mm: vmscan: do not OOM if aborting reclaim to start compaction
[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                                 unsigned int isolated_pages;
1200
1201                                 mem_cgroup_lru_del(cursor_page);
1202                                 list_move(&cursor_page->lru, dst);
1203                                 isolated_pages = hpage_nr_pages(cursor_page);
1204                                 nr_taken += isolated_pages;
1205                                 nr_lumpy_taken += isolated_pages;
1206                                 if (PageDirty(cursor_page))
1207                                         nr_lumpy_dirty += isolated_pages;
1208                                 scan++;
1209                                 pfn += isolated_pages - 1;
1210                         } else {
1211                                 /*
1212                                  * Check if the page is freed already.
1213                                  *
1214                                  * We can't use page_count() as that
1215                                  * requires compound_head and we don't
1216                                  * have a pin on the page here. If a
1217                                  * page is tail, we may or may not
1218                                  * have isolated the head, so assume
1219                                  * it's not free, it'd be tricky to
1220                                  * track the head status without a
1221                                  * page pin.
1222                                  */
1223                                 if (!PageTail(cursor_page) &&
1224                                     !atomic_read(&cursor_page->_count))
1225                                         continue;
1226                                 break;
1227                         }
1228                 }
1229
1230                 /* If we break out of the loop above, lumpy reclaim failed */
1231                 if (pfn < end_pfn)
1232                         nr_lumpy_failed++;
1233         }
1234
1235         *scanned = scan;
1236
1237         trace_mm_vmscan_lru_isolate(order,
1238                         nr_to_scan, scan,
1239                         nr_taken,
1240                         nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1241                         mode, file);
1242         return nr_taken;
1243 }
1244
1245 static unsigned long isolate_pages(unsigned long nr, struct mem_cgroup_zone *mz,
1246                                    struct list_head *dst,
1247                                    unsigned long *scanned, int order,
1248                                    isolate_mode_t mode, int active, int file)
1249 {
1250         struct lruvec *lruvec;
1251         int lru = LRU_BASE;
1252
1253         lruvec = mem_cgroup_zone_lruvec(mz->zone, mz->mem_cgroup);
1254         if (active)
1255                 lru += LRU_ACTIVE;
1256         if (file)
1257                 lru += LRU_FILE;
1258         return isolate_lru_pages(nr, &lruvec->lists[lru], dst,
1259                                  scanned, order, mode, file);
1260 }
1261
1262 /*
1263  * clear_active_flags() is a helper for shrink_active_list(), clearing
1264  * any active bits from the pages in the list.
1265  */
1266 static unsigned long clear_active_flags(struct list_head *page_list,
1267                                         unsigned int *count)
1268 {
1269         int nr_active = 0;
1270         int lru;
1271         struct page *page;
1272
1273         list_for_each_entry(page, page_list, lru) {
1274                 int numpages = hpage_nr_pages(page);
1275                 lru = page_lru_base_type(page);
1276                 if (PageActive(page)) {
1277                         lru += LRU_ACTIVE;
1278                         ClearPageActive(page);
1279                         nr_active += numpages;
1280                 }
1281                 if (count)
1282                         count[lru] += numpages;
1283         }
1284
1285         return nr_active;
1286 }
1287
1288 /**
1289  * isolate_lru_page - tries to isolate a page from its LRU list
1290  * @page: page to isolate from its LRU list
1291  *
1292  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1293  * vmstat statistic corresponding to whatever LRU list the page was on.
1294  *
1295  * Returns 0 if the page was removed from an LRU list.
1296  * Returns -EBUSY if the page was not on an LRU list.
1297  *
1298  * The returned page will have PageLRU() cleared.  If it was found on
1299  * the active list, it will have PageActive set.  If it was found on
1300  * the unevictable list, it will have the PageUnevictable bit set. That flag
1301  * may need to be cleared by the caller before letting the page go.
1302  *
1303  * The vmstat statistic corresponding to the list on which the page was
1304  * found will be decremented.
1305  *
1306  * Restrictions:
1307  * (1) Must be called with an elevated refcount on the page. This is a
1308  *     fundamentnal difference from isolate_lru_pages (which is called
1309  *     without a stable reference).
1310  * (2) the lru_lock must not be held.
1311  * (3) interrupts must be enabled.
1312  */
1313 int isolate_lru_page(struct page *page)
1314 {
1315         int ret = -EBUSY;
1316
1317         VM_BUG_ON(!page_count(page));
1318
1319         if (PageLRU(page)) {
1320                 struct zone *zone = page_zone(page);
1321
1322                 spin_lock_irq(&zone->lru_lock);
1323                 if (PageLRU(page)) {
1324                         int lru = page_lru(page);
1325                         ret = 0;
1326                         get_page(page);
1327                         ClearPageLRU(page);
1328
1329                         del_page_from_lru_list(zone, page, lru);
1330                 }
1331                 spin_unlock_irq(&zone->lru_lock);
1332         }
1333         return ret;
1334 }
1335
1336 /*
1337  * Are there way too many processes in the direct reclaim path already?
1338  */
1339 static int too_many_isolated(struct zone *zone, int file,
1340                 struct scan_control *sc)
1341 {
1342         unsigned long inactive, isolated;
1343
1344         if (current_is_kswapd())
1345                 return 0;
1346
1347         if (!global_reclaim(sc))
1348                 return 0;
1349
1350         if (file) {
1351                 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1352                 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1353         } else {
1354                 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1355                 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1356         }
1357
1358         return isolated > inactive;
1359 }
1360
1361 /*
1362  * TODO: Try merging with migrations version of putback_lru_pages
1363  */
1364 static noinline_for_stack void
1365 putback_lru_pages(struct mem_cgroup_zone *mz, struct scan_control *sc,
1366                   unsigned long nr_anon, unsigned long nr_file,
1367                   struct list_head *page_list)
1368 {
1369         struct page *page;
1370         struct pagevec pvec;
1371         struct zone *zone = mz->zone;
1372         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz);
1373
1374         pagevec_init(&pvec, 1);
1375
1376         /*
1377          * Put back any unfreeable pages.
1378          */
1379         spin_lock(&zone->lru_lock);
1380         while (!list_empty(page_list)) {
1381                 int lru;
1382                 page = lru_to_page(page_list);
1383                 VM_BUG_ON(PageLRU(page));
1384                 list_del(&page->lru);
1385                 if (unlikely(!page_evictable(page, NULL))) {
1386                         spin_unlock_irq(&zone->lru_lock);
1387                         putback_lru_page(page);
1388                         spin_lock_irq(&zone->lru_lock);
1389                         continue;
1390                 }
1391                 SetPageLRU(page);
1392                 lru = page_lru(page);
1393                 add_page_to_lru_list(zone, page, lru);
1394                 if (is_active_lru(lru)) {
1395                         int file = is_file_lru(lru);
1396                         int numpages = hpage_nr_pages(page);
1397                         reclaim_stat->recent_rotated[file] += numpages;
1398                 }
1399                 if (!pagevec_add(&pvec, page)) {
1400                         spin_unlock_irq(&zone->lru_lock);
1401                         __pagevec_release(&pvec);
1402                         spin_lock_irq(&zone->lru_lock);
1403                 }
1404         }
1405         __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1406         __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1407
1408         spin_unlock_irq(&zone->lru_lock);
1409         pagevec_release(&pvec);
1410 }
1411
1412 static noinline_for_stack void
1413 update_isolated_counts(struct mem_cgroup_zone *mz,
1414                        struct scan_control *sc,
1415                        unsigned long *nr_anon,
1416                        unsigned long *nr_file,
1417                        struct list_head *isolated_list)
1418 {
1419         unsigned long nr_active;
1420         struct zone *zone = mz->zone;
1421         unsigned int count[NR_LRU_LISTS] = { 0, };
1422         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz);
1423
1424         nr_active = clear_active_flags(isolated_list, count);
1425         __count_vm_events(PGDEACTIVATE, nr_active);
1426
1427         __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1428                               -count[LRU_ACTIVE_FILE]);
1429         __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1430                               -count[LRU_INACTIVE_FILE]);
1431         __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1432                               -count[LRU_ACTIVE_ANON]);
1433         __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1434                               -count[LRU_INACTIVE_ANON]);
1435
1436         *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1437         *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1438         __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1439         __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1440
1441         reclaim_stat->recent_scanned[0] += *nr_anon;
1442         reclaim_stat->recent_scanned[1] += *nr_file;
1443 }
1444
1445 /*
1446  * Returns true if a direct reclaim should wait on pages under writeback.
1447  *
1448  * If we are direct reclaiming for contiguous pages and we do not reclaim
1449  * everything in the list, try again and wait for writeback IO to complete.
1450  * This will stall high-order allocations noticeably. Only do that when really
1451  * need to free the pages under high memory pressure.
1452  */
1453 static inline bool should_reclaim_stall(unsigned long nr_taken,
1454                                         unsigned long nr_freed,
1455                                         int priority,
1456                                         struct scan_control *sc)
1457 {
1458         int lumpy_stall_priority;
1459
1460         /* kswapd should not stall on sync IO */
1461         if (current_is_kswapd())
1462                 return false;
1463
1464         /* Only stall on lumpy reclaim */
1465         if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
1466                 return false;
1467
1468         /* If we have reclaimed everything on the isolated list, no stall */
1469         if (nr_freed == nr_taken)
1470                 return false;
1471
1472         /*
1473          * For high-order allocations, there are two stall thresholds.
1474          * High-cost allocations stall immediately where as lower
1475          * order allocations such as stacks require the scanning
1476          * priority to be much higher before stalling.
1477          */
1478         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1479                 lumpy_stall_priority = DEF_PRIORITY;
1480         else
1481                 lumpy_stall_priority = DEF_PRIORITY / 3;
1482
1483         return priority <= lumpy_stall_priority;
1484 }
1485
1486 /*
1487  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1488  * of reclaimed pages
1489  */
1490 static noinline_for_stack unsigned long
1491 shrink_inactive_list(unsigned long nr_to_scan, struct mem_cgroup_zone *mz,
1492                      struct scan_control *sc, int priority, int file)
1493 {
1494         LIST_HEAD(page_list);
1495         unsigned long nr_scanned;
1496         unsigned long nr_reclaimed = 0;
1497         unsigned long nr_taken;
1498         unsigned long nr_anon;
1499         unsigned long nr_file;
1500         unsigned long nr_dirty = 0;
1501         unsigned long nr_writeback = 0;
1502         isolate_mode_t reclaim_mode = ISOLATE_INACTIVE;
1503         struct zone *zone = mz->zone;
1504
1505         while (unlikely(too_many_isolated(zone, file, sc))) {
1506                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1507
1508                 /* We are about to die and free our memory. Return now. */
1509                 if (fatal_signal_pending(current))
1510                         return SWAP_CLUSTER_MAX;
1511         }
1512
1513         set_reclaim_mode(priority, sc, false);
1514         if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
1515                 reclaim_mode |= ISOLATE_ACTIVE;
1516
1517         lru_add_drain();
1518
1519         if (!sc->may_unmap)
1520                 reclaim_mode |= ISOLATE_UNMAPPED;
1521         if (!sc->may_writepage)
1522                 reclaim_mode |= ISOLATE_CLEAN;
1523
1524         spin_lock_irq(&zone->lru_lock);
1525
1526         nr_taken = isolate_pages(nr_to_scan, mz, &page_list,
1527                                  &nr_scanned, sc->order,
1528                                  reclaim_mode, 0, file);
1529         if (global_reclaim(sc)) {
1530                 zone->pages_scanned += nr_scanned;
1531                 if (current_is_kswapd())
1532                         __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1533                                                nr_scanned);
1534                 else
1535                         __count_zone_vm_events(PGSCAN_DIRECT, zone,
1536                                                nr_scanned);
1537         }
1538
1539         if (nr_taken == 0) {
1540                 spin_unlock_irq(&zone->lru_lock);
1541                 return 0;
1542         }
1543
1544         update_isolated_counts(mz, sc, &nr_anon, &nr_file, &page_list);
1545
1546         spin_unlock_irq(&zone->lru_lock);
1547
1548         nr_reclaimed = shrink_page_list(&page_list, mz, sc, priority,
1549                                                 &nr_dirty, &nr_writeback);
1550
1551         /* Check if we should syncronously wait for writeback */
1552         if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1553                 set_reclaim_mode(priority, sc, true);
1554                 nr_reclaimed += shrink_page_list(&page_list, mz, sc,
1555                                         priority, &nr_dirty, &nr_writeback);
1556         }
1557
1558         local_irq_disable();
1559         if (current_is_kswapd())
1560                 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1561         __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1562
1563         putback_lru_pages(mz, sc, nr_anon, nr_file, &page_list);
1564
1565         /*
1566          * If reclaim is isolating dirty pages under writeback, it implies
1567          * that the long-lived page allocation rate is exceeding the page
1568          * laundering rate. Either the global limits are not being effective
1569          * at throttling processes due to the page distribution throughout
1570          * zones or there is heavy usage of a slow backing device. The
1571          * only option is to throttle from reclaim context which is not ideal
1572          * as there is no guarantee the dirtying process is throttled in the
1573          * same way balance_dirty_pages() manages.
1574          *
1575          * This scales the number of dirty pages that must be under writeback
1576          * before throttling depending on priority. It is a simple backoff
1577          * function that has the most effect in the range DEF_PRIORITY to
1578          * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1579          * in trouble and reclaim is considered to be in trouble.
1580          *
1581          * DEF_PRIORITY   100% isolated pages must be PageWriteback to throttle
1582          * DEF_PRIORITY-1  50% must be PageWriteback
1583          * DEF_PRIORITY-2  25% must be PageWriteback, kswapd in trouble
1584          * ...
1585          * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1586          *                     isolated page is PageWriteback
1587          */
1588         if (nr_writeback && nr_writeback >= (nr_taken >> (DEF_PRIORITY-priority)))
1589                 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1590
1591         trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1592                 zone_idx(zone),
1593                 nr_scanned, nr_reclaimed,
1594                 priority,
1595                 trace_shrink_flags(file, sc->reclaim_mode));
1596         return nr_reclaimed;
1597 }
1598
1599 /*
1600  * This moves pages from the active list to the inactive list.
1601  *
1602  * We move them the other way if the page is referenced by one or more
1603  * processes, from rmap.
1604  *
1605  * If the pages are mostly unmapped, the processing is fast and it is
1606  * appropriate to hold zone->lru_lock across the whole operation.  But if
1607  * the pages are mapped, the processing is slow (page_referenced()) so we
1608  * should drop zone->lru_lock around each page.  It's impossible to balance
1609  * this, so instead we remove the pages from the LRU while processing them.
1610  * It is safe to rely on PG_active against the non-LRU pages in here because
1611  * nobody will play with that bit on a non-LRU page.
1612  *
1613  * The downside is that we have to touch page->_count against each page.
1614  * But we had to alter page->flags anyway.
1615  */
1616
1617 static void move_active_pages_to_lru(struct zone *zone,
1618                                      struct list_head *list,
1619                                      enum lru_list lru)
1620 {
1621         unsigned long pgmoved = 0;
1622         struct pagevec pvec;
1623         struct page *page;
1624
1625         pagevec_init(&pvec, 1);
1626
1627         while (!list_empty(list)) {
1628                 struct lruvec *lruvec;
1629
1630                 page = lru_to_page(list);
1631
1632                 VM_BUG_ON(PageLRU(page));
1633                 SetPageLRU(page);
1634
1635                 lruvec = mem_cgroup_lru_add_list(zone, page, lru);
1636                 list_move(&page->lru, &lruvec->lists[lru]);
1637                 pgmoved += hpage_nr_pages(page);
1638
1639                 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1640                         spin_unlock_irq(&zone->lru_lock);
1641                         if (buffer_heads_over_limit)
1642                                 pagevec_strip(&pvec);
1643                         __pagevec_release(&pvec);
1644                         spin_lock_irq(&zone->lru_lock);
1645                 }
1646         }
1647         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1648         if (!is_active_lru(lru))
1649                 __count_vm_events(PGDEACTIVATE, pgmoved);
1650 }
1651
1652 static void shrink_active_list(unsigned long nr_pages,
1653                                struct mem_cgroup_zone *mz,
1654                                struct scan_control *sc,
1655                                int priority, int file)
1656 {
1657         unsigned long nr_taken;
1658         unsigned long pgscanned;
1659         unsigned long vm_flags;
1660         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1661         LIST_HEAD(l_active);
1662         LIST_HEAD(l_inactive);
1663         struct page *page;
1664         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz);
1665         unsigned long nr_rotated = 0;
1666         isolate_mode_t reclaim_mode = ISOLATE_ACTIVE;
1667         struct zone *zone = mz->zone;
1668
1669         lru_add_drain();
1670
1671         if (!sc->may_unmap)
1672                 reclaim_mode |= ISOLATE_UNMAPPED;
1673         if (!sc->may_writepage)
1674                 reclaim_mode |= ISOLATE_CLEAN;
1675
1676         spin_lock_irq(&zone->lru_lock);
1677
1678         nr_taken = isolate_pages(nr_pages, mz, &l_hold,
1679                                  &pgscanned, sc->order,
1680                                  reclaim_mode, 1, file);
1681
1682         if (global_reclaim(sc))
1683                 zone->pages_scanned += pgscanned;
1684
1685         reclaim_stat->recent_scanned[file] += nr_taken;
1686
1687         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1688         if (file)
1689                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1690         else
1691                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1692         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1693         spin_unlock_irq(&zone->lru_lock);
1694
1695         while (!list_empty(&l_hold)) {
1696                 cond_resched();
1697                 page = lru_to_page(&l_hold);
1698                 list_del(&page->lru);
1699
1700                 if (unlikely(!page_evictable(page, NULL))) {
1701                         putback_lru_page(page);
1702                         continue;
1703                 }
1704
1705                 if (page_referenced(page, 0, mz->mem_cgroup, &vm_flags)) {
1706                         nr_rotated += hpage_nr_pages(page);
1707                         /*
1708                          * Identify referenced, file-backed active pages and
1709                          * give them one more trip around the active list. So
1710                          * that executable code get better chances to stay in
1711                          * memory under moderate memory pressure.  Anon pages
1712                          * are not likely to be evicted by use-once streaming
1713                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1714                          * so we ignore them here.
1715                          */
1716                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1717                                 list_add(&page->lru, &l_active);
1718                                 continue;
1719                         }
1720                 }
1721
1722                 ClearPageActive(page);  /* we are de-activating */
1723                 list_add(&page->lru, &l_inactive);
1724         }
1725
1726         /*
1727          * Move pages back to the lru list.
1728          */
1729         spin_lock_irq(&zone->lru_lock);
1730         /*
1731          * Count referenced pages from currently used mappings as rotated,
1732          * even though only some of them are actually re-activated.  This
1733          * helps balance scan pressure between file and anonymous pages in
1734          * get_scan_ratio.
1735          */
1736         reclaim_stat->recent_rotated[file] += nr_rotated;
1737
1738         move_active_pages_to_lru(zone, &l_active,
1739                                                 LRU_ACTIVE + file * LRU_FILE);
1740         move_active_pages_to_lru(zone, &l_inactive,
1741                                                 LRU_BASE   + file * LRU_FILE);
1742         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1743         spin_unlock_irq(&zone->lru_lock);
1744 }
1745
1746 #ifdef CONFIG_SWAP
1747 static int inactive_anon_is_low_global(struct zone *zone)
1748 {
1749         unsigned long active, inactive;
1750
1751         active = zone_page_state(zone, NR_ACTIVE_ANON);
1752         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1753
1754         if (inactive * zone->inactive_ratio < active)
1755                 return 1;
1756
1757         return 0;
1758 }
1759
1760 /**
1761  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1762  * @zone: zone to check
1763  * @sc:   scan control of this context
1764  *
1765  * Returns true if the zone does not have enough inactive anon pages,
1766  * meaning some active anon pages need to be deactivated.
1767  */
1768 static int inactive_anon_is_low(struct mem_cgroup_zone *mz)
1769 {
1770         /*
1771          * If we don't have swap space, anonymous page deactivation
1772          * is pointless.
1773          */
1774         if (!total_swap_pages)
1775                 return 0;
1776
1777         if (!scanning_global_lru(mz))
1778                 return mem_cgroup_inactive_anon_is_low(mz->mem_cgroup,
1779                                                        mz->zone);
1780
1781         return inactive_anon_is_low_global(mz->zone);
1782 }
1783 #else
1784 static inline int inactive_anon_is_low(struct mem_cgroup_zone *mz)
1785 {
1786         return 0;
1787 }
1788 #endif
1789
1790 static int inactive_file_is_low_global(struct zone *zone)
1791 {
1792         unsigned long active, inactive;
1793
1794         active = zone_page_state(zone, NR_ACTIVE_FILE);
1795         inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1796
1797         return (active > inactive);
1798 }
1799
1800 /**
1801  * inactive_file_is_low - check if file pages need to be deactivated
1802  * @mz: memory cgroup and zone to check
1803  *
1804  * When the system is doing streaming IO, memory pressure here
1805  * ensures that active file pages get deactivated, until more
1806  * than half of the file pages are on the inactive list.
1807  *
1808  * Once we get to that situation, protect the system's working
1809  * set from being evicted by disabling active file page aging.
1810  *
1811  * This uses a different ratio than the anonymous pages, because
1812  * the page cache uses a use-once replacement algorithm.
1813  */
1814 static int inactive_file_is_low(struct mem_cgroup_zone *mz)
1815 {
1816         if (!scanning_global_lru(mz))
1817                 return mem_cgroup_inactive_file_is_low(mz->mem_cgroup,
1818                                                        mz->zone);
1819
1820         return inactive_file_is_low_global(mz->zone);
1821 }
1822
1823 static int inactive_list_is_low(struct mem_cgroup_zone *mz, int file)
1824 {
1825         if (file)
1826                 return inactive_file_is_low(mz);
1827         else
1828                 return inactive_anon_is_low(mz);
1829 }
1830
1831 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1832                                  struct mem_cgroup_zone *mz,
1833                                  struct scan_control *sc, int priority)
1834 {
1835         int file = is_file_lru(lru);
1836
1837         if (is_active_lru(lru)) {
1838                 if (inactive_list_is_low(mz, file))
1839                         shrink_active_list(nr_to_scan, mz, sc, priority, file);
1840                 return 0;
1841         }
1842
1843         return shrink_inactive_list(nr_to_scan, mz, sc, priority, file);
1844 }
1845
1846 static int vmscan_swappiness(struct mem_cgroup_zone *mz,
1847                              struct scan_control *sc)
1848 {
1849         if (global_reclaim(sc))
1850                 return vm_swappiness;
1851         return mem_cgroup_swappiness(mz->mem_cgroup);
1852 }
1853
1854 /*
1855  * Determine how aggressively the anon and file LRU lists should be
1856  * scanned.  The relative value of each set of LRU lists is determined
1857  * by looking at the fraction of the pages scanned we did rotate back
1858  * onto the active list instead of evict.
1859  *
1860  * nr[0] = anon pages to scan; nr[1] = file pages to scan
1861  */
1862 static void get_scan_count(struct mem_cgroup_zone *mz, struct scan_control *sc,
1863                            unsigned long *nr, int priority)
1864 {
1865         unsigned long anon, file, free;
1866         unsigned long anon_prio, file_prio;
1867         unsigned long ap, fp;
1868         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz);
1869         u64 fraction[2], denominator;
1870         enum lru_list l;
1871         int noswap = 0;
1872         bool force_scan = false;
1873
1874         /*
1875          * If the zone or memcg is small, nr[l] can be 0.  This
1876          * results in no scanning on this priority and a potential
1877          * priority drop.  Global direct reclaim can go to the next
1878          * zone and tends to have no problems. Global kswapd is for
1879          * zone balancing and it needs to scan a minimum amount. When
1880          * reclaiming for a memcg, a priority drop can cause high
1881          * latencies, so it's better to scan a minimum amount there as
1882          * well.
1883          */
1884         if (current_is_kswapd() && mz->zone->all_unreclaimable)
1885                 force_scan = true;
1886         if (!global_reclaim(sc))
1887                 force_scan = true;
1888
1889         /* If we have no swap space, do not bother scanning anon pages. */
1890         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1891                 noswap = 1;
1892                 fraction[0] = 0;
1893                 fraction[1] = 1;
1894                 denominator = 1;
1895                 goto out;
1896         }
1897
1898         anon  = zone_nr_lru_pages(mz, LRU_ACTIVE_ANON) +
1899                 zone_nr_lru_pages(mz, LRU_INACTIVE_ANON);
1900         file  = zone_nr_lru_pages(mz, LRU_ACTIVE_FILE) +
1901                 zone_nr_lru_pages(mz, LRU_INACTIVE_FILE);
1902
1903         if (global_reclaim(sc)) {
1904                 free  = zone_page_state(mz->zone, NR_FREE_PAGES);
1905                 /* If we have very few page cache pages,
1906                    force-scan anon pages. */
1907                 if (unlikely(file + free <= high_wmark_pages(mz->zone))) {
1908                         fraction[0] = 1;
1909                         fraction[1] = 0;
1910                         denominator = 1;
1911                         goto out;
1912                 }
1913         }
1914
1915         /*
1916          * With swappiness at 100, anonymous and file have the same priority.
1917          * This scanning priority is essentially the inverse of IO cost.
1918          */
1919         anon_prio = vmscan_swappiness(mz, sc);
1920         file_prio = 200 - vmscan_swappiness(mz, sc);
1921
1922         /*
1923          * OK, so we have swap space and a fair amount of page cache
1924          * pages.  We use the recently rotated / recently scanned
1925          * ratios to determine how valuable each cache is.
1926          *
1927          * Because workloads change over time (and to avoid overflow)
1928          * we keep these statistics as a floating average, which ends
1929          * up weighing recent references more than old ones.
1930          *
1931          * anon in [0], file in [1]
1932          */
1933         spin_lock_irq(&mz->zone->lru_lock);
1934         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1935                 reclaim_stat->recent_scanned[0] /= 2;
1936                 reclaim_stat->recent_rotated[0] /= 2;
1937         }
1938
1939         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1940                 reclaim_stat->recent_scanned[1] /= 2;
1941                 reclaim_stat->recent_rotated[1] /= 2;
1942         }
1943
1944         /*
1945          * The amount of pressure on anon vs file pages is inversely
1946          * proportional to the fraction of recently scanned pages on
1947          * each list that were recently referenced and in active use.
1948          */
1949         ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1950         ap /= reclaim_stat->recent_rotated[0] + 1;
1951
1952         fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1953         fp /= reclaim_stat->recent_rotated[1] + 1;
1954         spin_unlock_irq(&mz->zone->lru_lock);
1955
1956         fraction[0] = ap;
1957         fraction[1] = fp;
1958         denominator = ap + fp + 1;
1959 out:
1960         for_each_evictable_lru(l) {
1961                 int file = is_file_lru(l);
1962                 unsigned long scan;
1963
1964                 scan = zone_nr_lru_pages(mz, l);
1965                 if (priority || noswap) {
1966                         scan >>= priority;
1967                         if (!scan && force_scan)
1968                                 scan = SWAP_CLUSTER_MAX;
1969                         scan = div64_u64(scan * fraction[file], denominator);
1970                 }
1971                 nr[l] = scan;
1972         }
1973 }
1974
1975 /*
1976  * Reclaim/compaction depends on a number of pages being freed. To avoid
1977  * disruption to the system, a small number of order-0 pages continue to be
1978  * rotated and reclaimed in the normal fashion. However, by the time we get
1979  * back to the allocator and call try_to_compact_zone(), we ensure that
1980  * there are enough free pages for it to be likely successful
1981  */
1982 static inline bool should_continue_reclaim(struct mem_cgroup_zone *mz,
1983                                         unsigned long nr_reclaimed,
1984                                         unsigned long nr_scanned,
1985                                         struct scan_control *sc)
1986 {
1987         unsigned long pages_for_compaction;
1988         unsigned long inactive_lru_pages;
1989
1990         /* If not in reclaim/compaction mode, stop */
1991         if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
1992                 return false;
1993
1994         /* Consider stopping depending on scan and reclaim activity */
1995         if (sc->gfp_mask & __GFP_REPEAT) {
1996                 /*
1997                  * For __GFP_REPEAT allocations, stop reclaiming if the
1998                  * full LRU list has been scanned and we are still failing
1999                  * to reclaim pages. This full LRU scan is potentially
2000                  * expensive but a __GFP_REPEAT caller really wants to succeed
2001                  */
2002                 if (!nr_reclaimed && !nr_scanned)
2003                         return false;
2004         } else {
2005                 /*
2006                  * For non-__GFP_REPEAT allocations which can presumably
2007                  * fail without consequence, stop if we failed to reclaim
2008                  * any pages from the last SWAP_CLUSTER_MAX number of
2009                  * pages that were scanned. This will return to the
2010                  * caller faster at the risk reclaim/compaction and
2011                  * the resulting allocation attempt fails
2012                  */
2013                 if (!nr_reclaimed)
2014                         return false;
2015         }
2016
2017         /*
2018          * If we have not reclaimed enough pages for compaction and the
2019          * inactive lists are large enough, continue reclaiming
2020          */
2021         pages_for_compaction = (2UL << sc->order);
2022         inactive_lru_pages = zone_nr_lru_pages(mz, LRU_INACTIVE_FILE);
2023         if (nr_swap_pages > 0)
2024                 inactive_lru_pages += zone_nr_lru_pages(mz, LRU_INACTIVE_ANON);
2025         if (sc->nr_reclaimed < pages_for_compaction &&
2026                         inactive_lru_pages > pages_for_compaction)
2027                 return true;
2028
2029         /* If compaction would go ahead or the allocation would succeed, stop */
2030         switch (compaction_suitable(mz->zone, sc->order)) {
2031         case COMPACT_PARTIAL:
2032         case COMPACT_CONTINUE:
2033                 return false;
2034         default:
2035                 return true;
2036         }
2037 }
2038
2039 /*
2040  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
2041  */
2042 static void shrink_mem_cgroup_zone(int priority, struct mem_cgroup_zone *mz,
2043                                    struct scan_control *sc)
2044 {
2045         unsigned long nr[NR_LRU_LISTS];
2046         unsigned long nr_to_scan;
2047         enum lru_list l;
2048         unsigned long nr_reclaimed, nr_scanned;
2049         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2050         struct blk_plug plug;
2051
2052 restart:
2053         nr_reclaimed = 0;
2054         nr_scanned = sc->nr_scanned;
2055         get_scan_count(mz, sc, nr, priority);
2056
2057         blk_start_plug(&plug);
2058         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2059                                         nr[LRU_INACTIVE_FILE]) {
2060                 for_each_evictable_lru(l) {
2061                         if (nr[l]) {
2062                                 nr_to_scan = min_t(unsigned long,
2063                                                    nr[l], SWAP_CLUSTER_MAX);
2064                                 nr[l] -= nr_to_scan;
2065
2066                                 nr_reclaimed += shrink_list(l, nr_to_scan,
2067                                                             mz, sc, priority);
2068                         }
2069                 }
2070                 /*
2071                  * On large memory systems, scan >> priority can become
2072                  * really large. This is fine for the starting priority;
2073                  * we want to put equal scanning pressure on each zone.
2074                  * However, if the VM has a harder time of freeing pages,
2075                  * with multiple processes reclaiming pages, the total
2076                  * freeing target can get unreasonably large.
2077                  */
2078                 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
2079                         break;
2080         }
2081         blk_finish_plug(&plug);
2082         sc->nr_reclaimed += nr_reclaimed;
2083
2084         /*
2085          * Even if we did not try to evict anon pages at all, we want to
2086          * rebalance the anon lru active/inactive ratio.
2087          */
2088         if (inactive_anon_is_low(mz))
2089                 shrink_active_list(SWAP_CLUSTER_MAX, mz, sc, priority, 0);
2090
2091         /* reclaim/compaction might need reclaim to continue */
2092         if (should_continue_reclaim(mz, nr_reclaimed,
2093                                         sc->nr_scanned - nr_scanned, sc))
2094                 goto restart;
2095
2096         throttle_vm_writeout(sc->gfp_mask);
2097 }
2098
2099 static void shrink_zone(int priority, struct zone *zone,
2100                         struct scan_control *sc)
2101 {
2102         struct mem_cgroup *root = sc->target_mem_cgroup;
2103         struct mem_cgroup_reclaim_cookie reclaim = {
2104                 .zone = zone,
2105                 .priority = priority,
2106         };
2107         struct mem_cgroup *memcg;
2108
2109         memcg = mem_cgroup_iter(root, NULL, &reclaim);
2110         do {
2111                 struct mem_cgroup_zone mz = {
2112                         .mem_cgroup = memcg,
2113                         .zone = zone,
2114                 };
2115
2116                 shrink_mem_cgroup_zone(priority, &mz, sc);
2117                 /*
2118                  * Limit reclaim has historically picked one memcg and
2119                  * scanned it with decreasing priority levels until
2120                  * nr_to_reclaim had been reclaimed.  This priority
2121                  * cycle is thus over after a single memcg.
2122                  *
2123                  * Direct reclaim and kswapd, on the other hand, have
2124                  * to scan all memory cgroups to fulfill the overall
2125                  * scan target for the zone.
2126                  */
2127                 if (!global_reclaim(sc)) {
2128                         mem_cgroup_iter_break(root, memcg);
2129                         break;
2130                 }
2131                 memcg = mem_cgroup_iter(root, memcg, &reclaim);
2132         } while (memcg);
2133 }
2134
2135 /*
2136  * This is the direct reclaim path, for page-allocating processes.  We only
2137  * try to reclaim pages from zones which will satisfy the caller's allocation
2138  * request.
2139  *
2140  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2141  * Because:
2142  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2143  *    allocation or
2144  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2145  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2146  *    zone defense algorithm.
2147  *
2148  * If a zone is deemed to be full of pinned pages then just give it a light
2149  * scan then give up on it.
2150  *
2151  * This function returns true if a zone is being reclaimed for a costly
2152  * high-order allocation and compaction is either ready to begin or deferred.
2153  * This indicates to the caller that it should retry the allocation or fail.
2154  */
2155 static bool shrink_zones(int priority, struct zonelist *zonelist,
2156                                         struct scan_control *sc)
2157 {
2158         struct zoneref *z;
2159         struct zone *zone;
2160         unsigned long nr_soft_reclaimed;
2161         unsigned long nr_soft_scanned;
2162         bool should_abort_reclaim = false;
2163
2164         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2165                                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2166                 if (!populated_zone(zone))
2167                         continue;
2168                 /*
2169                  * Take care memory controller reclaiming has small influence
2170                  * to global LRU.
2171                  */
2172                 if (global_reclaim(sc)) {
2173                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2174                                 continue;
2175                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2176                                 continue;       /* Let kswapd poll it */
2177                         if (COMPACTION_BUILD) {
2178                                 /*
2179                                  * If we already have plenty of memory free for
2180                                  * compaction in this zone, don't free any more.
2181                                  * Even though compaction is invoked for any
2182                                  * non-zero order, only frequent costly order
2183                                  * reclamation is disruptive enough to become a
2184                                  * noticable problem, like transparent huge page
2185                                  * allocations.
2186                                  */
2187                                 if (sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2188                                         (compaction_suitable(zone, sc->order) ||
2189                                          compaction_deferred(zone))) {
2190                                         should_abort_reclaim = true;
2191                                         continue;
2192                                 }
2193                         }
2194                         /*
2195                          * This steals pages from memory cgroups over softlimit
2196                          * and returns the number of reclaimed pages and
2197                          * scanned pages. This works for global memory pressure
2198                          * and balancing, not for a memcg's limit.
2199                          */
2200                         nr_soft_scanned = 0;
2201                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2202                                                 sc->order, sc->gfp_mask,
2203                                                 &nr_soft_scanned);
2204                         sc->nr_reclaimed += nr_soft_reclaimed;
2205                         sc->nr_scanned += nr_soft_scanned;
2206                         /* need some check for avoid more shrink_zone() */
2207                 }
2208
2209                 shrink_zone(priority, zone, sc);
2210         }
2211
2212         return should_abort_reclaim;
2213 }
2214
2215 static bool zone_reclaimable(struct zone *zone)
2216 {
2217         return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2218 }
2219
2220 /* All zones in zonelist are unreclaimable? */
2221 static bool all_unreclaimable(struct zonelist *zonelist,
2222                 struct scan_control *sc)
2223 {
2224         struct zoneref *z;
2225         struct zone *zone;
2226
2227         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2228                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2229                 if (!populated_zone(zone))
2230                         continue;
2231                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2232                         continue;
2233                 if (!zone->all_unreclaimable)
2234                         return false;
2235         }
2236
2237         return true;
2238 }
2239
2240 /*
2241  * This is the main entry point to direct page reclaim.
2242  *
2243  * If a full scan of the inactive list fails to free enough memory then we
2244  * are "out of memory" and something needs to be killed.
2245  *
2246  * If the caller is !__GFP_FS then the probability of a failure is reasonably
2247  * high - the zone may be full of dirty or under-writeback pages, which this
2248  * caller can't do much about.  We kick the writeback threads and take explicit
2249  * naps in the hope that some of these pages can be written.  But if the
2250  * allocating task holds filesystem locks which prevent writeout this might not
2251  * work, and the allocation attempt will fail.
2252  *
2253  * returns:     0, if no pages reclaimed
2254  *              else, the number of pages reclaimed
2255  */
2256 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2257                                         struct scan_control *sc,
2258                                         struct shrink_control *shrink)
2259 {
2260         int priority;
2261         unsigned long total_scanned = 0;
2262         struct reclaim_state *reclaim_state = current->reclaim_state;
2263         struct zoneref *z;
2264         struct zone *zone;
2265         unsigned long writeback_threshold;
2266         bool should_abort_reclaim;
2267
2268         get_mems_allowed();
2269         delayacct_freepages_start();
2270
2271         if (global_reclaim(sc))
2272                 count_vm_event(ALLOCSTALL);
2273
2274         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2275                 sc->nr_scanned = 0;
2276                 if (!priority)
2277                         disable_swap_token(sc->target_mem_cgroup);
2278                 should_abort_reclaim = shrink_zones(priority, zonelist, sc);
2279                 if (should_abort_reclaim)
2280                         break;
2281
2282                 /*
2283                  * Don't shrink slabs when reclaiming memory from
2284                  * over limit cgroups
2285                  */
2286                 if (global_reclaim(sc)) {
2287                         unsigned long lru_pages = 0;
2288                         for_each_zone_zonelist(zone, z, zonelist,
2289                                         gfp_zone(sc->gfp_mask)) {
2290                                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2291                                         continue;
2292
2293                                 lru_pages += zone_reclaimable_pages(zone);
2294                         }
2295
2296                         shrink_slab(shrink, sc->nr_scanned, lru_pages);
2297                         if (reclaim_state) {
2298                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2299                                 reclaim_state->reclaimed_slab = 0;
2300                         }
2301                 }
2302                 total_scanned += sc->nr_scanned;
2303                 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2304                         goto out;
2305
2306                 /*
2307                  * Try to write back as many pages as we just scanned.  This
2308                  * tends to cause slow streaming writers to write data to the
2309                  * disk smoothly, at the dirtying rate, which is nice.   But
2310                  * that's undesirable in laptop mode, where we *want* lumpy
2311                  * writeout.  So in laptop mode, write out the whole world.
2312                  */
2313                 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2314                 if (total_scanned > writeback_threshold) {
2315                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2316                                                 WB_REASON_TRY_TO_FREE_PAGES);
2317                         sc->may_writepage = 1;
2318                 }
2319
2320                 /* Take a nap, wait for some writeback to complete */
2321                 if (!sc->hibernation_mode && sc->nr_scanned &&
2322                     priority < DEF_PRIORITY - 2) {
2323                         struct zone *preferred_zone;
2324
2325                         first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2326                                                 &cpuset_current_mems_allowed,
2327                                                 &preferred_zone);
2328                         wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2329                 }
2330         }
2331
2332 out:
2333         delayacct_freepages_end();
2334         put_mems_allowed();
2335
2336         if (sc->nr_reclaimed)
2337                 return sc->nr_reclaimed;
2338
2339         /*
2340          * As hibernation is going on, kswapd is freezed so that it can't mark
2341          * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2342          * check.
2343          */
2344         if (oom_killer_disabled)
2345                 return 0;
2346
2347         /* Aborting reclaim to try compaction? don't OOM, then */
2348         if (should_abort_reclaim)
2349                 return 1;
2350
2351         /* top priority shrink_zones still had more to do? don't OOM, then */
2352         if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
2353                 return 1;
2354
2355         return 0;
2356 }
2357
2358 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2359                                 gfp_t gfp_mask, nodemask_t *nodemask)
2360 {
2361         unsigned long nr_reclaimed;
2362         struct scan_control sc = {
2363                 .gfp_mask = gfp_mask,
2364                 .may_writepage = !laptop_mode,
2365                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2366                 .may_unmap = 1,
2367                 .may_swap = 1,
2368                 .order = order,
2369                 .target_mem_cgroup = NULL,
2370                 .nodemask = nodemask,
2371         };
2372         struct shrink_control shrink = {
2373                 .gfp_mask = sc.gfp_mask,
2374         };
2375
2376         trace_mm_vmscan_direct_reclaim_begin(order,
2377                                 sc.may_writepage,
2378                                 gfp_mask);
2379
2380         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2381
2382         trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2383
2384         return nr_reclaimed;
2385 }
2386
2387 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2388
2389 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
2390                                                 gfp_t gfp_mask, bool noswap,
2391                                                 struct zone *zone,
2392                                                 unsigned long *nr_scanned)
2393 {
2394         struct scan_control sc = {
2395                 .nr_scanned = 0,
2396                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2397                 .may_writepage = !laptop_mode,
2398                 .may_unmap = 1,
2399                 .may_swap = !noswap,
2400                 .order = 0,
2401                 .target_mem_cgroup = memcg,
2402         };
2403         struct mem_cgroup_zone mz = {
2404                 .mem_cgroup = memcg,
2405                 .zone = zone,
2406         };
2407
2408         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2409                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2410
2411         trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2412                                                       sc.may_writepage,
2413                                                       sc.gfp_mask);
2414
2415         /*
2416          * NOTE: Although we can get the priority field, using it
2417          * here is not a good idea, since it limits the pages we can scan.
2418          * if we don't reclaim here, the shrink_zone from balance_pgdat
2419          * will pick up pages from other mem cgroup's as well. We hack
2420          * the priority and make it zero.
2421          */
2422         shrink_mem_cgroup_zone(0, &mz, &sc);
2423
2424         trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2425
2426         *nr_scanned = sc.nr_scanned;
2427         return sc.nr_reclaimed;
2428 }
2429
2430 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2431                                            gfp_t gfp_mask,
2432                                            bool noswap)
2433 {
2434         struct zonelist *zonelist;
2435         unsigned long nr_reclaimed;
2436         int nid;
2437         struct scan_control sc = {
2438                 .may_writepage = !laptop_mode,
2439                 .may_unmap = 1,
2440                 .may_swap = !noswap,
2441                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2442                 .order = 0,
2443                 .target_mem_cgroup = memcg,
2444                 .nodemask = NULL, /* we don't care the placement */
2445                 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2446                                 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2447         };
2448         struct shrink_control shrink = {
2449                 .gfp_mask = sc.gfp_mask,
2450         };
2451
2452         /*
2453          * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2454          * take care of from where we get pages. So the node where we start the
2455          * scan does not need to be the current node.
2456          */
2457         nid = mem_cgroup_select_victim_node(memcg);
2458
2459         zonelist = NODE_DATA(nid)->node_zonelists;
2460
2461         trace_mm_vmscan_memcg_reclaim_begin(0,
2462                                             sc.may_writepage,
2463                                             sc.gfp_mask);
2464
2465         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2466
2467         trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2468
2469         return nr_reclaimed;
2470 }
2471 #endif
2472
2473 static void age_active_anon(struct zone *zone, struct scan_control *sc,
2474                             int priority)
2475 {
2476         struct mem_cgroup *memcg;
2477
2478         if (!total_swap_pages)
2479                 return;
2480
2481         memcg = mem_cgroup_iter(NULL, NULL, NULL);
2482         do {
2483                 struct mem_cgroup_zone mz = {
2484                         .mem_cgroup = memcg,
2485                         .zone = zone,
2486                 };
2487
2488                 if (inactive_anon_is_low(&mz))
2489                         shrink_active_list(SWAP_CLUSTER_MAX, &mz,
2490                                            sc, priority, 0);
2491
2492                 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2493         } while (memcg);
2494 }
2495
2496 /*
2497  * pgdat_balanced is used when checking if a node is balanced for high-order
2498  * allocations. Only zones that meet watermarks and are in a zone allowed
2499  * by the callers classzone_idx are added to balanced_pages. The total of
2500  * balanced pages must be at least 25% of the zones allowed by classzone_idx
2501  * for the node to be considered balanced. Forcing all zones to be balanced
2502  * for high orders can cause excessive reclaim when there are imbalanced zones.
2503  * The choice of 25% is due to
2504  *   o a 16M DMA zone that is balanced will not balance a zone on any
2505  *     reasonable sized machine
2506  *   o On all other machines, the top zone must be at least a reasonable
2507  *     percentage of the middle zones. For example, on 32-bit x86, highmem
2508  *     would need to be at least 256M for it to be balance a whole node.
2509  *     Similarly, on x86-64 the Normal zone would need to be at least 1G
2510  *     to balance a node on its own. These seemed like reasonable ratios.
2511  */
2512 static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2513                                                 int classzone_idx)
2514 {
2515         unsigned long present_pages = 0;
2516         int i;
2517
2518         for (i = 0; i <= classzone_idx; i++)
2519                 present_pages += pgdat->node_zones[i].present_pages;
2520
2521         /* A special case here: if zone has no page, we think it's balanced */
2522         return balanced_pages >= (present_pages >> 2);
2523 }
2524
2525 /* is kswapd sleeping prematurely? */
2526 static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2527                                         int classzone_idx)
2528 {
2529         int i;
2530         unsigned long balanced = 0;
2531         bool all_zones_ok = true;
2532
2533         /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2534         if (remaining)
2535                 return true;
2536
2537         /* Check the watermark levels */
2538         for (i = 0; i <= classzone_idx; i++) {
2539                 struct zone *zone = pgdat->node_zones + i;
2540
2541                 if (!populated_zone(zone))
2542                         continue;
2543
2544                 /*
2545                  * balance_pgdat() skips over all_unreclaimable after
2546                  * DEF_PRIORITY. Effectively, it considers them balanced so
2547                  * they must be considered balanced here as well if kswapd
2548                  * is to sleep
2549                  */
2550                 if (zone->all_unreclaimable) {
2551                         balanced += zone->present_pages;
2552                         continue;
2553                 }
2554
2555                 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2556                                                         i, 0))
2557                         all_zones_ok = false;
2558                 else
2559                         balanced += zone->present_pages;
2560         }
2561
2562         /*
2563          * For high-order requests, the balanced zones must contain at least
2564          * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2565          * must be balanced
2566          */
2567         if (order)
2568                 return !pgdat_balanced(pgdat, balanced, classzone_idx);
2569         else
2570                 return !all_zones_ok;
2571 }
2572
2573 /*
2574  * For kswapd, balance_pgdat() will work across all this node's zones until
2575  * they are all at high_wmark_pages(zone).
2576  *
2577  * Returns the final order kswapd was reclaiming at
2578  *
2579  * There is special handling here for zones which are full of pinned pages.
2580  * This can happen if the pages are all mlocked, or if they are all used by
2581  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
2582  * What we do is to detect the case where all pages in the zone have been
2583  * scanned twice and there has been zero successful reclaim.  Mark the zone as
2584  * dead and from now on, only perform a short scan.  Basically we're polling
2585  * the zone for when the problem goes away.
2586  *
2587  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
2588  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2589  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2590  * lower zones regardless of the number of free pages in the lower zones. This
2591  * interoperates with the page allocator fallback scheme to ensure that aging
2592  * of pages is balanced across the zones.
2593  */
2594 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2595                                                         int *classzone_idx)
2596 {
2597         int all_zones_ok;
2598         unsigned long balanced;
2599         int priority;
2600         int i;
2601         int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
2602         unsigned long total_scanned;
2603         struct reclaim_state *reclaim_state = current->reclaim_state;
2604         unsigned long nr_soft_reclaimed;
2605         unsigned long nr_soft_scanned;
2606         struct scan_control sc = {
2607                 .gfp_mask = GFP_KERNEL,
2608                 .may_unmap = 1,
2609                 .may_swap = 1,
2610                 /*
2611                  * kswapd doesn't want to be bailed out while reclaim. because
2612                  * we want to put equal scanning pressure on each zone.
2613                  */
2614                 .nr_to_reclaim = ULONG_MAX,
2615                 .order = order,
2616                 .target_mem_cgroup = NULL,
2617         };
2618         struct shrink_control shrink = {
2619                 .gfp_mask = sc.gfp_mask,
2620         };
2621 loop_again:
2622         total_scanned = 0;
2623         sc.nr_reclaimed = 0;
2624         sc.may_writepage = !laptop_mode;
2625         count_vm_event(PAGEOUTRUN);
2626
2627         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2628                 unsigned long lru_pages = 0;
2629                 int has_under_min_watermark_zone = 0;
2630
2631                 /* The swap token gets in the way of swapout... */
2632                 if (!priority)
2633                         disable_swap_token(NULL);
2634
2635                 all_zones_ok = 1;
2636                 balanced = 0;
2637
2638                 /*
2639                  * Scan in the highmem->dma direction for the highest
2640                  * zone which needs scanning
2641                  */
2642                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2643                         struct zone *zone = pgdat->node_zones + i;
2644
2645                         if (!populated_zone(zone))
2646                                 continue;
2647
2648                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2649                                 continue;
2650
2651                         /*
2652                          * Do some background aging of the anon list, to give
2653                          * pages a chance to be referenced before reclaiming.
2654                          */
2655                         age_active_anon(zone, &sc, priority);
2656
2657                         if (!zone_watermark_ok_safe(zone, order,
2658                                         high_wmark_pages(zone), 0, 0)) {
2659                                 end_zone = i;
2660                                 break;
2661                         } else {
2662                                 /* If balanced, clear the congested flag */
2663                                 zone_clear_flag(zone, ZONE_CONGESTED);
2664                         }
2665                 }
2666                 if (i < 0)
2667                         goto out;
2668
2669                 for (i = 0; i <= end_zone; i++) {
2670                         struct zone *zone = pgdat->node_zones + i;
2671
2672                         lru_pages += zone_reclaimable_pages(zone);
2673                 }
2674
2675                 /*
2676                  * Now scan the zone in the dma->highmem direction, stopping
2677                  * at the last zone which needs scanning.
2678                  *
2679                  * We do this because the page allocator works in the opposite
2680                  * direction.  This prevents the page allocator from allocating
2681                  * pages behind kswapd's direction of progress, which would
2682                  * cause too much scanning of the lower zones.
2683                  */
2684                 for (i = 0; i <= end_zone; i++) {
2685                         struct zone *zone = pgdat->node_zones + i;
2686                         int nr_slab;
2687                         unsigned long balance_gap;
2688
2689                         if (!populated_zone(zone))
2690                                 continue;
2691
2692                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2693                                 continue;
2694
2695                         sc.nr_scanned = 0;
2696
2697                         nr_soft_scanned = 0;
2698                         /*
2699                          * Call soft limit reclaim before calling shrink_zone.
2700                          */
2701                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2702                                                         order, sc.gfp_mask,
2703                                                         &nr_soft_scanned);
2704                         sc.nr_reclaimed += nr_soft_reclaimed;
2705                         total_scanned += nr_soft_scanned;
2706
2707                         /*
2708                          * We put equal pressure on every zone, unless
2709                          * one zone has way too many pages free
2710                          * already. The "too many pages" is defined
2711                          * as the high wmark plus a "gap" where the
2712                          * gap is either the low watermark or 1%
2713                          * of the zone, whichever is smaller.
2714                          */
2715                         balance_gap = min(low_wmark_pages(zone),
2716                                 (zone->present_pages +
2717                                         KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2718                                 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2719                         if (!zone_watermark_ok_safe(zone, order,
2720                                         high_wmark_pages(zone) + balance_gap,
2721                                         end_zone, 0)) {
2722                                 shrink_zone(priority, zone, &sc);
2723
2724                                 reclaim_state->reclaimed_slab = 0;
2725                                 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2726                                 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2727                                 total_scanned += sc.nr_scanned;
2728
2729                                 if (nr_slab == 0 && !zone_reclaimable(zone))
2730                                         zone->all_unreclaimable = 1;
2731                         }
2732
2733                         /*
2734                          * If we've done a decent amount of scanning and
2735                          * the reclaim ratio is low, start doing writepage
2736                          * even in laptop mode
2737                          */
2738                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2739                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2740                                 sc.may_writepage = 1;
2741
2742                         if (zone->all_unreclaimable) {
2743                                 if (end_zone && end_zone == i)
2744                                         end_zone--;
2745                                 continue;
2746                         }
2747
2748                         if (!zone_watermark_ok_safe(zone, order,
2749                                         high_wmark_pages(zone), end_zone, 0)) {
2750                                 all_zones_ok = 0;
2751                                 /*
2752                                  * We are still under min water mark.  This
2753                                  * means that we have a GFP_ATOMIC allocation
2754                                  * failure risk. Hurry up!
2755                                  */
2756                                 if (!zone_watermark_ok_safe(zone, order,
2757                                             min_wmark_pages(zone), end_zone, 0))
2758                                         has_under_min_watermark_zone = 1;
2759                         } else {
2760                                 /*
2761                                  * If a zone reaches its high watermark,
2762                                  * consider it to be no longer congested. It's
2763                                  * possible there are dirty pages backed by
2764                                  * congested BDIs but as pressure is relieved,
2765                                  * spectulatively avoid congestion waits
2766                                  */
2767                                 zone_clear_flag(zone, ZONE_CONGESTED);
2768                                 if (i <= *classzone_idx)
2769                                         balanced += zone->present_pages;
2770                         }
2771
2772                 }
2773                 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2774                         break;          /* kswapd: all done */
2775                 /*
2776                  * OK, kswapd is getting into trouble.  Take a nap, then take
2777                  * another pass across the zones.
2778                  */
2779                 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2780                         if (has_under_min_watermark_zone)
2781                                 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2782                         else
2783                                 congestion_wait(BLK_RW_ASYNC, HZ/10);
2784                 }
2785
2786                 /*
2787                  * We do this so kswapd doesn't build up large priorities for
2788                  * example when it is freeing in parallel with allocators. It
2789                  * matches the direct reclaim path behaviour in terms of impact
2790                  * on zone->*_priority.
2791                  */
2792                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2793                         break;
2794         }
2795 out:
2796
2797         /*
2798          * order-0: All zones must meet high watermark for a balanced node
2799          * high-order: Balanced zones must make up at least 25% of the node
2800          *             for the node to be balanced
2801          */
2802         if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2803                 cond_resched();
2804
2805                 try_to_freeze();
2806
2807                 /*
2808                  * Fragmentation may mean that the system cannot be
2809                  * rebalanced for high-order allocations in all zones.
2810                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2811                  * it means the zones have been fully scanned and are still
2812                  * not balanced. For high-order allocations, there is
2813                  * little point trying all over again as kswapd may
2814                  * infinite loop.
2815                  *
2816                  * Instead, recheck all watermarks at order-0 as they
2817                  * are the most important. If watermarks are ok, kswapd will go
2818                  * back to sleep. High-order users can still perform direct
2819                  * reclaim if they wish.
2820                  */
2821                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2822                         order = sc.order = 0;
2823
2824                 goto loop_again;
2825         }
2826
2827         /*
2828          * If kswapd was reclaiming at a higher order, it has the option of
2829          * sleeping without all zones being balanced. Before it does, it must
2830          * ensure that the watermarks for order-0 on *all* zones are met and
2831          * that the congestion flags are cleared. The congestion flag must
2832          * be cleared as kswapd is the only mechanism that clears the flag
2833          * and it is potentially going to sleep here.
2834          */
2835         if (order) {
2836                 for (i = 0; i <= end_zone; i++) {
2837                         struct zone *zone = pgdat->node_zones + i;
2838
2839                         if (!populated_zone(zone))
2840                                 continue;
2841
2842                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2843                                 continue;
2844
2845                         /* Confirm the zone is balanced for order-0 */
2846                         if (!zone_watermark_ok(zone, 0,
2847                                         high_wmark_pages(zone), 0, 0)) {
2848                                 order = sc.order = 0;
2849                                 goto loop_again;
2850                         }
2851
2852                         /* If balanced, clear the congested flag */
2853                         zone_clear_flag(zone, ZONE_CONGESTED);
2854                         if (i <= *classzone_idx)
2855                                 balanced += zone->present_pages;
2856                 }
2857         }
2858
2859         /*
2860          * Return the order we were reclaiming at so sleeping_prematurely()
2861          * makes a decision on the order we were last reclaiming at. However,
2862          * if another caller entered the allocator slow path while kswapd
2863          * was awake, order will remain at the higher level
2864          */
2865         *classzone_idx = end_zone;
2866         return order;
2867 }
2868
2869 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2870 {
2871         long remaining = 0;
2872         DEFINE_WAIT(wait);
2873
2874         if (freezing(current) || kthread_should_stop())
2875                 return;
2876
2877         prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2878
2879         /* Try to sleep for a short interval */
2880         if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2881                 remaining = schedule_timeout(HZ/10);
2882                 finish_wait(&pgdat->kswapd_wait, &wait);
2883                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2884         }
2885
2886         /*
2887          * After a short sleep, check if it was a premature sleep. If not, then
2888          * go fully to sleep until explicitly woken up.
2889          */
2890         if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2891                 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2892
2893                 /*
2894                  * vmstat counters are not perfectly accurate and the estimated
2895                  * value for counters such as NR_FREE_PAGES can deviate from the
2896                  * true value by nr_online_cpus * threshold. To avoid the zone
2897                  * watermarks being breached while under pressure, we reduce the
2898                  * per-cpu vmstat threshold while kswapd is awake and restore
2899                  * them before going back to sleep.
2900                  */
2901                 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2902                 schedule();
2903                 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2904         } else {
2905                 if (remaining)
2906                         count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2907                 else
2908                         count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2909         }
2910         finish_wait(&pgdat->kswapd_wait, &wait);
2911 }
2912
2913 /*
2914  * The background pageout daemon, started as a kernel thread
2915  * from the init process.
2916  *
2917  * This basically trickles out pages so that we have _some_
2918  * free memory available even if there is no other activity
2919  * that frees anything up. This is needed for things like routing
2920  * etc, where we otherwise might have all activity going on in
2921  * asynchronous contexts that cannot page things out.
2922  *
2923  * If there are applications that are active memory-allocators
2924  * (most normal use), this basically shouldn't matter.
2925  */
2926 static int kswapd(void *p)
2927 {
2928         unsigned long order, new_order;
2929         unsigned balanced_order;
2930         int classzone_idx, new_classzone_idx;
2931         int balanced_classzone_idx;
2932         pg_data_t *pgdat = (pg_data_t*)p;
2933         struct task_struct *tsk = current;
2934
2935         struct reclaim_state reclaim_state = {
2936                 .reclaimed_slab = 0,
2937         };
2938         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2939
2940         lockdep_set_current_reclaim_state(GFP_KERNEL);
2941
2942         if (!cpumask_empty(cpumask))
2943                 set_cpus_allowed_ptr(tsk, cpumask);
2944         current->reclaim_state = &reclaim_state;
2945
2946         /*
2947          * Tell the memory management that we're a "memory allocator",
2948          * and that if we need more memory we should get access to it
2949          * regardless (see "__alloc_pages()"). "kswapd" should
2950          * never get caught in the normal page freeing logic.
2951          *
2952          * (Kswapd normally doesn't need memory anyway, but sometimes
2953          * you need a small amount of memory in order to be able to
2954          * page out something else, and this flag essentially protects
2955          * us from recursively trying to free more memory as we're
2956          * trying to free the first piece of memory in the first place).
2957          */
2958         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2959         set_freezable();
2960
2961         order = new_order = 0;
2962         balanced_order = 0;
2963         classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
2964         balanced_classzone_idx = classzone_idx;
2965         for ( ; ; ) {
2966                 int ret;
2967
2968                 /*
2969                  * If the last balance_pgdat was unsuccessful it's unlikely a
2970                  * new request of a similar or harder type will succeed soon
2971                  * so consider going to sleep on the basis we reclaimed at
2972                  */
2973                 if (balanced_classzone_idx >= new_classzone_idx &&
2974                                         balanced_order == new_order) {
2975                         new_order = pgdat->kswapd_max_order;
2976                         new_classzone_idx = pgdat->classzone_idx;
2977                         pgdat->kswapd_max_order =  0;
2978                         pgdat->classzone_idx = pgdat->nr_zones - 1;
2979                 }
2980
2981                 if (order < new_order || classzone_idx > new_classzone_idx) {
2982                         /*
2983                          * Don't sleep if someone wants a larger 'order'
2984                          * allocation or has tigher zone constraints
2985                          */
2986                         order = new_order;
2987                         classzone_idx = new_classzone_idx;
2988                 } else {
2989                         kswapd_try_to_sleep(pgdat, balanced_order,
2990                                                 balanced_classzone_idx);
2991                         order = pgdat->kswapd_max_order;
2992                         classzone_idx = pgdat->classzone_idx;
2993                         new_order = order;
2994                         new_classzone_idx = classzone_idx;
2995                         pgdat->kswapd_max_order = 0;
2996                         pgdat->classzone_idx = pgdat->nr_zones - 1;
2997                 }
2998
2999                 ret = try_to_freeze();
3000                 if (kthread_should_stop())
3001                         break;
3002
3003                 /*
3004                  * We can speed up thawing tasks if we don't call balance_pgdat
3005                  * after returning from the refrigerator
3006                  */
3007                 if (!ret) {
3008                         trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
3009                         balanced_classzone_idx = classzone_idx;
3010                         balanced_order = balance_pgdat(pgdat, order,
3011                                                 &balanced_classzone_idx);
3012                 }
3013         }
3014         return 0;
3015 }
3016
3017 /*
3018  * A zone is low on free memory, so wake its kswapd task to service it.
3019  */
3020 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3021 {
3022         pg_data_t *pgdat;
3023
3024         if (!populated_zone(zone))
3025                 return;
3026
3027         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
3028                 return;
3029         pgdat = zone->zone_pgdat;
3030         if (pgdat->kswapd_max_order < order) {
3031                 pgdat->kswapd_max_order = order;
3032                 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3033         }
3034         if (!waitqueue_active(&pgdat->kswapd_wait))
3035                 return;
3036         if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
3037                 return;
3038
3039         trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3040         wake_up_interruptible(&pgdat->kswapd_wait);
3041 }
3042
3043 /*
3044  * The reclaimable count would be mostly accurate.
3045  * The less reclaimable pages may be
3046  * - mlocked pages, which will be moved to unevictable list when encountered
3047  * - mapped pages, which may require several travels to be reclaimed
3048  * - dirty pages, which is not "instantly" reclaimable
3049  */
3050 unsigned long global_reclaimable_pages(void)
3051 {
3052         int nr;
3053
3054         nr = global_page_state(NR_ACTIVE_FILE) +
3055              global_page_state(NR_INACTIVE_FILE);
3056
3057         if (nr_swap_pages > 0)
3058                 nr += global_page_state(NR_ACTIVE_ANON) +
3059                       global_page_state(NR_INACTIVE_ANON);
3060
3061         return nr;
3062 }
3063
3064 unsigned long zone_reclaimable_pages(struct zone *zone)
3065 {
3066         int nr;
3067
3068         nr = zone_page_state(zone, NR_ACTIVE_FILE) +
3069              zone_page_state(zone, NR_INACTIVE_FILE);
3070
3071         if (nr_swap_pages > 0)
3072                 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
3073                       zone_page_state(zone, NR_INACTIVE_ANON);
3074
3075         return nr;
3076 }
3077
3078 #ifdef CONFIG_HIBERNATION
3079 /*
3080  * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3081  * freed pages.
3082  *
3083  * Rather than trying to age LRUs the aim is to preserve the overall
3084  * LRU order by reclaiming preferentially
3085  * inactive > active > active referenced > active mapped
3086  */
3087 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3088 {
3089         struct reclaim_state reclaim_state;
3090         struct scan_control sc = {
3091                 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3092                 .may_swap = 1,
3093                 .may_unmap = 1,
3094                 .may_writepage = 1,
3095                 .nr_to_reclaim = nr_to_reclaim,
3096                 .hibernation_mode = 1,
3097                 .order = 0,
3098         };
3099         struct shrink_control shrink = {
3100                 .gfp_mask = sc.gfp_mask,
3101         };
3102         struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3103         struct task_struct *p = current;
3104         unsigned long nr_reclaimed;
3105
3106         p->flags |= PF_MEMALLOC;
3107         lockdep_set_current_reclaim_state(sc.gfp_mask);
3108         reclaim_state.reclaimed_slab = 0;
3109         p->reclaim_state = &reclaim_state;
3110
3111         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
3112
3113         p->reclaim_state = NULL;
3114         lockdep_clear_current_reclaim_state();
3115         p->flags &= ~PF_MEMALLOC;
3116
3117         return nr_reclaimed;
3118 }
3119 #endif /* CONFIG_HIBERNATION */
3120
3121 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3122    not required for correctness.  So if the last cpu in a node goes
3123    away, we get changed to run anywhere: as the first one comes back,
3124    restore their cpu bindings. */
3125 static int __devinit cpu_callback(struct notifier_block *nfb,
3126                                   unsigned long action, void *hcpu)
3127 {
3128         int nid;
3129
3130         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3131                 for_each_node_state(nid, N_HIGH_MEMORY) {
3132                         pg_data_t *pgdat = NODE_DATA(nid);
3133                         const struct cpumask *mask;
3134
3135                         mask = cpumask_of_node(pgdat->node_id);
3136
3137                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3138                                 /* One of our CPUs online: restore mask */
3139                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3140                 }
3141         }
3142         return NOTIFY_OK;
3143 }
3144
3145 /*
3146  * This kswapd start function will be called by init and node-hot-add.
3147  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3148  */
3149 int kswapd_run(int nid)
3150 {
3151         pg_data_t *pgdat = NODE_DATA(nid);
3152         int ret = 0;
3153
3154         if (pgdat->kswapd)
3155                 return 0;
3156
3157         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3158         if (IS_ERR(pgdat->kswapd)) {
3159                 /* failure at boot is fatal */
3160                 BUG_ON(system_state == SYSTEM_BOOTING);
3161                 printk("Failed to start kswapd on node %d\n",nid);
3162                 ret = -1;
3163         }
3164         return ret;
3165 }
3166
3167 /*
3168  * Called by memory hotplug when all memory in a node is offlined.
3169  */
3170 void kswapd_stop(int nid)
3171 {
3172         struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3173
3174         if (kswapd)
3175                 kthread_stop(kswapd);
3176 }
3177
3178 static int __init kswapd_init(void)
3179 {
3180         int nid;
3181
3182         swap_setup();
3183         for_each_node_state(nid, N_HIGH_MEMORY)
3184                 kswapd_run(nid);
3185         hotcpu_notifier(cpu_callback, 0);
3186         return 0;
3187 }
3188
3189 module_init(kswapd_init)
3190
3191 #ifdef CONFIG_NUMA
3192 /*
3193  * Zone reclaim mode
3194  *
3195  * If non-zero call zone_reclaim when the number of free pages falls below
3196  * the watermarks.
3197  */
3198 int zone_reclaim_mode __read_mostly;
3199
3200 #define RECLAIM_OFF 0
3201 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
3202 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
3203 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
3204
3205 /*
3206  * Priority for ZONE_RECLAIM. This determines the fraction of pages
3207  * of a node considered for each zone_reclaim. 4 scans 1/16th of
3208  * a zone.
3209  */
3210 #define ZONE_RECLAIM_PRIORITY 4
3211
3212 /*
3213  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3214  * occur.
3215  */
3216 int sysctl_min_unmapped_ratio = 1;
3217
3218 /*
3219  * If the number of slab pages in a zone grows beyond this percentage then
3220  * slab reclaim needs to occur.
3221  */
3222 int sysctl_min_slab_ratio = 5;
3223
3224 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3225 {
3226         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3227         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3228                 zone_page_state(zone, NR_ACTIVE_FILE);
3229
3230         /*
3231          * It's possible for there to be more file mapped pages than
3232          * accounted for by the pages on the file LRU lists because
3233          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3234          */
3235         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3236 }
3237
3238 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3239 static long zone_pagecache_reclaimable(struct zone *zone)
3240 {
3241         long nr_pagecache_reclaimable;
3242         long delta = 0;
3243
3244         /*
3245          * If RECLAIM_SWAP is set, then all file pages are considered
3246          * potentially reclaimable. Otherwise, we have to worry about
3247          * pages like swapcache and zone_unmapped_file_pages() provides
3248          * a better estimate
3249          */
3250         if (zone_reclaim_mode & RECLAIM_SWAP)
3251                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3252         else
3253                 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3254
3255         /* If we can't clean pages, remove dirty pages from consideration */
3256         if (!(zone_reclaim_mode & RECLAIM_WRITE))
3257                 delta += zone_page_state(zone, NR_FILE_DIRTY);
3258
3259         /* Watch for any possible underflows due to delta */
3260         if (unlikely(delta > nr_pagecache_reclaimable))
3261                 delta = nr_pagecache_reclaimable;
3262
3263         return nr_pagecache_reclaimable - delta;
3264 }
3265
3266 /*
3267  * Try to free up some pages from this zone through reclaim.
3268  */
3269 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3270 {
3271         /* Minimum pages needed in order to stay on node */
3272         const unsigned long nr_pages = 1 << order;
3273         struct task_struct *p = current;
3274         struct reclaim_state reclaim_state;
3275         int priority;
3276         struct scan_control sc = {
3277                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3278                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3279                 .may_swap = 1,
3280                 .nr_to_reclaim = max_t(unsigned long, nr_pages,
3281                                        SWAP_CLUSTER_MAX),
3282                 .gfp_mask = gfp_mask,
3283                 .order = order,
3284         };
3285         struct shrink_control shrink = {
3286                 .gfp_mask = sc.gfp_mask,
3287         };
3288         unsigned long nr_slab_pages0, nr_slab_pages1;
3289
3290         cond_resched();
3291         /*
3292          * We need to be able to allocate from the reserves for RECLAIM_SWAP
3293          * and we also need to be able to write out pages for RECLAIM_WRITE
3294          * and RECLAIM_SWAP.
3295          */
3296         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3297         lockdep_set_current_reclaim_state(gfp_mask);
3298         reclaim_state.reclaimed_slab = 0;
3299         p->reclaim_state = &reclaim_state;
3300
3301         if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3302                 /*
3303                  * Free memory by calling shrink zone with increasing
3304                  * priorities until we have enough memory freed.
3305                  */
3306                 priority = ZONE_RECLAIM_PRIORITY;
3307                 do {
3308                         shrink_zone(priority, zone, &sc);
3309                         priority--;
3310                 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
3311         }
3312
3313         nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3314         if (nr_slab_pages0 > zone->min_slab_pages) {
3315                 /*
3316                  * shrink_slab() does not currently allow us to determine how
3317                  * many pages were freed in this zone. So we take the current
3318                  * number of slab pages and shake the slab until it is reduced
3319                  * by the same nr_pages that we used for reclaiming unmapped
3320                  * pages.
3321                  *
3322                  * Note that shrink_slab will free memory on all zones and may
3323                  * take a long time.
3324                  */
3325                 for (;;) {
3326                         unsigned long lru_pages = zone_reclaimable_pages(zone);
3327
3328                         /* No reclaimable slab or very low memory pressure */
3329                         if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3330                                 break;
3331
3332                         /* Freed enough memory */
3333                         nr_slab_pages1 = zone_page_state(zone,
3334                                                         NR_SLAB_RECLAIMABLE);
3335                         if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3336                                 break;
3337                 }
3338
3339                 /*
3340                  * Update nr_reclaimed by the number of slab pages we
3341                  * reclaimed from this zone.
3342                  */
3343                 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3344                 if (nr_slab_pages1 < nr_slab_pages0)
3345                         sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3346         }
3347
3348         p->reclaim_state = NULL;
3349         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3350         lockdep_clear_current_reclaim_state();
3351         return sc.nr_reclaimed >= nr_pages;
3352 }
3353
3354 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3355 {
3356         int node_id;
3357         int ret;
3358
3359         /*
3360          * Zone reclaim reclaims unmapped file backed pages and
3361          * slab pages if we are over the defined limits.
3362          *
3363          * A small portion of unmapped file backed pages is needed for
3364          * file I/O otherwise pages read by file I/O will be immediately
3365          * thrown out if the zone is overallocated. So we do not reclaim
3366          * if less than a specified percentage of the zone is used by
3367          * unmapped file backed pages.
3368          */
3369         if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3370             zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3371                 return ZONE_RECLAIM_FULL;
3372
3373         if (zone->all_unreclaimable)
3374                 return ZONE_RECLAIM_FULL;
3375
3376         /*
3377          * Do not scan if the allocation should not be delayed.
3378          */
3379         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3380                 return ZONE_RECLAIM_NOSCAN;
3381
3382         /*
3383          * Only run zone reclaim on the local zone or on zones that do not
3384          * have associated processors. This will favor the local processor
3385          * over remote processors and spread off node memory allocations
3386          * as wide as possible.
3387          */
3388         node_id = zone_to_nid(zone);
3389         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3390                 return ZONE_RECLAIM_NOSCAN;
3391
3392         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3393                 return ZONE_RECLAIM_NOSCAN;
3394
3395         ret = __zone_reclaim(zone, gfp_mask, order);
3396         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3397
3398         if (!ret)
3399                 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3400
3401         return ret;
3402 }
3403 #endif
3404
3405 /*
3406  * page_evictable - test whether a page is evictable
3407  * @page: the page to test
3408  * @vma: the VMA in which the page is or will be mapped, may be NULL
3409  *
3410  * Test whether page is evictable--i.e., should be placed on active/inactive
3411  * lists vs unevictable list.  The vma argument is !NULL when called from the
3412  * fault path to determine how to instantate a new page.
3413  *
3414  * Reasons page might not be evictable:
3415  * (1) page's mapping marked unevictable
3416  * (2) page is part of an mlocked VMA
3417  *
3418  */
3419 int page_evictable(struct page *page, struct vm_area_struct *vma)
3420 {
3421
3422         if (mapping_unevictable(page_mapping(page)))
3423                 return 0;
3424
3425         if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
3426                 return 0;
3427
3428         return 1;
3429 }
3430
3431 /**
3432  * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
3433  * @page: page to check evictability and move to appropriate lru list
3434  * @zone: zone page is in
3435  *
3436  * Checks a page for evictability and moves the page to the appropriate
3437  * zone lru list.
3438  *
3439  * Restrictions: zone->lru_lock must be held, page must be on LRU and must
3440  * have PageUnevictable set.
3441  */
3442 static void check_move_unevictable_page(struct page *page, struct zone *zone)
3443 {
3444         struct lruvec *lruvec;
3445
3446         VM_BUG_ON(PageActive(page));
3447 retry:
3448         ClearPageUnevictable(page);
3449         if (page_evictable(page, NULL)) {
3450                 enum lru_list l = page_lru_base_type(page);
3451
3452                 __dec_zone_state(zone, NR_UNEVICTABLE);
3453                 lruvec = mem_cgroup_lru_move_lists(zone, page,
3454                                                    LRU_UNEVICTABLE, l);
3455                 list_move(&page->lru, &lruvec->lists[l]);
3456                 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
3457                 __count_vm_event(UNEVICTABLE_PGRESCUED);
3458         } else {
3459                 /*
3460                  * rotate unevictable list
3461                  */
3462                 SetPageUnevictable(page);
3463                 lruvec = mem_cgroup_lru_move_lists(zone, page, LRU_UNEVICTABLE,
3464                                                    LRU_UNEVICTABLE);
3465                 list_move(&page->lru, &lruvec->lists[LRU_UNEVICTABLE]);
3466                 if (page_evictable(page, NULL))
3467                         goto retry;
3468         }
3469 }
3470
3471 /**
3472  * scan_mapping_unevictable_pages - scan an address space for evictable pages
3473  * @mapping: struct address_space to scan for evictable pages
3474  *
3475  * Scan all pages in mapping.  Check unevictable pages for
3476  * evictability and move them to the appropriate zone lru list.
3477  */
3478 void scan_mapping_unevictable_pages(struct address_space *mapping)
3479 {
3480         pgoff_t next = 0;
3481         pgoff_t end   = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
3482                          PAGE_CACHE_SHIFT;
3483         struct zone *zone;
3484         struct pagevec pvec;
3485
3486         if (mapping->nrpages == 0)
3487                 return;
3488
3489         pagevec_init(&pvec, 0);
3490         while (next < end &&
3491                 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
3492                 int i;
3493                 int pg_scanned = 0;
3494
3495                 zone = NULL;
3496
3497                 for (i = 0; i < pagevec_count(&pvec); i++) {
3498                         struct page *page = pvec.pages[i];
3499                         pgoff_t page_index = page->index;
3500                         struct zone *pagezone = page_zone(page);
3501
3502                         pg_scanned++;
3503                         if (page_index > next)
3504                                 next = page_index;
3505                         next++;
3506
3507                         if (pagezone != zone) {
3508                                 if (zone)
3509                                         spin_unlock_irq(&zone->lru_lock);
3510                                 zone = pagezone;
3511                                 spin_lock_irq(&zone->lru_lock);
3512                         }
3513
3514                         if (PageLRU(page) && PageUnevictable(page))
3515                                 check_move_unevictable_page(page, zone);
3516                 }
3517                 if (zone)
3518                         spin_unlock_irq(&zone->lru_lock);
3519                 pagevec_release(&pvec);
3520
3521                 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
3522         }
3523
3524 }
3525
3526 static void warn_scan_unevictable_pages(void)
3527 {
3528         printk_once(KERN_WARNING
3529                     "%s: The scan_unevictable_pages sysctl/node-interface has been "
3530                     "disabled for lack of a legitimate use case.  If you have "
3531                     "one, please send an email to linux-mm@kvack.org.\n",
3532                     current->comm);
3533 }
3534
3535 /*
3536  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
3537  * all nodes' unevictable lists for evictable pages
3538  */
3539 unsigned long scan_unevictable_pages;
3540
3541 int scan_unevictable_handler(struct ctl_table *table, int write,
3542                            void __user *buffer,
3543                            size_t *length, loff_t *ppos)
3544 {
3545         warn_scan_unevictable_pages();
3546         proc_doulongvec_minmax(table, write, buffer, length, ppos);
3547         scan_unevictable_pages = 0;
3548         return 0;
3549 }
3550
3551 #ifdef CONFIG_NUMA
3552 /*
3553  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
3554  * a specified node's per zone unevictable lists for evictable pages.
3555  */
3556
3557 static ssize_t read_scan_unevictable_node(struct device *dev,
3558                                           struct device_attribute *attr,
3559                                           char *buf)
3560 {
3561         warn_scan_unevictable_pages();
3562         return sprintf(buf, "0\n");     /* always zero; should fit... */
3563 }
3564
3565 static ssize_t write_scan_unevictable_node(struct device *dev,
3566                                            struct device_attribute *attr,
3567                                         const char *buf, size_t count)
3568 {
3569         warn_scan_unevictable_pages();
3570         return 1;
3571 }
3572
3573
3574 static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3575                         read_scan_unevictable_node,
3576                         write_scan_unevictable_node);
3577
3578 int scan_unevictable_register_node(struct node *node)
3579 {
3580         return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
3581 }
3582
3583 void scan_unevictable_unregister_node(struct node *node)
3584 {
3585         device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);
3586 }
3587 #endif