mm: vmscan: fix misused nr_reclaimed in shrink_mem_cgroup_zone()
[linux-2.6.git] / mm / vmscan.c
1 /*
2  *  linux/mm/vmscan.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  *
6  *  Swap reorganised 29.12.95, Stephen Tweedie.
7  *  kswapd added: 7.1.96  sct
8  *  Removed kswapd_ctl limits, and swap out as many pages as needed
9  *  to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10  *  Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11  *  Multiqueue VM started 5.8.00, Rik van Riel.
12  */
13
14 #include <linux/mm.h>
15 #include <linux/module.h>
16 #include <linux/gfp.h>
17 #include <linux/kernel_stat.h>
18 #include <linux/swap.h>
19 #include <linux/pagemap.h>
20 #include <linux/init.h>
21 #include <linux/highmem.h>
22 #include <linux/vmstat.h>
23 #include <linux/file.h>
24 #include <linux/writeback.h>
25 #include <linux/blkdev.h>
26 #include <linux/buffer_head.h>  /* for try_to_release_page(),
27                                         buffer_heads_over_limit */
28 #include <linux/mm_inline.h>
29 #include <linux/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 /*
57  * reclaim_mode determines how the inactive list is shrunk
58  * RECLAIM_MODE_SINGLE: Reclaim only order-0 pages
59  * RECLAIM_MODE_ASYNC:  Do not block
60  * RECLAIM_MODE_SYNC:   Allow blocking e.g. call wait_on_page_writeback
61  * RECLAIM_MODE_LUMPYRECLAIM: For high-order allocations, take a reference
62  *                      page from the LRU and reclaim all pages within a
63  *                      naturally aligned range
64  * RECLAIM_MODE_COMPACTION: For high-order allocations, reclaim a number of
65  *                      order-0 pages and then compact the zone
66  */
67 typedef unsigned __bitwise__ reclaim_mode_t;
68 #define RECLAIM_MODE_SINGLE             ((__force reclaim_mode_t)0x01u)
69 #define RECLAIM_MODE_ASYNC              ((__force reclaim_mode_t)0x02u)
70 #define RECLAIM_MODE_SYNC               ((__force reclaim_mode_t)0x04u)
71 #define RECLAIM_MODE_LUMPYRECLAIM       ((__force reclaim_mode_t)0x08u)
72 #define RECLAIM_MODE_COMPACTION         ((__force reclaim_mode_t)0x10u)
73
74 struct scan_control {
75         /* Incremented by the number of inactive pages that were scanned */
76         unsigned long nr_scanned;
77
78         /* Number of pages freed so far during a call to shrink_zones() */
79         unsigned long nr_reclaimed;
80
81         /* How many pages shrink_list() should reclaim */
82         unsigned long nr_to_reclaim;
83
84         unsigned long hibernation_mode;
85
86         /* This context's GFP mask */
87         gfp_t gfp_mask;
88
89         int may_writepage;
90
91         /* Can mapped pages be reclaimed? */
92         int may_unmap;
93
94         /* Can pages be swapped as part of reclaim? */
95         int may_swap;
96
97         int order;
98
99         /*
100          * Intend to reclaim enough continuous memory rather than reclaim
101          * enough amount of memory. i.e, mode for high order allocation.
102          */
103         reclaim_mode_t reclaim_mode;
104
105         /*
106          * The memory cgroup that hit its limit and as a result is the
107          * primary target of this reclaim invocation.
108          */
109         struct mem_cgroup *target_mem_cgroup;
110
111         /*
112          * Nodemask of nodes allowed by the caller. If NULL, all nodes
113          * are scanned.
114          */
115         nodemask_t      *nodemask;
116 };
117
118 struct mem_cgroup_zone {
119         struct mem_cgroup *mem_cgroup;
120         struct zone *zone;
121 };
122
123 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
124
125 #ifdef ARCH_HAS_PREFETCH
126 #define prefetch_prev_lru_page(_page, _base, _field)                    \
127         do {                                                            \
128                 if ((_page)->lru.prev != _base) {                       \
129                         struct page *prev;                              \
130                                                                         \
131                         prev = lru_to_page(&(_page->lru));              \
132                         prefetch(&prev->_field);                        \
133                 }                                                       \
134         } while (0)
135 #else
136 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
137 #endif
138
139 #ifdef ARCH_HAS_PREFETCHW
140 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
141         do {                                                            \
142                 if ((_page)->lru.prev != _base) {                       \
143                         struct page *prev;                              \
144                                                                         \
145                         prev = lru_to_page(&(_page->lru));              \
146                         prefetchw(&prev->_field);                       \
147                 }                                                       \
148         } while (0)
149 #else
150 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
151 #endif
152
153 /*
154  * From 0 .. 100.  Higher means more swappy.
155  */
156 int vm_swappiness = 60;
157 long vm_total_pages;    /* The total number of pages which the VM controls */
158
159 static LIST_HEAD(shrinker_list);
160 static DECLARE_RWSEM(shrinker_rwsem);
161
162 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
163 static bool global_reclaim(struct scan_control *sc)
164 {
165         return !sc->target_mem_cgroup;
166 }
167
168 static bool scanning_global_lru(struct mem_cgroup_zone *mz)
169 {
170         return !mz->mem_cgroup;
171 }
172 #else
173 static bool global_reclaim(struct scan_control *sc)
174 {
175         return true;
176 }
177
178 static bool scanning_global_lru(struct mem_cgroup_zone *mz)
179 {
180         return true;
181 }
182 #endif
183
184 static struct zone_reclaim_stat *get_reclaim_stat(struct mem_cgroup_zone *mz)
185 {
186         if (!scanning_global_lru(mz))
187                 return mem_cgroup_get_reclaim_stat(mz->mem_cgroup, mz->zone);
188
189         return &mz->zone->reclaim_stat;
190 }
191
192 static unsigned long zone_nr_lru_pages(struct mem_cgroup_zone *mz,
193                                        enum lru_list lru)
194 {
195         if (!scanning_global_lru(mz))
196                 return mem_cgroup_zone_nr_lru_pages(mz->mem_cgroup,
197                                                     zone_to_nid(mz->zone),
198                                                     zone_idx(mz->zone),
199                                                     BIT(lru));
200
201         return zone_page_state(mz->zone, NR_LRU_BASE + lru);
202 }
203
204
205 /*
206  * Add a shrinker callback to be called from the vm
207  */
208 void register_shrinker(struct shrinker *shrinker)
209 {
210         atomic_long_set(&shrinker->nr_in_batch, 0);
211         down_write(&shrinker_rwsem);
212         list_add_tail(&shrinker->list, &shrinker_list);
213         up_write(&shrinker_rwsem);
214 }
215 EXPORT_SYMBOL(register_shrinker);
216
217 /*
218  * Remove one
219  */
220 void unregister_shrinker(struct shrinker *shrinker)
221 {
222         down_write(&shrinker_rwsem);
223         list_del(&shrinker->list);
224         up_write(&shrinker_rwsem);
225 }
226 EXPORT_SYMBOL(unregister_shrinker);
227
228 static inline int do_shrinker_shrink(struct shrinker *shrinker,
229                                      struct shrink_control *sc,
230                                      unsigned long nr_to_scan)
231 {
232         sc->nr_to_scan = nr_to_scan;
233         return (*shrinker->shrink)(shrinker, sc);
234 }
235
236 #define SHRINK_BATCH 128
237 /*
238  * Call the shrink functions to age shrinkable caches
239  *
240  * Here we assume it costs one seek to replace a lru page and that it also
241  * takes a seek to recreate a cache object.  With this in mind we age equal
242  * percentages of the lru and ageable caches.  This should balance the seeks
243  * generated by these structures.
244  *
245  * If the vm encountered mapped pages on the LRU it increase the pressure on
246  * slab to avoid swapping.
247  *
248  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
249  *
250  * `lru_pages' represents the number of on-LRU pages in all the zones which
251  * are eligible for the caller's allocation attempt.  It is used for balancing
252  * slab reclaim versus page reclaim.
253  *
254  * Returns the number of slab objects which we shrunk.
255  */
256 unsigned long shrink_slab(struct shrink_control *shrink,
257                           unsigned long nr_pages_scanned,
258                           unsigned long lru_pages)
259 {
260         struct shrinker *shrinker;
261         unsigned long ret = 0;
262
263         if (nr_pages_scanned == 0)
264                 nr_pages_scanned = SWAP_CLUSTER_MAX;
265
266         if (!down_read_trylock(&shrinker_rwsem)) {
267                 /* Assume we'll be able to shrink next time */
268                 ret = 1;
269                 goto out;
270         }
271
272         list_for_each_entry(shrinker, &shrinker_list, list) {
273                 unsigned long long delta;
274                 long total_scan;
275                 long max_pass;
276                 int shrink_ret = 0;
277                 long nr;
278                 long new_nr;
279                 long batch_size = shrinker->batch ? shrinker->batch
280                                                   : SHRINK_BATCH;
281
282                 max_pass = do_shrinker_shrink(shrinker, shrink, 0);
283                 if (max_pass <= 0)
284                         continue;
285
286                 /*
287                  * copy the current shrinker scan count into a local variable
288                  * and zero it so that other concurrent shrinker invocations
289                  * don't also do this scanning work.
290                  */
291                 nr = atomic_long_xchg(&shrinker->nr_in_batch, 0);
292
293                 total_scan = nr;
294                 delta = (4 * nr_pages_scanned) / shrinker->seeks;
295                 delta *= max_pass;
296                 do_div(delta, lru_pages + 1);
297                 total_scan += delta;
298                 if (total_scan < 0) {
299                         printk(KERN_ERR "shrink_slab: %pF negative objects to "
300                                "delete nr=%ld\n",
301                                shrinker->shrink, total_scan);
302                         total_scan = max_pass;
303                 }
304
305                 /*
306                  * We need to avoid excessive windup on filesystem shrinkers
307                  * due to large numbers of GFP_NOFS allocations causing the
308                  * shrinkers to return -1 all the time. This results in a large
309                  * nr being built up so when a shrink that can do some work
310                  * comes along it empties the entire cache due to nr >>>
311                  * max_pass.  This is bad for sustaining a working set in
312                  * memory.
313                  *
314                  * Hence only allow the shrinker to scan the entire cache when
315                  * a large delta change is calculated directly.
316                  */
317                 if (delta < max_pass / 4)
318                         total_scan = min(total_scan, max_pass / 2);
319
320                 /*
321                  * Avoid risking looping forever due to too large nr value:
322                  * never try to free more than twice the estimate number of
323                  * freeable entries.
324                  */
325                 if (total_scan > max_pass * 2)
326                         total_scan = max_pass * 2;
327
328                 trace_mm_shrink_slab_start(shrinker, shrink, nr,
329                                         nr_pages_scanned, lru_pages,
330                                         max_pass, delta, total_scan);
331
332                 while (total_scan >= batch_size) {
333                         int nr_before;
334
335                         nr_before = do_shrinker_shrink(shrinker, shrink, 0);
336                         shrink_ret = do_shrinker_shrink(shrinker, shrink,
337                                                         batch_size);
338                         if (shrink_ret == -1)
339                                 break;
340                         if (shrink_ret < nr_before)
341                                 ret += nr_before - shrink_ret;
342                         count_vm_events(SLABS_SCANNED, batch_size);
343                         total_scan -= batch_size;
344
345                         cond_resched();
346                 }
347
348                 /*
349                  * move the unused scan count back into the shrinker in a
350                  * manner that handles concurrent updates. If we exhausted the
351                  * scan, there is no need to do an update.
352                  */
353                 if (total_scan > 0)
354                         new_nr = atomic_long_add_return(total_scan,
355                                         &shrinker->nr_in_batch);
356                 else
357                         new_nr = atomic_long_read(&shrinker->nr_in_batch);
358
359                 trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr);
360         }
361         up_read(&shrinker_rwsem);
362 out:
363         cond_resched();
364         return ret;
365 }
366
367 static void set_reclaim_mode(int priority, struct scan_control *sc,
368                                    bool sync)
369 {
370         reclaim_mode_t syncmode = sync ? RECLAIM_MODE_SYNC : RECLAIM_MODE_ASYNC;
371
372         /*
373          * Initially assume we are entering either lumpy reclaim or
374          * reclaim/compaction.Depending on the order, we will either set the
375          * sync mode or just reclaim order-0 pages later.
376          */
377         if (COMPACTION_BUILD)
378                 sc->reclaim_mode = RECLAIM_MODE_COMPACTION;
379         else
380                 sc->reclaim_mode = RECLAIM_MODE_LUMPYRECLAIM;
381
382         /*
383          * Avoid using lumpy reclaim or reclaim/compaction if possible by
384          * restricting when its set to either costly allocations or when
385          * under memory pressure
386          */
387         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
388                 sc->reclaim_mode |= syncmode;
389         else if (sc->order && priority < DEF_PRIORITY - 2)
390                 sc->reclaim_mode |= syncmode;
391         else
392                 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
393 }
394
395 static void reset_reclaim_mode(struct scan_control *sc)
396 {
397         sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
398 }
399
400 static inline int is_page_cache_freeable(struct page *page)
401 {
402         /*
403          * A freeable page cache page is referenced only by the caller
404          * that isolated the page, the page cache radix tree and
405          * optional buffer heads at page->private.
406          */
407         return page_count(page) - page_has_private(page) == 2;
408 }
409
410 static int may_write_to_queue(struct backing_dev_info *bdi,
411                               struct scan_control *sc)
412 {
413         if (current->flags & PF_SWAPWRITE)
414                 return 1;
415         if (!bdi_write_congested(bdi))
416                 return 1;
417         if (bdi == current->backing_dev_info)
418                 return 1;
419
420         /* lumpy reclaim for hugepage often need a lot of write */
421         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
422                 return 1;
423         return 0;
424 }
425
426 /*
427  * We detected a synchronous write error writing a page out.  Probably
428  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
429  * fsync(), msync() or close().
430  *
431  * The tricky part is that after writepage we cannot touch the mapping: nothing
432  * prevents it from being freed up.  But we have a ref on the page and once
433  * that page is locked, the mapping is pinned.
434  *
435  * We're allowed to run sleeping lock_page() here because we know the caller has
436  * __GFP_FS.
437  */
438 static void handle_write_error(struct address_space *mapping,
439                                 struct page *page, int error)
440 {
441         lock_page(page);
442         if (page_mapping(page) == mapping)
443                 mapping_set_error(mapping, error);
444         unlock_page(page);
445 }
446
447 /* possible outcome of pageout() */
448 typedef enum {
449         /* failed to write page out, page is locked */
450         PAGE_KEEP,
451         /* move page to the active list, page is locked */
452         PAGE_ACTIVATE,
453         /* page has been sent to the disk successfully, page is unlocked */
454         PAGE_SUCCESS,
455         /* page is clean and locked */
456         PAGE_CLEAN,
457 } pageout_t;
458
459 /*
460  * pageout is called by shrink_page_list() for each dirty page.
461  * Calls ->writepage().
462  */
463 static pageout_t pageout(struct page *page, struct address_space *mapping,
464                          struct scan_control *sc)
465 {
466         /*
467          * If the page is dirty, only perform writeback if that write
468          * will be non-blocking.  To prevent this allocation from being
469          * stalled by pagecache activity.  But note that there may be
470          * stalls if we need to run get_block().  We could test
471          * PagePrivate for that.
472          *
473          * If this process is currently in __generic_file_aio_write() against
474          * this page's queue, we can perform writeback even if that
475          * will block.
476          *
477          * If the page is swapcache, write it back even if that would
478          * block, for some throttling. This happens by accident, because
479          * swap_backing_dev_info is bust: it doesn't reflect the
480          * congestion state of the swapdevs.  Easy to fix, if needed.
481          */
482         if (!is_page_cache_freeable(page))
483                 return PAGE_KEEP;
484         if (!mapping) {
485                 /*
486                  * Some data journaling orphaned pages can have
487                  * page->mapping == NULL while being dirty with clean buffers.
488                  */
489                 if (page_has_private(page)) {
490                         if (try_to_free_buffers(page)) {
491                                 ClearPageDirty(page);
492                                 printk("%s: orphaned page\n", __func__);
493                                 return PAGE_CLEAN;
494                         }
495                 }
496                 return PAGE_KEEP;
497         }
498         if (mapping->a_ops->writepage == NULL)
499                 return PAGE_ACTIVATE;
500         if (!may_write_to_queue(mapping->backing_dev_info, sc))
501                 return PAGE_KEEP;
502
503         if (clear_page_dirty_for_io(page)) {
504                 int res;
505                 struct writeback_control wbc = {
506                         .sync_mode = WB_SYNC_NONE,
507                         .nr_to_write = SWAP_CLUSTER_MAX,
508                         .range_start = 0,
509                         .range_end = LLONG_MAX,
510                         .for_reclaim = 1,
511                 };
512
513                 SetPageReclaim(page);
514                 res = mapping->a_ops->writepage(page, &wbc);
515                 if (res < 0)
516                         handle_write_error(mapping, page, res);
517                 if (res == AOP_WRITEPAGE_ACTIVATE) {
518                         ClearPageReclaim(page);
519                         return PAGE_ACTIVATE;
520                 }
521
522                 if (!PageWriteback(page)) {
523                         /* synchronous write or broken a_ops? */
524                         ClearPageReclaim(page);
525                 }
526                 trace_mm_vmscan_writepage(page,
527                         trace_reclaim_flags(page, sc->reclaim_mode));
528                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
529                 return PAGE_SUCCESS;
530         }
531
532         return PAGE_CLEAN;
533 }
534
535 /*
536  * Same as remove_mapping, but if the page is removed from the mapping, it
537  * gets returned with a refcount of 0.
538  */
539 static int __remove_mapping(struct address_space *mapping, struct page *page)
540 {
541         BUG_ON(!PageLocked(page));
542         BUG_ON(mapping != page_mapping(page));
543
544         spin_lock_irq(&mapping->tree_lock);
545         /*
546          * The non racy check for a busy page.
547          *
548          * Must be careful with the order of the tests. When someone has
549          * a ref to the page, it may be possible that they dirty it then
550          * drop the reference. So if PageDirty is tested before page_count
551          * here, then the following race may occur:
552          *
553          * get_user_pages(&page);
554          * [user mapping goes away]
555          * write_to(page);
556          *                              !PageDirty(page)    [good]
557          * SetPageDirty(page);
558          * put_page(page);
559          *                              !page_count(page)   [good, discard it]
560          *
561          * [oops, our write_to data is lost]
562          *
563          * Reversing the order of the tests ensures such a situation cannot
564          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
565          * load is not satisfied before that of page->_count.
566          *
567          * Note that if SetPageDirty is always performed via set_page_dirty,
568          * and thus under tree_lock, then this ordering is not required.
569          */
570         if (!page_freeze_refs(page, 2))
571                 goto cannot_free;
572         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
573         if (unlikely(PageDirty(page))) {
574                 page_unfreeze_refs(page, 2);
575                 goto cannot_free;
576         }
577
578         if (PageSwapCache(page)) {
579                 swp_entry_t swap = { .val = page_private(page) };
580                 __delete_from_swap_cache(page);
581                 spin_unlock_irq(&mapping->tree_lock);
582                 swapcache_free(swap, page);
583         } else {
584                 void (*freepage)(struct page *);
585
586                 freepage = mapping->a_ops->freepage;
587
588                 __delete_from_page_cache(page);
589                 spin_unlock_irq(&mapping->tree_lock);
590                 mem_cgroup_uncharge_cache_page(page);
591
592                 if (freepage != NULL)
593                         freepage(page);
594         }
595
596         return 1;
597
598 cannot_free:
599         spin_unlock_irq(&mapping->tree_lock);
600         return 0;
601 }
602
603 /*
604  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
605  * someone else has a ref on the page, abort and return 0.  If it was
606  * successfully detached, return 1.  Assumes the caller has a single ref on
607  * this page.
608  */
609 int remove_mapping(struct address_space *mapping, struct page *page)
610 {
611         if (__remove_mapping(mapping, page)) {
612                 /*
613                  * Unfreezing the refcount with 1 rather than 2 effectively
614                  * drops the pagecache ref for us without requiring another
615                  * atomic operation.
616                  */
617                 page_unfreeze_refs(page, 1);
618                 return 1;
619         }
620         return 0;
621 }
622
623 /**
624  * putback_lru_page - put previously isolated page onto appropriate LRU list
625  * @page: page to be put back to appropriate lru list
626  *
627  * Add previously isolated @page to appropriate LRU list.
628  * Page may still be unevictable for other reasons.
629  *
630  * lru_lock must not be held, interrupts must be enabled.
631  */
632 void putback_lru_page(struct page *page)
633 {
634         int lru;
635         int active = !!TestClearPageActive(page);
636         int was_unevictable = PageUnevictable(page);
637
638         VM_BUG_ON(PageLRU(page));
639
640 redo:
641         ClearPageUnevictable(page);
642
643         if (page_evictable(page, NULL)) {
644                 /*
645                  * For evictable pages, we can use the cache.
646                  * In event of a race, worst case is we end up with an
647                  * unevictable page on [in]active list.
648                  * We know how to handle that.
649                  */
650                 lru = active + page_lru_base_type(page);
651                 lru_cache_add_lru(page, lru);
652         } else {
653                 /*
654                  * Put unevictable pages directly on zone's unevictable
655                  * list.
656                  */
657                 lru = LRU_UNEVICTABLE;
658                 add_page_to_unevictable_list(page);
659                 /*
660                  * When racing with an mlock or AS_UNEVICTABLE clearing
661                  * (page is unlocked) make sure that if the other thread
662                  * does not observe our setting of PG_lru and fails
663                  * isolation/check_move_unevictable_pages,
664                  * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
665                  * the page back to the evictable list.
666                  *
667                  * The other side is TestClearPageMlocked() or shmem_lock().
668                  */
669                 smp_mb();
670         }
671
672         /*
673          * page's status can change while we move it among lru. If an evictable
674          * page is on unevictable list, it never be freed. To avoid that,
675          * check after we added it to the list, again.
676          */
677         if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
678                 if (!isolate_lru_page(page)) {
679                         put_page(page);
680                         goto redo;
681                 }
682                 /* This means someone else dropped this page from LRU
683                  * So, it will be freed or putback to LRU again. There is
684                  * nothing to do here.
685                  */
686         }
687
688         if (was_unevictable && lru != LRU_UNEVICTABLE)
689                 count_vm_event(UNEVICTABLE_PGRESCUED);
690         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
691                 count_vm_event(UNEVICTABLE_PGCULLED);
692
693         put_page(page);         /* drop ref from isolate */
694 }
695
696 enum page_references {
697         PAGEREF_RECLAIM,
698         PAGEREF_RECLAIM_CLEAN,
699         PAGEREF_KEEP,
700         PAGEREF_ACTIVATE,
701 };
702
703 static enum page_references page_check_references(struct page *page,
704                                                   struct mem_cgroup_zone *mz,
705                                                   struct scan_control *sc)
706 {
707         int referenced_ptes, referenced_page;
708         unsigned long vm_flags;
709
710         referenced_ptes = page_referenced(page, 1, mz->mem_cgroup, &vm_flags);
711         referenced_page = TestClearPageReferenced(page);
712
713         /* Lumpy reclaim - ignore references */
714         if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
715                 return PAGEREF_RECLAIM;
716
717         /*
718          * Mlock lost the isolation race with us.  Let try_to_unmap()
719          * move the page to the unevictable list.
720          */
721         if (vm_flags & VM_LOCKED)
722                 return PAGEREF_RECLAIM;
723
724         if (referenced_ptes) {
725                 if (PageAnon(page))
726                         return PAGEREF_ACTIVATE;
727                 /*
728                  * All mapped pages start out with page table
729                  * references from the instantiating fault, so we need
730                  * to look twice if a mapped file page is used more
731                  * than once.
732                  *
733                  * Mark it and spare it for another trip around the
734                  * inactive list.  Another page table reference will
735                  * lead to its activation.
736                  *
737                  * Note: the mark is set for activated pages as well
738                  * so that recently deactivated but used pages are
739                  * quickly recovered.
740                  */
741                 SetPageReferenced(page);
742
743                 if (referenced_page || referenced_ptes > 1)
744                         return PAGEREF_ACTIVATE;
745
746                 /*
747                  * Activate file-backed executable pages after first usage.
748                  */
749                 if (vm_flags & VM_EXEC)
750                         return PAGEREF_ACTIVATE;
751
752                 return PAGEREF_KEEP;
753         }
754
755         /* Reclaim if clean, defer dirty pages to writeback */
756         if (referenced_page && !PageSwapBacked(page))
757                 return PAGEREF_RECLAIM_CLEAN;
758
759         return PAGEREF_RECLAIM;
760 }
761
762 /*
763  * shrink_page_list() returns the number of reclaimed pages
764  */
765 static unsigned long shrink_page_list(struct list_head *page_list,
766                                       struct mem_cgroup_zone *mz,
767                                       struct scan_control *sc,
768                                       int priority,
769                                       unsigned long *ret_nr_dirty,
770                                       unsigned long *ret_nr_writeback)
771 {
772         LIST_HEAD(ret_pages);
773         LIST_HEAD(free_pages);
774         int pgactivate = 0;
775         unsigned long nr_dirty = 0;
776         unsigned long nr_congested = 0;
777         unsigned long nr_reclaimed = 0;
778         unsigned long nr_writeback = 0;
779
780         cond_resched();
781
782         while (!list_empty(page_list)) {
783                 enum page_references references;
784                 struct address_space *mapping;
785                 struct page *page;
786                 int may_enter_fs;
787
788                 cond_resched();
789
790                 page = lru_to_page(page_list);
791                 list_del(&page->lru);
792
793                 if (!trylock_page(page))
794                         goto keep;
795
796                 VM_BUG_ON(PageActive(page));
797                 VM_BUG_ON(page_zone(page) != mz->zone);
798
799                 sc->nr_scanned++;
800
801                 if (unlikely(!page_evictable(page, NULL)))
802                         goto cull_mlocked;
803
804                 if (!sc->may_unmap && page_mapped(page))
805                         goto keep_locked;
806
807                 /* Double the slab pressure for mapped and swapcache pages */
808                 if (page_mapped(page) || PageSwapCache(page))
809                         sc->nr_scanned++;
810
811                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
812                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
813
814                 if (PageWriteback(page)) {
815                         nr_writeback++;
816                         /*
817                          * Synchronous reclaim cannot queue pages for
818                          * writeback due to the possibility of stack overflow
819                          * but if it encounters a page under writeback, wait
820                          * for the IO to complete.
821                          */
822                         if ((sc->reclaim_mode & RECLAIM_MODE_SYNC) &&
823                             may_enter_fs)
824                                 wait_on_page_writeback(page);
825                         else {
826                                 unlock_page(page);
827                                 goto keep_lumpy;
828                         }
829                 }
830
831                 references = page_check_references(page, mz, sc);
832                 switch (references) {
833                 case PAGEREF_ACTIVATE:
834                         goto activate_locked;
835                 case PAGEREF_KEEP:
836                         goto keep_locked;
837                 case PAGEREF_RECLAIM:
838                 case PAGEREF_RECLAIM_CLEAN:
839                         ; /* try to reclaim the page below */
840                 }
841
842                 /*
843                  * Anonymous process memory has backing store?
844                  * Try to allocate it some swap space here.
845                  */
846                 if (PageAnon(page) && !PageSwapCache(page)) {
847                         if (!(sc->gfp_mask & __GFP_IO))
848                                 goto keep_locked;
849                         if (!add_to_swap(page))
850                                 goto activate_locked;
851                         may_enter_fs = 1;
852                 }
853
854                 mapping = page_mapping(page);
855
856                 /*
857                  * The page is mapped into the page tables of one or more
858                  * processes. Try to unmap it here.
859                  */
860                 if (page_mapped(page) && mapping) {
861                         switch (try_to_unmap(page, TTU_UNMAP)) {
862                         case SWAP_FAIL:
863                                 goto activate_locked;
864                         case SWAP_AGAIN:
865                                 goto keep_locked;
866                         case SWAP_MLOCK:
867                                 goto cull_mlocked;
868                         case SWAP_SUCCESS:
869                                 ; /* try to free the page below */
870                         }
871                 }
872
873                 if (PageDirty(page)) {
874                         nr_dirty++;
875
876                         /*
877                          * Only kswapd can writeback filesystem pages to
878                          * avoid risk of stack overflow but do not writeback
879                          * unless under significant pressure.
880                          */
881                         if (page_is_file_cache(page) &&
882                                         (!current_is_kswapd() || priority >= DEF_PRIORITY - 2)) {
883                                 /*
884                                  * Immediately reclaim when written back.
885                                  * Similar in principal to deactivate_page()
886                                  * except we already have the page isolated
887                                  * and know it's dirty
888                                  */
889                                 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
890                                 SetPageReclaim(page);
891
892                                 goto keep_locked;
893                         }
894
895                         if (references == PAGEREF_RECLAIM_CLEAN)
896                                 goto keep_locked;
897                         if (!may_enter_fs)
898                                 goto keep_locked;
899                         if (!sc->may_writepage)
900                                 goto keep_locked;
901
902                         /* Page is dirty, try to write it out here */
903                         switch (pageout(page, mapping, sc)) {
904                         case PAGE_KEEP:
905                                 nr_congested++;
906                                 goto keep_locked;
907                         case PAGE_ACTIVATE:
908                                 goto activate_locked;
909                         case PAGE_SUCCESS:
910                                 if (PageWriteback(page))
911                                         goto keep_lumpy;
912                                 if (PageDirty(page))
913                                         goto keep;
914
915                                 /*
916                                  * A synchronous write - probably a ramdisk.  Go
917                                  * ahead and try to reclaim the page.
918                                  */
919                                 if (!trylock_page(page))
920                                         goto keep;
921                                 if (PageDirty(page) || PageWriteback(page))
922                                         goto keep_locked;
923                                 mapping = page_mapping(page);
924                         case PAGE_CLEAN:
925                                 ; /* try to free the page below */
926                         }
927                 }
928
929                 /*
930                  * If the page has buffers, try to free the buffer mappings
931                  * associated with this page. If we succeed we try to free
932                  * the page as well.
933                  *
934                  * We do this even if the page is PageDirty().
935                  * try_to_release_page() does not perform I/O, but it is
936                  * possible for a page to have PageDirty set, but it is actually
937                  * clean (all its buffers are clean).  This happens if the
938                  * buffers were written out directly, with submit_bh(). ext3
939                  * will do this, as well as the blockdev mapping.
940                  * try_to_release_page() will discover that cleanness and will
941                  * drop the buffers and mark the page clean - it can be freed.
942                  *
943                  * Rarely, pages can have buffers and no ->mapping.  These are
944                  * the pages which were not successfully invalidated in
945                  * truncate_complete_page().  We try to drop those buffers here
946                  * and if that worked, and the page is no longer mapped into
947                  * process address space (page_count == 1) it can be freed.
948                  * Otherwise, leave the page on the LRU so it is swappable.
949                  */
950                 if (page_has_private(page)) {
951                         if (!try_to_release_page(page, sc->gfp_mask))
952                                 goto activate_locked;
953                         if (!mapping && page_count(page) == 1) {
954                                 unlock_page(page);
955                                 if (put_page_testzero(page))
956                                         goto free_it;
957                                 else {
958                                         /*
959                                          * rare race with speculative reference.
960                                          * the speculative reference will free
961                                          * this page shortly, so we may
962                                          * increment nr_reclaimed here (and
963                                          * leave it off the LRU).
964                                          */
965                                         nr_reclaimed++;
966                                         continue;
967                                 }
968                         }
969                 }
970
971                 if (!mapping || !__remove_mapping(mapping, page))
972                         goto keep_locked;
973
974                 /*
975                  * At this point, we have no other references and there is
976                  * no way to pick any more up (removed from LRU, removed
977                  * from pagecache). Can use non-atomic bitops now (and
978                  * we obviously don't have to worry about waking up a process
979                  * waiting on the page lock, because there are no references.
980                  */
981                 __clear_page_locked(page);
982 free_it:
983                 nr_reclaimed++;
984
985                 /*
986                  * Is there need to periodically free_page_list? It would
987                  * appear not as the counts should be low
988                  */
989                 list_add(&page->lru, &free_pages);
990                 continue;
991
992 cull_mlocked:
993                 if (PageSwapCache(page))
994                         try_to_free_swap(page);
995                 unlock_page(page);
996                 putback_lru_page(page);
997                 reset_reclaim_mode(sc);
998                 continue;
999
1000 activate_locked:
1001                 /* Not a candidate for swapping, so reclaim swap space. */
1002                 if (PageSwapCache(page) && vm_swap_full())
1003                         try_to_free_swap(page);
1004                 VM_BUG_ON(PageActive(page));
1005                 SetPageActive(page);
1006                 pgactivate++;
1007 keep_locked:
1008                 unlock_page(page);
1009 keep:
1010                 reset_reclaim_mode(sc);
1011 keep_lumpy:
1012                 list_add(&page->lru, &ret_pages);
1013                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
1014         }
1015
1016         /*
1017          * Tag a zone as congested if all the dirty pages encountered were
1018          * backed by a congested BDI. In this case, reclaimers should just
1019          * back off and wait for congestion to clear because further reclaim
1020          * will encounter the same problem
1021          */
1022         if (nr_dirty && nr_dirty == nr_congested && global_reclaim(sc))
1023                 zone_set_flag(mz->zone, ZONE_CONGESTED);
1024
1025         free_hot_cold_page_list(&free_pages, 1);
1026
1027         list_splice(&ret_pages, page_list);
1028         count_vm_events(PGACTIVATE, pgactivate);
1029         *ret_nr_dirty += nr_dirty;
1030         *ret_nr_writeback += nr_writeback;
1031         return nr_reclaimed;
1032 }
1033
1034 /*
1035  * Attempt to remove the specified page from its LRU.  Only take this page
1036  * if it is of the appropriate PageActive status.  Pages which are being
1037  * freed elsewhere are also ignored.
1038  *
1039  * page:        page to consider
1040  * mode:        one of the LRU isolation modes defined above
1041  *
1042  * returns 0 on success, -ve errno on failure.
1043  */
1044 int __isolate_lru_page(struct page *page, isolate_mode_t mode, int file)
1045 {
1046         bool all_lru_mode;
1047         int ret = -EINVAL;
1048
1049         /* Only take pages on the LRU. */
1050         if (!PageLRU(page))
1051                 return ret;
1052
1053         all_lru_mode = (mode & (ISOLATE_ACTIVE|ISOLATE_INACTIVE)) ==
1054                 (ISOLATE_ACTIVE|ISOLATE_INACTIVE);
1055
1056         /*
1057          * When checking the active state, we need to be sure we are
1058          * dealing with comparible boolean values.  Take the logical not
1059          * of each.
1060          */
1061         if (!all_lru_mode && !PageActive(page) != !(mode & ISOLATE_ACTIVE))
1062                 return ret;
1063
1064         if (!all_lru_mode && !!page_is_file_cache(page) != file)
1065                 return ret;
1066
1067         /*
1068          * When this function is being called for lumpy reclaim, we
1069          * initially look into all LRU pages, active, inactive and
1070          * unevictable; only give shrink_page_list evictable pages.
1071          */
1072         if (PageUnevictable(page))
1073                 return ret;
1074
1075         ret = -EBUSY;
1076
1077         /*
1078          * To minimise LRU disruption, the caller can indicate that it only
1079          * wants to isolate pages it will be able to operate on without
1080          * blocking - clean pages for the most part.
1081          *
1082          * ISOLATE_CLEAN means that only clean pages should be isolated. This
1083          * is used by reclaim when it is cannot write to backing storage
1084          *
1085          * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1086          * that it is possible to migrate without blocking
1087          */
1088         if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1089                 /* All the caller can do on PageWriteback is block */
1090                 if (PageWriteback(page))
1091                         return ret;
1092
1093                 if (PageDirty(page)) {
1094                         struct address_space *mapping;
1095
1096                         /* ISOLATE_CLEAN means only clean pages */
1097                         if (mode & ISOLATE_CLEAN)
1098                                 return ret;
1099
1100                         /*
1101                          * Only pages without mappings or that have a
1102                          * ->migratepage callback are possible to migrate
1103                          * without blocking
1104                          */
1105                         mapping = page_mapping(page);
1106                         if (mapping && !mapping->a_ops->migratepage)
1107                                 return ret;
1108                 }
1109         }
1110
1111         if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1112                 return ret;
1113
1114         if (likely(get_page_unless_zero(page))) {
1115                 /*
1116                  * Be careful not to clear PageLRU until after we're
1117                  * sure the page is not being freed elsewhere -- the
1118                  * page release code relies on it.
1119                  */
1120                 ClearPageLRU(page);
1121                 ret = 0;
1122         }
1123
1124         return ret;
1125 }
1126
1127 /*
1128  * zone->lru_lock is heavily contended.  Some of the functions that
1129  * shrink the lists perform better by taking out a batch of pages
1130  * and working on them outside the LRU lock.
1131  *
1132  * For pagecache intensive workloads, this function is the hottest
1133  * spot in the kernel (apart from copy_*_user functions).
1134  *
1135  * Appropriate locks must be held before calling this function.
1136  *
1137  * @nr_to_scan: The number of pages to look through on the list.
1138  * @mz:         The mem_cgroup_zone to pull pages from.
1139  * @dst:        The temp list to put pages on to.
1140  * @nr_scanned: The number of pages that were scanned.
1141  * @order:      The caller's attempted allocation order
1142  * @mode:       One of the LRU isolation modes
1143  * @active:     True [1] if isolating active pages
1144  * @file:       True [1] if isolating file [!anon] pages
1145  *
1146  * returns how many pages were moved onto *@dst.
1147  */
1148 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1149                 struct mem_cgroup_zone *mz, struct list_head *dst,
1150                 unsigned long *nr_scanned, int order, isolate_mode_t mode,
1151                 int active, int file)
1152 {
1153         struct lruvec *lruvec;
1154         struct list_head *src;
1155         unsigned long nr_taken = 0;
1156         unsigned long nr_lumpy_taken = 0;
1157         unsigned long nr_lumpy_dirty = 0;
1158         unsigned long nr_lumpy_failed = 0;
1159         unsigned long scan;
1160         int lru = LRU_BASE;
1161
1162         lruvec = mem_cgroup_zone_lruvec(mz->zone, mz->mem_cgroup);
1163         if (active)
1164                 lru += LRU_ACTIVE;
1165         if (file)
1166                 lru += LRU_FILE;
1167         src = &lruvec->lists[lru];
1168
1169         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1170                 struct page *page;
1171                 unsigned long pfn;
1172                 unsigned long end_pfn;
1173                 unsigned long page_pfn;
1174                 int zone_id;
1175
1176                 page = lru_to_page(src);
1177                 prefetchw_prev_lru_page(page, src, flags);
1178
1179                 VM_BUG_ON(!PageLRU(page));
1180
1181                 switch (__isolate_lru_page(page, mode, file)) {
1182                 case 0:
1183                         mem_cgroup_lru_del(page);
1184                         list_move(&page->lru, dst);
1185                         nr_taken += hpage_nr_pages(page);
1186                         break;
1187
1188                 case -EBUSY:
1189                         /* else it is being freed elsewhere */
1190                         list_move(&page->lru, src);
1191                         continue;
1192
1193                 default:
1194                         BUG();
1195                 }
1196
1197                 if (!order)
1198                         continue;
1199
1200                 /*
1201                  * Attempt to take all pages in the order aligned region
1202                  * surrounding the tag page.  Only take those pages of
1203                  * the same active state as that tag page.  We may safely
1204                  * round the target page pfn down to the requested order
1205                  * as the mem_map is guaranteed valid out to MAX_ORDER,
1206                  * where that page is in a different zone we will detect
1207                  * it from its zone id and abort this block scan.
1208                  */
1209                 zone_id = page_zone_id(page);
1210                 page_pfn = page_to_pfn(page);
1211                 pfn = page_pfn & ~((1 << order) - 1);
1212                 end_pfn = pfn + (1 << order);
1213                 for (; pfn < end_pfn; pfn++) {
1214                         struct page *cursor_page;
1215
1216                         /* The target page is in the block, ignore it. */
1217                         if (unlikely(pfn == page_pfn))
1218                                 continue;
1219
1220                         /* Avoid holes within the zone. */
1221                         if (unlikely(!pfn_valid_within(pfn)))
1222                                 break;
1223
1224                         cursor_page = pfn_to_page(pfn);
1225
1226                         /* Check that we have not crossed a zone boundary. */
1227                         if (unlikely(page_zone_id(cursor_page) != zone_id))
1228                                 break;
1229
1230                         /*
1231                          * If we don't have enough swap space, reclaiming of
1232                          * anon page which don't already have a swap slot is
1233                          * pointless.
1234                          */
1235                         if (nr_swap_pages <= 0 && PageSwapBacked(cursor_page) &&
1236                             !PageSwapCache(cursor_page))
1237                                 break;
1238
1239                         if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1240                                 unsigned int isolated_pages;
1241
1242                                 mem_cgroup_lru_del(cursor_page);
1243                                 list_move(&cursor_page->lru, dst);
1244                                 isolated_pages = hpage_nr_pages(cursor_page);
1245                                 nr_taken += isolated_pages;
1246                                 nr_lumpy_taken += isolated_pages;
1247                                 if (PageDirty(cursor_page))
1248                                         nr_lumpy_dirty += isolated_pages;
1249                                 scan++;
1250                                 pfn += isolated_pages - 1;
1251                         } else {
1252                                 /*
1253                                  * Check if the page is freed already.
1254                                  *
1255                                  * We can't use page_count() as that
1256                                  * requires compound_head and we don't
1257                                  * have a pin on the page here. If a
1258                                  * page is tail, we may or may not
1259                                  * have isolated the head, so assume
1260                                  * it's not free, it'd be tricky to
1261                                  * track the head status without a
1262                                  * page pin.
1263                                  */
1264                                 if (!PageTail(cursor_page) &&
1265                                     !atomic_read(&cursor_page->_count))
1266                                         continue;
1267                                 break;
1268                         }
1269                 }
1270
1271                 /* If we break out of the loop above, lumpy reclaim failed */
1272                 if (pfn < end_pfn)
1273                         nr_lumpy_failed++;
1274         }
1275
1276         *nr_scanned = scan;
1277
1278         trace_mm_vmscan_lru_isolate(order,
1279                         nr_to_scan, scan,
1280                         nr_taken,
1281                         nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1282                         mode, file);
1283         return nr_taken;
1284 }
1285
1286 /**
1287  * isolate_lru_page - tries to isolate a page from its LRU list
1288  * @page: page to isolate from its LRU list
1289  *
1290  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1291  * vmstat statistic corresponding to whatever LRU list the page was on.
1292  *
1293  * Returns 0 if the page was removed from an LRU list.
1294  * Returns -EBUSY if the page was not on an LRU list.
1295  *
1296  * The returned page will have PageLRU() cleared.  If it was found on
1297  * the active list, it will have PageActive set.  If it was found on
1298  * the unevictable list, it will have the PageUnevictable bit set. That flag
1299  * may need to be cleared by the caller before letting the page go.
1300  *
1301  * The vmstat statistic corresponding to the list on which the page was
1302  * found will be decremented.
1303  *
1304  * Restrictions:
1305  * (1) Must be called with an elevated refcount on the page. This is a
1306  *     fundamentnal difference from isolate_lru_pages (which is called
1307  *     without a stable reference).
1308  * (2) the lru_lock must not be held.
1309  * (3) interrupts must be enabled.
1310  */
1311 int isolate_lru_page(struct page *page)
1312 {
1313         int ret = -EBUSY;
1314
1315         VM_BUG_ON(!page_count(page));
1316
1317         if (PageLRU(page)) {
1318                 struct zone *zone = page_zone(page);
1319
1320                 spin_lock_irq(&zone->lru_lock);
1321                 if (PageLRU(page)) {
1322                         int lru = page_lru(page);
1323                         ret = 0;
1324                         get_page(page);
1325                         ClearPageLRU(page);
1326
1327                         del_page_from_lru_list(zone, page, lru);
1328                 }
1329                 spin_unlock_irq(&zone->lru_lock);
1330         }
1331         return ret;
1332 }
1333
1334 /*
1335  * Are there way too many processes in the direct reclaim path already?
1336  */
1337 static int too_many_isolated(struct zone *zone, int file,
1338                 struct scan_control *sc)
1339 {
1340         unsigned long inactive, isolated;
1341
1342         if (current_is_kswapd())
1343                 return 0;
1344
1345         if (!global_reclaim(sc))
1346                 return 0;
1347
1348         if (file) {
1349                 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1350                 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1351         } else {
1352                 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1353                 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1354         }
1355
1356         return isolated > inactive;
1357 }
1358
1359 static noinline_for_stack void
1360 putback_inactive_pages(struct mem_cgroup_zone *mz,
1361                        struct list_head *page_list)
1362 {
1363         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz);
1364         struct zone *zone = mz->zone;
1365         LIST_HEAD(pages_to_free);
1366
1367         /*
1368          * Put back any unfreeable pages.
1369          */
1370         while (!list_empty(page_list)) {
1371                 struct page *page = lru_to_page(page_list);
1372                 int lru;
1373
1374                 VM_BUG_ON(PageLRU(page));
1375                 list_del(&page->lru);
1376                 if (unlikely(!page_evictable(page, NULL))) {
1377                         spin_unlock_irq(&zone->lru_lock);
1378                         putback_lru_page(page);
1379                         spin_lock_irq(&zone->lru_lock);
1380                         continue;
1381                 }
1382                 SetPageLRU(page);
1383                 lru = page_lru(page);
1384                 add_page_to_lru_list(zone, page, lru);
1385                 if (is_active_lru(lru)) {
1386                         int file = is_file_lru(lru);
1387                         int numpages = hpage_nr_pages(page);
1388                         reclaim_stat->recent_rotated[file] += numpages;
1389                 }
1390                 if (put_page_testzero(page)) {
1391                         __ClearPageLRU(page);
1392                         __ClearPageActive(page);
1393                         del_page_from_lru_list(zone, page, lru);
1394
1395                         if (unlikely(PageCompound(page))) {
1396                                 spin_unlock_irq(&zone->lru_lock);
1397                                 (*get_compound_page_dtor(page))(page);
1398                                 spin_lock_irq(&zone->lru_lock);
1399                         } else
1400                                 list_add(&page->lru, &pages_to_free);
1401                 }
1402         }
1403
1404         /*
1405          * To save our caller's stack, now use input list for pages to free.
1406          */
1407         list_splice(&pages_to_free, page_list);
1408 }
1409
1410 static noinline_for_stack void
1411 update_isolated_counts(struct mem_cgroup_zone *mz,
1412                        struct list_head *page_list,
1413                        unsigned long *nr_anon,
1414                        unsigned long *nr_file)
1415 {
1416         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz);
1417         struct zone *zone = mz->zone;
1418         unsigned int count[NR_LRU_LISTS] = { 0, };
1419         unsigned long nr_active = 0;
1420         struct page *page;
1421         int lru;
1422
1423         /*
1424          * Count pages and clear active flags
1425          */
1426         list_for_each_entry(page, page_list, lru) {
1427                 int numpages = hpage_nr_pages(page);
1428                 lru = page_lru_base_type(page);
1429                 if (PageActive(page)) {
1430                         lru += LRU_ACTIVE;
1431                         ClearPageActive(page);
1432                         nr_active += numpages;
1433                 }
1434                 count[lru] += numpages;
1435         }
1436
1437         __count_vm_events(PGDEACTIVATE, nr_active);
1438
1439         __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1440                               -count[LRU_ACTIVE_FILE]);
1441         __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1442                               -count[LRU_INACTIVE_FILE]);
1443         __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1444                               -count[LRU_ACTIVE_ANON]);
1445         __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1446                               -count[LRU_INACTIVE_ANON]);
1447
1448         *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1449         *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1450
1451         reclaim_stat->recent_scanned[0] += *nr_anon;
1452         reclaim_stat->recent_scanned[1] += *nr_file;
1453 }
1454
1455 /*
1456  * Returns true if a direct reclaim should wait on pages under writeback.
1457  *
1458  * If we are direct reclaiming for contiguous pages and we do not reclaim
1459  * everything in the list, try again and wait for writeback IO to complete.
1460  * This will stall high-order allocations noticeably. Only do that when really
1461  * need to free the pages under high memory pressure.
1462  */
1463 static inline bool should_reclaim_stall(unsigned long nr_taken,
1464                                         unsigned long nr_freed,
1465                                         int priority,
1466                                         struct scan_control *sc)
1467 {
1468         int lumpy_stall_priority;
1469
1470         /* kswapd should not stall on sync IO */
1471         if (current_is_kswapd())
1472                 return false;
1473
1474         /* Only stall on lumpy reclaim */
1475         if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
1476                 return false;
1477
1478         /* If we have reclaimed everything on the isolated list, no stall */
1479         if (nr_freed == nr_taken)
1480                 return false;
1481
1482         /*
1483          * For high-order allocations, there are two stall thresholds.
1484          * High-cost allocations stall immediately where as lower
1485          * order allocations such as stacks require the scanning
1486          * priority to be much higher before stalling.
1487          */
1488         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1489                 lumpy_stall_priority = DEF_PRIORITY;
1490         else
1491                 lumpy_stall_priority = DEF_PRIORITY / 3;
1492
1493         return priority <= lumpy_stall_priority;
1494 }
1495
1496 /*
1497  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1498  * of reclaimed pages
1499  */
1500 static noinline_for_stack unsigned long
1501 shrink_inactive_list(unsigned long nr_to_scan, struct mem_cgroup_zone *mz,
1502                      struct scan_control *sc, int priority, int file)
1503 {
1504         LIST_HEAD(page_list);
1505         unsigned long nr_scanned;
1506         unsigned long nr_reclaimed = 0;
1507         unsigned long nr_taken;
1508         unsigned long nr_anon;
1509         unsigned long nr_file;
1510         unsigned long nr_dirty = 0;
1511         unsigned long nr_writeback = 0;
1512         isolate_mode_t isolate_mode = ISOLATE_INACTIVE;
1513         struct zone *zone = mz->zone;
1514
1515         while (unlikely(too_many_isolated(zone, file, sc))) {
1516                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1517
1518                 /* We are about to die and free our memory. Return now. */
1519                 if (fatal_signal_pending(current))
1520                         return SWAP_CLUSTER_MAX;
1521         }
1522
1523         set_reclaim_mode(priority, sc, false);
1524         if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
1525                 isolate_mode |= ISOLATE_ACTIVE;
1526
1527         lru_add_drain();
1528
1529         if (!sc->may_unmap)
1530                 isolate_mode |= ISOLATE_UNMAPPED;
1531         if (!sc->may_writepage)
1532                 isolate_mode |= ISOLATE_CLEAN;
1533
1534         spin_lock_irq(&zone->lru_lock);
1535
1536         nr_taken = isolate_lru_pages(nr_to_scan, mz, &page_list,
1537                                      &nr_scanned, sc->order,
1538                                      isolate_mode, 0, file);
1539         if (global_reclaim(sc)) {
1540                 zone->pages_scanned += nr_scanned;
1541                 if (current_is_kswapd())
1542                         __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1543                                                nr_scanned);
1544                 else
1545                         __count_zone_vm_events(PGSCAN_DIRECT, zone,
1546                                                nr_scanned);
1547         }
1548
1549         if (nr_taken == 0) {
1550                 spin_unlock_irq(&zone->lru_lock);
1551                 return 0;
1552         }
1553
1554         update_isolated_counts(mz, &page_list, &nr_anon, &nr_file);
1555
1556         __mod_zone_page_state(zone, NR_ISOLATED_ANON, nr_anon);
1557         __mod_zone_page_state(zone, NR_ISOLATED_FILE, nr_file);
1558
1559         spin_unlock_irq(&zone->lru_lock);
1560
1561         nr_reclaimed = shrink_page_list(&page_list, mz, sc, priority,
1562                                                 &nr_dirty, &nr_writeback);
1563
1564         /* Check if we should syncronously wait for writeback */
1565         if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1566                 set_reclaim_mode(priority, sc, true);
1567                 nr_reclaimed += shrink_page_list(&page_list, mz, sc,
1568                                         priority, &nr_dirty, &nr_writeback);
1569         }
1570
1571         spin_lock_irq(&zone->lru_lock);
1572
1573         if (current_is_kswapd())
1574                 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1575         __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1576
1577         putback_inactive_pages(mz, &page_list);
1578
1579         __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1580         __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1581
1582         spin_unlock_irq(&zone->lru_lock);
1583
1584         free_hot_cold_page_list(&page_list, 1);
1585
1586         /*
1587          * If reclaim is isolating dirty pages under writeback, it implies
1588          * that the long-lived page allocation rate is exceeding the page
1589          * laundering rate. Either the global limits are not being effective
1590          * at throttling processes due to the page distribution throughout
1591          * zones or there is heavy usage of a slow backing device. The
1592          * only option is to throttle from reclaim context which is not ideal
1593          * as there is no guarantee the dirtying process is throttled in the
1594          * same way balance_dirty_pages() manages.
1595          *
1596          * This scales the number of dirty pages that must be under writeback
1597          * before throttling depending on priority. It is a simple backoff
1598          * function that has the most effect in the range DEF_PRIORITY to
1599          * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1600          * in trouble and reclaim is considered to be in trouble.
1601          *
1602          * DEF_PRIORITY   100% isolated pages must be PageWriteback to throttle
1603          * DEF_PRIORITY-1  50% must be PageWriteback
1604          * DEF_PRIORITY-2  25% must be PageWriteback, kswapd in trouble
1605          * ...
1606          * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1607          *                     isolated page is PageWriteback
1608          */
1609         if (nr_writeback && nr_writeback >= (nr_taken >> (DEF_PRIORITY-priority)))
1610                 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1611
1612         trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1613                 zone_idx(zone),
1614                 nr_scanned, nr_reclaimed,
1615                 priority,
1616                 trace_shrink_flags(file, sc->reclaim_mode));
1617         return nr_reclaimed;
1618 }
1619
1620 /*
1621  * This moves pages from the active list to the inactive list.
1622  *
1623  * We move them the other way if the page is referenced by one or more
1624  * processes, from rmap.
1625  *
1626  * If the pages are mostly unmapped, the processing is fast and it is
1627  * appropriate to hold zone->lru_lock across the whole operation.  But if
1628  * the pages are mapped, the processing is slow (page_referenced()) so we
1629  * should drop zone->lru_lock around each page.  It's impossible to balance
1630  * this, so instead we remove the pages from the LRU while processing them.
1631  * It is safe to rely on PG_active against the non-LRU pages in here because
1632  * nobody will play with that bit on a non-LRU page.
1633  *
1634  * The downside is that we have to touch page->_count against each page.
1635  * But we had to alter page->flags anyway.
1636  */
1637
1638 static void move_active_pages_to_lru(struct zone *zone,
1639                                      struct list_head *list,
1640                                      struct list_head *pages_to_free,
1641                                      enum lru_list lru)
1642 {
1643         unsigned long pgmoved = 0;
1644         struct page *page;
1645
1646         if (buffer_heads_over_limit) {
1647                 spin_unlock_irq(&zone->lru_lock);
1648                 list_for_each_entry(page, list, lru) {
1649                         if (page_has_private(page) && trylock_page(page)) {
1650                                 if (page_has_private(page))
1651                                         try_to_release_page(page, 0);
1652                                 unlock_page(page);
1653                         }
1654                 }
1655                 spin_lock_irq(&zone->lru_lock);
1656         }
1657
1658         while (!list_empty(list)) {
1659                 struct lruvec *lruvec;
1660
1661                 page = lru_to_page(list);
1662
1663                 VM_BUG_ON(PageLRU(page));
1664                 SetPageLRU(page);
1665
1666                 lruvec = mem_cgroup_lru_add_list(zone, page, lru);
1667                 list_move(&page->lru, &lruvec->lists[lru]);
1668                 pgmoved += hpage_nr_pages(page);
1669
1670                 if (put_page_testzero(page)) {
1671                         __ClearPageLRU(page);
1672                         __ClearPageActive(page);
1673                         del_page_from_lru_list(zone, page, lru);
1674
1675                         if (unlikely(PageCompound(page))) {
1676                                 spin_unlock_irq(&zone->lru_lock);
1677                                 (*get_compound_page_dtor(page))(page);
1678                                 spin_lock_irq(&zone->lru_lock);
1679                         } else
1680                                 list_add(&page->lru, pages_to_free);
1681                 }
1682         }
1683         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1684         if (!is_active_lru(lru))
1685                 __count_vm_events(PGDEACTIVATE, pgmoved);
1686 }
1687
1688 static void shrink_active_list(unsigned long nr_to_scan,
1689                                struct mem_cgroup_zone *mz,
1690                                struct scan_control *sc,
1691                                int priority, int file)
1692 {
1693         unsigned long nr_taken;
1694         unsigned long nr_scanned;
1695         unsigned long vm_flags;
1696         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1697         LIST_HEAD(l_active);
1698         LIST_HEAD(l_inactive);
1699         struct page *page;
1700         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz);
1701         unsigned long nr_rotated = 0;
1702         isolate_mode_t isolate_mode = ISOLATE_ACTIVE;
1703         struct zone *zone = mz->zone;
1704
1705         lru_add_drain();
1706
1707         if (!sc->may_unmap)
1708                 isolate_mode |= ISOLATE_UNMAPPED;
1709         if (!sc->may_writepage)
1710                 isolate_mode |= ISOLATE_CLEAN;
1711
1712         spin_lock_irq(&zone->lru_lock);
1713
1714         nr_taken = isolate_lru_pages(nr_to_scan, mz, &l_hold,
1715                                      &nr_scanned, sc->order,
1716                                      isolate_mode, 1, file);
1717         if (global_reclaim(sc))
1718                 zone->pages_scanned += nr_scanned;
1719
1720         reclaim_stat->recent_scanned[file] += nr_taken;
1721
1722         __count_zone_vm_events(PGREFILL, zone, nr_scanned);
1723         if (file)
1724                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1725         else
1726                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1727         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1728         spin_unlock_irq(&zone->lru_lock);
1729
1730         while (!list_empty(&l_hold)) {
1731                 cond_resched();
1732                 page = lru_to_page(&l_hold);
1733                 list_del(&page->lru);
1734
1735                 if (unlikely(!page_evictable(page, NULL))) {
1736                         putback_lru_page(page);
1737                         continue;
1738                 }
1739
1740                 if (page_referenced(page, 0, mz->mem_cgroup, &vm_flags)) {
1741                         nr_rotated += hpage_nr_pages(page);
1742                         /*
1743                          * Identify referenced, file-backed active pages and
1744                          * give them one more trip around the active list. So
1745                          * that executable code get better chances to stay in
1746                          * memory under moderate memory pressure.  Anon pages
1747                          * are not likely to be evicted by use-once streaming
1748                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1749                          * so we ignore them here.
1750                          */
1751                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1752                                 list_add(&page->lru, &l_active);
1753                                 continue;
1754                         }
1755                 }
1756
1757                 ClearPageActive(page);  /* we are de-activating */
1758                 list_add(&page->lru, &l_inactive);
1759         }
1760
1761         /*
1762          * Move pages back to the lru list.
1763          */
1764         spin_lock_irq(&zone->lru_lock);
1765         /*
1766          * Count referenced pages from currently used mappings as rotated,
1767          * even though only some of them are actually re-activated.  This
1768          * helps balance scan pressure between file and anonymous pages in
1769          * get_scan_ratio.
1770          */
1771         reclaim_stat->recent_rotated[file] += nr_rotated;
1772
1773         move_active_pages_to_lru(zone, &l_active, &l_hold,
1774                                                 LRU_ACTIVE + file * LRU_FILE);
1775         move_active_pages_to_lru(zone, &l_inactive, &l_hold,
1776                                                 LRU_BASE   + file * LRU_FILE);
1777         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1778         spin_unlock_irq(&zone->lru_lock);
1779
1780         free_hot_cold_page_list(&l_hold, 1);
1781 }
1782
1783 #ifdef CONFIG_SWAP
1784 static int inactive_anon_is_low_global(struct zone *zone)
1785 {
1786         unsigned long active, inactive;
1787
1788         active = zone_page_state(zone, NR_ACTIVE_ANON);
1789         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1790
1791         if (inactive * zone->inactive_ratio < active)
1792                 return 1;
1793
1794         return 0;
1795 }
1796
1797 /**
1798  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1799  * @zone: zone to check
1800  * @sc:   scan control of this context
1801  *
1802  * Returns true if the zone does not have enough inactive anon pages,
1803  * meaning some active anon pages need to be deactivated.
1804  */
1805 static int inactive_anon_is_low(struct mem_cgroup_zone *mz)
1806 {
1807         /*
1808          * If we don't have swap space, anonymous page deactivation
1809          * is pointless.
1810          */
1811         if (!total_swap_pages)
1812                 return 0;
1813
1814         if (!scanning_global_lru(mz))
1815                 return mem_cgroup_inactive_anon_is_low(mz->mem_cgroup,
1816                                                        mz->zone);
1817
1818         return inactive_anon_is_low_global(mz->zone);
1819 }
1820 #else
1821 static inline int inactive_anon_is_low(struct mem_cgroup_zone *mz)
1822 {
1823         return 0;
1824 }
1825 #endif
1826
1827 static int inactive_file_is_low_global(struct zone *zone)
1828 {
1829         unsigned long active, inactive;
1830
1831         active = zone_page_state(zone, NR_ACTIVE_FILE);
1832         inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1833
1834         return (active > inactive);
1835 }
1836
1837 /**
1838  * inactive_file_is_low - check if file pages need to be deactivated
1839  * @mz: memory cgroup and zone to check
1840  *
1841  * When the system is doing streaming IO, memory pressure here
1842  * ensures that active file pages get deactivated, until more
1843  * than half of the file pages are on the inactive list.
1844  *
1845  * Once we get to that situation, protect the system's working
1846  * set from being evicted by disabling active file page aging.
1847  *
1848  * This uses a different ratio than the anonymous pages, because
1849  * the page cache uses a use-once replacement algorithm.
1850  */
1851 static int inactive_file_is_low(struct mem_cgroup_zone *mz)
1852 {
1853         if (!scanning_global_lru(mz))
1854                 return mem_cgroup_inactive_file_is_low(mz->mem_cgroup,
1855                                                        mz->zone);
1856
1857         return inactive_file_is_low_global(mz->zone);
1858 }
1859
1860 static int inactive_list_is_low(struct mem_cgroup_zone *mz, int file)
1861 {
1862         if (file)
1863                 return inactive_file_is_low(mz);
1864         else
1865                 return inactive_anon_is_low(mz);
1866 }
1867
1868 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1869                                  struct mem_cgroup_zone *mz,
1870                                  struct scan_control *sc, int priority)
1871 {
1872         int file = is_file_lru(lru);
1873
1874         if (is_active_lru(lru)) {
1875                 if (inactive_list_is_low(mz, file))
1876                         shrink_active_list(nr_to_scan, mz, sc, priority, file);
1877                 return 0;
1878         }
1879
1880         return shrink_inactive_list(nr_to_scan, mz, sc, priority, file);
1881 }
1882
1883 static int vmscan_swappiness(struct mem_cgroup_zone *mz,
1884                              struct scan_control *sc)
1885 {
1886         if (global_reclaim(sc))
1887                 return vm_swappiness;
1888         return mem_cgroup_swappiness(mz->mem_cgroup);
1889 }
1890
1891 /*
1892  * Determine how aggressively the anon and file LRU lists should be
1893  * scanned.  The relative value of each set of LRU lists is determined
1894  * by looking at the fraction of the pages scanned we did rotate back
1895  * onto the active list instead of evict.
1896  *
1897  * nr[0] = anon pages to scan; nr[1] = file pages to scan
1898  */
1899 static void get_scan_count(struct mem_cgroup_zone *mz, struct scan_control *sc,
1900                            unsigned long *nr, int priority)
1901 {
1902         unsigned long anon, file, free;
1903         unsigned long anon_prio, file_prio;
1904         unsigned long ap, fp;
1905         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz);
1906         u64 fraction[2], denominator;
1907         enum lru_list lru;
1908         int noswap = 0;
1909         bool force_scan = false;
1910
1911         /*
1912          * If the zone or memcg is small, nr[l] can be 0.  This
1913          * results in no scanning on this priority and a potential
1914          * priority drop.  Global direct reclaim can go to the next
1915          * zone and tends to have no problems. Global kswapd is for
1916          * zone balancing and it needs to scan a minimum amount. When
1917          * reclaiming for a memcg, a priority drop can cause high
1918          * latencies, so it's better to scan a minimum amount there as
1919          * well.
1920          */
1921         if (current_is_kswapd() && mz->zone->all_unreclaimable)
1922                 force_scan = true;
1923         if (!global_reclaim(sc))
1924                 force_scan = true;
1925
1926         /* If we have no swap space, do not bother scanning anon pages. */
1927         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1928                 noswap = 1;
1929                 fraction[0] = 0;
1930                 fraction[1] = 1;
1931                 denominator = 1;
1932                 goto out;
1933         }
1934
1935         anon  = zone_nr_lru_pages(mz, LRU_ACTIVE_ANON) +
1936                 zone_nr_lru_pages(mz, LRU_INACTIVE_ANON);
1937         file  = zone_nr_lru_pages(mz, LRU_ACTIVE_FILE) +
1938                 zone_nr_lru_pages(mz, LRU_INACTIVE_FILE);
1939
1940         if (global_reclaim(sc)) {
1941                 free  = zone_page_state(mz->zone, NR_FREE_PAGES);
1942                 /* If we have very few page cache pages,
1943                    force-scan anon pages. */
1944                 if (unlikely(file + free <= high_wmark_pages(mz->zone))) {
1945                         fraction[0] = 1;
1946                         fraction[1] = 0;
1947                         denominator = 1;
1948                         goto out;
1949                 }
1950         }
1951
1952         /*
1953          * With swappiness at 100, anonymous and file have the same priority.
1954          * This scanning priority is essentially the inverse of IO cost.
1955          */
1956         anon_prio = vmscan_swappiness(mz, sc);
1957         file_prio = 200 - vmscan_swappiness(mz, sc);
1958
1959         /*
1960          * OK, so we have swap space and a fair amount of page cache
1961          * pages.  We use the recently rotated / recently scanned
1962          * ratios to determine how valuable each cache is.
1963          *
1964          * Because workloads change over time (and to avoid overflow)
1965          * we keep these statistics as a floating average, which ends
1966          * up weighing recent references more than old ones.
1967          *
1968          * anon in [0], file in [1]
1969          */
1970         spin_lock_irq(&mz->zone->lru_lock);
1971         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1972                 reclaim_stat->recent_scanned[0] /= 2;
1973                 reclaim_stat->recent_rotated[0] /= 2;
1974         }
1975
1976         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1977                 reclaim_stat->recent_scanned[1] /= 2;
1978                 reclaim_stat->recent_rotated[1] /= 2;
1979         }
1980
1981         /*
1982          * The amount of pressure on anon vs file pages is inversely
1983          * proportional to the fraction of recently scanned pages on
1984          * each list that were recently referenced and in active use.
1985          */
1986         ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1987         ap /= reclaim_stat->recent_rotated[0] + 1;
1988
1989         fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1990         fp /= reclaim_stat->recent_rotated[1] + 1;
1991         spin_unlock_irq(&mz->zone->lru_lock);
1992
1993         fraction[0] = ap;
1994         fraction[1] = fp;
1995         denominator = ap + fp + 1;
1996 out:
1997         for_each_evictable_lru(lru) {
1998                 int file = is_file_lru(lru);
1999                 unsigned long scan;
2000
2001                 scan = zone_nr_lru_pages(mz, lru);
2002                 if (priority || noswap) {
2003                         scan >>= priority;
2004                         if (!scan && force_scan)
2005                                 scan = SWAP_CLUSTER_MAX;
2006                         scan = div64_u64(scan * fraction[file], denominator);
2007                 }
2008                 nr[lru] = scan;
2009         }
2010 }
2011
2012 /*
2013  * Reclaim/compaction depends on a number of pages being freed. To avoid
2014  * disruption to the system, a small number of order-0 pages continue to be
2015  * rotated and reclaimed in the normal fashion. However, by the time we get
2016  * back to the allocator and call try_to_compact_zone(), we ensure that
2017  * there are enough free pages for it to be likely successful
2018  */
2019 static inline bool should_continue_reclaim(struct mem_cgroup_zone *mz,
2020                                         unsigned long nr_reclaimed,
2021                                         unsigned long nr_scanned,
2022                                         struct scan_control *sc)
2023 {
2024         unsigned long pages_for_compaction;
2025         unsigned long inactive_lru_pages;
2026
2027         /* If not in reclaim/compaction mode, stop */
2028         if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
2029                 return false;
2030
2031         /* Consider stopping depending on scan and reclaim activity */
2032         if (sc->gfp_mask & __GFP_REPEAT) {
2033                 /*
2034                  * For __GFP_REPEAT allocations, stop reclaiming if the
2035                  * full LRU list has been scanned and we are still failing
2036                  * to reclaim pages. This full LRU scan is potentially
2037                  * expensive but a __GFP_REPEAT caller really wants to succeed
2038                  */
2039                 if (!nr_reclaimed && !nr_scanned)
2040                         return false;
2041         } else {
2042                 /*
2043                  * For non-__GFP_REPEAT allocations which can presumably
2044                  * fail without consequence, stop if we failed to reclaim
2045                  * any pages from the last SWAP_CLUSTER_MAX number of
2046                  * pages that were scanned. This will return to the
2047                  * caller faster at the risk reclaim/compaction and
2048                  * the resulting allocation attempt fails
2049                  */
2050                 if (!nr_reclaimed)
2051                         return false;
2052         }
2053
2054         /*
2055          * If we have not reclaimed enough pages for compaction and the
2056          * inactive lists are large enough, continue reclaiming
2057          */
2058         pages_for_compaction = (2UL << sc->order);
2059         inactive_lru_pages = zone_nr_lru_pages(mz, LRU_INACTIVE_FILE);
2060         if (nr_swap_pages > 0)
2061                 inactive_lru_pages += zone_nr_lru_pages(mz, LRU_INACTIVE_ANON);
2062         if (sc->nr_reclaimed < pages_for_compaction &&
2063                         inactive_lru_pages > pages_for_compaction)
2064                 return true;
2065
2066         /* If compaction would go ahead or the allocation would succeed, stop */
2067         switch (compaction_suitable(mz->zone, sc->order)) {
2068         case COMPACT_PARTIAL:
2069         case COMPACT_CONTINUE:
2070                 return false;
2071         default:
2072                 return true;
2073         }
2074 }
2075
2076 /*
2077  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
2078  */
2079 static void shrink_mem_cgroup_zone(int priority, struct mem_cgroup_zone *mz,
2080                                    struct scan_control *sc)
2081 {
2082         unsigned long nr[NR_LRU_LISTS];
2083         unsigned long nr_to_scan;
2084         enum lru_list lru;
2085         unsigned long nr_reclaimed, nr_scanned;
2086         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2087         struct blk_plug plug;
2088
2089 restart:
2090         nr_reclaimed = 0;
2091         nr_scanned = sc->nr_scanned;
2092         get_scan_count(mz, sc, nr, priority);
2093
2094         blk_start_plug(&plug);
2095         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2096                                         nr[LRU_INACTIVE_FILE]) {
2097                 for_each_evictable_lru(lru) {
2098                         if (nr[lru]) {
2099                                 nr_to_scan = min_t(unsigned long,
2100                                                    nr[lru], SWAP_CLUSTER_MAX);
2101                                 nr[lru] -= nr_to_scan;
2102
2103                                 nr_reclaimed += shrink_list(lru, nr_to_scan,
2104                                                             mz, sc, priority);
2105                         }
2106                 }
2107                 /*
2108                  * On large memory systems, scan >> priority can become
2109                  * really large. This is fine for the starting priority;
2110                  * we want to put equal scanning pressure on each zone.
2111                  * However, if the VM has a harder time of freeing pages,
2112                  * with multiple processes reclaiming pages, the total
2113                  * freeing target can get unreasonably large.
2114                  */
2115                 if (nr_reclaimed >= nr_to_reclaim)
2116                         nr_to_reclaim = 0;
2117                 else
2118                         nr_to_reclaim -= nr_reclaimed;
2119
2120                 if (!nr_to_reclaim && priority < DEF_PRIORITY)
2121                         break;
2122         }
2123         blk_finish_plug(&plug);
2124         sc->nr_reclaimed += nr_reclaimed;
2125
2126         /*
2127          * Even if we did not try to evict anon pages at all, we want to
2128          * rebalance the anon lru active/inactive ratio.
2129          */
2130         if (inactive_anon_is_low(mz))
2131                 shrink_active_list(SWAP_CLUSTER_MAX, mz, sc, priority, 0);
2132
2133         /* reclaim/compaction might need reclaim to continue */
2134         if (should_continue_reclaim(mz, nr_reclaimed,
2135                                         sc->nr_scanned - nr_scanned, sc))
2136                 goto restart;
2137
2138         throttle_vm_writeout(sc->gfp_mask);
2139 }
2140
2141 static void shrink_zone(int priority, struct zone *zone,
2142                         struct scan_control *sc)
2143 {
2144         struct mem_cgroup *root = sc->target_mem_cgroup;
2145         struct mem_cgroup_reclaim_cookie reclaim = {
2146                 .zone = zone,
2147                 .priority = priority,
2148         };
2149         struct mem_cgroup *memcg;
2150
2151         memcg = mem_cgroup_iter(root, NULL, &reclaim);
2152         do {
2153                 struct mem_cgroup_zone mz = {
2154                         .mem_cgroup = memcg,
2155                         .zone = zone,
2156                 };
2157
2158                 shrink_mem_cgroup_zone(priority, &mz, sc);
2159                 /*
2160                  * Limit reclaim has historically picked one memcg and
2161                  * scanned it with decreasing priority levels until
2162                  * nr_to_reclaim had been reclaimed.  This priority
2163                  * cycle is thus over after a single memcg.
2164                  *
2165                  * Direct reclaim and kswapd, on the other hand, have
2166                  * to scan all memory cgroups to fulfill the overall
2167                  * scan target for the zone.
2168                  */
2169                 if (!global_reclaim(sc)) {
2170                         mem_cgroup_iter_break(root, memcg);
2171                         break;
2172                 }
2173                 memcg = mem_cgroup_iter(root, memcg, &reclaim);
2174         } while (memcg);
2175 }
2176
2177 /* Returns true if compaction should go ahead for a high-order request */
2178 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2179 {
2180         unsigned long balance_gap, watermark;
2181         bool watermark_ok;
2182
2183         /* Do not consider compaction for orders reclaim is meant to satisfy */
2184         if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
2185                 return false;
2186
2187         /*
2188          * Compaction takes time to run and there are potentially other
2189          * callers using the pages just freed. Continue reclaiming until
2190          * there is a buffer of free pages available to give compaction
2191          * a reasonable chance of completing and allocating the page
2192          */
2193         balance_gap = min(low_wmark_pages(zone),
2194                 (zone->present_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2195                         KSWAPD_ZONE_BALANCE_GAP_RATIO);
2196         watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
2197         watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2198
2199         /*
2200          * If compaction is deferred, reclaim up to a point where
2201          * compaction will have a chance of success when re-enabled
2202          */
2203         if (compaction_deferred(zone))
2204                 return watermark_ok;
2205
2206         /* If compaction is not ready to start, keep reclaiming */
2207         if (!compaction_suitable(zone, sc->order))
2208                 return false;
2209
2210         return watermark_ok;
2211 }
2212
2213 /*
2214  * This is the direct reclaim path, for page-allocating processes.  We only
2215  * try to reclaim pages from zones which will satisfy the caller's allocation
2216  * request.
2217  *
2218  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2219  * Because:
2220  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2221  *    allocation or
2222  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2223  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2224  *    zone defense algorithm.
2225  *
2226  * If a zone is deemed to be full of pinned pages then just give it a light
2227  * scan then give up on it.
2228  *
2229  * This function returns true if a zone is being reclaimed for a costly
2230  * high-order allocation and compaction is ready to begin. This indicates to
2231  * the caller that it should consider retrying the allocation instead of
2232  * further reclaim.
2233  */
2234 static bool shrink_zones(int priority, struct zonelist *zonelist,
2235                                         struct scan_control *sc)
2236 {
2237         struct zoneref *z;
2238         struct zone *zone;
2239         unsigned long nr_soft_reclaimed;
2240         unsigned long nr_soft_scanned;
2241         bool aborted_reclaim = false;
2242
2243         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2244                                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2245                 if (!populated_zone(zone))
2246                         continue;
2247                 /*
2248                  * Take care memory controller reclaiming has small influence
2249                  * to global LRU.
2250                  */
2251                 if (global_reclaim(sc)) {
2252                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2253                                 continue;
2254                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2255                                 continue;       /* Let kswapd poll it */
2256                         if (COMPACTION_BUILD) {
2257                                 /*
2258                                  * If we already have plenty of memory free for
2259                                  * compaction in this zone, don't free any more.
2260                                  * Even though compaction is invoked for any
2261                                  * non-zero order, only frequent costly order
2262                                  * reclamation is disruptive enough to become a
2263                                  * noticable problem, like transparent huge page
2264                                  * allocations.
2265                                  */
2266                                 if (compaction_ready(zone, sc)) {
2267                                         aborted_reclaim = true;
2268                                         continue;
2269                                 }
2270                         }
2271                         /*
2272                          * This steals pages from memory cgroups over softlimit
2273                          * and returns the number of reclaimed pages and
2274                          * scanned pages. This works for global memory pressure
2275                          * and balancing, not for a memcg's limit.
2276                          */
2277                         nr_soft_scanned = 0;
2278                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2279                                                 sc->order, sc->gfp_mask,
2280                                                 &nr_soft_scanned);
2281                         sc->nr_reclaimed += nr_soft_reclaimed;
2282                         sc->nr_scanned += nr_soft_scanned;
2283                         /* need some check for avoid more shrink_zone() */
2284                 }
2285
2286                 shrink_zone(priority, zone, sc);
2287         }
2288
2289         return aborted_reclaim;
2290 }
2291
2292 static bool zone_reclaimable(struct zone *zone)
2293 {
2294         return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2295 }
2296
2297 /* All zones in zonelist are unreclaimable? */
2298 static bool all_unreclaimable(struct zonelist *zonelist,
2299                 struct scan_control *sc)
2300 {
2301         struct zoneref *z;
2302         struct zone *zone;
2303
2304         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2305                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2306                 if (!populated_zone(zone))
2307                         continue;
2308                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2309                         continue;
2310                 if (!zone->all_unreclaimable)
2311                         return false;
2312         }
2313
2314         return true;
2315 }
2316
2317 /*
2318  * This is the main entry point to direct page reclaim.
2319  *
2320  * If a full scan of the inactive list fails to free enough memory then we
2321  * are "out of memory" and something needs to be killed.
2322  *
2323  * If the caller is !__GFP_FS then the probability of a failure is reasonably
2324  * high - the zone may be full of dirty or under-writeback pages, which this
2325  * caller can't do much about.  We kick the writeback threads and take explicit
2326  * naps in the hope that some of these pages can be written.  But if the
2327  * allocating task holds filesystem locks which prevent writeout this might not
2328  * work, and the allocation attempt will fail.
2329  *
2330  * returns:     0, if no pages reclaimed
2331  *              else, the number of pages reclaimed
2332  */
2333 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2334                                         struct scan_control *sc,
2335                                         struct shrink_control *shrink)
2336 {
2337         int priority;
2338         unsigned long total_scanned = 0;
2339         struct reclaim_state *reclaim_state = current->reclaim_state;
2340         struct zoneref *z;
2341         struct zone *zone;
2342         unsigned long writeback_threshold;
2343         bool aborted_reclaim;
2344
2345         get_mems_allowed();
2346         delayacct_freepages_start();
2347
2348         if (global_reclaim(sc))
2349                 count_vm_event(ALLOCSTALL);
2350
2351         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2352                 sc->nr_scanned = 0;
2353                 if (!priority)
2354                         disable_swap_token(sc->target_mem_cgroup);
2355                 aborted_reclaim = shrink_zones(priority, zonelist, sc);
2356
2357                 /*
2358                  * Don't shrink slabs when reclaiming memory from
2359                  * over limit cgroups
2360                  */
2361                 if (global_reclaim(sc)) {
2362                         unsigned long lru_pages = 0;
2363                         for_each_zone_zonelist(zone, z, zonelist,
2364                                         gfp_zone(sc->gfp_mask)) {
2365                                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2366                                         continue;
2367
2368                                 lru_pages += zone_reclaimable_pages(zone);
2369                         }
2370
2371                         shrink_slab(shrink, sc->nr_scanned, lru_pages);
2372                         if (reclaim_state) {
2373                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2374                                 reclaim_state->reclaimed_slab = 0;
2375                         }
2376                 }
2377                 total_scanned += sc->nr_scanned;
2378                 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2379                         goto out;
2380
2381                 /*
2382                  * Try to write back as many pages as we just scanned.  This
2383                  * tends to cause slow streaming writers to write data to the
2384                  * disk smoothly, at the dirtying rate, which is nice.   But
2385                  * that's undesirable in laptop mode, where we *want* lumpy
2386                  * writeout.  So in laptop mode, write out the whole world.
2387                  */
2388                 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2389                 if (total_scanned > writeback_threshold) {
2390                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2391                                                 WB_REASON_TRY_TO_FREE_PAGES);
2392                         sc->may_writepage = 1;
2393                 }
2394
2395                 /* Take a nap, wait for some writeback to complete */
2396                 if (!sc->hibernation_mode && sc->nr_scanned &&
2397                     priority < DEF_PRIORITY - 2) {
2398                         struct zone *preferred_zone;
2399
2400                         first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2401                                                 &cpuset_current_mems_allowed,
2402                                                 &preferred_zone);
2403                         wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2404                 }
2405         }
2406
2407 out:
2408         delayacct_freepages_end();
2409         put_mems_allowed();
2410
2411         if (sc->nr_reclaimed)
2412                 return sc->nr_reclaimed;
2413
2414         /*
2415          * As hibernation is going on, kswapd is freezed so that it can't mark
2416          * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2417          * check.
2418          */
2419         if (oom_killer_disabled)
2420                 return 0;
2421
2422         /* Aborted reclaim to try compaction? don't OOM, then */
2423         if (aborted_reclaim)
2424                 return 1;
2425
2426         /* top priority shrink_zones still had more to do? don't OOM, then */
2427         if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
2428                 return 1;
2429
2430         return 0;
2431 }
2432
2433 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2434                                 gfp_t gfp_mask, nodemask_t *nodemask)
2435 {
2436         unsigned long nr_reclaimed;
2437         struct scan_control sc = {
2438                 .gfp_mask = gfp_mask,
2439                 .may_writepage = !laptop_mode,
2440                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2441                 .may_unmap = 1,
2442                 .may_swap = 1,
2443                 .order = order,
2444                 .target_mem_cgroup = NULL,
2445                 .nodemask = nodemask,
2446         };
2447         struct shrink_control shrink = {
2448                 .gfp_mask = sc.gfp_mask,
2449         };
2450
2451         trace_mm_vmscan_direct_reclaim_begin(order,
2452                                 sc.may_writepage,
2453                                 gfp_mask);
2454
2455         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2456
2457         trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2458
2459         return nr_reclaimed;
2460 }
2461
2462 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2463
2464 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
2465                                                 gfp_t gfp_mask, bool noswap,
2466                                                 struct zone *zone,
2467                                                 unsigned long *nr_scanned)
2468 {
2469         struct scan_control sc = {
2470                 .nr_scanned = 0,
2471                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2472                 .may_writepage = !laptop_mode,
2473                 .may_unmap = 1,
2474                 .may_swap = !noswap,
2475                 .order = 0,
2476                 .target_mem_cgroup = memcg,
2477         };
2478         struct mem_cgroup_zone mz = {
2479                 .mem_cgroup = memcg,
2480                 .zone = zone,
2481         };
2482
2483         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2484                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2485
2486         trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2487                                                       sc.may_writepage,
2488                                                       sc.gfp_mask);
2489
2490         /*
2491          * NOTE: Although we can get the priority field, using it
2492          * here is not a good idea, since it limits the pages we can scan.
2493          * if we don't reclaim here, the shrink_zone from balance_pgdat
2494          * will pick up pages from other mem cgroup's as well. We hack
2495          * the priority and make it zero.
2496          */
2497         shrink_mem_cgroup_zone(0, &mz, &sc);
2498
2499         trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2500
2501         *nr_scanned = sc.nr_scanned;
2502         return sc.nr_reclaimed;
2503 }
2504
2505 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2506                                            gfp_t gfp_mask,
2507                                            bool noswap)
2508 {
2509         struct zonelist *zonelist;
2510         unsigned long nr_reclaimed;
2511         int nid;
2512         struct scan_control sc = {
2513                 .may_writepage = !laptop_mode,
2514                 .may_unmap = 1,
2515                 .may_swap = !noswap,
2516                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2517                 .order = 0,
2518                 .target_mem_cgroup = memcg,
2519                 .nodemask = NULL, /* we don't care the placement */
2520                 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2521                                 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2522         };
2523         struct shrink_control shrink = {
2524                 .gfp_mask = sc.gfp_mask,
2525         };
2526
2527         /*
2528          * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2529          * take care of from where we get pages. So the node where we start the
2530          * scan does not need to be the current node.
2531          */
2532         nid = mem_cgroup_select_victim_node(memcg);
2533
2534         zonelist = NODE_DATA(nid)->node_zonelists;
2535
2536         trace_mm_vmscan_memcg_reclaim_begin(0,
2537                                             sc.may_writepage,
2538                                             sc.gfp_mask);
2539
2540         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2541
2542         trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2543
2544         return nr_reclaimed;
2545 }
2546 #endif
2547
2548 static void age_active_anon(struct zone *zone, struct scan_control *sc,
2549                             int priority)
2550 {
2551         struct mem_cgroup *memcg;
2552
2553         if (!total_swap_pages)
2554                 return;
2555
2556         memcg = mem_cgroup_iter(NULL, NULL, NULL);
2557         do {
2558                 struct mem_cgroup_zone mz = {
2559                         .mem_cgroup = memcg,
2560                         .zone = zone,
2561                 };
2562
2563                 if (inactive_anon_is_low(&mz))
2564                         shrink_active_list(SWAP_CLUSTER_MAX, &mz,
2565                                            sc, priority, 0);
2566
2567                 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2568         } while (memcg);
2569 }
2570
2571 /*
2572  * pgdat_balanced is used when checking if a node is balanced for high-order
2573  * allocations. Only zones that meet watermarks and are in a zone allowed
2574  * by the callers classzone_idx are added to balanced_pages. The total of
2575  * balanced pages must be at least 25% of the zones allowed by classzone_idx
2576  * for the node to be considered balanced. Forcing all zones to be balanced
2577  * for high orders can cause excessive reclaim when there are imbalanced zones.
2578  * The choice of 25% is due to
2579  *   o a 16M DMA zone that is balanced will not balance a zone on any
2580  *     reasonable sized machine
2581  *   o On all other machines, the top zone must be at least a reasonable
2582  *     percentage of the middle zones. For example, on 32-bit x86, highmem
2583  *     would need to be at least 256M for it to be balance a whole node.
2584  *     Similarly, on x86-64 the Normal zone would need to be at least 1G
2585  *     to balance a node on its own. These seemed like reasonable ratios.
2586  */
2587 static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2588                                                 int classzone_idx)
2589 {
2590         unsigned long present_pages = 0;
2591         int i;
2592
2593         for (i = 0; i <= classzone_idx; i++)
2594                 present_pages += pgdat->node_zones[i].present_pages;
2595
2596         /* A special case here: if zone has no page, we think it's balanced */
2597         return balanced_pages >= (present_pages >> 2);
2598 }
2599
2600 /* is kswapd sleeping prematurely? */
2601 static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2602                                         int classzone_idx)
2603 {
2604         int i;
2605         unsigned long balanced = 0;
2606         bool all_zones_ok = true;
2607
2608         /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2609         if (remaining)
2610                 return true;
2611
2612         /* Check the watermark levels */
2613         for (i = 0; i <= classzone_idx; i++) {
2614                 struct zone *zone = pgdat->node_zones + i;
2615
2616                 if (!populated_zone(zone))
2617                         continue;
2618
2619                 /*
2620                  * balance_pgdat() skips over all_unreclaimable after
2621                  * DEF_PRIORITY. Effectively, it considers them balanced so
2622                  * they must be considered balanced here as well if kswapd
2623                  * is to sleep
2624                  */
2625                 if (zone->all_unreclaimable) {
2626                         balanced += zone->present_pages;
2627                         continue;
2628                 }
2629
2630                 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2631                                                         i, 0))
2632                         all_zones_ok = false;
2633                 else
2634                         balanced += zone->present_pages;
2635         }
2636
2637         /*
2638          * For high-order requests, the balanced zones must contain at least
2639          * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2640          * must be balanced
2641          */
2642         if (order)
2643                 return !pgdat_balanced(pgdat, balanced, classzone_idx);
2644         else
2645                 return !all_zones_ok;
2646 }
2647
2648 /*
2649  * For kswapd, balance_pgdat() will work across all this node's zones until
2650  * they are all at high_wmark_pages(zone).
2651  *
2652  * Returns the final order kswapd was reclaiming at
2653  *
2654  * There is special handling here for zones which are full of pinned pages.
2655  * This can happen if the pages are all mlocked, or if they are all used by
2656  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
2657  * What we do is to detect the case where all pages in the zone have been
2658  * scanned twice and there has been zero successful reclaim.  Mark the zone as
2659  * dead and from now on, only perform a short scan.  Basically we're polling
2660  * the zone for when the problem goes away.
2661  *
2662  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
2663  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2664  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2665  * lower zones regardless of the number of free pages in the lower zones. This
2666  * interoperates with the page allocator fallback scheme to ensure that aging
2667  * of pages is balanced across the zones.
2668  */
2669 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2670                                                         int *classzone_idx)
2671 {
2672         int all_zones_ok;
2673         unsigned long balanced;
2674         int priority;
2675         int i;
2676         int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
2677         unsigned long total_scanned;
2678         struct reclaim_state *reclaim_state = current->reclaim_state;
2679         unsigned long nr_soft_reclaimed;
2680         unsigned long nr_soft_scanned;
2681         struct scan_control sc = {
2682                 .gfp_mask = GFP_KERNEL,
2683                 .may_unmap = 1,
2684                 .may_swap = 1,
2685                 /*
2686                  * kswapd doesn't want to be bailed out while reclaim. because
2687                  * we want to put equal scanning pressure on each zone.
2688                  */
2689                 .nr_to_reclaim = ULONG_MAX,
2690                 .order = order,
2691                 .target_mem_cgroup = NULL,
2692         };
2693         struct shrink_control shrink = {
2694                 .gfp_mask = sc.gfp_mask,
2695         };
2696 loop_again:
2697         total_scanned = 0;
2698         sc.nr_reclaimed = 0;
2699         sc.may_writepage = !laptop_mode;
2700         count_vm_event(PAGEOUTRUN);
2701
2702         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2703                 unsigned long lru_pages = 0;
2704                 int has_under_min_watermark_zone = 0;
2705
2706                 /* The swap token gets in the way of swapout... */
2707                 if (!priority)
2708                         disable_swap_token(NULL);
2709
2710                 all_zones_ok = 1;
2711                 balanced = 0;
2712
2713                 /*
2714                  * Scan in the highmem->dma direction for the highest
2715                  * zone which needs scanning
2716                  */
2717                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2718                         struct zone *zone = pgdat->node_zones + i;
2719
2720                         if (!populated_zone(zone))
2721                                 continue;
2722
2723                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2724                                 continue;
2725
2726                         /*
2727                          * Do some background aging of the anon list, to give
2728                          * pages a chance to be referenced before reclaiming.
2729                          */
2730                         age_active_anon(zone, &sc, priority);
2731
2732                         if (!zone_watermark_ok_safe(zone, order,
2733                                         high_wmark_pages(zone), 0, 0)) {
2734                                 end_zone = i;
2735                                 break;
2736                         } else {
2737                                 /* If balanced, clear the congested flag */
2738                                 zone_clear_flag(zone, ZONE_CONGESTED);
2739                         }
2740                 }
2741                 if (i < 0)
2742                         goto out;
2743
2744                 for (i = 0; i <= end_zone; i++) {
2745                         struct zone *zone = pgdat->node_zones + i;
2746
2747                         lru_pages += zone_reclaimable_pages(zone);
2748                 }
2749
2750                 /*
2751                  * Now scan the zone in the dma->highmem direction, stopping
2752                  * at the last zone which needs scanning.
2753                  *
2754                  * We do this because the page allocator works in the opposite
2755                  * direction.  This prevents the page allocator from allocating
2756                  * pages behind kswapd's direction of progress, which would
2757                  * cause too much scanning of the lower zones.
2758                  */
2759                 for (i = 0; i <= end_zone; i++) {
2760                         struct zone *zone = pgdat->node_zones + i;
2761                         int nr_slab;
2762                         unsigned long balance_gap;
2763
2764                         if (!populated_zone(zone))
2765                                 continue;
2766
2767                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2768                                 continue;
2769
2770                         sc.nr_scanned = 0;
2771
2772                         nr_soft_scanned = 0;
2773                         /*
2774                          * Call soft limit reclaim before calling shrink_zone.
2775                          */
2776                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2777                                                         order, sc.gfp_mask,
2778                                                         &nr_soft_scanned);
2779                         sc.nr_reclaimed += nr_soft_reclaimed;
2780                         total_scanned += nr_soft_scanned;
2781
2782                         /*
2783                          * We put equal pressure on every zone, unless
2784                          * one zone has way too many pages free
2785                          * already. The "too many pages" is defined
2786                          * as the high wmark plus a "gap" where the
2787                          * gap is either the low watermark or 1%
2788                          * of the zone, whichever is smaller.
2789                          */
2790                         balance_gap = min(low_wmark_pages(zone),
2791                                 (zone->present_pages +
2792                                         KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2793                                 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2794                         if (!zone_watermark_ok_safe(zone, order,
2795                                         high_wmark_pages(zone) + balance_gap,
2796                                         end_zone, 0)) {
2797                                 shrink_zone(priority, zone, &sc);
2798
2799                                 reclaim_state->reclaimed_slab = 0;
2800                                 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2801                                 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2802                                 total_scanned += sc.nr_scanned;
2803
2804                                 if (nr_slab == 0 && !zone_reclaimable(zone))
2805                                         zone->all_unreclaimable = 1;
2806                         }
2807
2808                         /*
2809                          * If we've done a decent amount of scanning and
2810                          * the reclaim ratio is low, start doing writepage
2811                          * even in laptop mode
2812                          */
2813                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2814                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2815                                 sc.may_writepage = 1;
2816
2817                         if (zone->all_unreclaimable) {
2818                                 if (end_zone && end_zone == i)
2819                                         end_zone--;
2820                                 continue;
2821                         }
2822
2823                         if (!zone_watermark_ok_safe(zone, order,
2824                                         high_wmark_pages(zone), end_zone, 0)) {
2825                                 all_zones_ok = 0;
2826                                 /*
2827                                  * We are still under min water mark.  This
2828                                  * means that we have a GFP_ATOMIC allocation
2829                                  * failure risk. Hurry up!
2830                                  */
2831                                 if (!zone_watermark_ok_safe(zone, order,
2832                                             min_wmark_pages(zone), end_zone, 0))
2833                                         has_under_min_watermark_zone = 1;
2834                         } else {
2835                                 /*
2836                                  * If a zone reaches its high watermark,
2837                                  * consider it to be no longer congested. It's
2838                                  * possible there are dirty pages backed by
2839                                  * congested BDIs but as pressure is relieved,
2840                                  * spectulatively avoid congestion waits
2841                                  */
2842                                 zone_clear_flag(zone, ZONE_CONGESTED);
2843                                 if (i <= *classzone_idx)
2844                                         balanced += zone->present_pages;
2845                         }
2846
2847                 }
2848                 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2849                         break;          /* kswapd: all done */
2850                 /*
2851                  * OK, kswapd is getting into trouble.  Take a nap, then take
2852                  * another pass across the zones.
2853                  */
2854                 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2855                         if (has_under_min_watermark_zone)
2856                                 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2857                         else
2858                                 congestion_wait(BLK_RW_ASYNC, HZ/10);
2859                 }
2860
2861                 /*
2862                  * We do this so kswapd doesn't build up large priorities for
2863                  * example when it is freeing in parallel with allocators. It
2864                  * matches the direct reclaim path behaviour in terms of impact
2865                  * on zone->*_priority.
2866                  */
2867                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2868                         break;
2869         }
2870 out:
2871
2872         /*
2873          * order-0: All zones must meet high watermark for a balanced node
2874          * high-order: Balanced zones must make up at least 25% of the node
2875          *             for the node to be balanced
2876          */
2877         if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2878                 cond_resched();
2879
2880                 try_to_freeze();
2881
2882                 /*
2883                  * Fragmentation may mean that the system cannot be
2884                  * rebalanced for high-order allocations in all zones.
2885                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2886                  * it means the zones have been fully scanned and are still
2887                  * not balanced. For high-order allocations, there is
2888                  * little point trying all over again as kswapd may
2889                  * infinite loop.
2890                  *
2891                  * Instead, recheck all watermarks at order-0 as they
2892                  * are the most important. If watermarks are ok, kswapd will go
2893                  * back to sleep. High-order users can still perform direct
2894                  * reclaim if they wish.
2895                  */
2896                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2897                         order = sc.order = 0;
2898
2899                 goto loop_again;
2900         }
2901
2902         /*
2903          * If kswapd was reclaiming at a higher order, it has the option of
2904          * sleeping without all zones being balanced. Before it does, it must
2905          * ensure that the watermarks for order-0 on *all* zones are met and
2906          * that the congestion flags are cleared. The congestion flag must
2907          * be cleared as kswapd is the only mechanism that clears the flag
2908          * and it is potentially going to sleep here.
2909          */
2910         if (order) {
2911                 for (i = 0; i <= end_zone; i++) {
2912                         struct zone *zone = pgdat->node_zones + i;
2913
2914                         if (!populated_zone(zone))
2915                                 continue;
2916
2917                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2918                                 continue;
2919
2920                         /* Confirm the zone is balanced for order-0 */
2921                         if (!zone_watermark_ok(zone, 0,
2922                                         high_wmark_pages(zone), 0, 0)) {
2923                                 order = sc.order = 0;
2924                                 goto loop_again;
2925                         }
2926
2927                         /* If balanced, clear the congested flag */
2928                         zone_clear_flag(zone, ZONE_CONGESTED);
2929                         if (i <= *classzone_idx)
2930                                 balanced += zone->present_pages;
2931                 }
2932         }
2933
2934         /*
2935          * Return the order we were reclaiming at so sleeping_prematurely()
2936          * makes a decision on the order we were last reclaiming at. However,
2937          * if another caller entered the allocator slow path while kswapd
2938          * was awake, order will remain at the higher level
2939          */
2940         *classzone_idx = end_zone;
2941         return order;
2942 }
2943
2944 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2945 {
2946         long remaining = 0;
2947         DEFINE_WAIT(wait);
2948
2949         if (freezing(current) || kthread_should_stop())
2950                 return;
2951
2952         prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2953
2954         /* Try to sleep for a short interval */
2955         if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2956                 remaining = schedule_timeout(HZ/10);
2957                 finish_wait(&pgdat->kswapd_wait, &wait);
2958                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2959         }
2960
2961         /*
2962          * After a short sleep, check if it was a premature sleep. If not, then
2963          * go fully to sleep until explicitly woken up.
2964          */
2965         if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2966                 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2967
2968                 /*
2969                  * vmstat counters are not perfectly accurate and the estimated
2970                  * value for counters such as NR_FREE_PAGES can deviate from the
2971                  * true value by nr_online_cpus * threshold. To avoid the zone
2972                  * watermarks being breached while under pressure, we reduce the
2973                  * per-cpu vmstat threshold while kswapd is awake and restore
2974                  * them before going back to sleep.
2975                  */
2976                 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2977                 schedule();
2978                 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2979         } else {
2980                 if (remaining)
2981                         count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2982                 else
2983                         count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2984         }
2985         finish_wait(&pgdat->kswapd_wait, &wait);
2986 }
2987
2988 /*
2989  * The background pageout daemon, started as a kernel thread
2990  * from the init process.
2991  *
2992  * This basically trickles out pages so that we have _some_
2993  * free memory available even if there is no other activity
2994  * that frees anything up. This is needed for things like routing
2995  * etc, where we otherwise might have all activity going on in
2996  * asynchronous contexts that cannot page things out.
2997  *
2998  * If there are applications that are active memory-allocators
2999  * (most normal use), this basically shouldn't matter.
3000  */
3001 static int kswapd(void *p)
3002 {
3003         unsigned long order, new_order;
3004         unsigned balanced_order;
3005         int classzone_idx, new_classzone_idx;
3006         int balanced_classzone_idx;
3007         pg_data_t *pgdat = (pg_data_t*)p;
3008         struct task_struct *tsk = current;
3009
3010         struct reclaim_state reclaim_state = {
3011                 .reclaimed_slab = 0,
3012         };
3013         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3014
3015         lockdep_set_current_reclaim_state(GFP_KERNEL);
3016
3017         if (!cpumask_empty(cpumask))
3018                 set_cpus_allowed_ptr(tsk, cpumask);
3019         current->reclaim_state = &reclaim_state;
3020
3021         /*
3022          * Tell the memory management that we're a "memory allocator",
3023          * and that if we need more memory we should get access to it
3024          * regardless (see "__alloc_pages()"). "kswapd" should
3025          * never get caught in the normal page freeing logic.
3026          *
3027          * (Kswapd normally doesn't need memory anyway, but sometimes
3028          * you need a small amount of memory in order to be able to
3029          * page out something else, and this flag essentially protects
3030          * us from recursively trying to free more memory as we're
3031          * trying to free the first piece of memory in the first place).
3032          */
3033         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3034         set_freezable();
3035
3036         order = new_order = 0;
3037         balanced_order = 0;
3038         classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
3039         balanced_classzone_idx = classzone_idx;
3040         for ( ; ; ) {
3041                 int ret;
3042
3043                 /*
3044                  * If the last balance_pgdat was unsuccessful it's unlikely a
3045                  * new request of a similar or harder type will succeed soon
3046                  * so consider going to sleep on the basis we reclaimed at
3047                  */
3048                 if (balanced_classzone_idx >= new_classzone_idx &&
3049                                         balanced_order == new_order) {
3050                         new_order = pgdat->kswapd_max_order;
3051                         new_classzone_idx = pgdat->classzone_idx;
3052                         pgdat->kswapd_max_order =  0;
3053                         pgdat->classzone_idx = pgdat->nr_zones - 1;
3054                 }
3055
3056                 if (order < new_order || classzone_idx > new_classzone_idx) {
3057                         /*
3058                          * Don't sleep if someone wants a larger 'order'
3059                          * allocation or has tigher zone constraints
3060                          */
3061                         order = new_order;
3062                         classzone_idx = new_classzone_idx;
3063                 } else {
3064                         kswapd_try_to_sleep(pgdat, balanced_order,
3065                                                 balanced_classzone_idx);
3066                         order = pgdat->kswapd_max_order;
3067                         classzone_idx = pgdat->classzone_idx;
3068                         new_order = order;
3069                         new_classzone_idx = classzone_idx;
3070                         pgdat->kswapd_max_order = 0;
3071                         pgdat->classzone_idx = pgdat->nr_zones - 1;
3072                 }
3073
3074                 ret = try_to_freeze();
3075                 if (kthread_should_stop())
3076                         break;
3077
3078                 /*
3079                  * We can speed up thawing tasks if we don't call balance_pgdat
3080                  * after returning from the refrigerator
3081                  */
3082                 if (!ret) {
3083                         trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
3084                         balanced_classzone_idx = classzone_idx;
3085                         balanced_order = balance_pgdat(pgdat, order,
3086                                                 &balanced_classzone_idx);
3087                 }
3088         }
3089         return 0;
3090 }
3091
3092 /*
3093  * A zone is low on free memory, so wake its kswapd task to service it.
3094  */
3095 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3096 {
3097         pg_data_t *pgdat;
3098
3099         if (!populated_zone(zone))
3100                 return;
3101
3102         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
3103                 return;
3104         pgdat = zone->zone_pgdat;
3105         if (pgdat->kswapd_max_order < order) {
3106                 pgdat->kswapd_max_order = order;
3107                 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3108         }
3109         if (!waitqueue_active(&pgdat->kswapd_wait))
3110                 return;
3111         if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
3112                 return;
3113
3114         trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3115         wake_up_interruptible(&pgdat->kswapd_wait);
3116 }
3117
3118 /*
3119  * The reclaimable count would be mostly accurate.
3120  * The less reclaimable pages may be
3121  * - mlocked pages, which will be moved to unevictable list when encountered
3122  * - mapped pages, which may require several travels to be reclaimed
3123  * - dirty pages, which is not "instantly" reclaimable
3124  */
3125 unsigned long global_reclaimable_pages(void)
3126 {
3127         int nr;
3128
3129         nr = global_page_state(NR_ACTIVE_FILE) +
3130              global_page_state(NR_INACTIVE_FILE);
3131
3132         if (nr_swap_pages > 0)
3133                 nr += global_page_state(NR_ACTIVE_ANON) +
3134                       global_page_state(NR_INACTIVE_ANON);
3135
3136         return nr;
3137 }
3138
3139 unsigned long zone_reclaimable_pages(struct zone *zone)
3140 {
3141         int nr;
3142
3143         nr = zone_page_state(zone, NR_ACTIVE_FILE) +
3144              zone_page_state(zone, NR_INACTIVE_FILE);
3145
3146         if (nr_swap_pages > 0)
3147                 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
3148                       zone_page_state(zone, NR_INACTIVE_ANON);
3149
3150         return nr;
3151 }
3152
3153 #ifdef CONFIG_HIBERNATION
3154 /*
3155  * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3156  * freed pages.
3157  *
3158  * Rather than trying to age LRUs the aim is to preserve the overall
3159  * LRU order by reclaiming preferentially
3160  * inactive > active > active referenced > active mapped
3161  */
3162 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3163 {
3164         struct reclaim_state reclaim_state;
3165         struct scan_control sc = {
3166                 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3167                 .may_swap = 1,
3168                 .may_unmap = 1,
3169                 .may_writepage = 1,
3170                 .nr_to_reclaim = nr_to_reclaim,
3171                 .hibernation_mode = 1,
3172                 .order = 0,
3173         };
3174         struct shrink_control shrink = {
3175                 .gfp_mask = sc.gfp_mask,
3176         };
3177         struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3178         struct task_struct *p = current;
3179         unsigned long nr_reclaimed;
3180
3181         p->flags |= PF_MEMALLOC;
3182         lockdep_set_current_reclaim_state(sc.gfp_mask);
3183         reclaim_state.reclaimed_slab = 0;
3184         p->reclaim_state = &reclaim_state;
3185
3186         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
3187
3188         p->reclaim_state = NULL;
3189         lockdep_clear_current_reclaim_state();
3190         p->flags &= ~PF_MEMALLOC;
3191
3192         return nr_reclaimed;
3193 }
3194 #endif /* CONFIG_HIBERNATION */
3195
3196 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3197    not required for correctness.  So if the last cpu in a node goes
3198    away, we get changed to run anywhere: as the first one comes back,
3199    restore their cpu bindings. */
3200 static int __devinit cpu_callback(struct notifier_block *nfb,
3201                                   unsigned long action, void *hcpu)
3202 {
3203         int nid;
3204
3205         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3206                 for_each_node_state(nid, N_HIGH_MEMORY) {
3207                         pg_data_t *pgdat = NODE_DATA(nid);
3208                         const struct cpumask *mask;
3209
3210                         mask = cpumask_of_node(pgdat->node_id);
3211
3212                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3213                                 /* One of our CPUs online: restore mask */
3214                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3215                 }
3216         }
3217         return NOTIFY_OK;
3218 }
3219
3220 /*
3221  * This kswapd start function will be called by init and node-hot-add.
3222  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3223  */
3224 int kswapd_run(int nid)
3225 {
3226         pg_data_t *pgdat = NODE_DATA(nid);
3227         int ret = 0;
3228
3229         if (pgdat->kswapd)
3230                 return 0;
3231
3232         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3233         if (IS_ERR(pgdat->kswapd)) {
3234                 /* failure at boot is fatal */
3235                 BUG_ON(system_state == SYSTEM_BOOTING);
3236                 printk("Failed to start kswapd on node %d\n",nid);
3237                 ret = -1;
3238         }
3239         return ret;
3240 }
3241
3242 /*
3243  * Called by memory hotplug when all memory in a node is offlined.
3244  */
3245 void kswapd_stop(int nid)
3246 {
3247         struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3248
3249         if (kswapd)
3250                 kthread_stop(kswapd);
3251 }
3252
3253 static int __init kswapd_init(void)
3254 {
3255         int nid;
3256
3257         swap_setup();
3258         for_each_node_state(nid, N_HIGH_MEMORY)
3259                 kswapd_run(nid);
3260         hotcpu_notifier(cpu_callback, 0);
3261         return 0;
3262 }
3263
3264 module_init(kswapd_init)
3265
3266 #ifdef CONFIG_NUMA
3267 /*
3268  * Zone reclaim mode
3269  *
3270  * If non-zero call zone_reclaim when the number of free pages falls below
3271  * the watermarks.
3272  */
3273 int zone_reclaim_mode __read_mostly;
3274
3275 #define RECLAIM_OFF 0
3276 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
3277 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
3278 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
3279
3280 /*
3281  * Priority for ZONE_RECLAIM. This determines the fraction of pages
3282  * of a node considered for each zone_reclaim. 4 scans 1/16th of
3283  * a zone.
3284  */
3285 #define ZONE_RECLAIM_PRIORITY 4
3286
3287 /*
3288  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3289  * occur.
3290  */
3291 int sysctl_min_unmapped_ratio = 1;
3292
3293 /*
3294  * If the number of slab pages in a zone grows beyond this percentage then
3295  * slab reclaim needs to occur.
3296  */
3297 int sysctl_min_slab_ratio = 5;
3298
3299 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3300 {
3301         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3302         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3303                 zone_page_state(zone, NR_ACTIVE_FILE);
3304
3305         /*
3306          * It's possible for there to be more file mapped pages than
3307          * accounted for by the pages on the file LRU lists because
3308          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3309          */
3310         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3311 }
3312
3313 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3314 static long zone_pagecache_reclaimable(struct zone *zone)
3315 {
3316         long nr_pagecache_reclaimable;
3317         long delta = 0;
3318
3319         /*
3320          * If RECLAIM_SWAP is set, then all file pages are considered
3321          * potentially reclaimable. Otherwise, we have to worry about
3322          * pages like swapcache and zone_unmapped_file_pages() provides
3323          * a better estimate
3324          */
3325         if (zone_reclaim_mode & RECLAIM_SWAP)
3326                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3327         else
3328                 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3329
3330         /* If we can't clean pages, remove dirty pages from consideration */
3331         if (!(zone_reclaim_mode & RECLAIM_WRITE))
3332                 delta += zone_page_state(zone, NR_FILE_DIRTY);
3333
3334         /* Watch for any possible underflows due to delta */
3335         if (unlikely(delta > nr_pagecache_reclaimable))
3336                 delta = nr_pagecache_reclaimable;
3337
3338         return nr_pagecache_reclaimable - delta;
3339 }
3340
3341 /*
3342  * Try to free up some pages from this zone through reclaim.
3343  */
3344 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3345 {
3346         /* Minimum pages needed in order to stay on node */
3347         const unsigned long nr_pages = 1 << order;
3348         struct task_struct *p = current;
3349         struct reclaim_state reclaim_state;
3350         int priority;
3351         struct scan_control sc = {
3352                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3353                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3354                 .may_swap = 1,
3355                 .nr_to_reclaim = max_t(unsigned long, nr_pages,
3356                                        SWAP_CLUSTER_MAX),
3357                 .gfp_mask = gfp_mask,
3358                 .order = order,
3359         };
3360         struct shrink_control shrink = {
3361                 .gfp_mask = sc.gfp_mask,
3362         };
3363         unsigned long nr_slab_pages0, nr_slab_pages1;
3364
3365         cond_resched();
3366         /*
3367          * We need to be able to allocate from the reserves for RECLAIM_SWAP
3368          * and we also need to be able to write out pages for RECLAIM_WRITE
3369          * and RECLAIM_SWAP.
3370          */
3371         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3372         lockdep_set_current_reclaim_state(gfp_mask);
3373         reclaim_state.reclaimed_slab = 0;
3374         p->reclaim_state = &reclaim_state;
3375
3376         if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3377                 /*
3378                  * Free memory by calling shrink zone with increasing
3379                  * priorities until we have enough memory freed.
3380                  */
3381                 priority = ZONE_RECLAIM_PRIORITY;
3382                 do {
3383                         shrink_zone(priority, zone, &sc);
3384                         priority--;
3385                 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
3386         }
3387
3388         nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3389         if (nr_slab_pages0 > zone->min_slab_pages) {
3390                 /*
3391                  * shrink_slab() does not currently allow us to determine how
3392                  * many pages were freed in this zone. So we take the current
3393                  * number of slab pages and shake the slab until it is reduced
3394                  * by the same nr_pages that we used for reclaiming unmapped
3395                  * pages.
3396                  *
3397                  * Note that shrink_slab will free memory on all zones and may
3398                  * take a long time.
3399                  */
3400                 for (;;) {
3401                         unsigned long lru_pages = zone_reclaimable_pages(zone);
3402
3403                         /* No reclaimable slab or very low memory pressure */
3404                         if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3405                                 break;
3406
3407                         /* Freed enough memory */
3408                         nr_slab_pages1 = zone_page_state(zone,
3409                                                         NR_SLAB_RECLAIMABLE);
3410                         if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3411                                 break;
3412                 }
3413
3414                 /*
3415                  * Update nr_reclaimed by the number of slab pages we
3416                  * reclaimed from this zone.
3417                  */
3418                 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3419                 if (nr_slab_pages1 < nr_slab_pages0)
3420                         sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3421         }
3422
3423         p->reclaim_state = NULL;
3424         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3425         lockdep_clear_current_reclaim_state();
3426         return sc.nr_reclaimed >= nr_pages;
3427 }
3428
3429 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3430 {
3431         int node_id;
3432         int ret;
3433
3434         /*
3435          * Zone reclaim reclaims unmapped file backed pages and
3436          * slab pages if we are over the defined limits.
3437          *
3438          * A small portion of unmapped file backed pages is needed for
3439          * file I/O otherwise pages read by file I/O will be immediately
3440          * thrown out if the zone is overallocated. So we do not reclaim
3441          * if less than a specified percentage of the zone is used by
3442          * unmapped file backed pages.
3443          */
3444         if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3445             zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3446                 return ZONE_RECLAIM_FULL;
3447
3448         if (zone->all_unreclaimable)
3449                 return ZONE_RECLAIM_FULL;
3450
3451         /*
3452          * Do not scan if the allocation should not be delayed.
3453          */
3454         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3455                 return ZONE_RECLAIM_NOSCAN;
3456
3457         /*
3458          * Only run zone reclaim on the local zone or on zones that do not
3459          * have associated processors. This will favor the local processor
3460          * over remote processors and spread off node memory allocations
3461          * as wide as possible.
3462          */
3463         node_id = zone_to_nid(zone);
3464         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3465                 return ZONE_RECLAIM_NOSCAN;
3466
3467         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3468                 return ZONE_RECLAIM_NOSCAN;
3469
3470         ret = __zone_reclaim(zone, gfp_mask, order);
3471         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3472
3473         if (!ret)
3474                 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3475
3476         return ret;
3477 }
3478 #endif
3479
3480 /*
3481  * page_evictable - test whether a page is evictable
3482  * @page: the page to test
3483  * @vma: the VMA in which the page is or will be mapped, may be NULL
3484  *
3485  * Test whether page is evictable--i.e., should be placed on active/inactive
3486  * lists vs unevictable list.  The vma argument is !NULL when called from the
3487  * fault path to determine how to instantate a new page.
3488  *
3489  * Reasons page might not be evictable:
3490  * (1) page's mapping marked unevictable
3491  * (2) page is part of an mlocked VMA
3492  *
3493  */
3494 int page_evictable(struct page *page, struct vm_area_struct *vma)
3495 {
3496
3497         if (mapping_unevictable(page_mapping(page)))
3498                 return 0;
3499
3500         if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
3501                 return 0;
3502
3503         return 1;
3504 }
3505
3506 #ifdef CONFIG_SHMEM
3507 /**
3508  * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3509  * @pages:      array of pages to check
3510  * @nr_pages:   number of pages to check
3511  *
3512  * Checks pages for evictability and moves them to the appropriate lru list.
3513  *
3514  * This function is only used for SysV IPC SHM_UNLOCK.
3515  */
3516 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3517 {
3518         struct lruvec *lruvec;
3519         struct zone *zone = NULL;
3520         int pgscanned = 0;
3521         int pgrescued = 0;
3522         int i;
3523
3524         for (i = 0; i < nr_pages; i++) {
3525                 struct page *page = pages[i];
3526                 struct zone *pagezone;
3527
3528                 pgscanned++;
3529                 pagezone = page_zone(page);
3530                 if (pagezone != zone) {
3531                         if (zone)
3532                                 spin_unlock_irq(&zone->lru_lock);
3533                         zone = pagezone;
3534                         spin_lock_irq(&zone->lru_lock);
3535                 }
3536
3537                 if (!PageLRU(page) || !PageUnevictable(page))
3538                         continue;
3539
3540                 if (page_evictable(page, NULL)) {
3541                         enum lru_list lru = page_lru_base_type(page);
3542
3543                         VM_BUG_ON(PageActive(page));
3544                         ClearPageUnevictable(page);
3545                         __dec_zone_state(zone, NR_UNEVICTABLE);
3546                         lruvec = mem_cgroup_lru_move_lists(zone, page,
3547                                                 LRU_UNEVICTABLE, lru);
3548                         list_move(&page->lru, &lruvec->lists[lru]);
3549                         __inc_zone_state(zone, NR_INACTIVE_ANON + lru);
3550                         pgrescued++;
3551                 }
3552         }
3553
3554         if (zone) {
3555                 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3556                 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3557                 spin_unlock_irq(&zone->lru_lock);
3558         }
3559 }
3560 #endif /* CONFIG_SHMEM */
3561
3562 static void warn_scan_unevictable_pages(void)
3563 {
3564         printk_once(KERN_WARNING
3565                     "%s: The scan_unevictable_pages sysctl/node-interface has been "
3566                     "disabled for lack of a legitimate use case.  If you have "
3567                     "one, please send an email to linux-mm@kvack.org.\n",
3568                     current->comm);
3569 }
3570
3571 /*
3572  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
3573  * all nodes' unevictable lists for evictable pages
3574  */
3575 unsigned long scan_unevictable_pages;
3576
3577 int scan_unevictable_handler(struct ctl_table *table, int write,
3578                            void __user *buffer,
3579                            size_t *length, loff_t *ppos)
3580 {
3581         warn_scan_unevictable_pages();
3582         proc_doulongvec_minmax(table, write, buffer, length, ppos);
3583         scan_unevictable_pages = 0;
3584         return 0;
3585 }
3586
3587 #ifdef CONFIG_NUMA
3588 /*
3589  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
3590  * a specified node's per zone unevictable lists for evictable pages.
3591  */
3592
3593 static ssize_t read_scan_unevictable_node(struct device *dev,
3594                                           struct device_attribute *attr,
3595                                           char *buf)
3596 {
3597         warn_scan_unevictable_pages();
3598         return sprintf(buf, "0\n");     /* always zero; should fit... */
3599 }
3600
3601 static ssize_t write_scan_unevictable_node(struct device *dev,
3602                                            struct device_attribute *attr,
3603                                         const char *buf, size_t count)
3604 {
3605         warn_scan_unevictable_pages();
3606         return 1;
3607 }
3608
3609
3610 static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3611                         read_scan_unevictable_node,
3612                         write_scan_unevictable_node);
3613
3614 int scan_unevictable_register_node(struct node *node)
3615 {
3616         return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
3617 }
3618
3619 void scan_unevictable_unregister_node(struct node *node)
3620 {
3621         device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);
3622 }
3623 #endif