vmscan/trace: Add 'file' info to trace_mm_vmscan_lru_isolate()
[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         if ((mode & ISOLATE_CLEAN) && (PageDirty(page) || PageWriteback(page)))
1079                 return ret;
1080
1081         if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1082                 return ret;
1083
1084         if (likely(get_page_unless_zero(page))) {
1085                 /*
1086                  * Be careful not to clear PageLRU until after we're
1087                  * sure the page is not being freed elsewhere -- the
1088                  * page release code relies on it.
1089                  */
1090                 ClearPageLRU(page);
1091                 ret = 0;
1092         }
1093
1094         return ret;
1095 }
1096
1097 /*
1098  * zone->lru_lock is heavily contended.  Some of the functions that
1099  * shrink the lists perform better by taking out a batch of pages
1100  * and working on them outside the LRU lock.
1101  *
1102  * For pagecache intensive workloads, this function is the hottest
1103  * spot in the kernel (apart from copy_*_user functions).
1104  *
1105  * Appropriate locks must be held before calling this function.
1106  *
1107  * @nr_to_scan: The number of pages to look through on the list.
1108  * @src:        The LRU list to pull pages off.
1109  * @dst:        The temp list to put pages on to.
1110  * @scanned:    The number of pages that were scanned.
1111  * @order:      The caller's attempted allocation order
1112  * @mode:       One of the LRU isolation modes
1113  * @file:       True [1] if isolating file [!anon] pages
1114  *
1115  * returns how many pages were moved onto *@dst.
1116  */
1117 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1118                 struct list_head *src, struct list_head *dst,
1119                 unsigned long *scanned, int order, isolate_mode_t mode,
1120                 int file)
1121 {
1122         unsigned long nr_taken = 0;
1123         unsigned long nr_lumpy_taken = 0;
1124         unsigned long nr_lumpy_dirty = 0;
1125         unsigned long nr_lumpy_failed = 0;
1126         unsigned long scan;
1127
1128         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1129                 struct page *page;
1130                 unsigned long pfn;
1131                 unsigned long end_pfn;
1132                 unsigned long page_pfn;
1133                 int zone_id;
1134
1135                 page = lru_to_page(src);
1136                 prefetchw_prev_lru_page(page, src, flags);
1137
1138                 VM_BUG_ON(!PageLRU(page));
1139
1140                 switch (__isolate_lru_page(page, mode, file)) {
1141                 case 0:
1142                         mem_cgroup_lru_del(page);
1143                         list_move(&page->lru, dst);
1144                         nr_taken += hpage_nr_pages(page);
1145                         break;
1146
1147                 case -EBUSY:
1148                         /* else it is being freed elsewhere */
1149                         list_move(&page->lru, src);
1150                         continue;
1151
1152                 default:
1153                         BUG();
1154                 }
1155
1156                 if (!order)
1157                         continue;
1158
1159                 /*
1160                  * Attempt to take all pages in the order aligned region
1161                  * surrounding the tag page.  Only take those pages of
1162                  * the same active state as that tag page.  We may safely
1163                  * round the target page pfn down to the requested order
1164                  * as the mem_map is guaranteed valid out to MAX_ORDER,
1165                  * where that page is in a different zone we will detect
1166                  * it from its zone id and abort this block scan.
1167                  */
1168                 zone_id = page_zone_id(page);
1169                 page_pfn = page_to_pfn(page);
1170                 pfn = page_pfn & ~((1 << order) - 1);
1171                 end_pfn = pfn + (1 << order);
1172                 for (; pfn < end_pfn; pfn++) {
1173                         struct page *cursor_page;
1174
1175                         /* The target page is in the block, ignore it. */
1176                         if (unlikely(pfn == page_pfn))
1177                                 continue;
1178
1179                         /* Avoid holes within the zone. */
1180                         if (unlikely(!pfn_valid_within(pfn)))
1181                                 break;
1182
1183                         cursor_page = pfn_to_page(pfn);
1184
1185                         /* Check that we have not crossed a zone boundary. */
1186                         if (unlikely(page_zone_id(cursor_page) != zone_id))
1187                                 break;
1188
1189                         /*
1190                          * If we don't have enough swap space, reclaiming of
1191                          * anon page which don't already have a swap slot is
1192                          * pointless.
1193                          */
1194                         if (nr_swap_pages <= 0 && PageSwapBacked(cursor_page) &&
1195                             !PageSwapCache(cursor_page))
1196                                 break;
1197
1198                         if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1199                                 mem_cgroup_lru_del(cursor_page);
1200                                 list_move(&cursor_page->lru, dst);
1201                                 nr_taken += hpage_nr_pages(cursor_page);
1202                                 nr_lumpy_taken++;
1203                                 if (PageDirty(cursor_page))
1204                                         nr_lumpy_dirty++;
1205                                 scan++;
1206                         } else {
1207                                 /*
1208                                  * Check if the page is freed already.
1209                                  *
1210                                  * We can't use page_count() as that
1211                                  * requires compound_head and we don't
1212                                  * have a pin on the page here. If a
1213                                  * page is tail, we may or may not
1214                                  * have isolated the head, so assume
1215                                  * it's not free, it'd be tricky to
1216                                  * track the head status without a
1217                                  * page pin.
1218                                  */
1219                                 if (!PageTail(cursor_page) &&
1220                                     !atomic_read(&cursor_page->_count))
1221                                         continue;
1222                                 break;
1223                         }
1224                 }
1225
1226                 /* If we break out of the loop above, lumpy reclaim failed */
1227                 if (pfn < end_pfn)
1228                         nr_lumpy_failed++;
1229         }
1230
1231         *scanned = scan;
1232
1233         trace_mm_vmscan_lru_isolate(order,
1234                         nr_to_scan, scan,
1235                         nr_taken,
1236                         nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1237                         mode, file);
1238         return nr_taken;
1239 }
1240
1241 static unsigned long isolate_pages(unsigned long nr, struct mem_cgroup_zone *mz,
1242                                    struct list_head *dst,
1243                                    unsigned long *scanned, int order,
1244                                    isolate_mode_t mode, int active, int file)
1245 {
1246         struct lruvec *lruvec;
1247         int lru = LRU_BASE;
1248
1249         lruvec = mem_cgroup_zone_lruvec(mz->zone, mz->mem_cgroup);
1250         if (active)
1251                 lru += LRU_ACTIVE;
1252         if (file)
1253                 lru += LRU_FILE;
1254         return isolate_lru_pages(nr, &lruvec->lists[lru], dst,
1255                                  scanned, order, mode, file);
1256 }
1257
1258 /*
1259  * clear_active_flags() is a helper for shrink_active_list(), clearing
1260  * any active bits from the pages in the list.
1261  */
1262 static unsigned long clear_active_flags(struct list_head *page_list,
1263                                         unsigned int *count)
1264 {
1265         int nr_active = 0;
1266         int lru;
1267         struct page *page;
1268
1269         list_for_each_entry(page, page_list, lru) {
1270                 int numpages = hpage_nr_pages(page);
1271                 lru = page_lru_base_type(page);
1272                 if (PageActive(page)) {
1273                         lru += LRU_ACTIVE;
1274                         ClearPageActive(page);
1275                         nr_active += numpages;
1276                 }
1277                 if (count)
1278                         count[lru] += numpages;
1279         }
1280
1281         return nr_active;
1282 }
1283
1284 /**
1285  * isolate_lru_page - tries to isolate a page from its LRU list
1286  * @page: page to isolate from its LRU list
1287  *
1288  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1289  * vmstat statistic corresponding to whatever LRU list the page was on.
1290  *
1291  * Returns 0 if the page was removed from an LRU list.
1292  * Returns -EBUSY if the page was not on an LRU list.
1293  *
1294  * The returned page will have PageLRU() cleared.  If it was found on
1295  * the active list, it will have PageActive set.  If it was found on
1296  * the unevictable list, it will have the PageUnevictable bit set. That flag
1297  * may need to be cleared by the caller before letting the page go.
1298  *
1299  * The vmstat statistic corresponding to the list on which the page was
1300  * found will be decremented.
1301  *
1302  * Restrictions:
1303  * (1) Must be called with an elevated refcount on the page. This is a
1304  *     fundamentnal difference from isolate_lru_pages (which is called
1305  *     without a stable reference).
1306  * (2) the lru_lock must not be held.
1307  * (3) interrupts must be enabled.
1308  */
1309 int isolate_lru_page(struct page *page)
1310 {
1311         int ret = -EBUSY;
1312
1313         VM_BUG_ON(!page_count(page));
1314
1315         if (PageLRU(page)) {
1316                 struct zone *zone = page_zone(page);
1317
1318                 spin_lock_irq(&zone->lru_lock);
1319                 if (PageLRU(page)) {
1320                         int lru = page_lru(page);
1321                         ret = 0;
1322                         get_page(page);
1323                         ClearPageLRU(page);
1324
1325                         del_page_from_lru_list(zone, page, lru);
1326                 }
1327                 spin_unlock_irq(&zone->lru_lock);
1328         }
1329         return ret;
1330 }
1331
1332 /*
1333  * Are there way too many processes in the direct reclaim path already?
1334  */
1335 static int too_many_isolated(struct zone *zone, int file,
1336                 struct scan_control *sc)
1337 {
1338         unsigned long inactive, isolated;
1339
1340         if (current_is_kswapd())
1341                 return 0;
1342
1343         if (!global_reclaim(sc))
1344                 return 0;
1345
1346         if (file) {
1347                 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1348                 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1349         } else {
1350                 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1351                 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1352         }
1353
1354         return isolated > inactive;
1355 }
1356
1357 /*
1358  * TODO: Try merging with migrations version of putback_lru_pages
1359  */
1360 static noinline_for_stack void
1361 putback_lru_pages(struct mem_cgroup_zone *mz, struct scan_control *sc,
1362                   unsigned long nr_anon, unsigned long nr_file,
1363                   struct list_head *page_list)
1364 {
1365         struct page *page;
1366         struct pagevec pvec;
1367         struct zone *zone = mz->zone;
1368         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz);
1369
1370         pagevec_init(&pvec, 1);
1371
1372         /*
1373          * Put back any unfreeable pages.
1374          */
1375         spin_lock(&zone->lru_lock);
1376         while (!list_empty(page_list)) {
1377                 int lru;
1378                 page = lru_to_page(page_list);
1379                 VM_BUG_ON(PageLRU(page));
1380                 list_del(&page->lru);
1381                 if (unlikely(!page_evictable(page, NULL))) {
1382                         spin_unlock_irq(&zone->lru_lock);
1383                         putback_lru_page(page);
1384                         spin_lock_irq(&zone->lru_lock);
1385                         continue;
1386                 }
1387                 SetPageLRU(page);
1388                 lru = page_lru(page);
1389                 add_page_to_lru_list(zone, page, lru);
1390                 if (is_active_lru(lru)) {
1391                         int file = is_file_lru(lru);
1392                         int numpages = hpage_nr_pages(page);
1393                         reclaim_stat->recent_rotated[file] += numpages;
1394                 }
1395                 if (!pagevec_add(&pvec, page)) {
1396                         spin_unlock_irq(&zone->lru_lock);
1397                         __pagevec_release(&pvec);
1398                         spin_lock_irq(&zone->lru_lock);
1399                 }
1400         }
1401         __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1402         __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1403
1404         spin_unlock_irq(&zone->lru_lock);
1405         pagevec_release(&pvec);
1406 }
1407
1408 static noinline_for_stack void
1409 update_isolated_counts(struct mem_cgroup_zone *mz,
1410                        struct scan_control *sc,
1411                        unsigned long *nr_anon,
1412                        unsigned long *nr_file,
1413                        struct list_head *isolated_list)
1414 {
1415         unsigned long nr_active;
1416         struct zone *zone = mz->zone;
1417         unsigned int count[NR_LRU_LISTS] = { 0, };
1418         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz);
1419
1420         nr_active = clear_active_flags(isolated_list, count);
1421         __count_vm_events(PGDEACTIVATE, nr_active);
1422
1423         __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1424                               -count[LRU_ACTIVE_FILE]);
1425         __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1426                               -count[LRU_INACTIVE_FILE]);
1427         __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1428                               -count[LRU_ACTIVE_ANON]);
1429         __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1430                               -count[LRU_INACTIVE_ANON]);
1431
1432         *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1433         *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1434         __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1435         __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1436
1437         reclaim_stat->recent_scanned[0] += *nr_anon;
1438         reclaim_stat->recent_scanned[1] += *nr_file;
1439 }
1440
1441 /*
1442  * Returns true if a direct reclaim should wait on pages under writeback.
1443  *
1444  * If we are direct reclaiming for contiguous pages and we do not reclaim
1445  * everything in the list, try again and wait for writeback IO to complete.
1446  * This will stall high-order allocations noticeably. Only do that when really
1447  * need to free the pages under high memory pressure.
1448  */
1449 static inline bool should_reclaim_stall(unsigned long nr_taken,
1450                                         unsigned long nr_freed,
1451                                         int priority,
1452                                         struct scan_control *sc)
1453 {
1454         int lumpy_stall_priority;
1455
1456         /* kswapd should not stall on sync IO */
1457         if (current_is_kswapd())
1458                 return false;
1459
1460         /* Only stall on lumpy reclaim */
1461         if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
1462                 return false;
1463
1464         /* If we have reclaimed everything on the isolated list, no stall */
1465         if (nr_freed == nr_taken)
1466                 return false;
1467
1468         /*
1469          * For high-order allocations, there are two stall thresholds.
1470          * High-cost allocations stall immediately where as lower
1471          * order allocations such as stacks require the scanning
1472          * priority to be much higher before stalling.
1473          */
1474         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1475                 lumpy_stall_priority = DEF_PRIORITY;
1476         else
1477                 lumpy_stall_priority = DEF_PRIORITY / 3;
1478
1479         return priority <= lumpy_stall_priority;
1480 }
1481
1482 /*
1483  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1484  * of reclaimed pages
1485  */
1486 static noinline_for_stack unsigned long
1487 shrink_inactive_list(unsigned long nr_to_scan, struct mem_cgroup_zone *mz,
1488                      struct scan_control *sc, int priority, int file)
1489 {
1490         LIST_HEAD(page_list);
1491         unsigned long nr_scanned;
1492         unsigned long nr_reclaimed = 0;
1493         unsigned long nr_taken;
1494         unsigned long nr_anon;
1495         unsigned long nr_file;
1496         unsigned long nr_dirty = 0;
1497         unsigned long nr_writeback = 0;
1498         isolate_mode_t reclaim_mode = ISOLATE_INACTIVE;
1499         struct zone *zone = mz->zone;
1500
1501         while (unlikely(too_many_isolated(zone, file, sc))) {
1502                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1503
1504                 /* We are about to die and free our memory. Return now. */
1505                 if (fatal_signal_pending(current))
1506                         return SWAP_CLUSTER_MAX;
1507         }
1508
1509         set_reclaim_mode(priority, sc, false);
1510         if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
1511                 reclaim_mode |= ISOLATE_ACTIVE;
1512
1513         lru_add_drain();
1514
1515         if (!sc->may_unmap)
1516                 reclaim_mode |= ISOLATE_UNMAPPED;
1517         if (!sc->may_writepage)
1518                 reclaim_mode |= ISOLATE_CLEAN;
1519
1520         spin_lock_irq(&zone->lru_lock);
1521
1522         nr_taken = isolate_pages(nr_to_scan, mz, &page_list,
1523                                  &nr_scanned, sc->order,
1524                                  reclaim_mode, 0, file);
1525         if (global_reclaim(sc)) {
1526                 zone->pages_scanned += nr_scanned;
1527                 if (current_is_kswapd())
1528                         __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1529                                                nr_scanned);
1530                 else
1531                         __count_zone_vm_events(PGSCAN_DIRECT, zone,
1532                                                nr_scanned);
1533         }
1534
1535         if (nr_taken == 0) {
1536                 spin_unlock_irq(&zone->lru_lock);
1537                 return 0;
1538         }
1539
1540         update_isolated_counts(mz, sc, &nr_anon, &nr_file, &page_list);
1541
1542         spin_unlock_irq(&zone->lru_lock);
1543
1544         nr_reclaimed = shrink_page_list(&page_list, mz, sc, priority,
1545                                                 &nr_dirty, &nr_writeback);
1546
1547         /* Check if we should syncronously wait for writeback */
1548         if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1549                 set_reclaim_mode(priority, sc, true);
1550                 nr_reclaimed += shrink_page_list(&page_list, mz, sc,
1551                                         priority, &nr_dirty, &nr_writeback);
1552         }
1553
1554         local_irq_disable();
1555         if (current_is_kswapd())
1556                 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1557         __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1558
1559         putback_lru_pages(mz, sc, nr_anon, nr_file, &page_list);
1560
1561         /*
1562          * If reclaim is isolating dirty pages under writeback, it implies
1563          * that the long-lived page allocation rate is exceeding the page
1564          * laundering rate. Either the global limits are not being effective
1565          * at throttling processes due to the page distribution throughout
1566          * zones or there is heavy usage of a slow backing device. The
1567          * only option is to throttle from reclaim context which is not ideal
1568          * as there is no guarantee the dirtying process is throttled in the
1569          * same way balance_dirty_pages() manages.
1570          *
1571          * This scales the number of dirty pages that must be under writeback
1572          * before throttling depending on priority. It is a simple backoff
1573          * function that has the most effect in the range DEF_PRIORITY to
1574          * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1575          * in trouble and reclaim is considered to be in trouble.
1576          *
1577          * DEF_PRIORITY   100% isolated pages must be PageWriteback to throttle
1578          * DEF_PRIORITY-1  50% must be PageWriteback
1579          * DEF_PRIORITY-2  25% must be PageWriteback, kswapd in trouble
1580          * ...
1581          * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1582          *                     isolated page is PageWriteback
1583          */
1584         if (nr_writeback && nr_writeback >= (nr_taken >> (DEF_PRIORITY-priority)))
1585                 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1586
1587         trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1588                 zone_idx(zone),
1589                 nr_scanned, nr_reclaimed,
1590                 priority,
1591                 trace_shrink_flags(file, sc->reclaim_mode));
1592         return nr_reclaimed;
1593 }
1594
1595 /*
1596  * This moves pages from the active list to the inactive list.
1597  *
1598  * We move them the other way if the page is referenced by one or more
1599  * processes, from rmap.
1600  *
1601  * If the pages are mostly unmapped, the processing is fast and it is
1602  * appropriate to hold zone->lru_lock across the whole operation.  But if
1603  * the pages are mapped, the processing is slow (page_referenced()) so we
1604  * should drop zone->lru_lock around each page.  It's impossible to balance
1605  * this, so instead we remove the pages from the LRU while processing them.
1606  * It is safe to rely on PG_active against the non-LRU pages in here because
1607  * nobody will play with that bit on a non-LRU page.
1608  *
1609  * The downside is that we have to touch page->_count against each page.
1610  * But we had to alter page->flags anyway.
1611  */
1612
1613 static void move_active_pages_to_lru(struct zone *zone,
1614                                      struct list_head *list,
1615                                      enum lru_list lru)
1616 {
1617         unsigned long pgmoved = 0;
1618         struct pagevec pvec;
1619         struct page *page;
1620
1621         pagevec_init(&pvec, 1);
1622
1623         while (!list_empty(list)) {
1624                 struct lruvec *lruvec;
1625
1626                 page = lru_to_page(list);
1627
1628                 VM_BUG_ON(PageLRU(page));
1629                 SetPageLRU(page);
1630
1631                 lruvec = mem_cgroup_lru_add_list(zone, page, lru);
1632                 list_move(&page->lru, &lruvec->lists[lru]);
1633                 pgmoved += hpage_nr_pages(page);
1634
1635                 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1636                         spin_unlock_irq(&zone->lru_lock);
1637                         if (buffer_heads_over_limit)
1638                                 pagevec_strip(&pvec);
1639                         __pagevec_release(&pvec);
1640                         spin_lock_irq(&zone->lru_lock);
1641                 }
1642         }
1643         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1644         if (!is_active_lru(lru))
1645                 __count_vm_events(PGDEACTIVATE, pgmoved);
1646 }
1647
1648 static void shrink_active_list(unsigned long nr_pages,
1649                                struct mem_cgroup_zone *mz,
1650                                struct scan_control *sc,
1651                                int priority, int file)
1652 {
1653         unsigned long nr_taken;
1654         unsigned long pgscanned;
1655         unsigned long vm_flags;
1656         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1657         LIST_HEAD(l_active);
1658         LIST_HEAD(l_inactive);
1659         struct page *page;
1660         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz);
1661         unsigned long nr_rotated = 0;
1662         isolate_mode_t reclaim_mode = ISOLATE_ACTIVE;
1663         struct zone *zone = mz->zone;
1664
1665         lru_add_drain();
1666
1667         if (!sc->may_unmap)
1668                 reclaim_mode |= ISOLATE_UNMAPPED;
1669         if (!sc->may_writepage)
1670                 reclaim_mode |= ISOLATE_CLEAN;
1671
1672         spin_lock_irq(&zone->lru_lock);
1673
1674         nr_taken = isolate_pages(nr_pages, mz, &l_hold,
1675                                  &pgscanned, sc->order,
1676                                  reclaim_mode, 1, file);
1677
1678         if (global_reclaim(sc))
1679                 zone->pages_scanned += pgscanned;
1680
1681         reclaim_stat->recent_scanned[file] += nr_taken;
1682
1683         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1684         if (file)
1685                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1686         else
1687                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1688         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1689         spin_unlock_irq(&zone->lru_lock);
1690
1691         while (!list_empty(&l_hold)) {
1692                 cond_resched();
1693                 page = lru_to_page(&l_hold);
1694                 list_del(&page->lru);
1695
1696                 if (unlikely(!page_evictable(page, NULL))) {
1697                         putback_lru_page(page);
1698                         continue;
1699                 }
1700
1701                 if (page_referenced(page, 0, mz->mem_cgroup, &vm_flags)) {
1702                         nr_rotated += hpage_nr_pages(page);
1703                         /*
1704                          * Identify referenced, file-backed active pages and
1705                          * give them one more trip around the active list. So
1706                          * that executable code get better chances to stay in
1707                          * memory under moderate memory pressure.  Anon pages
1708                          * are not likely to be evicted by use-once streaming
1709                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1710                          * so we ignore them here.
1711                          */
1712                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1713                                 list_add(&page->lru, &l_active);
1714                                 continue;
1715                         }
1716                 }
1717
1718                 ClearPageActive(page);  /* we are de-activating */
1719                 list_add(&page->lru, &l_inactive);
1720         }
1721
1722         /*
1723          * Move pages back to the lru list.
1724          */
1725         spin_lock_irq(&zone->lru_lock);
1726         /*
1727          * Count referenced pages from currently used mappings as rotated,
1728          * even though only some of them are actually re-activated.  This
1729          * helps balance scan pressure between file and anonymous pages in
1730          * get_scan_ratio.
1731          */
1732         reclaim_stat->recent_rotated[file] += nr_rotated;
1733
1734         move_active_pages_to_lru(zone, &l_active,
1735                                                 LRU_ACTIVE + file * LRU_FILE);
1736         move_active_pages_to_lru(zone, &l_inactive,
1737                                                 LRU_BASE   + file * LRU_FILE);
1738         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1739         spin_unlock_irq(&zone->lru_lock);
1740 }
1741
1742 #ifdef CONFIG_SWAP
1743 static int inactive_anon_is_low_global(struct zone *zone)
1744 {
1745         unsigned long active, inactive;
1746
1747         active = zone_page_state(zone, NR_ACTIVE_ANON);
1748         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1749
1750         if (inactive * zone->inactive_ratio < active)
1751                 return 1;
1752
1753         return 0;
1754 }
1755
1756 /**
1757  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1758  * @zone: zone to check
1759  * @sc:   scan control of this context
1760  *
1761  * Returns true if the zone does not have enough inactive anon pages,
1762  * meaning some active anon pages need to be deactivated.
1763  */
1764 static int inactive_anon_is_low(struct mem_cgroup_zone *mz)
1765 {
1766         /*
1767          * If we don't have swap space, anonymous page deactivation
1768          * is pointless.
1769          */
1770         if (!total_swap_pages)
1771                 return 0;
1772
1773         if (!scanning_global_lru(mz))
1774                 return mem_cgroup_inactive_anon_is_low(mz->mem_cgroup,
1775                                                        mz->zone);
1776
1777         return inactive_anon_is_low_global(mz->zone);
1778 }
1779 #else
1780 static inline int inactive_anon_is_low(struct mem_cgroup_zone *mz)
1781 {
1782         return 0;
1783 }
1784 #endif
1785
1786 static int inactive_file_is_low_global(struct zone *zone)
1787 {
1788         unsigned long active, inactive;
1789
1790         active = zone_page_state(zone, NR_ACTIVE_FILE);
1791         inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1792
1793         return (active > inactive);
1794 }
1795
1796 /**
1797  * inactive_file_is_low - check if file pages need to be deactivated
1798  * @mz: memory cgroup and zone to check
1799  *
1800  * When the system is doing streaming IO, memory pressure here
1801  * ensures that active file pages get deactivated, until more
1802  * than half of the file pages are on the inactive list.
1803  *
1804  * Once we get to that situation, protect the system's working
1805  * set from being evicted by disabling active file page aging.
1806  *
1807  * This uses a different ratio than the anonymous pages, because
1808  * the page cache uses a use-once replacement algorithm.
1809  */
1810 static int inactive_file_is_low(struct mem_cgroup_zone *mz)
1811 {
1812         if (!scanning_global_lru(mz))
1813                 return mem_cgroup_inactive_file_is_low(mz->mem_cgroup,
1814                                                        mz->zone);
1815
1816         return inactive_file_is_low_global(mz->zone);
1817 }
1818
1819 static int inactive_list_is_low(struct mem_cgroup_zone *mz, int file)
1820 {
1821         if (file)
1822                 return inactive_file_is_low(mz);
1823         else
1824                 return inactive_anon_is_low(mz);
1825 }
1826
1827 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1828                                  struct mem_cgroup_zone *mz,
1829                                  struct scan_control *sc, int priority)
1830 {
1831         int file = is_file_lru(lru);
1832
1833         if (is_active_lru(lru)) {
1834                 if (inactive_list_is_low(mz, file))
1835                         shrink_active_list(nr_to_scan, mz, sc, priority, file);
1836                 return 0;
1837         }
1838
1839         return shrink_inactive_list(nr_to_scan, mz, sc, priority, file);
1840 }
1841
1842 static int vmscan_swappiness(struct mem_cgroup_zone *mz,
1843                              struct scan_control *sc)
1844 {
1845         if (global_reclaim(sc))
1846                 return vm_swappiness;
1847         return mem_cgroup_swappiness(mz->mem_cgroup);
1848 }
1849
1850 /*
1851  * Determine how aggressively the anon and file LRU lists should be
1852  * scanned.  The relative value of each set of LRU lists is determined
1853  * by looking at the fraction of the pages scanned we did rotate back
1854  * onto the active list instead of evict.
1855  *
1856  * nr[0] = anon pages to scan; nr[1] = file pages to scan
1857  */
1858 static void get_scan_count(struct mem_cgroup_zone *mz, struct scan_control *sc,
1859                            unsigned long *nr, int priority)
1860 {
1861         unsigned long anon, file, free;
1862         unsigned long anon_prio, file_prio;
1863         unsigned long ap, fp;
1864         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz);
1865         u64 fraction[2], denominator;
1866         enum lru_list l;
1867         int noswap = 0;
1868         bool force_scan = false;
1869
1870         /*
1871          * If the zone or memcg is small, nr[l] can be 0.  This
1872          * results in no scanning on this priority and a potential
1873          * priority drop.  Global direct reclaim can go to the next
1874          * zone and tends to have no problems. Global kswapd is for
1875          * zone balancing and it needs to scan a minimum amount. When
1876          * reclaiming for a memcg, a priority drop can cause high
1877          * latencies, so it's better to scan a minimum amount there as
1878          * well.
1879          */
1880         if (current_is_kswapd() && mz->zone->all_unreclaimable)
1881                 force_scan = true;
1882         if (!global_reclaim(sc))
1883                 force_scan = true;
1884
1885         /* If we have no swap space, do not bother scanning anon pages. */
1886         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1887                 noswap = 1;
1888                 fraction[0] = 0;
1889                 fraction[1] = 1;
1890                 denominator = 1;
1891                 goto out;
1892         }
1893
1894         anon  = zone_nr_lru_pages(mz, LRU_ACTIVE_ANON) +
1895                 zone_nr_lru_pages(mz, LRU_INACTIVE_ANON);
1896         file  = zone_nr_lru_pages(mz, LRU_ACTIVE_FILE) +
1897                 zone_nr_lru_pages(mz, LRU_INACTIVE_FILE);
1898
1899         if (global_reclaim(sc)) {
1900                 free  = zone_page_state(mz->zone, NR_FREE_PAGES);
1901                 /* If we have very few page cache pages,
1902                    force-scan anon pages. */
1903                 if (unlikely(file + free <= high_wmark_pages(mz->zone))) {
1904                         fraction[0] = 1;
1905                         fraction[1] = 0;
1906                         denominator = 1;
1907                         goto out;
1908                 }
1909         }
1910
1911         /*
1912          * With swappiness at 100, anonymous and file have the same priority.
1913          * This scanning priority is essentially the inverse of IO cost.
1914          */
1915         anon_prio = vmscan_swappiness(mz, sc);
1916         file_prio = 200 - vmscan_swappiness(mz, sc);
1917
1918         /*
1919          * OK, so we have swap space and a fair amount of page cache
1920          * pages.  We use the recently rotated / recently scanned
1921          * ratios to determine how valuable each cache is.
1922          *
1923          * Because workloads change over time (and to avoid overflow)
1924          * we keep these statistics as a floating average, which ends
1925          * up weighing recent references more than old ones.
1926          *
1927          * anon in [0], file in [1]
1928          */
1929         spin_lock_irq(&mz->zone->lru_lock);
1930         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1931                 reclaim_stat->recent_scanned[0] /= 2;
1932                 reclaim_stat->recent_rotated[0] /= 2;
1933         }
1934
1935         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1936                 reclaim_stat->recent_scanned[1] /= 2;
1937                 reclaim_stat->recent_rotated[1] /= 2;
1938         }
1939
1940         /*
1941          * The amount of pressure on anon vs file pages is inversely
1942          * proportional to the fraction of recently scanned pages on
1943          * each list that were recently referenced and in active use.
1944          */
1945         ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1946         ap /= reclaim_stat->recent_rotated[0] + 1;
1947
1948         fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1949         fp /= reclaim_stat->recent_rotated[1] + 1;
1950         spin_unlock_irq(&mz->zone->lru_lock);
1951
1952         fraction[0] = ap;
1953         fraction[1] = fp;
1954         denominator = ap + fp + 1;
1955 out:
1956         for_each_evictable_lru(l) {
1957                 int file = is_file_lru(l);
1958                 unsigned long scan;
1959
1960                 scan = zone_nr_lru_pages(mz, l);
1961                 if (priority || noswap) {
1962                         scan >>= priority;
1963                         if (!scan && force_scan)
1964                                 scan = SWAP_CLUSTER_MAX;
1965                         scan = div64_u64(scan * fraction[file], denominator);
1966                 }
1967                 nr[l] = scan;
1968         }
1969 }
1970
1971 /*
1972  * Reclaim/compaction depends on a number of pages being freed. To avoid
1973  * disruption to the system, a small number of order-0 pages continue to be
1974  * rotated and reclaimed in the normal fashion. However, by the time we get
1975  * back to the allocator and call try_to_compact_zone(), we ensure that
1976  * there are enough free pages for it to be likely successful
1977  */
1978 static inline bool should_continue_reclaim(struct mem_cgroup_zone *mz,
1979                                         unsigned long nr_reclaimed,
1980                                         unsigned long nr_scanned,
1981                                         struct scan_control *sc)
1982 {
1983         unsigned long pages_for_compaction;
1984         unsigned long inactive_lru_pages;
1985
1986         /* If not in reclaim/compaction mode, stop */
1987         if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
1988                 return false;
1989
1990         /* Consider stopping depending on scan and reclaim activity */
1991         if (sc->gfp_mask & __GFP_REPEAT) {
1992                 /*
1993                  * For __GFP_REPEAT allocations, stop reclaiming if the
1994                  * full LRU list has been scanned and we are still failing
1995                  * to reclaim pages. This full LRU scan is potentially
1996                  * expensive but a __GFP_REPEAT caller really wants to succeed
1997                  */
1998                 if (!nr_reclaimed && !nr_scanned)
1999                         return false;
2000         } else {
2001                 /*
2002                  * For non-__GFP_REPEAT allocations which can presumably
2003                  * fail without consequence, stop if we failed to reclaim
2004                  * any pages from the last SWAP_CLUSTER_MAX number of
2005                  * pages that were scanned. This will return to the
2006                  * caller faster at the risk reclaim/compaction and
2007                  * the resulting allocation attempt fails
2008                  */
2009                 if (!nr_reclaimed)
2010                         return false;
2011         }
2012
2013         /*
2014          * If we have not reclaimed enough pages for compaction and the
2015          * inactive lists are large enough, continue reclaiming
2016          */
2017         pages_for_compaction = (2UL << sc->order);
2018         inactive_lru_pages = zone_nr_lru_pages(mz, LRU_INACTIVE_FILE);
2019         if (nr_swap_pages > 0)
2020                 inactive_lru_pages += zone_nr_lru_pages(mz, LRU_INACTIVE_ANON);
2021         if (sc->nr_reclaimed < pages_for_compaction &&
2022                         inactive_lru_pages > pages_for_compaction)
2023                 return true;
2024
2025         /* If compaction would go ahead or the allocation would succeed, stop */
2026         switch (compaction_suitable(mz->zone, sc->order)) {
2027         case COMPACT_PARTIAL:
2028         case COMPACT_CONTINUE:
2029                 return false;
2030         default:
2031                 return true;
2032         }
2033 }
2034
2035 /*
2036  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
2037  */
2038 static void shrink_mem_cgroup_zone(int priority, struct mem_cgroup_zone *mz,
2039                                    struct scan_control *sc)
2040 {
2041         unsigned long nr[NR_LRU_LISTS];
2042         unsigned long nr_to_scan;
2043         enum lru_list l;
2044         unsigned long nr_reclaimed, nr_scanned;
2045         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2046         struct blk_plug plug;
2047
2048 restart:
2049         nr_reclaimed = 0;
2050         nr_scanned = sc->nr_scanned;
2051         get_scan_count(mz, sc, nr, priority);
2052
2053         blk_start_plug(&plug);
2054         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2055                                         nr[LRU_INACTIVE_FILE]) {
2056                 for_each_evictable_lru(l) {
2057                         if (nr[l]) {
2058                                 nr_to_scan = min_t(unsigned long,
2059                                                    nr[l], SWAP_CLUSTER_MAX);
2060                                 nr[l] -= nr_to_scan;
2061
2062                                 nr_reclaimed += shrink_list(l, nr_to_scan,
2063                                                             mz, sc, priority);
2064                         }
2065                 }
2066                 /*
2067                  * On large memory systems, scan >> priority can become
2068                  * really large. This is fine for the starting priority;
2069                  * we want to put equal scanning pressure on each zone.
2070                  * However, if the VM has a harder time of freeing pages,
2071                  * with multiple processes reclaiming pages, the total
2072                  * freeing target can get unreasonably large.
2073                  */
2074                 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
2075                         break;
2076         }
2077         blk_finish_plug(&plug);
2078         sc->nr_reclaimed += nr_reclaimed;
2079
2080         /*
2081          * Even if we did not try to evict anon pages at all, we want to
2082          * rebalance the anon lru active/inactive ratio.
2083          */
2084         if (inactive_anon_is_low(mz))
2085                 shrink_active_list(SWAP_CLUSTER_MAX, mz, sc, priority, 0);
2086
2087         /* reclaim/compaction might need reclaim to continue */
2088         if (should_continue_reclaim(mz, nr_reclaimed,
2089                                         sc->nr_scanned - nr_scanned, sc))
2090                 goto restart;
2091
2092         throttle_vm_writeout(sc->gfp_mask);
2093 }
2094
2095 static void shrink_zone(int priority, struct zone *zone,
2096                         struct scan_control *sc)
2097 {
2098         struct mem_cgroup *root = sc->target_mem_cgroup;
2099         struct mem_cgroup_reclaim_cookie reclaim = {
2100                 .zone = zone,
2101                 .priority = priority,
2102         };
2103         struct mem_cgroup *memcg;
2104
2105         memcg = mem_cgroup_iter(root, NULL, &reclaim);
2106         do {
2107                 struct mem_cgroup_zone mz = {
2108                         .mem_cgroup = memcg,
2109                         .zone = zone,
2110                 };
2111
2112                 shrink_mem_cgroup_zone(priority, &mz, sc);
2113                 /*
2114                  * Limit reclaim has historically picked one memcg and
2115                  * scanned it with decreasing priority levels until
2116                  * nr_to_reclaim had been reclaimed.  This priority
2117                  * cycle is thus over after a single memcg.
2118                  *
2119                  * Direct reclaim and kswapd, on the other hand, have
2120                  * to scan all memory cgroups to fulfill the overall
2121                  * scan target for the zone.
2122                  */
2123                 if (!global_reclaim(sc)) {
2124                         mem_cgroup_iter_break(root, memcg);
2125                         break;
2126                 }
2127                 memcg = mem_cgroup_iter(root, memcg, &reclaim);
2128         } while (memcg);
2129 }
2130
2131 /*
2132  * This is the direct reclaim path, for page-allocating processes.  We only
2133  * try to reclaim pages from zones which will satisfy the caller's allocation
2134  * request.
2135  *
2136  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2137  * Because:
2138  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2139  *    allocation or
2140  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2141  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2142  *    zone defense algorithm.
2143  *
2144  * If a zone is deemed to be full of pinned pages then just give it a light
2145  * scan then give up on it.
2146  *
2147  * This function returns true if a zone is being reclaimed for a costly
2148  * high-order allocation and compaction is either ready to begin or deferred.
2149  * This indicates to the caller that it should retry the allocation or fail.
2150  */
2151 static bool shrink_zones(int priority, struct zonelist *zonelist,
2152                                         struct scan_control *sc)
2153 {
2154         struct zoneref *z;
2155         struct zone *zone;
2156         unsigned long nr_soft_reclaimed;
2157         unsigned long nr_soft_scanned;
2158         bool should_abort_reclaim = false;
2159
2160         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2161                                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2162                 if (!populated_zone(zone))
2163                         continue;
2164                 /*
2165                  * Take care memory controller reclaiming has small influence
2166                  * to global LRU.
2167                  */
2168                 if (global_reclaim(sc)) {
2169                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2170                                 continue;
2171                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2172                                 continue;       /* Let kswapd poll it */
2173                         if (COMPACTION_BUILD) {
2174                                 /*
2175                                  * If we already have plenty of memory free for
2176                                  * compaction in this zone, don't free any more.
2177                                  * Even though compaction is invoked for any
2178                                  * non-zero order, only frequent costly order
2179                                  * reclamation is disruptive enough to become a
2180                                  * noticable problem, like transparent huge page
2181                                  * allocations.
2182                                  */
2183                                 if (sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2184                                         (compaction_suitable(zone, sc->order) ||
2185                                          compaction_deferred(zone))) {
2186                                         should_abort_reclaim = true;
2187                                         continue;
2188                                 }
2189                         }
2190                         /*
2191                          * This steals pages from memory cgroups over softlimit
2192                          * and returns the number of reclaimed pages and
2193                          * scanned pages. This works for global memory pressure
2194                          * and balancing, not for a memcg's limit.
2195                          */
2196                         nr_soft_scanned = 0;
2197                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2198                                                 sc->order, sc->gfp_mask,
2199                                                 &nr_soft_scanned);
2200                         sc->nr_reclaimed += nr_soft_reclaimed;
2201                         sc->nr_scanned += nr_soft_scanned;
2202                         /* need some check for avoid more shrink_zone() */
2203                 }
2204
2205                 shrink_zone(priority, zone, sc);
2206         }
2207
2208         return should_abort_reclaim;
2209 }
2210
2211 static bool zone_reclaimable(struct zone *zone)
2212 {
2213         return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2214 }
2215
2216 /* All zones in zonelist are unreclaimable? */
2217 static bool all_unreclaimable(struct zonelist *zonelist,
2218                 struct scan_control *sc)
2219 {
2220         struct zoneref *z;
2221         struct zone *zone;
2222
2223         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2224                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2225                 if (!populated_zone(zone))
2226                         continue;
2227                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2228                         continue;
2229                 if (!zone->all_unreclaimable)
2230                         return false;
2231         }
2232
2233         return true;
2234 }
2235
2236 /*
2237  * This is the main entry point to direct page reclaim.
2238  *
2239  * If a full scan of the inactive list fails to free enough memory then we
2240  * are "out of memory" and something needs to be killed.
2241  *
2242  * If the caller is !__GFP_FS then the probability of a failure is reasonably
2243  * high - the zone may be full of dirty or under-writeback pages, which this
2244  * caller can't do much about.  We kick the writeback threads and take explicit
2245  * naps in the hope that some of these pages can be written.  But if the
2246  * allocating task holds filesystem locks which prevent writeout this might not
2247  * work, and the allocation attempt will fail.
2248  *
2249  * returns:     0, if no pages reclaimed
2250  *              else, the number of pages reclaimed
2251  */
2252 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2253                                         struct scan_control *sc,
2254                                         struct shrink_control *shrink)
2255 {
2256         int priority;
2257         unsigned long total_scanned = 0;
2258         struct reclaim_state *reclaim_state = current->reclaim_state;
2259         struct zoneref *z;
2260         struct zone *zone;
2261         unsigned long writeback_threshold;
2262
2263         get_mems_allowed();
2264         delayacct_freepages_start();
2265
2266         if (global_reclaim(sc))
2267                 count_vm_event(ALLOCSTALL);
2268
2269         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2270                 sc->nr_scanned = 0;
2271                 if (!priority)
2272                         disable_swap_token(sc->target_mem_cgroup);
2273                 if (shrink_zones(priority, zonelist, sc))
2274                         break;
2275
2276                 /*
2277                  * Don't shrink slabs when reclaiming memory from
2278                  * over limit cgroups
2279                  */
2280                 if (global_reclaim(sc)) {
2281                         unsigned long lru_pages = 0;
2282                         for_each_zone_zonelist(zone, z, zonelist,
2283                                         gfp_zone(sc->gfp_mask)) {
2284                                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2285                                         continue;
2286
2287                                 lru_pages += zone_reclaimable_pages(zone);
2288                         }
2289
2290                         shrink_slab(shrink, sc->nr_scanned, lru_pages);
2291                         if (reclaim_state) {
2292                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2293                                 reclaim_state->reclaimed_slab = 0;
2294                         }
2295                 }
2296                 total_scanned += sc->nr_scanned;
2297                 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2298                         goto out;
2299
2300                 /*
2301                  * Try to write back as many pages as we just scanned.  This
2302                  * tends to cause slow streaming writers to write data to the
2303                  * disk smoothly, at the dirtying rate, which is nice.   But
2304                  * that's undesirable in laptop mode, where we *want* lumpy
2305                  * writeout.  So in laptop mode, write out the whole world.
2306                  */
2307                 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2308                 if (total_scanned > writeback_threshold) {
2309                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2310                                                 WB_REASON_TRY_TO_FREE_PAGES);
2311                         sc->may_writepage = 1;
2312                 }
2313
2314                 /* Take a nap, wait for some writeback to complete */
2315                 if (!sc->hibernation_mode && sc->nr_scanned &&
2316                     priority < DEF_PRIORITY - 2) {
2317                         struct zone *preferred_zone;
2318
2319                         first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2320                                                 &cpuset_current_mems_allowed,
2321                                                 &preferred_zone);
2322                         wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2323                 }
2324         }
2325
2326 out:
2327         delayacct_freepages_end();
2328         put_mems_allowed();
2329
2330         if (sc->nr_reclaimed)
2331                 return sc->nr_reclaimed;
2332
2333         /*
2334          * As hibernation is going on, kswapd is freezed so that it can't mark
2335          * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2336          * check.
2337          */
2338         if (oom_killer_disabled)
2339                 return 0;
2340
2341         /* top priority shrink_zones still had more to do? don't OOM, then */
2342         if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
2343                 return 1;
2344
2345         return 0;
2346 }
2347
2348 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2349                                 gfp_t gfp_mask, nodemask_t *nodemask)
2350 {
2351         unsigned long nr_reclaimed;
2352         struct scan_control sc = {
2353                 .gfp_mask = gfp_mask,
2354                 .may_writepage = !laptop_mode,
2355                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2356                 .may_unmap = 1,
2357                 .may_swap = 1,
2358                 .order = order,
2359                 .target_mem_cgroup = NULL,
2360                 .nodemask = nodemask,
2361         };
2362         struct shrink_control shrink = {
2363                 .gfp_mask = sc.gfp_mask,
2364         };
2365
2366         trace_mm_vmscan_direct_reclaim_begin(order,
2367                                 sc.may_writepage,
2368                                 gfp_mask);
2369
2370         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2371
2372         trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2373
2374         return nr_reclaimed;
2375 }
2376
2377 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2378
2379 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
2380                                                 gfp_t gfp_mask, bool noswap,
2381                                                 struct zone *zone,
2382                                                 unsigned long *nr_scanned)
2383 {
2384         struct scan_control sc = {
2385                 .nr_scanned = 0,
2386                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2387                 .may_writepage = !laptop_mode,
2388                 .may_unmap = 1,
2389                 .may_swap = !noswap,
2390                 .order = 0,
2391                 .target_mem_cgroup = memcg,
2392         };
2393         struct mem_cgroup_zone mz = {
2394                 .mem_cgroup = memcg,
2395                 .zone = zone,
2396         };
2397
2398         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2399                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2400
2401         trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2402                                                       sc.may_writepage,
2403                                                       sc.gfp_mask);
2404
2405         /*
2406          * NOTE: Although we can get the priority field, using it
2407          * here is not a good idea, since it limits the pages we can scan.
2408          * if we don't reclaim here, the shrink_zone from balance_pgdat
2409          * will pick up pages from other mem cgroup's as well. We hack
2410          * the priority and make it zero.
2411          */
2412         shrink_mem_cgroup_zone(0, &mz, &sc);
2413
2414         trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2415
2416         *nr_scanned = sc.nr_scanned;
2417         return sc.nr_reclaimed;
2418 }
2419
2420 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2421                                            gfp_t gfp_mask,
2422                                            bool noswap)
2423 {
2424         struct zonelist *zonelist;
2425         unsigned long nr_reclaimed;
2426         int nid;
2427         struct scan_control sc = {
2428                 .may_writepage = !laptop_mode,
2429                 .may_unmap = 1,
2430                 .may_swap = !noswap,
2431                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2432                 .order = 0,
2433                 .target_mem_cgroup = memcg,
2434                 .nodemask = NULL, /* we don't care the placement */
2435                 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2436                                 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2437         };
2438         struct shrink_control shrink = {
2439                 .gfp_mask = sc.gfp_mask,
2440         };
2441
2442         /*
2443          * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2444          * take care of from where we get pages. So the node where we start the
2445          * scan does not need to be the current node.
2446          */
2447         nid = mem_cgroup_select_victim_node(memcg);
2448
2449         zonelist = NODE_DATA(nid)->node_zonelists;
2450
2451         trace_mm_vmscan_memcg_reclaim_begin(0,
2452                                             sc.may_writepage,
2453                                             sc.gfp_mask);
2454
2455         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2456
2457         trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2458
2459         return nr_reclaimed;
2460 }
2461 #endif
2462
2463 static void age_active_anon(struct zone *zone, struct scan_control *sc,
2464                             int priority)
2465 {
2466         struct mem_cgroup *memcg;
2467
2468         if (!total_swap_pages)
2469                 return;
2470
2471         memcg = mem_cgroup_iter(NULL, NULL, NULL);
2472         do {
2473                 struct mem_cgroup_zone mz = {
2474                         .mem_cgroup = memcg,
2475                         .zone = zone,
2476                 };
2477
2478                 if (inactive_anon_is_low(&mz))
2479                         shrink_active_list(SWAP_CLUSTER_MAX, &mz,
2480                                            sc, priority, 0);
2481
2482                 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2483         } while (memcg);
2484 }
2485
2486 /*
2487  * pgdat_balanced is used when checking if a node is balanced for high-order
2488  * allocations. Only zones that meet watermarks and are in a zone allowed
2489  * by the callers classzone_idx are added to balanced_pages. The total of
2490  * balanced pages must be at least 25% of the zones allowed by classzone_idx
2491  * for the node to be considered balanced. Forcing all zones to be balanced
2492  * for high orders can cause excessive reclaim when there are imbalanced zones.
2493  * The choice of 25% is due to
2494  *   o a 16M DMA zone that is balanced will not balance a zone on any
2495  *     reasonable sized machine
2496  *   o On all other machines, the top zone must be at least a reasonable
2497  *     percentage of the middle zones. For example, on 32-bit x86, highmem
2498  *     would need to be at least 256M for it to be balance a whole node.
2499  *     Similarly, on x86-64 the Normal zone would need to be at least 1G
2500  *     to balance a node on its own. These seemed like reasonable ratios.
2501  */
2502 static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2503                                                 int classzone_idx)
2504 {
2505         unsigned long present_pages = 0;
2506         int i;
2507
2508         for (i = 0; i <= classzone_idx; i++)
2509                 present_pages += pgdat->node_zones[i].present_pages;
2510
2511         /* A special case here: if zone has no page, we think it's balanced */
2512         return balanced_pages >= (present_pages >> 2);
2513 }
2514
2515 /* is kswapd sleeping prematurely? */
2516 static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2517                                         int classzone_idx)
2518 {
2519         int i;
2520         unsigned long balanced = 0;
2521         bool all_zones_ok = true;
2522
2523         /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2524         if (remaining)
2525                 return true;
2526
2527         /* Check the watermark levels */
2528         for (i = 0; i <= classzone_idx; i++) {
2529                 struct zone *zone = pgdat->node_zones + i;
2530
2531                 if (!populated_zone(zone))
2532                         continue;
2533
2534                 /*
2535                  * balance_pgdat() skips over all_unreclaimable after
2536                  * DEF_PRIORITY. Effectively, it considers them balanced so
2537                  * they must be considered balanced here as well if kswapd
2538                  * is to sleep
2539                  */
2540                 if (zone->all_unreclaimable) {
2541                         balanced += zone->present_pages;
2542                         continue;
2543                 }
2544
2545                 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2546                                                         i, 0))
2547                         all_zones_ok = false;
2548                 else
2549                         balanced += zone->present_pages;
2550         }
2551
2552         /*
2553          * For high-order requests, the balanced zones must contain at least
2554          * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2555          * must be balanced
2556          */
2557         if (order)
2558                 return !pgdat_balanced(pgdat, balanced, classzone_idx);
2559         else
2560                 return !all_zones_ok;
2561 }
2562
2563 /*
2564  * For kswapd, balance_pgdat() will work across all this node's zones until
2565  * they are all at high_wmark_pages(zone).
2566  *
2567  * Returns the final order kswapd was reclaiming at
2568  *
2569  * There is special handling here for zones which are full of pinned pages.
2570  * This can happen if the pages are all mlocked, or if they are all used by
2571  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
2572  * What we do is to detect the case where all pages in the zone have been
2573  * scanned twice and there has been zero successful reclaim.  Mark the zone as
2574  * dead and from now on, only perform a short scan.  Basically we're polling
2575  * the zone for when the problem goes away.
2576  *
2577  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
2578  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2579  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2580  * lower zones regardless of the number of free pages in the lower zones. This
2581  * interoperates with the page allocator fallback scheme to ensure that aging
2582  * of pages is balanced across the zones.
2583  */
2584 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2585                                                         int *classzone_idx)
2586 {
2587         int all_zones_ok;
2588         unsigned long balanced;
2589         int priority;
2590         int i;
2591         int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
2592         unsigned long total_scanned;
2593         struct reclaim_state *reclaim_state = current->reclaim_state;
2594         unsigned long nr_soft_reclaimed;
2595         unsigned long nr_soft_scanned;
2596         struct scan_control sc = {
2597                 .gfp_mask = GFP_KERNEL,
2598                 .may_unmap = 1,
2599                 .may_swap = 1,
2600                 /*
2601                  * kswapd doesn't want to be bailed out while reclaim. because
2602                  * we want to put equal scanning pressure on each zone.
2603                  */
2604                 .nr_to_reclaim = ULONG_MAX,
2605                 .order = order,
2606                 .target_mem_cgroup = NULL,
2607         };
2608         struct shrink_control shrink = {
2609                 .gfp_mask = sc.gfp_mask,
2610         };
2611 loop_again:
2612         total_scanned = 0;
2613         sc.nr_reclaimed = 0;
2614         sc.may_writepage = !laptop_mode;
2615         count_vm_event(PAGEOUTRUN);
2616
2617         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2618                 unsigned long lru_pages = 0;
2619                 int has_under_min_watermark_zone = 0;
2620
2621                 /* The swap token gets in the way of swapout... */
2622                 if (!priority)
2623                         disable_swap_token(NULL);
2624
2625                 all_zones_ok = 1;
2626                 balanced = 0;
2627
2628                 /*
2629                  * Scan in the highmem->dma direction for the highest
2630                  * zone which needs scanning
2631                  */
2632                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2633                         struct zone *zone = pgdat->node_zones + i;
2634
2635                         if (!populated_zone(zone))
2636                                 continue;
2637
2638                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2639                                 continue;
2640
2641                         /*
2642                          * Do some background aging of the anon list, to give
2643                          * pages a chance to be referenced before reclaiming.
2644                          */
2645                         age_active_anon(zone, &sc, priority);
2646
2647                         if (!zone_watermark_ok_safe(zone, order,
2648                                         high_wmark_pages(zone), 0, 0)) {
2649                                 end_zone = i;
2650                                 break;
2651                         } else {
2652                                 /* If balanced, clear the congested flag */
2653                                 zone_clear_flag(zone, ZONE_CONGESTED);
2654                         }
2655                 }
2656                 if (i < 0)
2657                         goto out;
2658
2659                 for (i = 0; i <= end_zone; i++) {
2660                         struct zone *zone = pgdat->node_zones + i;
2661
2662                         lru_pages += zone_reclaimable_pages(zone);
2663                 }
2664
2665                 /*
2666                  * Now scan the zone in the dma->highmem direction, stopping
2667                  * at the last zone which needs scanning.
2668                  *
2669                  * We do this because the page allocator works in the opposite
2670                  * direction.  This prevents the page allocator from allocating
2671                  * pages behind kswapd's direction of progress, which would
2672                  * cause too much scanning of the lower zones.
2673                  */
2674                 for (i = 0; i <= end_zone; i++) {
2675                         struct zone *zone = pgdat->node_zones + i;
2676                         int nr_slab;
2677                         unsigned long balance_gap;
2678
2679                         if (!populated_zone(zone))
2680                                 continue;
2681
2682                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2683                                 continue;
2684
2685                         sc.nr_scanned = 0;
2686
2687                         nr_soft_scanned = 0;
2688                         /*
2689                          * Call soft limit reclaim before calling shrink_zone.
2690                          */
2691                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2692                                                         order, sc.gfp_mask,
2693                                                         &nr_soft_scanned);
2694                         sc.nr_reclaimed += nr_soft_reclaimed;
2695                         total_scanned += nr_soft_scanned;
2696
2697                         /*
2698                          * We put equal pressure on every zone, unless
2699                          * one zone has way too many pages free
2700                          * already. The "too many pages" is defined
2701                          * as the high wmark plus a "gap" where the
2702                          * gap is either the low watermark or 1%
2703                          * of the zone, whichever is smaller.
2704                          */
2705                         balance_gap = min(low_wmark_pages(zone),
2706                                 (zone->present_pages +
2707                                         KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2708                                 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2709                         if (!zone_watermark_ok_safe(zone, order,
2710                                         high_wmark_pages(zone) + balance_gap,
2711                                         end_zone, 0)) {
2712                                 shrink_zone(priority, zone, &sc);
2713
2714                                 reclaim_state->reclaimed_slab = 0;
2715                                 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2716                                 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2717                                 total_scanned += sc.nr_scanned;
2718
2719                                 if (nr_slab == 0 && !zone_reclaimable(zone))
2720                                         zone->all_unreclaimable = 1;
2721                         }
2722
2723                         /*
2724                          * If we've done a decent amount of scanning and
2725                          * the reclaim ratio is low, start doing writepage
2726                          * even in laptop mode
2727                          */
2728                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2729                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2730                                 sc.may_writepage = 1;
2731
2732                         if (zone->all_unreclaimable) {
2733                                 if (end_zone && end_zone == i)
2734                                         end_zone--;
2735                                 continue;
2736                         }
2737
2738                         if (!zone_watermark_ok_safe(zone, order,
2739                                         high_wmark_pages(zone), end_zone, 0)) {
2740                                 all_zones_ok = 0;
2741                                 /*
2742                                  * We are still under min water mark.  This
2743                                  * means that we have a GFP_ATOMIC allocation
2744                                  * failure risk. Hurry up!
2745                                  */
2746                                 if (!zone_watermark_ok_safe(zone, order,
2747                                             min_wmark_pages(zone), end_zone, 0))
2748                                         has_under_min_watermark_zone = 1;
2749                         } else {
2750                                 /*
2751                                  * If a zone reaches its high watermark,
2752                                  * consider it to be no longer congested. It's
2753                                  * possible there are dirty pages backed by
2754                                  * congested BDIs but as pressure is relieved,
2755                                  * spectulatively avoid congestion waits
2756                                  */
2757                                 zone_clear_flag(zone, ZONE_CONGESTED);
2758                                 if (i <= *classzone_idx)
2759                                         balanced += zone->present_pages;
2760                         }
2761
2762                 }
2763                 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2764                         break;          /* kswapd: all done */
2765                 /*
2766                  * OK, kswapd is getting into trouble.  Take a nap, then take
2767                  * another pass across the zones.
2768                  */
2769                 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2770                         if (has_under_min_watermark_zone)
2771                                 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2772                         else
2773                                 congestion_wait(BLK_RW_ASYNC, HZ/10);
2774                 }
2775
2776                 /*
2777                  * We do this so kswapd doesn't build up large priorities for
2778                  * example when it is freeing in parallel with allocators. It
2779                  * matches the direct reclaim path behaviour in terms of impact
2780                  * on zone->*_priority.
2781                  */
2782                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2783                         break;
2784         }
2785 out:
2786
2787         /*
2788          * order-0: All zones must meet high watermark for a balanced node
2789          * high-order: Balanced zones must make up at least 25% of the node
2790          *             for the node to be balanced
2791          */
2792         if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2793                 cond_resched();
2794
2795                 try_to_freeze();
2796
2797                 /*
2798                  * Fragmentation may mean that the system cannot be
2799                  * rebalanced for high-order allocations in all zones.
2800                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2801                  * it means the zones have been fully scanned and are still
2802                  * not balanced. For high-order allocations, there is
2803                  * little point trying all over again as kswapd may
2804                  * infinite loop.
2805                  *
2806                  * Instead, recheck all watermarks at order-0 as they
2807                  * are the most important. If watermarks are ok, kswapd will go
2808                  * back to sleep. High-order users can still perform direct
2809                  * reclaim if they wish.
2810                  */
2811                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2812                         order = sc.order = 0;
2813
2814                 goto loop_again;
2815         }
2816
2817         /*
2818          * If kswapd was reclaiming at a higher order, it has the option of
2819          * sleeping without all zones being balanced. Before it does, it must
2820          * ensure that the watermarks for order-0 on *all* zones are met and
2821          * that the congestion flags are cleared. The congestion flag must
2822          * be cleared as kswapd is the only mechanism that clears the flag
2823          * and it is potentially going to sleep here.
2824          */
2825         if (order) {
2826                 for (i = 0; i <= end_zone; i++) {
2827                         struct zone *zone = pgdat->node_zones + i;
2828
2829                         if (!populated_zone(zone))
2830                                 continue;
2831
2832                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2833                                 continue;
2834
2835                         /* Confirm the zone is balanced for order-0 */
2836                         if (!zone_watermark_ok(zone, 0,
2837                                         high_wmark_pages(zone), 0, 0)) {
2838                                 order = sc.order = 0;
2839                                 goto loop_again;
2840                         }
2841
2842                         /* If balanced, clear the congested flag */
2843                         zone_clear_flag(zone, ZONE_CONGESTED);
2844                         if (i <= *classzone_idx)
2845                                 balanced += zone->present_pages;
2846                 }
2847         }
2848
2849         /*
2850          * Return the order we were reclaiming at so sleeping_prematurely()
2851          * makes a decision on the order we were last reclaiming at. However,
2852          * if another caller entered the allocator slow path while kswapd
2853          * was awake, order will remain at the higher level
2854          */
2855         *classzone_idx = end_zone;
2856         return order;
2857 }
2858
2859 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2860 {
2861         long remaining = 0;
2862         DEFINE_WAIT(wait);
2863
2864         if (freezing(current) || kthread_should_stop())
2865                 return;
2866
2867         prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2868
2869         /* Try to sleep for a short interval */
2870         if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2871                 remaining = schedule_timeout(HZ/10);
2872                 finish_wait(&pgdat->kswapd_wait, &wait);
2873                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2874         }
2875
2876         /*
2877          * After a short sleep, check if it was a premature sleep. If not, then
2878          * go fully to sleep until explicitly woken up.
2879          */
2880         if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2881                 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2882
2883                 /*
2884                  * vmstat counters are not perfectly accurate and the estimated
2885                  * value for counters such as NR_FREE_PAGES can deviate from the
2886                  * true value by nr_online_cpus * threshold. To avoid the zone
2887                  * watermarks being breached while under pressure, we reduce the
2888                  * per-cpu vmstat threshold while kswapd is awake and restore
2889                  * them before going back to sleep.
2890                  */
2891                 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2892                 schedule();
2893                 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2894         } else {
2895                 if (remaining)
2896                         count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2897                 else
2898                         count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2899         }
2900         finish_wait(&pgdat->kswapd_wait, &wait);
2901 }
2902
2903 /*
2904  * The background pageout daemon, started as a kernel thread
2905  * from the init process.
2906  *
2907  * This basically trickles out pages so that we have _some_
2908  * free memory available even if there is no other activity
2909  * that frees anything up. This is needed for things like routing
2910  * etc, where we otherwise might have all activity going on in
2911  * asynchronous contexts that cannot page things out.
2912  *
2913  * If there are applications that are active memory-allocators
2914  * (most normal use), this basically shouldn't matter.
2915  */
2916 static int kswapd(void *p)
2917 {
2918         unsigned long order, new_order;
2919         unsigned balanced_order;
2920         int classzone_idx, new_classzone_idx;
2921         int balanced_classzone_idx;
2922         pg_data_t *pgdat = (pg_data_t*)p;
2923         struct task_struct *tsk = current;
2924
2925         struct reclaim_state reclaim_state = {
2926                 .reclaimed_slab = 0,
2927         };
2928         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2929
2930         lockdep_set_current_reclaim_state(GFP_KERNEL);
2931
2932         if (!cpumask_empty(cpumask))
2933                 set_cpus_allowed_ptr(tsk, cpumask);
2934         current->reclaim_state = &reclaim_state;
2935
2936         /*
2937          * Tell the memory management that we're a "memory allocator",
2938          * and that if we need more memory we should get access to it
2939          * regardless (see "__alloc_pages()"). "kswapd" should
2940          * never get caught in the normal page freeing logic.
2941          *
2942          * (Kswapd normally doesn't need memory anyway, but sometimes
2943          * you need a small amount of memory in order to be able to
2944          * page out something else, and this flag essentially protects
2945          * us from recursively trying to free more memory as we're
2946          * trying to free the first piece of memory in the first place).
2947          */
2948         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2949         set_freezable();
2950
2951         order = new_order = 0;
2952         balanced_order = 0;
2953         classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
2954         balanced_classzone_idx = classzone_idx;
2955         for ( ; ; ) {
2956                 int ret;
2957
2958                 /*
2959                  * If the last balance_pgdat was unsuccessful it's unlikely a
2960                  * new request of a similar or harder type will succeed soon
2961                  * so consider going to sleep on the basis we reclaimed at
2962                  */
2963                 if (balanced_classzone_idx >= new_classzone_idx &&
2964                                         balanced_order == new_order) {
2965                         new_order = pgdat->kswapd_max_order;
2966                         new_classzone_idx = pgdat->classzone_idx;
2967                         pgdat->kswapd_max_order =  0;
2968                         pgdat->classzone_idx = pgdat->nr_zones - 1;
2969                 }
2970
2971                 if (order < new_order || classzone_idx > new_classzone_idx) {
2972                         /*
2973                          * Don't sleep if someone wants a larger 'order'
2974                          * allocation or has tigher zone constraints
2975                          */
2976                         order = new_order;
2977                         classzone_idx = new_classzone_idx;
2978                 } else {
2979                         kswapd_try_to_sleep(pgdat, balanced_order,
2980                                                 balanced_classzone_idx);
2981                         order = pgdat->kswapd_max_order;
2982                         classzone_idx = pgdat->classzone_idx;
2983                         new_order = order;
2984                         new_classzone_idx = classzone_idx;
2985                         pgdat->kswapd_max_order = 0;
2986                         pgdat->classzone_idx = pgdat->nr_zones - 1;
2987                 }
2988
2989                 ret = try_to_freeze();
2990                 if (kthread_should_stop())
2991                         break;
2992
2993                 /*
2994                  * We can speed up thawing tasks if we don't call balance_pgdat
2995                  * after returning from the refrigerator
2996                  */
2997                 if (!ret) {
2998                         trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2999                         balanced_classzone_idx = classzone_idx;
3000                         balanced_order = balance_pgdat(pgdat, order,
3001                                                 &balanced_classzone_idx);
3002                 }
3003         }
3004         return 0;
3005 }
3006
3007 /*
3008  * A zone is low on free memory, so wake its kswapd task to service it.
3009  */
3010 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3011 {
3012         pg_data_t *pgdat;
3013
3014         if (!populated_zone(zone))
3015                 return;
3016
3017         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
3018                 return;
3019         pgdat = zone->zone_pgdat;
3020         if (pgdat->kswapd_max_order < order) {
3021                 pgdat->kswapd_max_order = order;
3022                 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3023         }
3024         if (!waitqueue_active(&pgdat->kswapd_wait))
3025                 return;
3026         if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
3027                 return;
3028
3029         trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3030         wake_up_interruptible(&pgdat->kswapd_wait);
3031 }
3032
3033 /*
3034  * The reclaimable count would be mostly accurate.
3035  * The less reclaimable pages may be
3036  * - mlocked pages, which will be moved to unevictable list when encountered
3037  * - mapped pages, which may require several travels to be reclaimed
3038  * - dirty pages, which is not "instantly" reclaimable
3039  */
3040 unsigned long global_reclaimable_pages(void)
3041 {
3042         int nr;
3043
3044         nr = global_page_state(NR_ACTIVE_FILE) +
3045              global_page_state(NR_INACTIVE_FILE);
3046
3047         if (nr_swap_pages > 0)
3048                 nr += global_page_state(NR_ACTIVE_ANON) +
3049                       global_page_state(NR_INACTIVE_ANON);
3050
3051         return nr;
3052 }
3053
3054 unsigned long zone_reclaimable_pages(struct zone *zone)
3055 {
3056         int nr;
3057
3058         nr = zone_page_state(zone, NR_ACTIVE_FILE) +
3059              zone_page_state(zone, NR_INACTIVE_FILE);
3060
3061         if (nr_swap_pages > 0)
3062                 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
3063                       zone_page_state(zone, NR_INACTIVE_ANON);
3064
3065         return nr;
3066 }
3067
3068 #ifdef CONFIG_HIBERNATION
3069 /*
3070  * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3071  * freed pages.
3072  *
3073  * Rather than trying to age LRUs the aim is to preserve the overall
3074  * LRU order by reclaiming preferentially
3075  * inactive > active > active referenced > active mapped
3076  */
3077 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3078 {
3079         struct reclaim_state reclaim_state;
3080         struct scan_control sc = {
3081                 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3082                 .may_swap = 1,
3083                 .may_unmap = 1,
3084                 .may_writepage = 1,
3085                 .nr_to_reclaim = nr_to_reclaim,
3086                 .hibernation_mode = 1,
3087                 .order = 0,
3088         };
3089         struct shrink_control shrink = {
3090                 .gfp_mask = sc.gfp_mask,
3091         };
3092         struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3093         struct task_struct *p = current;
3094         unsigned long nr_reclaimed;
3095
3096         p->flags |= PF_MEMALLOC;
3097         lockdep_set_current_reclaim_state(sc.gfp_mask);
3098         reclaim_state.reclaimed_slab = 0;
3099         p->reclaim_state = &reclaim_state;
3100
3101         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
3102
3103         p->reclaim_state = NULL;
3104         lockdep_clear_current_reclaim_state();
3105         p->flags &= ~PF_MEMALLOC;
3106
3107         return nr_reclaimed;
3108 }
3109 #endif /* CONFIG_HIBERNATION */
3110
3111 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3112    not required for correctness.  So if the last cpu in a node goes
3113    away, we get changed to run anywhere: as the first one comes back,
3114    restore their cpu bindings. */
3115 static int __devinit cpu_callback(struct notifier_block *nfb,
3116                                   unsigned long action, void *hcpu)
3117 {
3118         int nid;
3119
3120         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3121                 for_each_node_state(nid, N_HIGH_MEMORY) {
3122                         pg_data_t *pgdat = NODE_DATA(nid);
3123                         const struct cpumask *mask;
3124
3125                         mask = cpumask_of_node(pgdat->node_id);
3126
3127                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3128                                 /* One of our CPUs online: restore mask */
3129                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3130                 }
3131         }
3132         return NOTIFY_OK;
3133 }
3134
3135 /*
3136  * This kswapd start function will be called by init and node-hot-add.
3137  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3138  */
3139 int kswapd_run(int nid)
3140 {
3141         pg_data_t *pgdat = NODE_DATA(nid);
3142         int ret = 0;
3143
3144         if (pgdat->kswapd)
3145                 return 0;
3146
3147         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3148         if (IS_ERR(pgdat->kswapd)) {
3149                 /* failure at boot is fatal */
3150                 BUG_ON(system_state == SYSTEM_BOOTING);
3151                 printk("Failed to start kswapd on node %d\n",nid);
3152                 ret = -1;
3153         }
3154         return ret;
3155 }
3156
3157 /*
3158  * Called by memory hotplug when all memory in a node is offlined.
3159  */
3160 void kswapd_stop(int nid)
3161 {
3162         struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3163
3164         if (kswapd)
3165                 kthread_stop(kswapd);
3166 }
3167
3168 static int __init kswapd_init(void)
3169 {
3170         int nid;
3171
3172         swap_setup();
3173         for_each_node_state(nid, N_HIGH_MEMORY)
3174                 kswapd_run(nid);
3175         hotcpu_notifier(cpu_callback, 0);
3176         return 0;
3177 }
3178
3179 module_init(kswapd_init)
3180
3181 #ifdef CONFIG_NUMA
3182 /*
3183  * Zone reclaim mode
3184  *
3185  * If non-zero call zone_reclaim when the number of free pages falls below
3186  * the watermarks.
3187  */
3188 int zone_reclaim_mode __read_mostly;
3189
3190 #define RECLAIM_OFF 0
3191 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
3192 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
3193 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
3194
3195 /*
3196  * Priority for ZONE_RECLAIM. This determines the fraction of pages
3197  * of a node considered for each zone_reclaim. 4 scans 1/16th of
3198  * a zone.
3199  */
3200 #define ZONE_RECLAIM_PRIORITY 4
3201
3202 /*
3203  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3204  * occur.
3205  */
3206 int sysctl_min_unmapped_ratio = 1;
3207
3208 /*
3209  * If the number of slab pages in a zone grows beyond this percentage then
3210  * slab reclaim needs to occur.
3211  */
3212 int sysctl_min_slab_ratio = 5;
3213
3214 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3215 {
3216         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3217         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3218                 zone_page_state(zone, NR_ACTIVE_FILE);
3219
3220         /*
3221          * It's possible for there to be more file mapped pages than
3222          * accounted for by the pages on the file LRU lists because
3223          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3224          */
3225         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3226 }
3227
3228 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3229 static long zone_pagecache_reclaimable(struct zone *zone)
3230 {
3231         long nr_pagecache_reclaimable;
3232         long delta = 0;
3233
3234         /*
3235          * If RECLAIM_SWAP is set, then all file pages are considered
3236          * potentially reclaimable. Otherwise, we have to worry about
3237          * pages like swapcache and zone_unmapped_file_pages() provides
3238          * a better estimate
3239          */
3240         if (zone_reclaim_mode & RECLAIM_SWAP)
3241                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3242         else
3243                 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3244
3245         /* If we can't clean pages, remove dirty pages from consideration */
3246         if (!(zone_reclaim_mode & RECLAIM_WRITE))
3247                 delta += zone_page_state(zone, NR_FILE_DIRTY);
3248
3249         /* Watch for any possible underflows due to delta */
3250         if (unlikely(delta > nr_pagecache_reclaimable))
3251                 delta = nr_pagecache_reclaimable;
3252
3253         return nr_pagecache_reclaimable - delta;
3254 }
3255
3256 /*
3257  * Try to free up some pages from this zone through reclaim.
3258  */
3259 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3260 {
3261         /* Minimum pages needed in order to stay on node */
3262         const unsigned long nr_pages = 1 << order;
3263         struct task_struct *p = current;
3264         struct reclaim_state reclaim_state;
3265         int priority;
3266         struct scan_control sc = {
3267                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3268                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3269                 .may_swap = 1,
3270                 .nr_to_reclaim = max_t(unsigned long, nr_pages,
3271                                        SWAP_CLUSTER_MAX),
3272                 .gfp_mask = gfp_mask,
3273                 .order = order,
3274         };
3275         struct shrink_control shrink = {
3276                 .gfp_mask = sc.gfp_mask,
3277         };
3278         unsigned long nr_slab_pages0, nr_slab_pages1;
3279
3280         cond_resched();
3281         /*
3282          * We need to be able to allocate from the reserves for RECLAIM_SWAP
3283          * and we also need to be able to write out pages for RECLAIM_WRITE
3284          * and RECLAIM_SWAP.
3285          */
3286         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3287         lockdep_set_current_reclaim_state(gfp_mask);
3288         reclaim_state.reclaimed_slab = 0;
3289         p->reclaim_state = &reclaim_state;
3290
3291         if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3292                 /*
3293                  * Free memory by calling shrink zone with increasing
3294                  * priorities until we have enough memory freed.
3295                  */
3296                 priority = ZONE_RECLAIM_PRIORITY;
3297                 do {
3298                         shrink_zone(priority, zone, &sc);
3299                         priority--;
3300                 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
3301         }
3302
3303         nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3304         if (nr_slab_pages0 > zone->min_slab_pages) {
3305                 /*
3306                  * shrink_slab() does not currently allow us to determine how
3307                  * many pages were freed in this zone. So we take the current
3308                  * number of slab pages and shake the slab until it is reduced
3309                  * by the same nr_pages that we used for reclaiming unmapped
3310                  * pages.
3311                  *
3312                  * Note that shrink_slab will free memory on all zones and may
3313                  * take a long time.
3314                  */
3315                 for (;;) {
3316                         unsigned long lru_pages = zone_reclaimable_pages(zone);
3317
3318                         /* No reclaimable slab or very low memory pressure */
3319                         if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3320                                 break;
3321
3322                         /* Freed enough memory */
3323                         nr_slab_pages1 = zone_page_state(zone,
3324                                                         NR_SLAB_RECLAIMABLE);
3325                         if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3326                                 break;
3327                 }
3328
3329                 /*
3330                  * Update nr_reclaimed by the number of slab pages we
3331                  * reclaimed from this zone.
3332                  */
3333                 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3334                 if (nr_slab_pages1 < nr_slab_pages0)
3335                         sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3336         }
3337
3338         p->reclaim_state = NULL;
3339         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3340         lockdep_clear_current_reclaim_state();
3341         return sc.nr_reclaimed >= nr_pages;
3342 }
3343
3344 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3345 {
3346         int node_id;
3347         int ret;
3348
3349         /*
3350          * Zone reclaim reclaims unmapped file backed pages and
3351          * slab pages if we are over the defined limits.
3352          *
3353          * A small portion of unmapped file backed pages is needed for
3354          * file I/O otherwise pages read by file I/O will be immediately
3355          * thrown out if the zone is overallocated. So we do not reclaim
3356          * if less than a specified percentage of the zone is used by
3357          * unmapped file backed pages.
3358          */
3359         if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3360             zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3361                 return ZONE_RECLAIM_FULL;
3362
3363         if (zone->all_unreclaimable)
3364                 return ZONE_RECLAIM_FULL;
3365
3366         /*
3367          * Do not scan if the allocation should not be delayed.
3368          */
3369         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3370                 return ZONE_RECLAIM_NOSCAN;
3371
3372         /*
3373          * Only run zone reclaim on the local zone or on zones that do not
3374          * have associated processors. This will favor the local processor
3375          * over remote processors and spread off node memory allocations
3376          * as wide as possible.
3377          */
3378         node_id = zone_to_nid(zone);
3379         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3380                 return ZONE_RECLAIM_NOSCAN;
3381
3382         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3383                 return ZONE_RECLAIM_NOSCAN;
3384
3385         ret = __zone_reclaim(zone, gfp_mask, order);
3386         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3387
3388         if (!ret)
3389                 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3390
3391         return ret;
3392 }
3393 #endif
3394
3395 /*
3396  * page_evictable - test whether a page is evictable
3397  * @page: the page to test
3398  * @vma: the VMA in which the page is or will be mapped, may be NULL
3399  *
3400  * Test whether page is evictable--i.e., should be placed on active/inactive
3401  * lists vs unevictable list.  The vma argument is !NULL when called from the
3402  * fault path to determine how to instantate a new page.
3403  *
3404  * Reasons page might not be evictable:
3405  * (1) page's mapping marked unevictable
3406  * (2) page is part of an mlocked VMA
3407  *
3408  */
3409 int page_evictable(struct page *page, struct vm_area_struct *vma)
3410 {
3411
3412         if (mapping_unevictable(page_mapping(page)))
3413                 return 0;
3414
3415         if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
3416                 return 0;
3417
3418         return 1;
3419 }
3420
3421 /**
3422  * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
3423  * @page: page to check evictability and move to appropriate lru list
3424  * @zone: zone page is in
3425  *
3426  * Checks a page for evictability and moves the page to the appropriate
3427  * zone lru list.
3428  *
3429  * Restrictions: zone->lru_lock must be held, page must be on LRU and must
3430  * have PageUnevictable set.
3431  */
3432 static void check_move_unevictable_page(struct page *page, struct zone *zone)
3433 {
3434         struct lruvec *lruvec;
3435
3436         VM_BUG_ON(PageActive(page));
3437 retry:
3438         ClearPageUnevictable(page);
3439         if (page_evictable(page, NULL)) {
3440                 enum lru_list l = page_lru_base_type(page);
3441
3442                 __dec_zone_state(zone, NR_UNEVICTABLE);
3443                 lruvec = mem_cgroup_lru_move_lists(zone, page,
3444                                                    LRU_UNEVICTABLE, l);
3445                 list_move(&page->lru, &lruvec->lists[l]);
3446                 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
3447                 __count_vm_event(UNEVICTABLE_PGRESCUED);
3448         } else {
3449                 /*
3450                  * rotate unevictable list
3451                  */
3452                 SetPageUnevictable(page);
3453                 lruvec = mem_cgroup_lru_move_lists(zone, page, LRU_UNEVICTABLE,
3454                                                    LRU_UNEVICTABLE);
3455                 list_move(&page->lru, &lruvec->lists[LRU_UNEVICTABLE]);
3456                 if (page_evictable(page, NULL))
3457                         goto retry;
3458         }
3459 }
3460
3461 /**
3462  * scan_mapping_unevictable_pages - scan an address space for evictable pages
3463  * @mapping: struct address_space to scan for evictable pages
3464  *
3465  * Scan all pages in mapping.  Check unevictable pages for
3466  * evictability and move them to the appropriate zone lru list.
3467  */
3468 void scan_mapping_unevictable_pages(struct address_space *mapping)
3469 {
3470         pgoff_t next = 0;
3471         pgoff_t end   = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
3472                          PAGE_CACHE_SHIFT;
3473         struct zone *zone;
3474         struct pagevec pvec;
3475
3476         if (mapping->nrpages == 0)
3477                 return;
3478
3479         pagevec_init(&pvec, 0);
3480         while (next < end &&
3481                 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
3482                 int i;
3483                 int pg_scanned = 0;
3484
3485                 zone = NULL;
3486
3487                 for (i = 0; i < pagevec_count(&pvec); i++) {
3488                         struct page *page = pvec.pages[i];
3489                         pgoff_t page_index = page->index;
3490                         struct zone *pagezone = page_zone(page);
3491
3492                         pg_scanned++;
3493                         if (page_index > next)
3494                                 next = page_index;
3495                         next++;
3496
3497                         if (pagezone != zone) {
3498                                 if (zone)
3499                                         spin_unlock_irq(&zone->lru_lock);
3500                                 zone = pagezone;
3501                                 spin_lock_irq(&zone->lru_lock);
3502                         }
3503
3504                         if (PageLRU(page) && PageUnevictable(page))
3505                                 check_move_unevictable_page(page, zone);
3506                 }
3507                 if (zone)
3508                         spin_unlock_irq(&zone->lru_lock);
3509                 pagevec_release(&pvec);
3510
3511                 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
3512         }
3513
3514 }
3515
3516 static void warn_scan_unevictable_pages(void)
3517 {
3518         printk_once(KERN_WARNING
3519                     "%s: The scan_unevictable_pages sysctl/node-interface has been "
3520                     "disabled for lack of a legitimate use case.  If you have "
3521                     "one, please send an email to linux-mm@kvack.org.\n",
3522                     current->comm);
3523 }
3524
3525 /*
3526  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
3527  * all nodes' unevictable lists for evictable pages
3528  */
3529 unsigned long scan_unevictable_pages;
3530
3531 int scan_unevictable_handler(struct ctl_table *table, int write,
3532                            void __user *buffer,
3533                            size_t *length, loff_t *ppos)
3534 {
3535         warn_scan_unevictable_pages();
3536         proc_doulongvec_minmax(table, write, buffer, length, ppos);
3537         scan_unevictable_pages = 0;
3538         return 0;
3539 }
3540
3541 #ifdef CONFIG_NUMA
3542 /*
3543  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
3544  * a specified node's per zone unevictable lists for evictable pages.
3545  */
3546
3547 static ssize_t read_scan_unevictable_node(struct device *dev,
3548                                           struct device_attribute *attr,
3549                                           char *buf)
3550 {
3551         warn_scan_unevictable_pages();
3552         return sprintf(buf, "0\n");     /* always zero; should fit... */
3553 }
3554
3555 static ssize_t write_scan_unevictable_node(struct device *dev,
3556                                            struct device_attribute *attr,
3557                                         const char *buf, size_t count)
3558 {
3559         warn_scan_unevictable_pages();
3560         return 1;
3561 }
3562
3563
3564 static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3565                         read_scan_unevictable_node,
3566                         write_scan_unevictable_node);
3567
3568 int scan_unevictable_register_node(struct node *node)
3569 {
3570         return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
3571 }
3572
3573 void scan_unevictable_unregister_node(struct node *node)
3574 {
3575         device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);
3576 }
3577 #endif