mm: vmscan: when reclaiming for compaction, ensure there are sufficient free pages...
[linux-2.6.git] / mm / vmscan.c
1 /*
2  *  linux/mm/vmscan.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  *
6  *  Swap reorganised 29.12.95, Stephen Tweedie.
7  *  kswapd added: 7.1.96  sct
8  *  Removed kswapd_ctl limits, and swap out as many pages as needed
9  *  to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10  *  Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11  *  Multiqueue VM started 5.8.00, Rik van Riel.
12  */
13
14 #include <linux/mm.h>
15 #include <linux/module.h>
16 #include <linux/gfp.h>
17 #include <linux/kernel_stat.h>
18 #include <linux/swap.h>
19 #include <linux/pagemap.h>
20 #include <linux/init.h>
21 #include <linux/highmem.h>
22 #include <linux/vmstat.h>
23 #include <linux/file.h>
24 #include <linux/writeback.h>
25 #include <linux/blkdev.h>
26 #include <linux/buffer_head.h>  /* for try_to_release_page(),
27                                         buffer_heads_over_limit */
28 #include <linux/mm_inline.h>
29 #include <linux/pagevec.h>
30 #include <linux/backing-dev.h>
31 #include <linux/rmap.h>
32 #include <linux/topology.h>
33 #include <linux/cpu.h>
34 #include <linux/cpuset.h>
35 #include <linux/compaction.h>
36 #include <linux/notifier.h>
37 #include <linux/rwsem.h>
38 #include <linux/delay.h>
39 #include <linux/kthread.h>
40 #include <linux/freezer.h>
41 #include <linux/memcontrol.h>
42 #include <linux/delayacct.h>
43 #include <linux/sysctl.h>
44 #include <linux/oom.h>
45 #include <linux/prefetch.h>
46
47 #include <asm/tlbflush.h>
48 #include <asm/div64.h>
49
50 #include <linux/swapops.h>
51
52 #include "internal.h"
53
54 #define CREATE_TRACE_POINTS
55 #include <trace/events/vmscan.h>
56
57 /*
58  * reclaim_mode determines how the inactive list is shrunk
59  * RECLAIM_MODE_SINGLE: Reclaim only order-0 pages
60  * RECLAIM_MODE_ASYNC:  Do not block
61  * RECLAIM_MODE_SYNC:   Allow blocking e.g. call wait_on_page_writeback
62  * RECLAIM_MODE_LUMPYRECLAIM: For high-order allocations, take a reference
63  *                      page from the LRU and reclaim all pages within a
64  *                      naturally aligned range
65  * RECLAIM_MODE_COMPACTION: For high-order allocations, reclaim a number of
66  *                      order-0 pages and then compact the zone
67  */
68 typedef unsigned __bitwise__ reclaim_mode_t;
69 #define RECLAIM_MODE_SINGLE             ((__force reclaim_mode_t)0x01u)
70 #define RECLAIM_MODE_ASYNC              ((__force reclaim_mode_t)0x02u)
71 #define RECLAIM_MODE_SYNC               ((__force reclaim_mode_t)0x04u)
72 #define RECLAIM_MODE_LUMPYRECLAIM       ((__force reclaim_mode_t)0x08u)
73 #define RECLAIM_MODE_COMPACTION         ((__force reclaim_mode_t)0x10u)
74
75 struct scan_control {
76         /* Incremented by the number of inactive pages that were scanned */
77         unsigned long nr_scanned;
78
79         /* Number of pages freed so far during a call to shrink_zones() */
80         unsigned long nr_reclaimed;
81
82         /* How many pages shrink_list() should reclaim */
83         unsigned long nr_to_reclaim;
84
85         unsigned long hibernation_mode;
86
87         /* This context's GFP mask */
88         gfp_t gfp_mask;
89
90         int may_writepage;
91
92         /* Can mapped pages be reclaimed? */
93         int may_unmap;
94
95         /* Can pages be swapped as part of reclaim? */
96         int may_swap;
97
98         int order;
99
100         /*
101          * Intend to reclaim enough continuous memory rather than reclaim
102          * enough amount of memory. i.e, mode for high order allocation.
103          */
104         reclaim_mode_t reclaim_mode;
105
106         /*
107          * The memory cgroup that hit its limit and as a result is the
108          * primary target of this reclaim invocation.
109          */
110         struct mem_cgroup *target_mem_cgroup;
111
112         /*
113          * Nodemask of nodes allowed by the caller. If NULL, all nodes
114          * are scanned.
115          */
116         nodemask_t      *nodemask;
117 };
118
119 struct mem_cgroup_zone {
120         struct mem_cgroup *mem_cgroup;
121         struct zone *zone;
122 };
123
124 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
125
126 #ifdef ARCH_HAS_PREFETCH
127 #define prefetch_prev_lru_page(_page, _base, _field)                    \
128         do {                                                            \
129                 if ((_page)->lru.prev != _base) {                       \
130                         struct page *prev;                              \
131                                                                         \
132                         prev = lru_to_page(&(_page->lru));              \
133                         prefetch(&prev->_field);                        \
134                 }                                                       \
135         } while (0)
136 #else
137 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
138 #endif
139
140 #ifdef ARCH_HAS_PREFETCHW
141 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
142         do {                                                            \
143                 if ((_page)->lru.prev != _base) {                       \
144                         struct page *prev;                              \
145                                                                         \
146                         prev = lru_to_page(&(_page->lru));              \
147                         prefetchw(&prev->_field);                       \
148                 }                                                       \
149         } while (0)
150 #else
151 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
152 #endif
153
154 /*
155  * From 0 .. 100.  Higher means more swappy.
156  */
157 int vm_swappiness = 60;
158 long vm_total_pages;    /* The total number of pages which the VM controls */
159
160 static LIST_HEAD(shrinker_list);
161 static DECLARE_RWSEM(shrinker_rwsem);
162
163 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
164 static bool global_reclaim(struct scan_control *sc)
165 {
166         return !sc->target_mem_cgroup;
167 }
168
169 static bool scanning_global_lru(struct mem_cgroup_zone *mz)
170 {
171         return !mz->mem_cgroup;
172 }
173 #else
174 static bool global_reclaim(struct scan_control *sc)
175 {
176         return true;
177 }
178
179 static bool scanning_global_lru(struct mem_cgroup_zone *mz)
180 {
181         return true;
182 }
183 #endif
184
185 static struct zone_reclaim_stat *get_reclaim_stat(struct mem_cgroup_zone *mz)
186 {
187         if (!scanning_global_lru(mz))
188                 return mem_cgroup_get_reclaim_stat(mz->mem_cgroup, mz->zone);
189
190         return &mz->zone->reclaim_stat;
191 }
192
193 static unsigned long zone_nr_lru_pages(struct mem_cgroup_zone *mz,
194                                        enum lru_list lru)
195 {
196         if (!scanning_global_lru(mz))
197                 return mem_cgroup_zone_nr_lru_pages(mz->mem_cgroup,
198                                                     zone_to_nid(mz->zone),
199                                                     zone_idx(mz->zone),
200                                                     BIT(lru));
201
202         return zone_page_state(mz->zone, NR_LRU_BASE + lru);
203 }
204
205
206 /*
207  * Add a shrinker callback to be called from the vm
208  */
209 void register_shrinker(struct shrinker *shrinker)
210 {
211         atomic_long_set(&shrinker->nr_in_batch, 0);
212         down_write(&shrinker_rwsem);
213         list_add_tail(&shrinker->list, &shrinker_list);
214         up_write(&shrinker_rwsem);
215 }
216 EXPORT_SYMBOL(register_shrinker);
217
218 /*
219  * Remove one
220  */
221 void unregister_shrinker(struct shrinker *shrinker)
222 {
223         down_write(&shrinker_rwsem);
224         list_del(&shrinker->list);
225         up_write(&shrinker_rwsem);
226 }
227 EXPORT_SYMBOL(unregister_shrinker);
228
229 static inline int do_shrinker_shrink(struct shrinker *shrinker,
230                                      struct shrink_control *sc,
231                                      unsigned long nr_to_scan)
232 {
233         sc->nr_to_scan = nr_to_scan;
234         return (*shrinker->shrink)(shrinker, sc);
235 }
236
237 #define SHRINK_BATCH 128
238 /*
239  * Call the shrink functions to age shrinkable caches
240  *
241  * Here we assume it costs one seek to replace a lru page and that it also
242  * takes a seek to recreate a cache object.  With this in mind we age equal
243  * percentages of the lru and ageable caches.  This should balance the seeks
244  * generated by these structures.
245  *
246  * If the vm encountered mapped pages on the LRU it increase the pressure on
247  * slab to avoid swapping.
248  *
249  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
250  *
251  * `lru_pages' represents the number of on-LRU pages in all the zones which
252  * are eligible for the caller's allocation attempt.  It is used for balancing
253  * slab reclaim versus page reclaim.
254  *
255  * Returns the number of slab objects which we shrunk.
256  */
257 unsigned long shrink_slab(struct shrink_control *shrink,
258                           unsigned long nr_pages_scanned,
259                           unsigned long lru_pages)
260 {
261         struct shrinker *shrinker;
262         unsigned long ret = 0;
263
264         if (nr_pages_scanned == 0)
265                 nr_pages_scanned = SWAP_CLUSTER_MAX;
266
267         if (!down_read_trylock(&shrinker_rwsem)) {
268                 /* Assume we'll be able to shrink next time */
269                 ret = 1;
270                 goto out;
271         }
272
273         list_for_each_entry(shrinker, &shrinker_list, list) {
274                 unsigned long long delta;
275                 long total_scan;
276                 long max_pass;
277                 int shrink_ret = 0;
278                 long nr;
279                 long new_nr;
280                 long batch_size = shrinker->batch ? shrinker->batch
281                                                   : SHRINK_BATCH;
282
283                 max_pass = do_shrinker_shrink(shrinker, shrink, 0);
284                 if (max_pass <= 0)
285                         continue;
286
287                 /*
288                  * copy the current shrinker scan count into a local variable
289                  * and zero it so that other concurrent shrinker invocations
290                  * don't also do this scanning work.
291                  */
292                 nr = atomic_long_xchg(&shrinker->nr_in_batch, 0);
293
294                 total_scan = nr;
295                 delta = (4 * nr_pages_scanned) / shrinker->seeks;
296                 delta *= max_pass;
297                 do_div(delta, lru_pages + 1);
298                 total_scan += delta;
299                 if (total_scan < 0) {
300                         printk(KERN_ERR "shrink_slab: %pF negative objects to "
301                                "delete nr=%ld\n",
302                                shrinker->shrink, total_scan);
303                         total_scan = max_pass;
304                 }
305
306                 /*
307                  * We need to avoid excessive windup on filesystem shrinkers
308                  * due to large numbers of GFP_NOFS allocations causing the
309                  * shrinkers to return -1 all the time. This results in a large
310                  * nr being built up so when a shrink that can do some work
311                  * comes along it empties the entire cache due to nr >>>
312                  * max_pass.  This is bad for sustaining a working set in
313                  * memory.
314                  *
315                  * Hence only allow the shrinker to scan the entire cache when
316                  * a large delta change is calculated directly.
317                  */
318                 if (delta < max_pass / 4)
319                         total_scan = min(total_scan, max_pass / 2);
320
321                 /*
322                  * Avoid risking looping forever due to too large nr value:
323                  * never try to free more than twice the estimate number of
324                  * freeable entries.
325                  */
326                 if (total_scan > max_pass * 2)
327                         total_scan = max_pass * 2;
328
329                 trace_mm_shrink_slab_start(shrinker, shrink, nr,
330                                         nr_pages_scanned, lru_pages,
331                                         max_pass, delta, total_scan);
332
333                 while (total_scan >= batch_size) {
334                         int nr_before;
335
336                         nr_before = do_shrinker_shrink(shrinker, shrink, 0);
337                         shrink_ret = do_shrinker_shrink(shrinker, shrink,
338                                                         batch_size);
339                         if (shrink_ret == -1)
340                                 break;
341                         if (shrink_ret < nr_before)
342                                 ret += nr_before - shrink_ret;
343                         count_vm_events(SLABS_SCANNED, batch_size);
344                         total_scan -= batch_size;
345
346                         cond_resched();
347                 }
348
349                 /*
350                  * move the unused scan count back into the shrinker in a
351                  * manner that handles concurrent updates. If we exhausted the
352                  * scan, there is no need to do an update.
353                  */
354                 if (total_scan > 0)
355                         new_nr = atomic_long_add_return(total_scan,
356                                         &shrinker->nr_in_batch);
357                 else
358                         new_nr = atomic_long_read(&shrinker->nr_in_batch);
359
360                 trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr);
361         }
362         up_read(&shrinker_rwsem);
363 out:
364         cond_resched();
365         return ret;
366 }
367
368 static void set_reclaim_mode(int priority, struct scan_control *sc,
369                                    bool sync)
370 {
371         reclaim_mode_t syncmode = sync ? RECLAIM_MODE_SYNC : RECLAIM_MODE_ASYNC;
372
373         /*
374          * Initially assume we are entering either lumpy reclaim or
375          * reclaim/compaction.Depending on the order, we will either set the
376          * sync mode or just reclaim order-0 pages later.
377          */
378         if (COMPACTION_BUILD)
379                 sc->reclaim_mode = RECLAIM_MODE_COMPACTION;
380         else
381                 sc->reclaim_mode = RECLAIM_MODE_LUMPYRECLAIM;
382
383         /*
384          * Avoid using lumpy reclaim or reclaim/compaction if possible by
385          * restricting when its set to either costly allocations or when
386          * under memory pressure
387          */
388         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
389                 sc->reclaim_mode |= syncmode;
390         else if (sc->order && priority < DEF_PRIORITY - 2)
391                 sc->reclaim_mode |= syncmode;
392         else
393                 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
394 }
395
396 static void reset_reclaim_mode(struct scan_control *sc)
397 {
398         sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
399 }
400
401 static inline int is_page_cache_freeable(struct page *page)
402 {
403         /*
404          * A freeable page cache page is referenced only by the caller
405          * that isolated the page, the page cache radix tree and
406          * optional buffer heads at page->private.
407          */
408         return page_count(page) - page_has_private(page) == 2;
409 }
410
411 static int may_write_to_queue(struct backing_dev_info *bdi,
412                               struct scan_control *sc)
413 {
414         if (current->flags & PF_SWAPWRITE)
415                 return 1;
416         if (!bdi_write_congested(bdi))
417                 return 1;
418         if (bdi == current->backing_dev_info)
419                 return 1;
420
421         /* lumpy reclaim for hugepage often need a lot of write */
422         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
423                 return 1;
424         return 0;
425 }
426
427 /*
428  * We detected a synchronous write error writing a page out.  Probably
429  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
430  * fsync(), msync() or close().
431  *
432  * The tricky part is that after writepage we cannot touch the mapping: nothing
433  * prevents it from being freed up.  But we have a ref on the page and once
434  * that page is locked, the mapping is pinned.
435  *
436  * We're allowed to run sleeping lock_page() here because we know the caller has
437  * __GFP_FS.
438  */
439 static void handle_write_error(struct address_space *mapping,
440                                 struct page *page, int error)
441 {
442         lock_page(page);
443         if (page_mapping(page) == mapping)
444                 mapping_set_error(mapping, error);
445         unlock_page(page);
446 }
447
448 /* possible outcome of pageout() */
449 typedef enum {
450         /* failed to write page out, page is locked */
451         PAGE_KEEP,
452         /* move page to the active list, page is locked */
453         PAGE_ACTIVATE,
454         /* page has been sent to the disk successfully, page is unlocked */
455         PAGE_SUCCESS,
456         /* page is clean and locked */
457         PAGE_CLEAN,
458 } pageout_t;
459
460 /*
461  * pageout is called by shrink_page_list() for each dirty page.
462  * Calls ->writepage().
463  */
464 static pageout_t pageout(struct page *page, struct address_space *mapping,
465                          struct scan_control *sc)
466 {
467         /*
468          * If the page is dirty, only perform writeback if that write
469          * will be non-blocking.  To prevent this allocation from being
470          * stalled by pagecache activity.  But note that there may be
471          * stalls if we need to run get_block().  We could test
472          * PagePrivate for that.
473          *
474          * If this process is currently in __generic_file_aio_write() against
475          * this page's queue, we can perform writeback even if that
476          * will block.
477          *
478          * If the page is swapcache, write it back even if that would
479          * block, for some throttling. This happens by accident, because
480          * swap_backing_dev_info is bust: it doesn't reflect the
481          * congestion state of the swapdevs.  Easy to fix, if needed.
482          */
483         if (!is_page_cache_freeable(page))
484                 return PAGE_KEEP;
485         if (!mapping) {
486                 /*
487                  * Some data journaling orphaned pages can have
488                  * page->mapping == NULL while being dirty with clean buffers.
489                  */
490                 if (page_has_private(page)) {
491                         if (try_to_free_buffers(page)) {
492                                 ClearPageDirty(page);
493                                 printk("%s: orphaned page\n", __func__);
494                                 return PAGE_CLEAN;
495                         }
496                 }
497                 return PAGE_KEEP;
498         }
499         if (mapping->a_ops->writepage == NULL)
500                 return PAGE_ACTIVATE;
501         if (!may_write_to_queue(mapping->backing_dev_info, sc))
502                 return PAGE_KEEP;
503
504         if (clear_page_dirty_for_io(page)) {
505                 int res;
506                 struct writeback_control wbc = {
507                         .sync_mode = WB_SYNC_NONE,
508                         .nr_to_write = SWAP_CLUSTER_MAX,
509                         .range_start = 0,
510                         .range_end = LLONG_MAX,
511                         .for_reclaim = 1,
512                 };
513
514                 SetPageReclaim(page);
515                 res = mapping->a_ops->writepage(page, &wbc);
516                 if (res < 0)
517                         handle_write_error(mapping, page, res);
518                 if (res == AOP_WRITEPAGE_ACTIVATE) {
519                         ClearPageReclaim(page);
520                         return PAGE_ACTIVATE;
521                 }
522
523                 if (!PageWriteback(page)) {
524                         /* synchronous write or broken a_ops? */
525                         ClearPageReclaim(page);
526                 }
527                 trace_mm_vmscan_writepage(page,
528                         trace_reclaim_flags(page, sc->reclaim_mode));
529                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
530                 return PAGE_SUCCESS;
531         }
532
533         return PAGE_CLEAN;
534 }
535
536 /*
537  * Same as remove_mapping, but if the page is removed from the mapping, it
538  * gets returned with a refcount of 0.
539  */
540 static int __remove_mapping(struct address_space *mapping, struct page *page)
541 {
542         BUG_ON(!PageLocked(page));
543         BUG_ON(mapping != page_mapping(page));
544
545         spin_lock_irq(&mapping->tree_lock);
546         /*
547          * The non racy check for a busy page.
548          *
549          * Must be careful with the order of the tests. When someone has
550          * a ref to the page, it may be possible that they dirty it then
551          * drop the reference. So if PageDirty is tested before page_count
552          * here, then the following race may occur:
553          *
554          * get_user_pages(&page);
555          * [user mapping goes away]
556          * write_to(page);
557          *                              !PageDirty(page)    [good]
558          * SetPageDirty(page);
559          * put_page(page);
560          *                              !page_count(page)   [good, discard it]
561          *
562          * [oops, our write_to data is lost]
563          *
564          * Reversing the order of the tests ensures such a situation cannot
565          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
566          * load is not satisfied before that of page->_count.
567          *
568          * Note that if SetPageDirty is always performed via set_page_dirty,
569          * and thus under tree_lock, then this ordering is not required.
570          */
571         if (!page_freeze_refs(page, 2))
572                 goto cannot_free;
573         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
574         if (unlikely(PageDirty(page))) {
575                 page_unfreeze_refs(page, 2);
576                 goto cannot_free;
577         }
578
579         if (PageSwapCache(page)) {
580                 swp_entry_t swap = { .val = page_private(page) };
581                 __delete_from_swap_cache(page);
582                 spin_unlock_irq(&mapping->tree_lock);
583                 swapcache_free(swap, page);
584         } else {
585                 void (*freepage)(struct page *);
586
587                 freepage = mapping->a_ops->freepage;
588
589                 __delete_from_page_cache(page);
590                 spin_unlock_irq(&mapping->tree_lock);
591                 mem_cgroup_uncharge_cache_page(page);
592
593                 if (freepage != NULL)
594                         freepage(page);
595         }
596
597         return 1;
598
599 cannot_free:
600         spin_unlock_irq(&mapping->tree_lock);
601         return 0;
602 }
603
604 /*
605  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
606  * someone else has a ref on the page, abort and return 0.  If it was
607  * successfully detached, return 1.  Assumes the caller has a single ref on
608  * this page.
609  */
610 int remove_mapping(struct address_space *mapping, struct page *page)
611 {
612         if (__remove_mapping(mapping, page)) {
613                 /*
614                  * Unfreezing the refcount with 1 rather than 2 effectively
615                  * drops the pagecache ref for us without requiring another
616                  * atomic operation.
617                  */
618                 page_unfreeze_refs(page, 1);
619                 return 1;
620         }
621         return 0;
622 }
623
624 /**
625  * putback_lru_page - put previously isolated page onto appropriate LRU list
626  * @page: page to be put back to appropriate lru list
627  *
628  * Add previously isolated @page to appropriate LRU list.
629  * Page may still be unevictable for other reasons.
630  *
631  * lru_lock must not be held, interrupts must be enabled.
632  */
633 void putback_lru_page(struct page *page)
634 {
635         int lru;
636         int active = !!TestClearPageActive(page);
637         int was_unevictable = PageUnevictable(page);
638
639         VM_BUG_ON(PageLRU(page));
640
641 redo:
642         ClearPageUnevictable(page);
643
644         if (page_evictable(page, NULL)) {
645                 /*
646                  * For evictable pages, we can use the cache.
647                  * In event of a race, worst case is we end up with an
648                  * unevictable page on [in]active list.
649                  * We know how to handle that.
650                  */
651                 lru = active + page_lru_base_type(page);
652                 lru_cache_add_lru(page, lru);
653         } else {
654                 /*
655                  * Put unevictable pages directly on zone's unevictable
656                  * list.
657                  */
658                 lru = LRU_UNEVICTABLE;
659                 add_page_to_unevictable_list(page);
660                 /*
661                  * When racing with an mlock or AS_UNEVICTABLE clearing
662                  * (page is unlocked) make sure that if the other thread
663                  * does not observe our setting of PG_lru and fails
664                  * isolation/check_move_unevictable_page,
665                  * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
666                  * the page back to the evictable list.
667                  *
668                  * The other side is TestClearPageMlocked() or shmem_lock().
669                  */
670                 smp_mb();
671         }
672
673         /*
674          * page's status can change while we move it among lru. If an evictable
675          * page is on unevictable list, it never be freed. To avoid that,
676          * check after we added it to the list, again.
677          */
678         if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
679                 if (!isolate_lru_page(page)) {
680                         put_page(page);
681                         goto redo;
682                 }
683                 /* This means someone else dropped this page from LRU
684                  * So, it will be freed or putback to LRU again. There is
685                  * nothing to do here.
686                  */
687         }
688
689         if (was_unevictable && lru != LRU_UNEVICTABLE)
690                 count_vm_event(UNEVICTABLE_PGRESCUED);
691         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
692                 count_vm_event(UNEVICTABLE_PGCULLED);
693
694         put_page(page);         /* drop ref from isolate */
695 }
696
697 enum page_references {
698         PAGEREF_RECLAIM,
699         PAGEREF_RECLAIM_CLEAN,
700         PAGEREF_KEEP,
701         PAGEREF_ACTIVATE,
702 };
703
704 static enum page_references page_check_references(struct page *page,
705                                                   struct mem_cgroup_zone *mz,
706                                                   struct scan_control *sc)
707 {
708         int referenced_ptes, referenced_page;
709         unsigned long vm_flags;
710
711         referenced_ptes = page_referenced(page, 1, mz->mem_cgroup, &vm_flags);
712         referenced_page = TestClearPageReferenced(page);
713
714         /* Lumpy reclaim - ignore references */
715         if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
716                 return PAGEREF_RECLAIM;
717
718         /*
719          * Mlock lost the isolation race with us.  Let try_to_unmap()
720          * move the page to the unevictable list.
721          */
722         if (vm_flags & VM_LOCKED)
723                 return PAGEREF_RECLAIM;
724
725         if (referenced_ptes) {
726                 if (PageAnon(page))
727                         return PAGEREF_ACTIVATE;
728                 /*
729                  * All mapped pages start out with page table
730                  * references from the instantiating fault, so we need
731                  * to look twice if a mapped file page is used more
732                  * than once.
733                  *
734                  * Mark it and spare it for another trip around the
735                  * inactive list.  Another page table reference will
736                  * lead to its activation.
737                  *
738                  * Note: the mark is set for activated pages as well
739                  * so that recently deactivated but used pages are
740                  * quickly recovered.
741                  */
742                 SetPageReferenced(page);
743
744                 if (referenced_page || referenced_ptes > 1)
745                         return PAGEREF_ACTIVATE;
746
747                 /*
748                  * Activate file-backed executable pages after first usage.
749                  */
750                 if (vm_flags & VM_EXEC)
751                         return PAGEREF_ACTIVATE;
752
753                 return PAGEREF_KEEP;
754         }
755
756         /* Reclaim if clean, defer dirty pages to writeback */
757         if (referenced_page && !PageSwapBacked(page))
758                 return PAGEREF_RECLAIM_CLEAN;
759
760         return PAGEREF_RECLAIM;
761 }
762
763 /*
764  * shrink_page_list() returns the number of reclaimed pages
765  */
766 static unsigned long shrink_page_list(struct list_head *page_list,
767                                       struct mem_cgroup_zone *mz,
768                                       struct scan_control *sc,
769                                       int priority,
770                                       unsigned long *ret_nr_dirty,
771                                       unsigned long *ret_nr_writeback)
772 {
773         LIST_HEAD(ret_pages);
774         LIST_HEAD(free_pages);
775         int pgactivate = 0;
776         unsigned long nr_dirty = 0;
777         unsigned long nr_congested = 0;
778         unsigned long nr_reclaimed = 0;
779         unsigned long nr_writeback = 0;
780
781         cond_resched();
782
783         while (!list_empty(page_list)) {
784                 enum page_references references;
785                 struct address_space *mapping;
786                 struct page *page;
787                 int may_enter_fs;
788
789                 cond_resched();
790
791                 page = lru_to_page(page_list);
792                 list_del(&page->lru);
793
794                 if (!trylock_page(page))
795                         goto keep;
796
797                 VM_BUG_ON(PageActive(page));
798                 VM_BUG_ON(page_zone(page) != mz->zone);
799
800                 sc->nr_scanned++;
801
802                 if (unlikely(!page_evictable(page, NULL)))
803                         goto cull_mlocked;
804
805                 if (!sc->may_unmap && page_mapped(page))
806                         goto keep_locked;
807
808                 /* Double the slab pressure for mapped and swapcache pages */
809                 if (page_mapped(page) || PageSwapCache(page))
810                         sc->nr_scanned++;
811
812                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
813                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
814
815                 if (PageWriteback(page)) {
816                         nr_writeback++;
817                         /*
818                          * Synchronous reclaim cannot queue pages for
819                          * writeback due to the possibility of stack overflow
820                          * but if it encounters a page under writeback, wait
821                          * for the IO to complete.
822                          */
823                         if ((sc->reclaim_mode & RECLAIM_MODE_SYNC) &&
824                             may_enter_fs)
825                                 wait_on_page_writeback(page);
826                         else {
827                                 unlock_page(page);
828                                 goto keep_lumpy;
829                         }
830                 }
831
832                 references = page_check_references(page, mz, sc);
833                 switch (references) {
834                 case PAGEREF_ACTIVATE:
835                         goto activate_locked;
836                 case PAGEREF_KEEP:
837                         goto keep_locked;
838                 case PAGEREF_RECLAIM:
839                 case PAGEREF_RECLAIM_CLEAN:
840                         ; /* try to reclaim the page below */
841                 }
842
843                 /*
844                  * Anonymous process memory has backing store?
845                  * Try to allocate it some swap space here.
846                  */
847                 if (PageAnon(page) && !PageSwapCache(page)) {
848                         if (!(sc->gfp_mask & __GFP_IO))
849                                 goto keep_locked;
850                         if (!add_to_swap(page))
851                                 goto activate_locked;
852                         may_enter_fs = 1;
853                 }
854
855                 mapping = page_mapping(page);
856
857                 /*
858                  * The page is mapped into the page tables of one or more
859                  * processes. Try to unmap it here.
860                  */
861                 if (page_mapped(page) && mapping) {
862                         switch (try_to_unmap(page, TTU_UNMAP)) {
863                         case SWAP_FAIL:
864                                 goto activate_locked;
865                         case SWAP_AGAIN:
866                                 goto keep_locked;
867                         case SWAP_MLOCK:
868                                 goto cull_mlocked;
869                         case SWAP_SUCCESS:
870                                 ; /* try to free the page below */
871                         }
872                 }
873
874                 if (PageDirty(page)) {
875                         nr_dirty++;
876
877                         /*
878                          * Only kswapd can writeback filesystem pages to
879                          * avoid risk of stack overflow but do not writeback
880                          * unless under significant pressure.
881                          */
882                         if (page_is_file_cache(page) &&
883                                         (!current_is_kswapd() || priority >= DEF_PRIORITY - 2)) {
884                                 /*
885                                  * Immediately reclaim when written back.
886                                  * Similar in principal to deactivate_page()
887                                  * except we already have the page isolated
888                                  * and know it's dirty
889                                  */
890                                 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
891                                 SetPageReclaim(page);
892
893                                 goto keep_locked;
894                         }
895
896                         if (references == PAGEREF_RECLAIM_CLEAN)
897                                 goto keep_locked;
898                         if (!may_enter_fs)
899                                 goto keep_locked;
900                         if (!sc->may_writepage)
901                                 goto keep_locked;
902
903                         /* Page is dirty, try to write it out here */
904                         switch (pageout(page, mapping, sc)) {
905                         case PAGE_KEEP:
906                                 nr_congested++;
907                                 goto keep_locked;
908                         case PAGE_ACTIVATE:
909                                 goto activate_locked;
910                         case PAGE_SUCCESS:
911                                 if (PageWriteback(page))
912                                         goto keep_lumpy;
913                                 if (PageDirty(page))
914                                         goto keep;
915
916                                 /*
917                                  * A synchronous write - probably a ramdisk.  Go
918                                  * ahead and try to reclaim the page.
919                                  */
920                                 if (!trylock_page(page))
921                                         goto keep;
922                                 if (PageDirty(page) || PageWriteback(page))
923                                         goto keep_locked;
924                                 mapping = page_mapping(page);
925                         case PAGE_CLEAN:
926                                 ; /* try to free the page below */
927                         }
928                 }
929
930                 /*
931                  * If the page has buffers, try to free the buffer mappings
932                  * associated with this page. If we succeed we try to free
933                  * the page as well.
934                  *
935                  * We do this even if the page is PageDirty().
936                  * try_to_release_page() does not perform I/O, but it is
937                  * possible for a page to have PageDirty set, but it is actually
938                  * clean (all its buffers are clean).  This happens if the
939                  * buffers were written out directly, with submit_bh(). ext3
940                  * will do this, as well as the blockdev mapping.
941                  * try_to_release_page() will discover that cleanness and will
942                  * drop the buffers and mark the page clean - it can be freed.
943                  *
944                  * Rarely, pages can have buffers and no ->mapping.  These are
945                  * the pages which were not successfully invalidated in
946                  * truncate_complete_page().  We try to drop those buffers here
947                  * and if that worked, and the page is no longer mapped into
948                  * process address space (page_count == 1) it can be freed.
949                  * Otherwise, leave the page on the LRU so it is swappable.
950                  */
951                 if (page_has_private(page)) {
952                         if (!try_to_release_page(page, sc->gfp_mask))
953                                 goto activate_locked;
954                         if (!mapping && page_count(page) == 1) {
955                                 unlock_page(page);
956                                 if (put_page_testzero(page))
957                                         goto free_it;
958                                 else {
959                                         /*
960                                          * rare race with speculative reference.
961                                          * the speculative reference will free
962                                          * this page shortly, so we may
963                                          * increment nr_reclaimed here (and
964                                          * leave it off the LRU).
965                                          */
966                                         nr_reclaimed++;
967                                         continue;
968                                 }
969                         }
970                 }
971
972                 if (!mapping || !__remove_mapping(mapping, page))
973                         goto keep_locked;
974
975                 /*
976                  * At this point, we have no other references and there is
977                  * no way to pick any more up (removed from LRU, removed
978                  * from pagecache). Can use non-atomic bitops now (and
979                  * we obviously don't have to worry about waking up a process
980                  * waiting on the page lock, because there are no references.
981                  */
982                 __clear_page_locked(page);
983 free_it:
984                 nr_reclaimed++;
985
986                 /*
987                  * Is there need to periodically free_page_list? It would
988                  * appear not as the counts should be low
989                  */
990                 list_add(&page->lru, &free_pages);
991                 continue;
992
993 cull_mlocked:
994                 if (PageSwapCache(page))
995                         try_to_free_swap(page);
996                 unlock_page(page);
997                 putback_lru_page(page);
998                 reset_reclaim_mode(sc);
999                 continue;
1000
1001 activate_locked:
1002                 /* Not a candidate for swapping, so reclaim swap space. */
1003                 if (PageSwapCache(page) && vm_swap_full())
1004                         try_to_free_swap(page);
1005                 VM_BUG_ON(PageActive(page));
1006                 SetPageActive(page);
1007                 pgactivate++;
1008 keep_locked:
1009                 unlock_page(page);
1010 keep:
1011                 reset_reclaim_mode(sc);
1012 keep_lumpy:
1013                 list_add(&page->lru, &ret_pages);
1014                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
1015         }
1016
1017         /*
1018          * Tag a zone as congested if all the dirty pages encountered were
1019          * backed by a congested BDI. In this case, reclaimers should just
1020          * back off and wait for congestion to clear because further reclaim
1021          * will encounter the same problem
1022          */
1023         if (nr_dirty && nr_dirty == nr_congested && global_reclaim(sc))
1024                 zone_set_flag(mz->zone, ZONE_CONGESTED);
1025
1026         free_hot_cold_page_list(&free_pages, 1);
1027
1028         list_splice(&ret_pages, page_list);
1029         count_vm_events(PGACTIVATE, pgactivate);
1030         *ret_nr_dirty += nr_dirty;
1031         *ret_nr_writeback += nr_writeback;
1032         return nr_reclaimed;
1033 }
1034
1035 /*
1036  * Attempt to remove the specified page from its LRU.  Only take this page
1037  * if it is of the appropriate PageActive status.  Pages which are being
1038  * freed elsewhere are also ignored.
1039  *
1040  * page:        page to consider
1041  * mode:        one of the LRU isolation modes defined above
1042  *
1043  * returns 0 on success, -ve errno on failure.
1044  */
1045 int __isolate_lru_page(struct page *page, isolate_mode_t mode, int file)
1046 {
1047         bool all_lru_mode;
1048         int ret = -EINVAL;
1049
1050         /* Only take pages on the LRU. */
1051         if (!PageLRU(page))
1052                 return ret;
1053
1054         all_lru_mode = (mode & (ISOLATE_ACTIVE|ISOLATE_INACTIVE)) ==
1055                 (ISOLATE_ACTIVE|ISOLATE_INACTIVE);
1056
1057         /*
1058          * When checking the active state, we need to be sure we are
1059          * dealing with comparible boolean values.  Take the logical not
1060          * of each.
1061          */
1062         if (!all_lru_mode && !PageActive(page) != !(mode & ISOLATE_ACTIVE))
1063                 return ret;
1064
1065         if (!all_lru_mode && !!page_is_file_cache(page) != file)
1066                 return ret;
1067
1068         /*
1069          * When this function is being called for lumpy reclaim, we
1070          * initially look into all LRU pages, active, inactive and
1071          * unevictable; only give shrink_page_list evictable pages.
1072          */
1073         if (PageUnevictable(page))
1074                 return ret;
1075
1076         ret = -EBUSY;
1077
1078         /*
1079          * To minimise LRU disruption, the caller can indicate that it only
1080          * wants to isolate pages it will be able to operate on without
1081          * blocking - clean pages for the most part.
1082          *
1083          * ISOLATE_CLEAN means that only clean pages should be isolated. This
1084          * is used by reclaim when it is cannot write to backing storage
1085          *
1086          * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1087          * that it is possible to migrate without blocking
1088          */
1089         if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1090                 /* All the caller can do on PageWriteback is block */
1091                 if (PageWriteback(page))
1092                         return ret;
1093
1094                 if (PageDirty(page)) {
1095                         struct address_space *mapping;
1096
1097                         /* ISOLATE_CLEAN means only clean pages */
1098                         if (mode & ISOLATE_CLEAN)
1099                                 return ret;
1100
1101                         /*
1102                          * Only pages without mappings or that have a
1103                          * ->migratepage callback are possible to migrate
1104                          * without blocking
1105                          */
1106                         mapping = page_mapping(page);
1107                         if (mapping && !mapping->a_ops->migratepage)
1108                                 return ret;
1109                 }
1110         }
1111
1112         if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1113                 return ret;
1114
1115         if (likely(get_page_unless_zero(page))) {
1116                 /*
1117                  * Be careful not to clear PageLRU until after we're
1118                  * sure the page is not being freed elsewhere -- the
1119                  * page release code relies on it.
1120                  */
1121                 ClearPageLRU(page);
1122                 ret = 0;
1123         }
1124
1125         return ret;
1126 }
1127
1128 /*
1129  * zone->lru_lock is heavily contended.  Some of the functions that
1130  * shrink the lists perform better by taking out a batch of pages
1131  * and working on them outside the LRU lock.
1132  *
1133  * For pagecache intensive workloads, this function is the hottest
1134  * spot in the kernel (apart from copy_*_user functions).
1135  *
1136  * Appropriate locks must be held before calling this function.
1137  *
1138  * @nr_to_scan: The number of pages to look through on the list.
1139  * @src:        The LRU list to pull pages off.
1140  * @dst:        The temp list to put pages on to.
1141  * @scanned:    The number of pages that were scanned.
1142  * @order:      The caller's attempted allocation order
1143  * @mode:       One of the LRU isolation modes
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 list_head *src, struct list_head *dst,
1150                 unsigned long *scanned, int order, isolate_mode_t mode,
1151                 int file)
1152 {
1153         unsigned long nr_taken = 0;
1154         unsigned long nr_lumpy_taken = 0;
1155         unsigned long nr_lumpy_dirty = 0;
1156         unsigned long nr_lumpy_failed = 0;
1157         unsigned long scan;
1158
1159         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1160                 struct page *page;
1161                 unsigned long pfn;
1162                 unsigned long end_pfn;
1163                 unsigned long page_pfn;
1164                 int zone_id;
1165
1166                 page = lru_to_page(src);
1167                 prefetchw_prev_lru_page(page, src, flags);
1168
1169                 VM_BUG_ON(!PageLRU(page));
1170
1171                 switch (__isolate_lru_page(page, mode, file)) {
1172                 case 0:
1173                         mem_cgroup_lru_del(page);
1174                         list_move(&page->lru, dst);
1175                         nr_taken += hpage_nr_pages(page);
1176                         break;
1177
1178                 case -EBUSY:
1179                         /* else it is being freed elsewhere */
1180                         list_move(&page->lru, src);
1181                         continue;
1182
1183                 default:
1184                         BUG();
1185                 }
1186
1187                 if (!order)
1188                         continue;
1189
1190                 /*
1191                  * Attempt to take all pages in the order aligned region
1192                  * surrounding the tag page.  Only take those pages of
1193                  * the same active state as that tag page.  We may safely
1194                  * round the target page pfn down to the requested order
1195                  * as the mem_map is guaranteed valid out to MAX_ORDER,
1196                  * where that page is in a different zone we will detect
1197                  * it from its zone id and abort this block scan.
1198                  */
1199                 zone_id = page_zone_id(page);
1200                 page_pfn = page_to_pfn(page);
1201                 pfn = page_pfn & ~((1 << order) - 1);
1202                 end_pfn = pfn + (1 << order);
1203                 for (; pfn < end_pfn; pfn++) {
1204                         struct page *cursor_page;
1205
1206                         /* The target page is in the block, ignore it. */
1207                         if (unlikely(pfn == page_pfn))
1208                                 continue;
1209
1210                         /* Avoid holes within the zone. */
1211                         if (unlikely(!pfn_valid_within(pfn)))
1212                                 break;
1213
1214                         cursor_page = pfn_to_page(pfn);
1215
1216                         /* Check that we have not crossed a zone boundary. */
1217                         if (unlikely(page_zone_id(cursor_page) != zone_id))
1218                                 break;
1219
1220                         /*
1221                          * If we don't have enough swap space, reclaiming of
1222                          * anon page which don't already have a swap slot is
1223                          * pointless.
1224                          */
1225                         if (nr_swap_pages <= 0 && PageSwapBacked(cursor_page) &&
1226                             !PageSwapCache(cursor_page))
1227                                 break;
1228
1229                         if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1230                                 unsigned int isolated_pages;
1231
1232                                 mem_cgroup_lru_del(cursor_page);
1233                                 list_move(&cursor_page->lru, dst);
1234                                 isolated_pages = hpage_nr_pages(cursor_page);
1235                                 nr_taken += isolated_pages;
1236                                 nr_lumpy_taken += isolated_pages;
1237                                 if (PageDirty(cursor_page))
1238                                         nr_lumpy_dirty += isolated_pages;
1239                                 scan++;
1240                                 pfn += isolated_pages - 1;
1241                         } else {
1242                                 /*
1243                                  * Check if the page is freed already.
1244                                  *
1245                                  * We can't use page_count() as that
1246                                  * requires compound_head and we don't
1247                                  * have a pin on the page here. If a
1248                                  * page is tail, we may or may not
1249                                  * have isolated the head, so assume
1250                                  * it's not free, it'd be tricky to
1251                                  * track the head status without a
1252                                  * page pin.
1253                                  */
1254                                 if (!PageTail(cursor_page) &&
1255                                     !atomic_read(&cursor_page->_count))
1256                                         continue;
1257                                 break;
1258                         }
1259                 }
1260
1261                 /* If we break out of the loop above, lumpy reclaim failed */
1262                 if (pfn < end_pfn)
1263                         nr_lumpy_failed++;
1264         }
1265
1266         *scanned = scan;
1267
1268         trace_mm_vmscan_lru_isolate(order,
1269                         nr_to_scan, scan,
1270                         nr_taken,
1271                         nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1272                         mode, file);
1273         return nr_taken;
1274 }
1275
1276 static unsigned long isolate_pages(unsigned long nr, struct mem_cgroup_zone *mz,
1277                                    struct list_head *dst,
1278                                    unsigned long *scanned, int order,
1279                                    isolate_mode_t mode, int active, int file)
1280 {
1281         struct lruvec *lruvec;
1282         int lru = LRU_BASE;
1283
1284         lruvec = mem_cgroup_zone_lruvec(mz->zone, mz->mem_cgroup);
1285         if (active)
1286                 lru += LRU_ACTIVE;
1287         if (file)
1288                 lru += LRU_FILE;
1289         return isolate_lru_pages(nr, &lruvec->lists[lru], dst,
1290                                  scanned, order, mode, file);
1291 }
1292
1293 /*
1294  * clear_active_flags() is a helper for shrink_active_list(), clearing
1295  * any active bits from the pages in the list.
1296  */
1297 static unsigned long clear_active_flags(struct list_head *page_list,
1298                                         unsigned int *count)
1299 {
1300         int nr_active = 0;
1301         int lru;
1302         struct page *page;
1303
1304         list_for_each_entry(page, page_list, lru) {
1305                 int numpages = hpage_nr_pages(page);
1306                 lru = page_lru_base_type(page);
1307                 if (PageActive(page)) {
1308                         lru += LRU_ACTIVE;
1309                         ClearPageActive(page);
1310                         nr_active += numpages;
1311                 }
1312                 if (count)
1313                         count[lru] += numpages;
1314         }
1315
1316         return nr_active;
1317 }
1318
1319 /**
1320  * isolate_lru_page - tries to isolate a page from its LRU list
1321  * @page: page to isolate from its LRU list
1322  *
1323  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1324  * vmstat statistic corresponding to whatever LRU list the page was on.
1325  *
1326  * Returns 0 if the page was removed from an LRU list.
1327  * Returns -EBUSY if the page was not on an LRU list.
1328  *
1329  * The returned page will have PageLRU() cleared.  If it was found on
1330  * the active list, it will have PageActive set.  If it was found on
1331  * the unevictable list, it will have the PageUnevictable bit set. That flag
1332  * may need to be cleared by the caller before letting the page go.
1333  *
1334  * The vmstat statistic corresponding to the list on which the page was
1335  * found will be decremented.
1336  *
1337  * Restrictions:
1338  * (1) Must be called with an elevated refcount on the page. This is a
1339  *     fundamentnal difference from isolate_lru_pages (which is called
1340  *     without a stable reference).
1341  * (2) the lru_lock must not be held.
1342  * (3) interrupts must be enabled.
1343  */
1344 int isolate_lru_page(struct page *page)
1345 {
1346         int ret = -EBUSY;
1347
1348         VM_BUG_ON(!page_count(page));
1349
1350         if (PageLRU(page)) {
1351                 struct zone *zone = page_zone(page);
1352
1353                 spin_lock_irq(&zone->lru_lock);
1354                 if (PageLRU(page)) {
1355                         int lru = page_lru(page);
1356                         ret = 0;
1357                         get_page(page);
1358                         ClearPageLRU(page);
1359
1360                         del_page_from_lru_list(zone, page, lru);
1361                 }
1362                 spin_unlock_irq(&zone->lru_lock);
1363         }
1364         return ret;
1365 }
1366
1367 /*
1368  * Are there way too many processes in the direct reclaim path already?
1369  */
1370 static int too_many_isolated(struct zone *zone, int file,
1371                 struct scan_control *sc)
1372 {
1373         unsigned long inactive, isolated;
1374
1375         if (current_is_kswapd())
1376                 return 0;
1377
1378         if (!global_reclaim(sc))
1379                 return 0;
1380
1381         if (file) {
1382                 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1383                 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1384         } else {
1385                 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1386                 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1387         }
1388
1389         return isolated > inactive;
1390 }
1391
1392 /*
1393  * TODO: Try merging with migrations version of putback_lru_pages
1394  */
1395 static noinline_for_stack void
1396 putback_lru_pages(struct mem_cgroup_zone *mz, struct scan_control *sc,
1397                   unsigned long nr_anon, unsigned long nr_file,
1398                   struct list_head *page_list)
1399 {
1400         struct page *page;
1401         struct pagevec pvec;
1402         struct zone *zone = mz->zone;
1403         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz);
1404
1405         pagevec_init(&pvec, 1);
1406
1407         /*
1408          * Put back any unfreeable pages.
1409          */
1410         spin_lock(&zone->lru_lock);
1411         while (!list_empty(page_list)) {
1412                 int lru;
1413                 page = lru_to_page(page_list);
1414                 VM_BUG_ON(PageLRU(page));
1415                 list_del(&page->lru);
1416                 if (unlikely(!page_evictable(page, NULL))) {
1417                         spin_unlock_irq(&zone->lru_lock);
1418                         putback_lru_page(page);
1419                         spin_lock_irq(&zone->lru_lock);
1420                         continue;
1421                 }
1422                 SetPageLRU(page);
1423                 lru = page_lru(page);
1424                 add_page_to_lru_list(zone, page, lru);
1425                 if (is_active_lru(lru)) {
1426                         int file = is_file_lru(lru);
1427                         int numpages = hpage_nr_pages(page);
1428                         reclaim_stat->recent_rotated[file] += numpages;
1429                 }
1430                 if (!pagevec_add(&pvec, page)) {
1431                         spin_unlock_irq(&zone->lru_lock);
1432                         __pagevec_release(&pvec);
1433                         spin_lock_irq(&zone->lru_lock);
1434                 }
1435         }
1436         __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1437         __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1438
1439         spin_unlock_irq(&zone->lru_lock);
1440         pagevec_release(&pvec);
1441 }
1442
1443 static noinline_for_stack void
1444 update_isolated_counts(struct mem_cgroup_zone *mz,
1445                        struct scan_control *sc,
1446                        unsigned long *nr_anon,
1447                        unsigned long *nr_file,
1448                        struct list_head *isolated_list)
1449 {
1450         unsigned long nr_active;
1451         struct zone *zone = mz->zone;
1452         unsigned int count[NR_LRU_LISTS] = { 0, };
1453         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz);
1454
1455         nr_active = clear_active_flags(isolated_list, count);
1456         __count_vm_events(PGDEACTIVATE, nr_active);
1457
1458         __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1459                               -count[LRU_ACTIVE_FILE]);
1460         __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1461                               -count[LRU_INACTIVE_FILE]);
1462         __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1463                               -count[LRU_ACTIVE_ANON]);
1464         __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1465                               -count[LRU_INACTIVE_ANON]);
1466
1467         *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1468         *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1469         __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1470         __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1471
1472         reclaim_stat->recent_scanned[0] += *nr_anon;
1473         reclaim_stat->recent_scanned[1] += *nr_file;
1474 }
1475
1476 /*
1477  * Returns true if a direct reclaim should wait on pages under writeback.
1478  *
1479  * If we are direct reclaiming for contiguous pages and we do not reclaim
1480  * everything in the list, try again and wait for writeback IO to complete.
1481  * This will stall high-order allocations noticeably. Only do that when really
1482  * need to free the pages under high memory pressure.
1483  */
1484 static inline bool should_reclaim_stall(unsigned long nr_taken,
1485                                         unsigned long nr_freed,
1486                                         int priority,
1487                                         struct scan_control *sc)
1488 {
1489         int lumpy_stall_priority;
1490
1491         /* kswapd should not stall on sync IO */
1492         if (current_is_kswapd())
1493                 return false;
1494
1495         /* Only stall on lumpy reclaim */
1496         if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
1497                 return false;
1498
1499         /* If we have reclaimed everything on the isolated list, no stall */
1500         if (nr_freed == nr_taken)
1501                 return false;
1502
1503         /*
1504          * For high-order allocations, there are two stall thresholds.
1505          * High-cost allocations stall immediately where as lower
1506          * order allocations such as stacks require the scanning
1507          * priority to be much higher before stalling.
1508          */
1509         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1510                 lumpy_stall_priority = DEF_PRIORITY;
1511         else
1512                 lumpy_stall_priority = DEF_PRIORITY / 3;
1513
1514         return priority <= lumpy_stall_priority;
1515 }
1516
1517 /*
1518  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1519  * of reclaimed pages
1520  */
1521 static noinline_for_stack unsigned long
1522 shrink_inactive_list(unsigned long nr_to_scan, struct mem_cgroup_zone *mz,
1523                      struct scan_control *sc, int priority, int file)
1524 {
1525         LIST_HEAD(page_list);
1526         unsigned long nr_scanned;
1527         unsigned long nr_reclaimed = 0;
1528         unsigned long nr_taken;
1529         unsigned long nr_anon;
1530         unsigned long nr_file;
1531         unsigned long nr_dirty = 0;
1532         unsigned long nr_writeback = 0;
1533         isolate_mode_t reclaim_mode = ISOLATE_INACTIVE;
1534         struct zone *zone = mz->zone;
1535
1536         while (unlikely(too_many_isolated(zone, file, sc))) {
1537                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1538
1539                 /* We are about to die and free our memory. Return now. */
1540                 if (fatal_signal_pending(current))
1541                         return SWAP_CLUSTER_MAX;
1542         }
1543
1544         set_reclaim_mode(priority, sc, false);
1545         if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
1546                 reclaim_mode |= ISOLATE_ACTIVE;
1547
1548         lru_add_drain();
1549
1550         if (!sc->may_unmap)
1551                 reclaim_mode |= ISOLATE_UNMAPPED;
1552         if (!sc->may_writepage)
1553                 reclaim_mode |= ISOLATE_CLEAN;
1554
1555         spin_lock_irq(&zone->lru_lock);
1556
1557         nr_taken = isolate_pages(nr_to_scan, mz, &page_list,
1558                                  &nr_scanned, sc->order,
1559                                  reclaim_mode, 0, file);
1560         if (global_reclaim(sc)) {
1561                 zone->pages_scanned += nr_scanned;
1562                 if (current_is_kswapd())
1563                         __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1564                                                nr_scanned);
1565                 else
1566                         __count_zone_vm_events(PGSCAN_DIRECT, zone,
1567                                                nr_scanned);
1568         }
1569
1570         if (nr_taken == 0) {
1571                 spin_unlock_irq(&zone->lru_lock);
1572                 return 0;
1573         }
1574
1575         update_isolated_counts(mz, sc, &nr_anon, &nr_file, &page_list);
1576
1577         spin_unlock_irq(&zone->lru_lock);
1578
1579         nr_reclaimed = shrink_page_list(&page_list, mz, sc, priority,
1580                                                 &nr_dirty, &nr_writeback);
1581
1582         /* Check if we should syncronously wait for writeback */
1583         if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1584                 set_reclaim_mode(priority, sc, true);
1585                 nr_reclaimed += shrink_page_list(&page_list, mz, sc,
1586                                         priority, &nr_dirty, &nr_writeback);
1587         }
1588
1589         local_irq_disable();
1590         if (current_is_kswapd())
1591                 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1592         __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1593
1594         putback_lru_pages(mz, sc, nr_anon, nr_file, &page_list);
1595
1596         /*
1597          * If reclaim is isolating dirty pages under writeback, it implies
1598          * that the long-lived page allocation rate is exceeding the page
1599          * laundering rate. Either the global limits are not being effective
1600          * at throttling processes due to the page distribution throughout
1601          * zones or there is heavy usage of a slow backing device. The
1602          * only option is to throttle from reclaim context which is not ideal
1603          * as there is no guarantee the dirtying process is throttled in the
1604          * same way balance_dirty_pages() manages.
1605          *
1606          * This scales the number of dirty pages that must be under writeback
1607          * before throttling depending on priority. It is a simple backoff
1608          * function that has the most effect in the range DEF_PRIORITY to
1609          * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1610          * in trouble and reclaim is considered to be in trouble.
1611          *
1612          * DEF_PRIORITY   100% isolated pages must be PageWriteback to throttle
1613          * DEF_PRIORITY-1  50% must be PageWriteback
1614          * DEF_PRIORITY-2  25% must be PageWriteback, kswapd in trouble
1615          * ...
1616          * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1617          *                     isolated page is PageWriteback
1618          */
1619         if (nr_writeback && nr_writeback >= (nr_taken >> (DEF_PRIORITY-priority)))
1620                 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1621
1622         trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1623                 zone_idx(zone),
1624                 nr_scanned, nr_reclaimed,
1625                 priority,
1626                 trace_shrink_flags(file, sc->reclaim_mode));
1627         return nr_reclaimed;
1628 }
1629
1630 /*
1631  * This moves pages from the active list to the inactive list.
1632  *
1633  * We move them the other way if the page is referenced by one or more
1634  * processes, from rmap.
1635  *
1636  * If the pages are mostly unmapped, the processing is fast and it is
1637  * appropriate to hold zone->lru_lock across the whole operation.  But if
1638  * the pages are mapped, the processing is slow (page_referenced()) so we
1639  * should drop zone->lru_lock around each page.  It's impossible to balance
1640  * this, so instead we remove the pages from the LRU while processing them.
1641  * It is safe to rely on PG_active against the non-LRU pages in here because
1642  * nobody will play with that bit on a non-LRU page.
1643  *
1644  * The downside is that we have to touch page->_count against each page.
1645  * But we had to alter page->flags anyway.
1646  */
1647
1648 static void move_active_pages_to_lru(struct zone *zone,
1649                                      struct list_head *list,
1650                                      enum lru_list lru)
1651 {
1652         unsigned long pgmoved = 0;
1653         struct pagevec pvec;
1654         struct page *page;
1655
1656         pagevec_init(&pvec, 1);
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 (!pagevec_add(&pvec, page) || list_empty(list)) {
1671                         spin_unlock_irq(&zone->lru_lock);
1672                         if (buffer_heads_over_limit)
1673                                 pagevec_strip(&pvec);
1674                         __pagevec_release(&pvec);
1675                         spin_lock_irq(&zone->lru_lock);
1676                 }
1677         }
1678         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1679         if (!is_active_lru(lru))
1680                 __count_vm_events(PGDEACTIVATE, pgmoved);
1681 }
1682
1683 static void shrink_active_list(unsigned long nr_pages,
1684                                struct mem_cgroup_zone *mz,
1685                                struct scan_control *sc,
1686                                int priority, int file)
1687 {
1688         unsigned long nr_taken;
1689         unsigned long pgscanned;
1690         unsigned long vm_flags;
1691         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1692         LIST_HEAD(l_active);
1693         LIST_HEAD(l_inactive);
1694         struct page *page;
1695         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz);
1696         unsigned long nr_rotated = 0;
1697         isolate_mode_t reclaim_mode = ISOLATE_ACTIVE;
1698         struct zone *zone = mz->zone;
1699
1700         lru_add_drain();
1701
1702         if (!sc->may_unmap)
1703                 reclaim_mode |= ISOLATE_UNMAPPED;
1704         if (!sc->may_writepage)
1705                 reclaim_mode |= ISOLATE_CLEAN;
1706
1707         spin_lock_irq(&zone->lru_lock);
1708
1709         nr_taken = isolate_pages(nr_pages, mz, &l_hold,
1710                                  &pgscanned, sc->order,
1711                                  reclaim_mode, 1, file);
1712
1713         if (global_reclaim(sc))
1714                 zone->pages_scanned += pgscanned;
1715
1716         reclaim_stat->recent_scanned[file] += nr_taken;
1717
1718         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1719         if (file)
1720                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1721         else
1722                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1723         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1724         spin_unlock_irq(&zone->lru_lock);
1725
1726         while (!list_empty(&l_hold)) {
1727                 cond_resched();
1728                 page = lru_to_page(&l_hold);
1729                 list_del(&page->lru);
1730
1731                 if (unlikely(!page_evictable(page, NULL))) {
1732                         putback_lru_page(page);
1733                         continue;
1734                 }
1735
1736                 if (page_referenced(page, 0, mz->mem_cgroup, &vm_flags)) {
1737                         nr_rotated += hpage_nr_pages(page);
1738                         /*
1739                          * Identify referenced, file-backed active pages and
1740                          * give them one more trip around the active list. So
1741                          * that executable code get better chances to stay in
1742                          * memory under moderate memory pressure.  Anon pages
1743                          * are not likely to be evicted by use-once streaming
1744                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1745                          * so we ignore them here.
1746                          */
1747                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1748                                 list_add(&page->lru, &l_active);
1749                                 continue;
1750                         }
1751                 }
1752
1753                 ClearPageActive(page);  /* we are de-activating */
1754                 list_add(&page->lru, &l_inactive);
1755         }
1756
1757         /*
1758          * Move pages back to the lru list.
1759          */
1760         spin_lock_irq(&zone->lru_lock);
1761         /*
1762          * Count referenced pages from currently used mappings as rotated,
1763          * even though only some of them are actually re-activated.  This
1764          * helps balance scan pressure between file and anonymous pages in
1765          * get_scan_ratio.
1766          */
1767         reclaim_stat->recent_rotated[file] += nr_rotated;
1768
1769         move_active_pages_to_lru(zone, &l_active,
1770                                                 LRU_ACTIVE + file * LRU_FILE);
1771         move_active_pages_to_lru(zone, &l_inactive,
1772                                                 LRU_BASE   + file * LRU_FILE);
1773         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1774         spin_unlock_irq(&zone->lru_lock);
1775 }
1776
1777 #ifdef CONFIG_SWAP
1778 static int inactive_anon_is_low_global(struct zone *zone)
1779 {
1780         unsigned long active, inactive;
1781
1782         active = zone_page_state(zone, NR_ACTIVE_ANON);
1783         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1784
1785         if (inactive * zone->inactive_ratio < active)
1786                 return 1;
1787
1788         return 0;
1789 }
1790
1791 /**
1792  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1793  * @zone: zone to check
1794  * @sc:   scan control of this context
1795  *
1796  * Returns true if the zone does not have enough inactive anon pages,
1797  * meaning some active anon pages need to be deactivated.
1798  */
1799 static int inactive_anon_is_low(struct mem_cgroup_zone *mz)
1800 {
1801         /*
1802          * If we don't have swap space, anonymous page deactivation
1803          * is pointless.
1804          */
1805         if (!total_swap_pages)
1806                 return 0;
1807
1808         if (!scanning_global_lru(mz))
1809                 return mem_cgroup_inactive_anon_is_low(mz->mem_cgroup,
1810                                                        mz->zone);
1811
1812         return inactive_anon_is_low_global(mz->zone);
1813 }
1814 #else
1815 static inline int inactive_anon_is_low(struct mem_cgroup_zone *mz)
1816 {
1817         return 0;
1818 }
1819 #endif
1820
1821 static int inactive_file_is_low_global(struct zone *zone)
1822 {
1823         unsigned long active, inactive;
1824
1825         active = zone_page_state(zone, NR_ACTIVE_FILE);
1826         inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1827
1828         return (active > inactive);
1829 }
1830
1831 /**
1832  * inactive_file_is_low - check if file pages need to be deactivated
1833  * @mz: memory cgroup and zone to check
1834  *
1835  * When the system is doing streaming IO, memory pressure here
1836  * ensures that active file pages get deactivated, until more
1837  * than half of the file pages are on the inactive list.
1838  *
1839  * Once we get to that situation, protect the system's working
1840  * set from being evicted by disabling active file page aging.
1841  *
1842  * This uses a different ratio than the anonymous pages, because
1843  * the page cache uses a use-once replacement algorithm.
1844  */
1845 static int inactive_file_is_low(struct mem_cgroup_zone *mz)
1846 {
1847         if (!scanning_global_lru(mz))
1848                 return mem_cgroup_inactive_file_is_low(mz->mem_cgroup,
1849                                                        mz->zone);
1850
1851         return inactive_file_is_low_global(mz->zone);
1852 }
1853
1854 static int inactive_list_is_low(struct mem_cgroup_zone *mz, int file)
1855 {
1856         if (file)
1857                 return inactive_file_is_low(mz);
1858         else
1859                 return inactive_anon_is_low(mz);
1860 }
1861
1862 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1863                                  struct mem_cgroup_zone *mz,
1864                                  struct scan_control *sc, int priority)
1865 {
1866         int file = is_file_lru(lru);
1867
1868         if (is_active_lru(lru)) {
1869                 if (inactive_list_is_low(mz, file))
1870                         shrink_active_list(nr_to_scan, mz, sc, priority, file);
1871                 return 0;
1872         }
1873
1874         return shrink_inactive_list(nr_to_scan, mz, sc, priority, file);
1875 }
1876
1877 static int vmscan_swappiness(struct mem_cgroup_zone *mz,
1878                              struct scan_control *sc)
1879 {
1880         if (global_reclaim(sc))
1881                 return vm_swappiness;
1882         return mem_cgroup_swappiness(mz->mem_cgroup);
1883 }
1884
1885 /*
1886  * Determine how aggressively the anon and file LRU lists should be
1887  * scanned.  The relative value of each set of LRU lists is determined
1888  * by looking at the fraction of the pages scanned we did rotate back
1889  * onto the active list instead of evict.
1890  *
1891  * nr[0] = anon pages to scan; nr[1] = file pages to scan
1892  */
1893 static void get_scan_count(struct mem_cgroup_zone *mz, struct scan_control *sc,
1894                            unsigned long *nr, int priority)
1895 {
1896         unsigned long anon, file, free;
1897         unsigned long anon_prio, file_prio;
1898         unsigned long ap, fp;
1899         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz);
1900         u64 fraction[2], denominator;
1901         enum lru_list l;
1902         int noswap = 0;
1903         bool force_scan = false;
1904
1905         /*
1906          * If the zone or memcg is small, nr[l] can be 0.  This
1907          * results in no scanning on this priority and a potential
1908          * priority drop.  Global direct reclaim can go to the next
1909          * zone and tends to have no problems. Global kswapd is for
1910          * zone balancing and it needs to scan a minimum amount. When
1911          * reclaiming for a memcg, a priority drop can cause high
1912          * latencies, so it's better to scan a minimum amount there as
1913          * well.
1914          */
1915         if (current_is_kswapd() && mz->zone->all_unreclaimable)
1916                 force_scan = true;
1917         if (!global_reclaim(sc))
1918                 force_scan = true;
1919
1920         /* If we have no swap space, do not bother scanning anon pages. */
1921         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1922                 noswap = 1;
1923                 fraction[0] = 0;
1924                 fraction[1] = 1;
1925                 denominator = 1;
1926                 goto out;
1927         }
1928
1929         anon  = zone_nr_lru_pages(mz, LRU_ACTIVE_ANON) +
1930                 zone_nr_lru_pages(mz, LRU_INACTIVE_ANON);
1931         file  = zone_nr_lru_pages(mz, LRU_ACTIVE_FILE) +
1932                 zone_nr_lru_pages(mz, LRU_INACTIVE_FILE);
1933
1934         if (global_reclaim(sc)) {
1935                 free  = zone_page_state(mz->zone, NR_FREE_PAGES);
1936                 /* If we have very few page cache pages,
1937                    force-scan anon pages. */
1938                 if (unlikely(file + free <= high_wmark_pages(mz->zone))) {
1939                         fraction[0] = 1;
1940                         fraction[1] = 0;
1941                         denominator = 1;
1942                         goto out;
1943                 }
1944         }
1945
1946         /*
1947          * With swappiness at 100, anonymous and file have the same priority.
1948          * This scanning priority is essentially the inverse of IO cost.
1949          */
1950         anon_prio = vmscan_swappiness(mz, sc);
1951         file_prio = 200 - vmscan_swappiness(mz, sc);
1952
1953         /*
1954          * OK, so we have swap space and a fair amount of page cache
1955          * pages.  We use the recently rotated / recently scanned
1956          * ratios to determine how valuable each cache is.
1957          *
1958          * Because workloads change over time (and to avoid overflow)
1959          * we keep these statistics as a floating average, which ends
1960          * up weighing recent references more than old ones.
1961          *
1962          * anon in [0], file in [1]
1963          */
1964         spin_lock_irq(&mz->zone->lru_lock);
1965         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1966                 reclaim_stat->recent_scanned[0] /= 2;
1967                 reclaim_stat->recent_rotated[0] /= 2;
1968         }
1969
1970         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1971                 reclaim_stat->recent_scanned[1] /= 2;
1972                 reclaim_stat->recent_rotated[1] /= 2;
1973         }
1974
1975         /*
1976          * The amount of pressure on anon vs file pages is inversely
1977          * proportional to the fraction of recently scanned pages on
1978          * each list that were recently referenced and in active use.
1979          */
1980         ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1981         ap /= reclaim_stat->recent_rotated[0] + 1;
1982
1983         fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1984         fp /= reclaim_stat->recent_rotated[1] + 1;
1985         spin_unlock_irq(&mz->zone->lru_lock);
1986
1987         fraction[0] = ap;
1988         fraction[1] = fp;
1989         denominator = ap + fp + 1;
1990 out:
1991         for_each_evictable_lru(l) {
1992                 int file = is_file_lru(l);
1993                 unsigned long scan;
1994
1995                 scan = zone_nr_lru_pages(mz, l);
1996                 if (priority || noswap) {
1997                         scan >>= priority;
1998                         if (!scan && force_scan)
1999                                 scan = SWAP_CLUSTER_MAX;
2000                         scan = div64_u64(scan * fraction[file], denominator);
2001                 }
2002                 nr[l] = scan;
2003         }
2004 }
2005
2006 /*
2007  * Reclaim/compaction depends on a number of pages being freed. To avoid
2008  * disruption to the system, a small number of order-0 pages continue to be
2009  * rotated and reclaimed in the normal fashion. However, by the time we get
2010  * back to the allocator and call try_to_compact_zone(), we ensure that
2011  * there are enough free pages for it to be likely successful
2012  */
2013 static inline bool should_continue_reclaim(struct mem_cgroup_zone *mz,
2014                                         unsigned long nr_reclaimed,
2015                                         unsigned long nr_scanned,
2016                                         struct scan_control *sc)
2017 {
2018         unsigned long pages_for_compaction;
2019         unsigned long inactive_lru_pages;
2020
2021         /* If not in reclaim/compaction mode, stop */
2022         if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
2023                 return false;
2024
2025         /* Consider stopping depending on scan and reclaim activity */
2026         if (sc->gfp_mask & __GFP_REPEAT) {
2027                 /*
2028                  * For __GFP_REPEAT allocations, stop reclaiming if the
2029                  * full LRU list has been scanned and we are still failing
2030                  * to reclaim pages. This full LRU scan is potentially
2031                  * expensive but a __GFP_REPEAT caller really wants to succeed
2032                  */
2033                 if (!nr_reclaimed && !nr_scanned)
2034                         return false;
2035         } else {
2036                 /*
2037                  * For non-__GFP_REPEAT allocations which can presumably
2038                  * fail without consequence, stop if we failed to reclaim
2039                  * any pages from the last SWAP_CLUSTER_MAX number of
2040                  * pages that were scanned. This will return to the
2041                  * caller faster at the risk reclaim/compaction and
2042                  * the resulting allocation attempt fails
2043                  */
2044                 if (!nr_reclaimed)
2045                         return false;
2046         }
2047
2048         /*
2049          * If we have not reclaimed enough pages for compaction and the
2050          * inactive lists are large enough, continue reclaiming
2051          */
2052         pages_for_compaction = (2UL << sc->order);
2053         inactive_lru_pages = zone_nr_lru_pages(mz, LRU_INACTIVE_FILE);
2054         if (nr_swap_pages > 0)
2055                 inactive_lru_pages += zone_nr_lru_pages(mz, LRU_INACTIVE_ANON);
2056         if (sc->nr_reclaimed < pages_for_compaction &&
2057                         inactive_lru_pages > pages_for_compaction)
2058                 return true;
2059
2060         /* If compaction would go ahead or the allocation would succeed, stop */
2061         switch (compaction_suitable(mz->zone, sc->order)) {
2062         case COMPACT_PARTIAL:
2063         case COMPACT_CONTINUE:
2064                 return false;
2065         default:
2066                 return true;
2067         }
2068 }
2069
2070 /*
2071  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
2072  */
2073 static void shrink_mem_cgroup_zone(int priority, struct mem_cgroup_zone *mz,
2074                                    struct scan_control *sc)
2075 {
2076         unsigned long nr[NR_LRU_LISTS];
2077         unsigned long nr_to_scan;
2078         enum lru_list l;
2079         unsigned long nr_reclaimed, nr_scanned;
2080         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2081         struct blk_plug plug;
2082
2083 restart:
2084         nr_reclaimed = 0;
2085         nr_scanned = sc->nr_scanned;
2086         get_scan_count(mz, sc, nr, priority);
2087
2088         blk_start_plug(&plug);
2089         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2090                                         nr[LRU_INACTIVE_FILE]) {
2091                 for_each_evictable_lru(l) {
2092                         if (nr[l]) {
2093                                 nr_to_scan = min_t(unsigned long,
2094                                                    nr[l], SWAP_CLUSTER_MAX);
2095                                 nr[l] -= nr_to_scan;
2096
2097                                 nr_reclaimed += shrink_list(l, nr_to_scan,
2098                                                             mz, sc, priority);
2099                         }
2100                 }
2101                 /*
2102                  * On large memory systems, scan >> priority can become
2103                  * really large. This is fine for the starting priority;
2104                  * we want to put equal scanning pressure on each zone.
2105                  * However, if the VM has a harder time of freeing pages,
2106                  * with multiple processes reclaiming pages, the total
2107                  * freeing target can get unreasonably large.
2108                  */
2109                 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
2110                         break;
2111         }
2112         blk_finish_plug(&plug);
2113         sc->nr_reclaimed += nr_reclaimed;
2114
2115         /*
2116          * Even if we did not try to evict anon pages at all, we want to
2117          * rebalance the anon lru active/inactive ratio.
2118          */
2119         if (inactive_anon_is_low(mz))
2120                 shrink_active_list(SWAP_CLUSTER_MAX, mz, sc, priority, 0);
2121
2122         /* reclaim/compaction might need reclaim to continue */
2123         if (should_continue_reclaim(mz, nr_reclaimed,
2124                                         sc->nr_scanned - nr_scanned, sc))
2125                 goto restart;
2126
2127         throttle_vm_writeout(sc->gfp_mask);
2128 }
2129
2130 static void shrink_zone(int priority, struct zone *zone,
2131                         struct scan_control *sc)
2132 {
2133         struct mem_cgroup *root = sc->target_mem_cgroup;
2134         struct mem_cgroup_reclaim_cookie reclaim = {
2135                 .zone = zone,
2136                 .priority = priority,
2137         };
2138         struct mem_cgroup *memcg;
2139
2140         memcg = mem_cgroup_iter(root, NULL, &reclaim);
2141         do {
2142                 struct mem_cgroup_zone mz = {
2143                         .mem_cgroup = memcg,
2144                         .zone = zone,
2145                 };
2146
2147                 shrink_mem_cgroup_zone(priority, &mz, sc);
2148                 /*
2149                  * Limit reclaim has historically picked one memcg and
2150                  * scanned it with decreasing priority levels until
2151                  * nr_to_reclaim had been reclaimed.  This priority
2152                  * cycle is thus over after a single memcg.
2153                  *
2154                  * Direct reclaim and kswapd, on the other hand, have
2155                  * to scan all memory cgroups to fulfill the overall
2156                  * scan target for the zone.
2157                  */
2158                 if (!global_reclaim(sc)) {
2159                         mem_cgroup_iter_break(root, memcg);
2160                         break;
2161                 }
2162                 memcg = mem_cgroup_iter(root, memcg, &reclaim);
2163         } while (memcg);
2164 }
2165
2166 /* Returns true if compaction should go ahead for a high-order request */
2167 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2168 {
2169         unsigned long balance_gap, watermark;
2170         bool watermark_ok;
2171
2172         /* Do not consider compaction for orders reclaim is meant to satisfy */
2173         if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
2174                 return false;
2175
2176         /*
2177          * Compaction takes time to run and there are potentially other
2178          * callers using the pages just freed. Continue reclaiming until
2179          * there is a buffer of free pages available to give compaction
2180          * a reasonable chance of completing and allocating the page
2181          */
2182         balance_gap = min(low_wmark_pages(zone),
2183                 (zone->present_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2184                         KSWAPD_ZONE_BALANCE_GAP_RATIO);
2185         watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
2186         watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2187
2188         /*
2189          * If compaction is deferred, reclaim up to a point where
2190          * compaction will have a chance of success when re-enabled
2191          */
2192         if (compaction_deferred(zone))
2193                 return watermark_ok;
2194
2195         /* If compaction is not ready to start, keep reclaiming */
2196         if (!compaction_suitable(zone, sc->order))
2197                 return false;
2198
2199         return watermark_ok;
2200 }
2201
2202 /*
2203  * This is the direct reclaim path, for page-allocating processes.  We only
2204  * try to reclaim pages from zones which will satisfy the caller's allocation
2205  * request.
2206  *
2207  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2208  * Because:
2209  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2210  *    allocation or
2211  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2212  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2213  *    zone defense algorithm.
2214  *
2215  * If a zone is deemed to be full of pinned pages then just give it a light
2216  * scan then give up on it.
2217  *
2218  * This function returns true if a zone is being reclaimed for a costly
2219  * high-order allocation and compaction is ready to begin. This indicates to
2220  * the caller that it should retry the allocation or fail.
2221  */
2222 static bool shrink_zones(int priority, struct zonelist *zonelist,
2223                                         struct scan_control *sc)
2224 {
2225         struct zoneref *z;
2226         struct zone *zone;
2227         unsigned long nr_soft_reclaimed;
2228         unsigned long nr_soft_scanned;
2229         bool should_abort_reclaim = false;
2230
2231         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2232                                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2233                 if (!populated_zone(zone))
2234                         continue;
2235                 /*
2236                  * Take care memory controller reclaiming has small influence
2237                  * to global LRU.
2238                  */
2239                 if (global_reclaim(sc)) {
2240                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2241                                 continue;
2242                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2243                                 continue;       /* Let kswapd poll it */
2244                         if (COMPACTION_BUILD) {
2245                                 /*
2246                                  * If we already have plenty of memory free for
2247                                  * compaction in this zone, don't free any more.
2248                                  * Even though compaction is invoked for any
2249                                  * non-zero order, only frequent costly order
2250                                  * reclamation is disruptive enough to become a
2251                                  * noticable problem, like transparent huge page
2252                                  * allocations.
2253                                  */
2254                                 if (compaction_ready(zone, sc)) {
2255                                         should_abort_reclaim = true;
2256                                         continue;
2257                                 }
2258                         }
2259                         /*
2260                          * This steals pages from memory cgroups over softlimit
2261                          * and returns the number of reclaimed pages and
2262                          * scanned pages. This works for global memory pressure
2263                          * and balancing, not for a memcg's limit.
2264                          */
2265                         nr_soft_scanned = 0;
2266                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2267                                                 sc->order, sc->gfp_mask,
2268                                                 &nr_soft_scanned);
2269                         sc->nr_reclaimed += nr_soft_reclaimed;
2270                         sc->nr_scanned += nr_soft_scanned;
2271                         /* need some check for avoid more shrink_zone() */
2272                 }
2273
2274                 shrink_zone(priority, zone, sc);
2275         }
2276
2277         return should_abort_reclaim;
2278 }
2279
2280 static bool zone_reclaimable(struct zone *zone)
2281 {
2282         return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2283 }
2284
2285 /* All zones in zonelist are unreclaimable? */
2286 static bool all_unreclaimable(struct zonelist *zonelist,
2287                 struct scan_control *sc)
2288 {
2289         struct zoneref *z;
2290         struct zone *zone;
2291
2292         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2293                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2294                 if (!populated_zone(zone))
2295                         continue;
2296                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2297                         continue;
2298                 if (!zone->all_unreclaimable)
2299                         return false;
2300         }
2301
2302         return true;
2303 }
2304
2305 /*
2306  * This is the main entry point to direct page reclaim.
2307  *
2308  * If a full scan of the inactive list fails to free enough memory then we
2309  * are "out of memory" and something needs to be killed.
2310  *
2311  * If the caller is !__GFP_FS then the probability of a failure is reasonably
2312  * high - the zone may be full of dirty or under-writeback pages, which this
2313  * caller can't do much about.  We kick the writeback threads and take explicit
2314  * naps in the hope that some of these pages can be written.  But if the
2315  * allocating task holds filesystem locks which prevent writeout this might not
2316  * work, and the allocation attempt will fail.
2317  *
2318  * returns:     0, if no pages reclaimed
2319  *              else, the number of pages reclaimed
2320  */
2321 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2322                                         struct scan_control *sc,
2323                                         struct shrink_control *shrink)
2324 {
2325         int priority;
2326         unsigned long total_scanned = 0;
2327         struct reclaim_state *reclaim_state = current->reclaim_state;
2328         struct zoneref *z;
2329         struct zone *zone;
2330         unsigned long writeback_threshold;
2331         bool should_abort_reclaim;
2332
2333         get_mems_allowed();
2334         delayacct_freepages_start();
2335
2336         if (global_reclaim(sc))
2337                 count_vm_event(ALLOCSTALL);
2338
2339         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2340                 sc->nr_scanned = 0;
2341                 if (!priority)
2342                         disable_swap_token(sc->target_mem_cgroup);
2343                 should_abort_reclaim = shrink_zones(priority, zonelist, sc);
2344                 if (should_abort_reclaim)
2345                         break;
2346
2347                 /*
2348                  * Don't shrink slabs when reclaiming memory from
2349                  * over limit cgroups
2350                  */
2351                 if (global_reclaim(sc)) {
2352                         unsigned long lru_pages = 0;
2353                         for_each_zone_zonelist(zone, z, zonelist,
2354                                         gfp_zone(sc->gfp_mask)) {
2355                                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2356                                         continue;
2357
2358                                 lru_pages += zone_reclaimable_pages(zone);
2359                         }
2360
2361                         shrink_slab(shrink, sc->nr_scanned, lru_pages);
2362                         if (reclaim_state) {
2363                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2364                                 reclaim_state->reclaimed_slab = 0;
2365                         }
2366                 }
2367                 total_scanned += sc->nr_scanned;
2368                 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2369                         goto out;
2370
2371                 /*
2372                  * Try to write back as many pages as we just scanned.  This
2373                  * tends to cause slow streaming writers to write data to the
2374                  * disk smoothly, at the dirtying rate, which is nice.   But
2375                  * that's undesirable in laptop mode, where we *want* lumpy
2376                  * writeout.  So in laptop mode, write out the whole world.
2377                  */
2378                 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2379                 if (total_scanned > writeback_threshold) {
2380                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2381                                                 WB_REASON_TRY_TO_FREE_PAGES);
2382                         sc->may_writepage = 1;
2383                 }
2384
2385                 /* Take a nap, wait for some writeback to complete */
2386                 if (!sc->hibernation_mode && sc->nr_scanned &&
2387                     priority < DEF_PRIORITY - 2) {
2388                         struct zone *preferred_zone;
2389
2390                         first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2391                                                 &cpuset_current_mems_allowed,
2392                                                 &preferred_zone);
2393                         wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2394                 }
2395         }
2396
2397 out:
2398         delayacct_freepages_end();
2399         put_mems_allowed();
2400
2401         if (sc->nr_reclaimed)
2402                 return sc->nr_reclaimed;
2403
2404         /*
2405          * As hibernation is going on, kswapd is freezed so that it can't mark
2406          * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2407          * check.
2408          */
2409         if (oom_killer_disabled)
2410                 return 0;
2411
2412         /* Aborting reclaim to try compaction? don't OOM, then */
2413         if (should_abort_reclaim)
2414                 return 1;
2415
2416         /* top priority shrink_zones still had more to do? don't OOM, then */
2417         if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
2418                 return 1;
2419
2420         return 0;
2421 }
2422
2423 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2424                                 gfp_t gfp_mask, nodemask_t *nodemask)
2425 {
2426         unsigned long nr_reclaimed;
2427         struct scan_control sc = {
2428                 .gfp_mask = gfp_mask,
2429                 .may_writepage = !laptop_mode,
2430                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2431                 .may_unmap = 1,
2432                 .may_swap = 1,
2433                 .order = order,
2434                 .target_mem_cgroup = NULL,
2435                 .nodemask = nodemask,
2436         };
2437         struct shrink_control shrink = {
2438                 .gfp_mask = sc.gfp_mask,
2439         };
2440
2441         trace_mm_vmscan_direct_reclaim_begin(order,
2442                                 sc.may_writepage,
2443                                 gfp_mask);
2444
2445         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2446
2447         trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2448
2449         return nr_reclaimed;
2450 }
2451
2452 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2453
2454 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
2455                                                 gfp_t gfp_mask, bool noswap,
2456                                                 struct zone *zone,
2457                                                 unsigned long *nr_scanned)
2458 {
2459         struct scan_control sc = {
2460                 .nr_scanned = 0,
2461                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2462                 .may_writepage = !laptop_mode,
2463                 .may_unmap = 1,
2464                 .may_swap = !noswap,
2465                 .order = 0,
2466                 .target_mem_cgroup = memcg,
2467         };
2468         struct mem_cgroup_zone mz = {
2469                 .mem_cgroup = memcg,
2470                 .zone = zone,
2471         };
2472
2473         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2474                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2475
2476         trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2477                                                       sc.may_writepage,
2478                                                       sc.gfp_mask);
2479
2480         /*
2481          * NOTE: Although we can get the priority field, using it
2482          * here is not a good idea, since it limits the pages we can scan.
2483          * if we don't reclaim here, the shrink_zone from balance_pgdat
2484          * will pick up pages from other mem cgroup's as well. We hack
2485          * the priority and make it zero.
2486          */
2487         shrink_mem_cgroup_zone(0, &mz, &sc);
2488
2489         trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2490
2491         *nr_scanned = sc.nr_scanned;
2492         return sc.nr_reclaimed;
2493 }
2494
2495 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2496                                            gfp_t gfp_mask,
2497                                            bool noswap)
2498 {
2499         struct zonelist *zonelist;
2500         unsigned long nr_reclaimed;
2501         int nid;
2502         struct scan_control sc = {
2503                 .may_writepage = !laptop_mode,
2504                 .may_unmap = 1,
2505                 .may_swap = !noswap,
2506                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2507                 .order = 0,
2508                 .target_mem_cgroup = memcg,
2509                 .nodemask = NULL, /* we don't care the placement */
2510                 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2511                                 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2512         };
2513         struct shrink_control shrink = {
2514                 .gfp_mask = sc.gfp_mask,
2515         };
2516
2517         /*
2518          * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2519          * take care of from where we get pages. So the node where we start the
2520          * scan does not need to be the current node.
2521          */
2522         nid = mem_cgroup_select_victim_node(memcg);
2523
2524         zonelist = NODE_DATA(nid)->node_zonelists;
2525
2526         trace_mm_vmscan_memcg_reclaim_begin(0,
2527                                             sc.may_writepage,
2528                                             sc.gfp_mask);
2529
2530         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2531
2532         trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2533
2534         return nr_reclaimed;
2535 }
2536 #endif
2537
2538 static void age_active_anon(struct zone *zone, struct scan_control *sc,
2539                             int priority)
2540 {
2541         struct mem_cgroup *memcg;
2542
2543         if (!total_swap_pages)
2544                 return;
2545
2546         memcg = mem_cgroup_iter(NULL, NULL, NULL);
2547         do {
2548                 struct mem_cgroup_zone mz = {
2549                         .mem_cgroup = memcg,
2550                         .zone = zone,
2551                 };
2552
2553                 if (inactive_anon_is_low(&mz))
2554                         shrink_active_list(SWAP_CLUSTER_MAX, &mz,
2555                                            sc, priority, 0);
2556
2557                 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2558         } while (memcg);
2559 }
2560
2561 /*
2562  * pgdat_balanced is used when checking if a node is balanced for high-order
2563  * allocations. Only zones that meet watermarks and are in a zone allowed
2564  * by the callers classzone_idx are added to balanced_pages. The total of
2565  * balanced pages must be at least 25% of the zones allowed by classzone_idx
2566  * for the node to be considered balanced. Forcing all zones to be balanced
2567  * for high orders can cause excessive reclaim when there are imbalanced zones.
2568  * The choice of 25% is due to
2569  *   o a 16M DMA zone that is balanced will not balance a zone on any
2570  *     reasonable sized machine
2571  *   o On all other machines, the top zone must be at least a reasonable
2572  *     percentage of the middle zones. For example, on 32-bit x86, highmem
2573  *     would need to be at least 256M for it to be balance a whole node.
2574  *     Similarly, on x86-64 the Normal zone would need to be at least 1G
2575  *     to balance a node on its own. These seemed like reasonable ratios.
2576  */
2577 static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2578                                                 int classzone_idx)
2579 {
2580         unsigned long present_pages = 0;
2581         int i;
2582
2583         for (i = 0; i <= classzone_idx; i++)
2584                 present_pages += pgdat->node_zones[i].present_pages;
2585
2586         /* A special case here: if zone has no page, we think it's balanced */
2587         return balanced_pages >= (present_pages >> 2);
2588 }
2589
2590 /* is kswapd sleeping prematurely? */
2591 static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2592                                         int classzone_idx)
2593 {
2594         int i;
2595         unsigned long balanced = 0;
2596         bool all_zones_ok = true;
2597
2598         /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2599         if (remaining)
2600                 return true;
2601
2602         /* Check the watermark levels */
2603         for (i = 0; i <= classzone_idx; i++) {
2604                 struct zone *zone = pgdat->node_zones + i;
2605
2606                 if (!populated_zone(zone))
2607                         continue;
2608
2609                 /*
2610                  * balance_pgdat() skips over all_unreclaimable after
2611                  * DEF_PRIORITY. Effectively, it considers them balanced so
2612                  * they must be considered balanced here as well if kswapd
2613                  * is to sleep
2614                  */
2615                 if (zone->all_unreclaimable) {
2616                         balanced += zone->present_pages;
2617                         continue;
2618                 }
2619
2620                 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2621                                                         i, 0))
2622                         all_zones_ok = false;
2623                 else
2624                         balanced += zone->present_pages;
2625         }
2626
2627         /*
2628          * For high-order requests, the balanced zones must contain at least
2629          * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2630          * must be balanced
2631          */
2632         if (order)
2633                 return !pgdat_balanced(pgdat, balanced, classzone_idx);
2634         else
2635                 return !all_zones_ok;
2636 }
2637
2638 /*
2639  * For kswapd, balance_pgdat() will work across all this node's zones until
2640  * they are all at high_wmark_pages(zone).
2641  *
2642  * Returns the final order kswapd was reclaiming at
2643  *
2644  * There is special handling here for zones which are full of pinned pages.
2645  * This can happen if the pages are all mlocked, or if they are all used by
2646  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
2647  * What we do is to detect the case where all pages in the zone have been
2648  * scanned twice and there has been zero successful reclaim.  Mark the zone as
2649  * dead and from now on, only perform a short scan.  Basically we're polling
2650  * the zone for when the problem goes away.
2651  *
2652  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
2653  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2654  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2655  * lower zones regardless of the number of free pages in the lower zones. This
2656  * interoperates with the page allocator fallback scheme to ensure that aging
2657  * of pages is balanced across the zones.
2658  */
2659 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2660                                                         int *classzone_idx)
2661 {
2662         int all_zones_ok;
2663         unsigned long balanced;
2664         int priority;
2665         int i;
2666         int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
2667         unsigned long total_scanned;
2668         struct reclaim_state *reclaim_state = current->reclaim_state;
2669         unsigned long nr_soft_reclaimed;
2670         unsigned long nr_soft_scanned;
2671         struct scan_control sc = {
2672                 .gfp_mask = GFP_KERNEL,
2673                 .may_unmap = 1,
2674                 .may_swap = 1,
2675                 /*
2676                  * kswapd doesn't want to be bailed out while reclaim. because
2677                  * we want to put equal scanning pressure on each zone.
2678                  */
2679                 .nr_to_reclaim = ULONG_MAX,
2680                 .order = order,
2681                 .target_mem_cgroup = NULL,
2682         };
2683         struct shrink_control shrink = {
2684                 .gfp_mask = sc.gfp_mask,
2685         };
2686 loop_again:
2687         total_scanned = 0;
2688         sc.nr_reclaimed = 0;
2689         sc.may_writepage = !laptop_mode;
2690         count_vm_event(PAGEOUTRUN);
2691
2692         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2693                 unsigned long lru_pages = 0;
2694                 int has_under_min_watermark_zone = 0;
2695
2696                 /* The swap token gets in the way of swapout... */
2697                 if (!priority)
2698                         disable_swap_token(NULL);
2699
2700                 all_zones_ok = 1;
2701                 balanced = 0;
2702
2703                 /*
2704                  * Scan in the highmem->dma direction for the highest
2705                  * zone which needs scanning
2706                  */
2707                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2708                         struct zone *zone = pgdat->node_zones + i;
2709
2710                         if (!populated_zone(zone))
2711                                 continue;
2712
2713                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2714                                 continue;
2715
2716                         /*
2717                          * Do some background aging of the anon list, to give
2718                          * pages a chance to be referenced before reclaiming.
2719                          */
2720                         age_active_anon(zone, &sc, priority);
2721
2722                         if (!zone_watermark_ok_safe(zone, order,
2723                                         high_wmark_pages(zone), 0, 0)) {
2724                                 end_zone = i;
2725                                 break;
2726                         } else {
2727                                 /* If balanced, clear the congested flag */
2728                                 zone_clear_flag(zone, ZONE_CONGESTED);
2729                         }
2730                 }
2731                 if (i < 0)
2732                         goto out;
2733
2734                 for (i = 0; i <= end_zone; i++) {
2735                         struct zone *zone = pgdat->node_zones + i;
2736
2737                         lru_pages += zone_reclaimable_pages(zone);
2738                 }
2739
2740                 /*
2741                  * Now scan the zone in the dma->highmem direction, stopping
2742                  * at the last zone which needs scanning.
2743                  *
2744                  * We do this because the page allocator works in the opposite
2745                  * direction.  This prevents the page allocator from allocating
2746                  * pages behind kswapd's direction of progress, which would
2747                  * cause too much scanning of the lower zones.
2748                  */
2749                 for (i = 0; i <= end_zone; i++) {
2750                         struct zone *zone = pgdat->node_zones + i;
2751                         int nr_slab;
2752                         unsigned long balance_gap;
2753
2754                         if (!populated_zone(zone))
2755                                 continue;
2756
2757                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2758                                 continue;
2759
2760                         sc.nr_scanned = 0;
2761
2762                         nr_soft_scanned = 0;
2763                         /*
2764                          * Call soft limit reclaim before calling shrink_zone.
2765                          */
2766                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2767                                                         order, sc.gfp_mask,
2768                                                         &nr_soft_scanned);
2769                         sc.nr_reclaimed += nr_soft_reclaimed;
2770                         total_scanned += nr_soft_scanned;
2771
2772                         /*
2773                          * We put equal pressure on every zone, unless
2774                          * one zone has way too many pages free
2775                          * already. The "too many pages" is defined
2776                          * as the high wmark plus a "gap" where the
2777                          * gap is either the low watermark or 1%
2778                          * of the zone, whichever is smaller.
2779                          */
2780                         balance_gap = min(low_wmark_pages(zone),
2781                                 (zone->present_pages +
2782                                         KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2783                                 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2784                         if (!zone_watermark_ok_safe(zone, order,
2785                                         high_wmark_pages(zone) + balance_gap,
2786                                         end_zone, 0)) {
2787                                 shrink_zone(priority, zone, &sc);
2788
2789                                 reclaim_state->reclaimed_slab = 0;
2790                                 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2791                                 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2792                                 total_scanned += sc.nr_scanned;
2793
2794                                 if (nr_slab == 0 && !zone_reclaimable(zone))
2795                                         zone->all_unreclaimable = 1;
2796                         }
2797
2798                         /*
2799                          * If we've done a decent amount of scanning and
2800                          * the reclaim ratio is low, start doing writepage
2801                          * even in laptop mode
2802                          */
2803                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2804                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2805                                 sc.may_writepage = 1;
2806
2807                         if (zone->all_unreclaimable) {
2808                                 if (end_zone && end_zone == i)
2809                                         end_zone--;
2810                                 continue;
2811                         }
2812
2813                         if (!zone_watermark_ok_safe(zone, order,
2814                                         high_wmark_pages(zone), end_zone, 0)) {
2815                                 all_zones_ok = 0;
2816                                 /*
2817                                  * We are still under min water mark.  This
2818                                  * means that we have a GFP_ATOMIC allocation
2819                                  * failure risk. Hurry up!
2820                                  */
2821                                 if (!zone_watermark_ok_safe(zone, order,
2822                                             min_wmark_pages(zone), end_zone, 0))
2823                                         has_under_min_watermark_zone = 1;
2824                         } else {
2825                                 /*
2826                                  * If a zone reaches its high watermark,
2827                                  * consider it to be no longer congested. It's
2828                                  * possible there are dirty pages backed by
2829                                  * congested BDIs but as pressure is relieved,
2830                                  * spectulatively avoid congestion waits
2831                                  */
2832                                 zone_clear_flag(zone, ZONE_CONGESTED);
2833                                 if (i <= *classzone_idx)
2834                                         balanced += zone->present_pages;
2835                         }
2836
2837                 }
2838                 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2839                         break;          /* kswapd: all done */
2840                 /*
2841                  * OK, kswapd is getting into trouble.  Take a nap, then take
2842                  * another pass across the zones.
2843                  */
2844                 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2845                         if (has_under_min_watermark_zone)
2846                                 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2847                         else
2848                                 congestion_wait(BLK_RW_ASYNC, HZ/10);
2849                 }
2850
2851                 /*
2852                  * We do this so kswapd doesn't build up large priorities for
2853                  * example when it is freeing in parallel with allocators. It
2854                  * matches the direct reclaim path behaviour in terms of impact
2855                  * on zone->*_priority.
2856                  */
2857                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2858                         break;
2859         }
2860 out:
2861
2862         /*
2863          * order-0: All zones must meet high watermark for a balanced node
2864          * high-order: Balanced zones must make up at least 25% of the node
2865          *             for the node to be balanced
2866          */
2867         if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2868                 cond_resched();
2869
2870                 try_to_freeze();
2871
2872                 /*
2873                  * Fragmentation may mean that the system cannot be
2874                  * rebalanced for high-order allocations in all zones.
2875                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2876                  * it means the zones have been fully scanned and are still
2877                  * not balanced. For high-order allocations, there is
2878                  * little point trying all over again as kswapd may
2879                  * infinite loop.
2880                  *
2881                  * Instead, recheck all watermarks at order-0 as they
2882                  * are the most important. If watermarks are ok, kswapd will go
2883                  * back to sleep. High-order users can still perform direct
2884                  * reclaim if they wish.
2885                  */
2886                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2887                         order = sc.order = 0;
2888
2889                 goto loop_again;
2890         }
2891
2892         /*
2893          * If kswapd was reclaiming at a higher order, it has the option of
2894          * sleeping without all zones being balanced. Before it does, it must
2895          * ensure that the watermarks for order-0 on *all* zones are met and
2896          * that the congestion flags are cleared. The congestion flag must
2897          * be cleared as kswapd is the only mechanism that clears the flag
2898          * and it is potentially going to sleep here.
2899          */
2900         if (order) {
2901                 for (i = 0; i <= end_zone; i++) {
2902                         struct zone *zone = pgdat->node_zones + i;
2903
2904                         if (!populated_zone(zone))
2905                                 continue;
2906
2907                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2908                                 continue;
2909
2910                         /* Confirm the zone is balanced for order-0 */
2911                         if (!zone_watermark_ok(zone, 0,
2912                                         high_wmark_pages(zone), 0, 0)) {
2913                                 order = sc.order = 0;
2914                                 goto loop_again;
2915                         }
2916
2917                         /* If balanced, clear the congested flag */
2918                         zone_clear_flag(zone, ZONE_CONGESTED);
2919                         if (i <= *classzone_idx)
2920                                 balanced += zone->present_pages;
2921                 }
2922         }
2923
2924         /*
2925          * Return the order we were reclaiming at so sleeping_prematurely()
2926          * makes a decision on the order we were last reclaiming at. However,
2927          * if another caller entered the allocator slow path while kswapd
2928          * was awake, order will remain at the higher level
2929          */
2930         *classzone_idx = end_zone;
2931         return order;
2932 }
2933
2934 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2935 {
2936         long remaining = 0;
2937         DEFINE_WAIT(wait);
2938
2939         if (freezing(current) || kthread_should_stop())
2940                 return;
2941
2942         prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2943
2944         /* Try to sleep for a short interval */
2945         if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2946                 remaining = schedule_timeout(HZ/10);
2947                 finish_wait(&pgdat->kswapd_wait, &wait);
2948                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2949         }
2950
2951         /*
2952          * After a short sleep, check if it was a premature sleep. If not, then
2953          * go fully to sleep until explicitly woken up.
2954          */
2955         if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2956                 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2957
2958                 /*
2959                  * vmstat counters are not perfectly accurate and the estimated
2960                  * value for counters such as NR_FREE_PAGES can deviate from the
2961                  * true value by nr_online_cpus * threshold. To avoid the zone
2962                  * watermarks being breached while under pressure, we reduce the
2963                  * per-cpu vmstat threshold while kswapd is awake and restore
2964                  * them before going back to sleep.
2965                  */
2966                 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2967                 schedule();
2968                 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2969         } else {
2970                 if (remaining)
2971                         count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2972                 else
2973                         count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2974         }
2975         finish_wait(&pgdat->kswapd_wait, &wait);
2976 }
2977
2978 /*
2979  * The background pageout daemon, started as a kernel thread
2980  * from the init process.
2981  *
2982  * This basically trickles out pages so that we have _some_
2983  * free memory available even if there is no other activity
2984  * that frees anything up. This is needed for things like routing
2985  * etc, where we otherwise might have all activity going on in
2986  * asynchronous contexts that cannot page things out.
2987  *
2988  * If there are applications that are active memory-allocators
2989  * (most normal use), this basically shouldn't matter.
2990  */
2991 static int kswapd(void *p)
2992 {
2993         unsigned long order, new_order;
2994         unsigned balanced_order;
2995         int classzone_idx, new_classzone_idx;
2996         int balanced_classzone_idx;
2997         pg_data_t *pgdat = (pg_data_t*)p;
2998         struct task_struct *tsk = current;
2999
3000         struct reclaim_state reclaim_state = {
3001                 .reclaimed_slab = 0,
3002         };
3003         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3004
3005         lockdep_set_current_reclaim_state(GFP_KERNEL);
3006
3007         if (!cpumask_empty(cpumask))
3008                 set_cpus_allowed_ptr(tsk, cpumask);
3009         current->reclaim_state = &reclaim_state;
3010
3011         /*
3012          * Tell the memory management that we're a "memory allocator",
3013          * and that if we need more memory we should get access to it
3014          * regardless (see "__alloc_pages()"). "kswapd" should
3015          * never get caught in the normal page freeing logic.
3016          *
3017          * (Kswapd normally doesn't need memory anyway, but sometimes
3018          * you need a small amount of memory in order to be able to
3019          * page out something else, and this flag essentially protects
3020          * us from recursively trying to free more memory as we're
3021          * trying to free the first piece of memory in the first place).
3022          */
3023         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3024         set_freezable();
3025
3026         order = new_order = 0;
3027         balanced_order = 0;
3028         classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
3029         balanced_classzone_idx = classzone_idx;
3030         for ( ; ; ) {
3031                 int ret;
3032
3033                 /*
3034                  * If the last balance_pgdat was unsuccessful it's unlikely a
3035                  * new request of a similar or harder type will succeed soon
3036                  * so consider going to sleep on the basis we reclaimed at
3037                  */
3038                 if (balanced_classzone_idx >= new_classzone_idx &&
3039                                         balanced_order == new_order) {
3040                         new_order = pgdat->kswapd_max_order;
3041                         new_classzone_idx = pgdat->classzone_idx;
3042                         pgdat->kswapd_max_order =  0;
3043                         pgdat->classzone_idx = pgdat->nr_zones - 1;
3044                 }
3045
3046                 if (order < new_order || classzone_idx > new_classzone_idx) {
3047                         /*
3048                          * Don't sleep if someone wants a larger 'order'
3049                          * allocation or has tigher zone constraints
3050                          */
3051                         order = new_order;
3052                         classzone_idx = new_classzone_idx;
3053                 } else {
3054                         kswapd_try_to_sleep(pgdat, balanced_order,
3055                                                 balanced_classzone_idx);
3056                         order = pgdat->kswapd_max_order;
3057                         classzone_idx = pgdat->classzone_idx;
3058                         new_order = order;
3059                         new_classzone_idx = classzone_idx;
3060                         pgdat->kswapd_max_order = 0;
3061                         pgdat->classzone_idx = pgdat->nr_zones - 1;
3062                 }
3063
3064                 ret = try_to_freeze();
3065                 if (kthread_should_stop())
3066                         break;
3067
3068                 /*
3069                  * We can speed up thawing tasks if we don't call balance_pgdat
3070                  * after returning from the refrigerator
3071                  */
3072                 if (!ret) {
3073                         trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
3074                         balanced_classzone_idx = classzone_idx;
3075                         balanced_order = balance_pgdat(pgdat, order,
3076                                                 &balanced_classzone_idx);
3077                 }
3078         }
3079         return 0;
3080 }
3081
3082 /*
3083  * A zone is low on free memory, so wake its kswapd task to service it.
3084  */
3085 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3086 {
3087         pg_data_t *pgdat;
3088
3089         if (!populated_zone(zone))
3090                 return;
3091
3092         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
3093                 return;
3094         pgdat = zone->zone_pgdat;
3095         if (pgdat->kswapd_max_order < order) {
3096                 pgdat->kswapd_max_order = order;
3097                 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3098         }
3099         if (!waitqueue_active(&pgdat->kswapd_wait))
3100                 return;
3101         if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
3102                 return;
3103
3104         trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3105         wake_up_interruptible(&pgdat->kswapd_wait);
3106 }
3107
3108 /*
3109  * The reclaimable count would be mostly accurate.
3110  * The less reclaimable pages may be
3111  * - mlocked pages, which will be moved to unevictable list when encountered
3112  * - mapped pages, which may require several travels to be reclaimed
3113  * - dirty pages, which is not "instantly" reclaimable
3114  */
3115 unsigned long global_reclaimable_pages(void)
3116 {
3117         int nr;
3118
3119         nr = global_page_state(NR_ACTIVE_FILE) +
3120              global_page_state(NR_INACTIVE_FILE);
3121
3122         if (nr_swap_pages > 0)
3123                 nr += global_page_state(NR_ACTIVE_ANON) +
3124                       global_page_state(NR_INACTIVE_ANON);
3125
3126         return nr;
3127 }
3128
3129 unsigned long zone_reclaimable_pages(struct zone *zone)
3130 {
3131         int nr;
3132
3133         nr = zone_page_state(zone, NR_ACTIVE_FILE) +
3134              zone_page_state(zone, NR_INACTIVE_FILE);
3135
3136         if (nr_swap_pages > 0)
3137                 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
3138                       zone_page_state(zone, NR_INACTIVE_ANON);
3139
3140         return nr;
3141 }
3142
3143 #ifdef CONFIG_HIBERNATION
3144 /*
3145  * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3146  * freed pages.
3147  *
3148  * Rather than trying to age LRUs the aim is to preserve the overall
3149  * LRU order by reclaiming preferentially
3150  * inactive > active > active referenced > active mapped
3151  */
3152 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3153 {
3154         struct reclaim_state reclaim_state;
3155         struct scan_control sc = {
3156                 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3157                 .may_swap = 1,
3158                 .may_unmap = 1,
3159                 .may_writepage = 1,
3160                 .nr_to_reclaim = nr_to_reclaim,
3161                 .hibernation_mode = 1,
3162                 .order = 0,
3163         };
3164         struct shrink_control shrink = {
3165                 .gfp_mask = sc.gfp_mask,
3166         };
3167         struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3168         struct task_struct *p = current;
3169         unsigned long nr_reclaimed;
3170
3171         p->flags |= PF_MEMALLOC;
3172         lockdep_set_current_reclaim_state(sc.gfp_mask);
3173         reclaim_state.reclaimed_slab = 0;
3174         p->reclaim_state = &reclaim_state;
3175
3176         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
3177
3178         p->reclaim_state = NULL;
3179         lockdep_clear_current_reclaim_state();
3180         p->flags &= ~PF_MEMALLOC;
3181
3182         return nr_reclaimed;
3183 }
3184 #endif /* CONFIG_HIBERNATION */
3185
3186 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3187    not required for correctness.  So if the last cpu in a node goes
3188    away, we get changed to run anywhere: as the first one comes back,
3189    restore their cpu bindings. */
3190 static int __devinit cpu_callback(struct notifier_block *nfb,
3191                                   unsigned long action, void *hcpu)
3192 {
3193         int nid;
3194
3195         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3196                 for_each_node_state(nid, N_HIGH_MEMORY) {
3197                         pg_data_t *pgdat = NODE_DATA(nid);
3198                         const struct cpumask *mask;
3199
3200                         mask = cpumask_of_node(pgdat->node_id);
3201
3202                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3203                                 /* One of our CPUs online: restore mask */
3204                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3205                 }
3206         }
3207         return NOTIFY_OK;
3208 }
3209
3210 /*
3211  * This kswapd start function will be called by init and node-hot-add.
3212  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3213  */
3214 int kswapd_run(int nid)
3215 {
3216         pg_data_t *pgdat = NODE_DATA(nid);
3217         int ret = 0;
3218
3219         if (pgdat->kswapd)
3220                 return 0;
3221
3222         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3223         if (IS_ERR(pgdat->kswapd)) {
3224                 /* failure at boot is fatal */
3225                 BUG_ON(system_state == SYSTEM_BOOTING);
3226                 printk("Failed to start kswapd on node %d\n",nid);
3227                 ret = -1;
3228         }
3229         return ret;
3230 }
3231
3232 /*
3233  * Called by memory hotplug when all memory in a node is offlined.
3234  */
3235 void kswapd_stop(int nid)
3236 {
3237         struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3238
3239         if (kswapd)
3240                 kthread_stop(kswapd);
3241 }
3242
3243 static int __init kswapd_init(void)
3244 {
3245         int nid;
3246
3247         swap_setup();
3248         for_each_node_state(nid, N_HIGH_MEMORY)
3249                 kswapd_run(nid);
3250         hotcpu_notifier(cpu_callback, 0);
3251         return 0;
3252 }
3253
3254 module_init(kswapd_init)
3255
3256 #ifdef CONFIG_NUMA
3257 /*
3258  * Zone reclaim mode
3259  *
3260  * If non-zero call zone_reclaim when the number of free pages falls below
3261  * the watermarks.
3262  */
3263 int zone_reclaim_mode __read_mostly;
3264
3265 #define RECLAIM_OFF 0
3266 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
3267 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
3268 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
3269
3270 /*
3271  * Priority for ZONE_RECLAIM. This determines the fraction of pages
3272  * of a node considered for each zone_reclaim. 4 scans 1/16th of
3273  * a zone.
3274  */
3275 #define ZONE_RECLAIM_PRIORITY 4
3276
3277 /*
3278  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3279  * occur.
3280  */
3281 int sysctl_min_unmapped_ratio = 1;
3282
3283 /*
3284  * If the number of slab pages in a zone grows beyond this percentage then
3285  * slab reclaim needs to occur.
3286  */
3287 int sysctl_min_slab_ratio = 5;
3288
3289 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3290 {
3291         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3292         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3293                 zone_page_state(zone, NR_ACTIVE_FILE);
3294
3295         /*
3296          * It's possible for there to be more file mapped pages than
3297          * accounted for by the pages on the file LRU lists because
3298          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3299          */
3300         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3301 }
3302
3303 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3304 static long zone_pagecache_reclaimable(struct zone *zone)
3305 {
3306         long nr_pagecache_reclaimable;
3307         long delta = 0;
3308
3309         /*
3310          * If RECLAIM_SWAP is set, then all file pages are considered
3311          * potentially reclaimable. Otherwise, we have to worry about
3312          * pages like swapcache and zone_unmapped_file_pages() provides
3313          * a better estimate
3314          */
3315         if (zone_reclaim_mode & RECLAIM_SWAP)
3316                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3317         else
3318                 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3319
3320         /* If we can't clean pages, remove dirty pages from consideration */
3321         if (!(zone_reclaim_mode & RECLAIM_WRITE))
3322                 delta += zone_page_state(zone, NR_FILE_DIRTY);
3323
3324         /* Watch for any possible underflows due to delta */
3325         if (unlikely(delta > nr_pagecache_reclaimable))
3326                 delta = nr_pagecache_reclaimable;
3327
3328         return nr_pagecache_reclaimable - delta;
3329 }
3330
3331 /*
3332  * Try to free up some pages from this zone through reclaim.
3333  */
3334 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3335 {
3336         /* Minimum pages needed in order to stay on node */
3337         const unsigned long nr_pages = 1 << order;
3338         struct task_struct *p = current;
3339         struct reclaim_state reclaim_state;
3340         int priority;
3341         struct scan_control sc = {
3342                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3343                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3344                 .may_swap = 1,
3345                 .nr_to_reclaim = max_t(unsigned long, nr_pages,
3346                                        SWAP_CLUSTER_MAX),
3347                 .gfp_mask = gfp_mask,
3348                 .order = order,
3349         };
3350         struct shrink_control shrink = {
3351                 .gfp_mask = sc.gfp_mask,
3352         };
3353         unsigned long nr_slab_pages0, nr_slab_pages1;
3354
3355         cond_resched();
3356         /*
3357          * We need to be able to allocate from the reserves for RECLAIM_SWAP
3358          * and we also need to be able to write out pages for RECLAIM_WRITE
3359          * and RECLAIM_SWAP.
3360          */
3361         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3362         lockdep_set_current_reclaim_state(gfp_mask);
3363         reclaim_state.reclaimed_slab = 0;
3364         p->reclaim_state = &reclaim_state;
3365
3366         if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3367                 /*
3368                  * Free memory by calling shrink zone with increasing
3369                  * priorities until we have enough memory freed.
3370                  */
3371                 priority = ZONE_RECLAIM_PRIORITY;
3372                 do {
3373                         shrink_zone(priority, zone, &sc);
3374                         priority--;
3375                 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
3376         }
3377
3378         nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3379         if (nr_slab_pages0 > zone->min_slab_pages) {
3380                 /*
3381                  * shrink_slab() does not currently allow us to determine how
3382                  * many pages were freed in this zone. So we take the current
3383                  * number of slab pages and shake the slab until it is reduced
3384                  * by the same nr_pages that we used for reclaiming unmapped
3385                  * pages.
3386                  *
3387                  * Note that shrink_slab will free memory on all zones and may
3388                  * take a long time.
3389                  */
3390                 for (;;) {
3391                         unsigned long lru_pages = zone_reclaimable_pages(zone);
3392
3393                         /* No reclaimable slab or very low memory pressure */
3394                         if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3395                                 break;
3396
3397                         /* Freed enough memory */
3398                         nr_slab_pages1 = zone_page_state(zone,
3399                                                         NR_SLAB_RECLAIMABLE);
3400                         if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3401                                 break;
3402                 }
3403
3404                 /*
3405                  * Update nr_reclaimed by the number of slab pages we
3406                  * reclaimed from this zone.
3407                  */
3408                 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3409                 if (nr_slab_pages1 < nr_slab_pages0)
3410                         sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3411         }
3412
3413         p->reclaim_state = NULL;
3414         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3415         lockdep_clear_current_reclaim_state();
3416         return sc.nr_reclaimed >= nr_pages;
3417 }
3418
3419 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3420 {
3421         int node_id;
3422         int ret;
3423
3424         /*
3425          * Zone reclaim reclaims unmapped file backed pages and
3426          * slab pages if we are over the defined limits.
3427          *
3428          * A small portion of unmapped file backed pages is needed for
3429          * file I/O otherwise pages read by file I/O will be immediately
3430          * thrown out if the zone is overallocated. So we do not reclaim
3431          * if less than a specified percentage of the zone is used by
3432          * unmapped file backed pages.
3433          */
3434         if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3435             zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3436                 return ZONE_RECLAIM_FULL;
3437
3438         if (zone->all_unreclaimable)
3439                 return ZONE_RECLAIM_FULL;
3440
3441         /*
3442          * Do not scan if the allocation should not be delayed.
3443          */
3444         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3445                 return ZONE_RECLAIM_NOSCAN;
3446
3447         /*
3448          * Only run zone reclaim on the local zone or on zones that do not
3449          * have associated processors. This will favor the local processor
3450          * over remote processors and spread off node memory allocations
3451          * as wide as possible.
3452          */
3453         node_id = zone_to_nid(zone);
3454         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3455                 return ZONE_RECLAIM_NOSCAN;
3456
3457         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3458                 return ZONE_RECLAIM_NOSCAN;
3459
3460         ret = __zone_reclaim(zone, gfp_mask, order);
3461         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3462
3463         if (!ret)
3464                 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3465
3466         return ret;
3467 }
3468 #endif
3469
3470 /*
3471  * page_evictable - test whether a page is evictable
3472  * @page: the page to test
3473  * @vma: the VMA in which the page is or will be mapped, may be NULL
3474  *
3475  * Test whether page is evictable--i.e., should be placed on active/inactive
3476  * lists vs unevictable list.  The vma argument is !NULL when called from the
3477  * fault path to determine how to instantate a new page.
3478  *
3479  * Reasons page might not be evictable:
3480  * (1) page's mapping marked unevictable
3481  * (2) page is part of an mlocked VMA
3482  *
3483  */
3484 int page_evictable(struct page *page, struct vm_area_struct *vma)
3485 {
3486
3487         if (mapping_unevictable(page_mapping(page)))
3488                 return 0;
3489
3490         if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
3491                 return 0;
3492
3493         return 1;
3494 }
3495
3496 /**
3497  * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
3498  * @page: page to check evictability and move to appropriate lru list
3499  * @zone: zone page is in
3500  *
3501  * Checks a page for evictability and moves the page to the appropriate
3502  * zone lru list.
3503  *
3504  * Restrictions: zone->lru_lock must be held, page must be on LRU and must
3505  * have PageUnevictable set.
3506  */
3507 static void check_move_unevictable_page(struct page *page, struct zone *zone)
3508 {
3509         struct lruvec *lruvec;
3510
3511         VM_BUG_ON(PageActive(page));
3512 retry:
3513         ClearPageUnevictable(page);
3514         if (page_evictable(page, NULL)) {
3515                 enum lru_list l = page_lru_base_type(page);
3516
3517                 __dec_zone_state(zone, NR_UNEVICTABLE);
3518                 lruvec = mem_cgroup_lru_move_lists(zone, page,
3519                                                    LRU_UNEVICTABLE, l);
3520                 list_move(&page->lru, &lruvec->lists[l]);
3521                 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
3522                 __count_vm_event(UNEVICTABLE_PGRESCUED);
3523         } else {
3524                 /*
3525                  * rotate unevictable list
3526                  */
3527                 SetPageUnevictable(page);
3528                 lruvec = mem_cgroup_lru_move_lists(zone, page, LRU_UNEVICTABLE,
3529                                                    LRU_UNEVICTABLE);
3530                 list_move(&page->lru, &lruvec->lists[LRU_UNEVICTABLE]);
3531                 if (page_evictable(page, NULL))
3532                         goto retry;
3533         }
3534 }
3535
3536 /**
3537  * scan_mapping_unevictable_pages - scan an address space for evictable pages
3538  * @mapping: struct address_space to scan for evictable pages
3539  *
3540  * Scan all pages in mapping.  Check unevictable pages for
3541  * evictability and move them to the appropriate zone lru list.
3542  */
3543 void scan_mapping_unevictable_pages(struct address_space *mapping)
3544 {
3545         pgoff_t next = 0;
3546         pgoff_t end   = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
3547                          PAGE_CACHE_SHIFT;
3548         struct zone *zone;
3549         struct pagevec pvec;
3550
3551         if (mapping->nrpages == 0)
3552                 return;
3553
3554         pagevec_init(&pvec, 0);
3555         while (next < end &&
3556                 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
3557                 int i;
3558                 int pg_scanned = 0;
3559
3560                 zone = NULL;
3561
3562                 for (i = 0; i < pagevec_count(&pvec); i++) {
3563                         struct page *page = pvec.pages[i];
3564                         pgoff_t page_index = page->index;
3565                         struct zone *pagezone = page_zone(page);
3566
3567                         pg_scanned++;
3568                         if (page_index > next)
3569                                 next = page_index;
3570                         next++;
3571
3572                         if (pagezone != zone) {
3573                                 if (zone)
3574                                         spin_unlock_irq(&zone->lru_lock);
3575                                 zone = pagezone;
3576                                 spin_lock_irq(&zone->lru_lock);
3577                         }
3578
3579                         if (PageLRU(page) && PageUnevictable(page))
3580                                 check_move_unevictable_page(page, zone);
3581                 }
3582                 if (zone)
3583                         spin_unlock_irq(&zone->lru_lock);
3584                 pagevec_release(&pvec);
3585
3586                 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
3587         }
3588
3589 }
3590
3591 static void warn_scan_unevictable_pages(void)
3592 {
3593         printk_once(KERN_WARNING
3594                     "%s: The scan_unevictable_pages sysctl/node-interface has been "
3595                     "disabled for lack of a legitimate use case.  If you have "
3596                     "one, please send an email to linux-mm@kvack.org.\n",
3597                     current->comm);
3598 }
3599
3600 /*
3601  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
3602  * all nodes' unevictable lists for evictable pages
3603  */
3604 unsigned long scan_unevictable_pages;
3605
3606 int scan_unevictable_handler(struct ctl_table *table, int write,
3607                            void __user *buffer,
3608                            size_t *length, loff_t *ppos)
3609 {
3610         warn_scan_unevictable_pages();
3611         proc_doulongvec_minmax(table, write, buffer, length, ppos);
3612         scan_unevictable_pages = 0;
3613         return 0;
3614 }
3615
3616 #ifdef CONFIG_NUMA
3617 /*
3618  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
3619  * a specified node's per zone unevictable lists for evictable pages.
3620  */
3621
3622 static ssize_t read_scan_unevictable_node(struct device *dev,
3623                                           struct device_attribute *attr,
3624                                           char *buf)
3625 {
3626         warn_scan_unevictable_pages();
3627         return sprintf(buf, "0\n");     /* always zero; should fit... */
3628 }
3629
3630 static ssize_t write_scan_unevictable_node(struct device *dev,
3631                                            struct device_attribute *attr,
3632                                         const char *buf, size_t count)
3633 {
3634         warn_scan_unevictable_pages();
3635         return 1;
3636 }
3637
3638
3639 static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3640                         read_scan_unevictable_node,
3641                         write_scan_unevictable_node);
3642
3643 int scan_unevictable_register_node(struct node *node)
3644 {
3645         return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
3646 }
3647
3648 void scan_unevictable_unregister_node(struct node *node)
3649 {
3650         device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);
3651 }
3652 #endif