap_setup();
3130         for_each_node_state(nid, N_HIGH_MEMORY)
3131                 kswapd_run(nid);
3132         hotcpu_notifier(cpu_callback, 0);
3133         return 0;
3134 }
3135
3136 module_init(kswapd_init)
3137
3138 #ifdef CONFIG_NUMA
3139 /*
3140  * Zone reclaim mode
3141  *
3142  * If non-zero call zone_reclaim when the number of free pages falls below
3143  * the watermarks.
3144  */
3145 int zone_reclaim_mode __read_mostly;
3146
3147 #define RECLAIM_OFF 0
3148 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
3149 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
3150 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
3151
3152 /*
3153  * Priority for ZONE_RECLAIM. This determines the fraction of pages
3154  * of a node considered for each zone_reclaim. 4 scans 1/16th of
3155  * a zone.
3156  */
3157 #define ZONE_RECLAIM_PRIORITY 4
3158
3159 /*
3160  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3161  * occur.
3162  */
3163 int sysctl_min_unmapped_ratio = 1;
3164
3165 /*
3166  * If the number of slab pages in a zone grows beyond this percentage then
3167  * slab reclaim needs to occur.
3168  */
3169 int sysctl_min_slab_ratio = 5;
3170
3171 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3172 {
3173         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3174         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3175                 zone_page_state(zone, NR_ACTIVE_FILE);
3176
3177         /*
3178          * It's possible for there to be more file mapped pages than
3179          * accounted for by the pages on the file LRU lists because
3180          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3181          */
3182         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3183 }
3184
3185 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3186 static long zone_pagecache_reclaimable(struct zone *zone)
3187 {
3188         long nr_pagecache_reclaimable;
3189         long delta = 0;
3190
3191         /*
3192          * If RECLAIM_SWAP is set, then all file pages are considered
3193          * potentially reclaimable. Otherwise, we have to worry about
3194          * pages like swapcache and zone_unmapped_file_pages() provides
3195          * a better estimate
3196          */
3197         if (zone_reclaim_mode & RECLAIM_SWAP)
3198                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3199         else
3200                 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3201
3202         /* If we can't clean pages, remove dirty pages from consideration */
3203         if (!(zone_reclaim_mode & RECLAIM_WRITE))
3204                 delta += zone_page_state(zone, NR_FILE_DIRTY);
3205
3206         /* Watch for any possible underflows due to delta */
3207         if (unlikely(delta > nr_pagecache_reclaimable))
3208                 delta = nr_pagecache_reclaimable;
3209
3210         return nr_pagecache_reclaimable - delta;
3211 }
3212
3213 /*
3214  * Try to free up some pages from this zone through reclaim.
3215  */
3216 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3217 {
3218         /* Minimum pages needed in order to stay on node */
3219         const unsigned long nr_pages = 1 << order;
3220         struct task_struct *p = current;
3221         struct reclaim_state reclaim_state;
3222         struct scan_control sc = {
3223                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3224                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3225                 .may_swap = 1,
3226                 .nr_to_reclaim = max_t(unsigned long, nr_pages,
3227                                        SWAP_CLUSTER_MAX),
3228                 .gfp_mask = gfp_mask,
3229                 .order = order,
3230                 .priority = ZONE_RECLAIM_PRIORITY,
3231         };
3232         struct shrink_control shrink = {
3233                 .gfp_mask = sc.gfp_mask,
3234         };
3235         unsigned long nr_slab_pages0, nr_slab_pages1;
3236
3237         cond_resched();
3238         /*
3239          * We need to be able to allocate from the reserves for RECLAIM_SWAP
3240          * and we also need to be able to write out pages for RECLAIM_WRITE
3241          * and RECLAIM_SWAP.
3242          */
3243         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3244         lockdep_set_current_reclaim_state(gfp_mask);
3245         reclaim_state.reclaimed_slab = 0;
3246         p->reclaim_state = &reclaim_state;
3247
3248         if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3249                 /*
3250                  * Free memory by calling shrink zone with increasing
3251                  * priorities until we have enough memory freed.
3252                  */
3253                 do {
3254                         shrink_zone(zone, &sc);
3255                 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3256         }
3257
3258         nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3259         if (nr_slab_pages0 > zone->min_slab_pages) {
3260                 /*
3261                  * shrink_slab() does not currently allow us to determine how
3262                  * many pages were freed in this zone. So we take the current
3263                  * number of slab pages and shake the slab until it is reduced
3264                  * by the same nr_pages that we used for reclaiming unmapped
3265                  * pages.
3266                  *
3267                  * Note that shrink_slab will free memory on all zones and may
3268                  * take a long time.
3269                  */
3270                 for (;;) {
3271                         unsigned long lru_pages = zone_reclaimable_pages(zone);
3272
3273                         /* No reclaimable slab or very low memory pressure */
3274                         if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3275                                 break;
3276
3277                         /* Freed enough memory */
3278                         nr_slab_pages1 = zone_page_state(zone,
3279                                                         NR_SLAB_RECLAIMABLE);
3280                         if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3281                                 break;
3282                 }
3283
3284                 /*
3285                  * Update nr_reclaimed by the number of slab pages we
3286                  * reclaimed from this zone.
3287                  */
3288                 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3289                 if (nr_slab_pages1 < nr_slab_pages0)
3290                         sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3291         }
3292
3293         p->reclaim_state = NULL;
3294         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3295         lockdep_clear_current_reclaim_state();
3296         return sc.nr_reclaimed >= nr_pages;
3297 }
3298
3299 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3300 {
3301         int node_id;
3302         int ret;
3303
3304         /*
3305          * Zone reclaim reclaims unmapped file backed pages and
3306          * slab pages if we are over the defined limits.
3307          *
3308          * A small portion of unmapped file backed pages is needed for
3309          * file I/O otherwise pages read by file I/O will be immediately
3310          * thrown out if the zone is overallocated. So we do not reclaim
3311          * if less than a specified percentage of the zone is used by
3312          * unmapped file backed pages.
3313          */
3314         if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3315             zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3316                 return ZONE_RECLAIM_FULL;
3317
3318         if (zone->all_unreclaimable)
3319                 return ZONE_RECLAIM_FULL;
3320
3321         /*
3322          * Do not scan if the allocation should not be delayed.
3323          */
3324         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3325                 return ZONE_RECLAIM_NOSCAN;
3326
3327         /*
3328          * Only run zone reclaim on the local zone or on zones that do not
3329          * have associated processors. This will favor the local processor
3330          * over remote processors and spread off node memory allocations
3331          * as wide as possible.
3332          */
3333         node_id = zone_to_nid(zone);
3334         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3335                 return ZONE_RECLAIM_NOSCAN;
3336
3337         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3338                 return ZONE_RECLAIM_NOSCAN;
3339
3340         ret = __zone_reclaim(zone, gfp_mask, order);
3341         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3342
3343         if (!ret)
3344                 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3345
3346         return ret;
3347 }
3348 #endif
3349
3350 /*
3351  * page_evictable - test whether a page is evictable
3352  * @page: the page to test
3353  * @vma: the VMA in which the page is or will be mapped, may be NULL
3354  *
3355  * Test whether page is evictable--i.e., should be placed on active/inactive
3356  * lists vs unevictable list.  The vma argument is !NULL when called from the
3357  * fault path to determine how to instantate a new page.
3358  *
3359  * Reasons page might not be evictable:
3360  * (1) page's mapping marked unevictable
3361  * (2) page is part of an mlocked VMA
3362  *
3363  */
3364 int page_evictable(struct page *page, struct vm_area_struct *vma)
3365 {
3366
3367         if (mapping_unevictable(page_mapping(page)))
3368                 return 0;
3369
3370         if (PageMlocked(page) || (vma && mlocked_vma_newpage(vma, page)))
3371                 return 0;
3372
3373         return 1;
3374 }
3375
3376 #ifdef CONFIG_SHMEM
3377 /**
3378  * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3379  * @pages:      array of pages to check
3380  * @nr_pages:   number of pages to check
3381  *
3382  * Checks pages for evictability and moves them to the appropriate lru list.
3383  *
3384  * This function is only used for SysV IPC SHM_UNLOCK.
3385  */
3386 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3387 {
3388         struct lruvec *lruvec;
3389         struct zone *zone = NULL;
3390         int pgscanned = 0;
3391         int pgrescued = 0;
3392         int i;
3393
3394         for (i = 0; i < nr_pages; i++) {
3395                 struct page *page = pages[i];
3396                 struct zone *pagezone;
3397
3398                 pgscanned++;
3399                 pagezone = page_zone(page);
3400                 if (pagezone != zone) {
3401                         if (zone)
3402                                 spin_unlock_irq(&zone->lru_lock);
3403                         zone = pagezone;
3404                         spin_lock_irq(&zone->lru_lock);
3405                 }
3406                 lruvec = mem_cgroup_page_lruvec(page, zone);
3407
3408                 if (!PageLRU(page) || !PageUnevictable(page))
3409                         continue;
3410
3411                 if (page_evictable(page, NULL)) {
3412                         enum lru_list lru = page_lru_base_type(page);
3413
3414                         VM_BUG_ON(PageActive(page));
3415                         ClearPageUnevictable(page);
3416                         del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3417                         add_page_to_lru_list(page, lruvec, lru);
3418                         pgrescued++;
3419                 }
3420         }
3421
3422         if (zone) {
3423                 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3424                 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3425                 spin_unlock_irq(&zone->lru_lock);
3426         }
3427 }
3428 #endif /* CONFIG_SHMEM */
3429
3430 static void warn_scan_unevictable_pages(void)
3431 {
3432         printk_once(KERN_WARNING
3433                     "%s: The scan_unevictable_pages sysctl/node-interface has been "
3434                     "disabled for lack of a legitimate use case.  If you have "
3435                     "one, please send an email to linux-mm@kvack.org.\n",
3436                     current->comm);
3437 }
3438
3439 /*
3440  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
3441  * all nodes' unevictable lists for evictable pages
3442  */
3443 unsigned long scan_unevictable_pages;
3444
3445 int scan_unevictable_handler(struct ctl_table *table, int write,
3446                            void __user *buffer,
3447                            size_t *length, loff_t *ppos)
3448 {
3449         warn_scan_unevictable_pages();
3450         proc_doulongvec_minmax(table, write, buffer, length, ppos);
3451         scan_unevictable_pages = 0;
3452         return 0;
3453 }
3454
3455 #ifdef CONFIG_NUMA
3456 /*
3457  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
3458  * a specified node's per zone unevictable lists for evictable pages.
3459  */
3460
3461 static ssize_t read_scan_unevictable_node(struct device *dev,
3462                                           struct device_attribute *attr,
3463                                           char *buf)
3464 {
3465         warn_scan_unevictable_pages();
3466         return sprintf(buf, "0\n");     /* always zero; should fit... */
3467 }
3468
3469 static ssize_t write_scan_unevictable_node(struct device *dev,
3470                                            struct device_attribute *attr,
3471                                         const char *buf, size_t count)
3472 {
3473         warn_scan_unevictable_pages();
3474         return 1;
3475 }
3476
3477
3478 static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3479                         read_scan_unevictable_node,
3480                         write_scan_unevictable_node);
3481
3482 int scan_unevictable_register_node(struct node *node)
3483 {
3484         return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
3485 }
3486
3487 void scan_unevictable_unregister_node(struct node *node)
3488 {
3489         device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);
3490 }
3491 #endif