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