vmscan: count pages into balanced for zone with good watermark
[linux-2.6.git] / mm / vmscan.c
1 /*
2  *  linux/mm/vmscan.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  *
6  *  Swap reorganised 29.12.95, Stephen Tweedie.
7  *  kswapd added: 7.1.96  sct
8  *  Removed kswapd_ctl limits, and swap out as many pages as needed
9  *  to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10  *  Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11  *  Multiqueue VM started 5.8.00, Rik van Riel.
12  */
13
14 #include <linux/mm.h>
15 #include <linux/module.h>
16 #include <linux/gfp.h>
17 #include <linux/kernel_stat.h>
18 #include <linux/swap.h>
19 #include <linux/pagemap.h>
20 #include <linux/init.h>
21 #include <linux/highmem.h>
22 #include <linux/vmstat.h>
23 #include <linux/file.h>
24 #include <linux/writeback.h>
25 #include <linux/blkdev.h>
26 #include <linux/buffer_head.h>  /* for try_to_release_page(),
27                                         buffer_heads_over_limit */
28 #include <linux/mm_inline.h>
29 #include <linux/pagevec.h>
30 #include <linux/backing-dev.h>
31 #include <linux/rmap.h>
32 #include <linux/topology.h>
33 #include <linux/cpu.h>
34 #include <linux/cpuset.h>
35 #include <linux/compaction.h>
36 #include <linux/notifier.h>
37 #include <linux/rwsem.h>
38 #include <linux/delay.h>
39 #include <linux/kthread.h>
40 #include <linux/freezer.h>
41 #include <linux/memcontrol.h>
42 #include <linux/delayacct.h>
43 #include <linux/sysctl.h>
44 #include <linux/oom.h>
45 #include <linux/prefetch.h>
46
47 #include <asm/tlbflush.h>
48 #include <asm/div64.h>
49
50 #include <linux/swapops.h>
51
52 #include "internal.h"
53
54 #define CREATE_TRACE_POINTS
55 #include <trace/events/vmscan.h>
56
57 /*
58  * reclaim_mode determines how the inactive list is shrunk
59  * RECLAIM_MODE_SINGLE: Reclaim only order-0 pages
60  * RECLAIM_MODE_ASYNC:  Do not block
61  * RECLAIM_MODE_SYNC:   Allow blocking e.g. call wait_on_page_writeback
62  * RECLAIM_MODE_LUMPYRECLAIM: For high-order allocations, take a reference
63  *                      page from the LRU and reclaim all pages within a
64  *                      naturally aligned range
65  * RECLAIM_MODE_COMPACTION: For high-order allocations, reclaim a number of
66  *                      order-0 pages and then compact the zone
67  */
68 typedef unsigned __bitwise__ reclaim_mode_t;
69 #define RECLAIM_MODE_SINGLE             ((__force reclaim_mode_t)0x01u)
70 #define RECLAIM_MODE_ASYNC              ((__force reclaim_mode_t)0x02u)
71 #define RECLAIM_MODE_SYNC               ((__force reclaim_mode_t)0x04u)
72 #define RECLAIM_MODE_LUMPYRECLAIM       ((__force reclaim_mode_t)0x08u)
73 #define RECLAIM_MODE_COMPACTION         ((__force reclaim_mode_t)0x10u)
74
75 struct scan_control {
76         /* Incremented by the number of inactive pages that were scanned */
77         unsigned long nr_scanned;
78
79         /* Number of pages freed so far during a call to shrink_zones() */
80         unsigned long nr_reclaimed;
81
82         /* How many pages shrink_list() should reclaim */
83         unsigned long nr_to_reclaim;
84
85         unsigned long hibernation_mode;
86
87         /* This context's GFP mask */
88         gfp_t gfp_mask;
89
90         int may_writepage;
91
92         /* Can mapped pages be reclaimed? */
93         int may_unmap;
94
95         /* Can pages be swapped as part of reclaim? */
96         int may_swap;
97
98         int order;
99
100         /*
101          * Intend to reclaim enough continuous memory rather than reclaim
102          * enough amount of memory. i.e, mode for high order allocation.
103          */
104         reclaim_mode_t reclaim_mode;
105
106         /* 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         shrinker->nr = 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                 unsigned long total_scan;
251                 unsigned 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                 /*
259                  * copy the current shrinker scan count into a local variable
260                  * and zero it so that other concurrent shrinker invocations
261                  * don't also do this scanning work.
262                  */
263                 do {
264                         nr = shrinker->nr;
265                 } while (cmpxchg(&shrinker->nr, nr, 0) != nr);
266
267                 total_scan = nr;
268                 max_pass = do_shrinker_shrink(shrinker, shrink, 0);
269                 delta = (4 * nr_pages_scanned) / shrinker->seeks;
270                 delta *= max_pass;
271                 do_div(delta, lru_pages + 1);
272                 total_scan += delta;
273                 if (total_scan < 0) {
274                         printk(KERN_ERR "shrink_slab: %pF negative objects to "
275                                "delete nr=%ld\n",
276                                shrinker->shrink, total_scan);
277                         total_scan = max_pass;
278                 }
279
280                 /*
281                  * We need to avoid excessive windup on filesystem shrinkers
282                  * due to large numbers of GFP_NOFS allocations causing the
283                  * shrinkers to return -1 all the time. This results in a large
284                  * nr being built up so when a shrink that can do some work
285                  * comes along it empties the entire cache due to nr >>>
286                  * max_pass.  This is bad for sustaining a working set in
287                  * memory.
288                  *
289                  * Hence only allow the shrinker to scan the entire cache when
290                  * a large delta change is calculated directly.
291                  */
292                 if (delta < max_pass / 4)
293                         total_scan = min(total_scan, max_pass / 2);
294
295                 /*
296                  * Avoid risking looping forever due to too large nr value:
297                  * never try to free more than twice the estimate number of
298                  * freeable entries.
299                  */
300                 if (total_scan > max_pass * 2)
301                         total_scan = max_pass * 2;
302
303                 trace_mm_shrink_slab_start(shrinker, shrink, nr,
304                                         nr_pages_scanned, lru_pages,
305                                         max_pass, delta, total_scan);
306
307                 while (total_scan >= batch_size) {
308                         int nr_before;
309
310                         nr_before = do_shrinker_shrink(shrinker, shrink, 0);
311                         shrink_ret = do_shrinker_shrink(shrinker, shrink,
312                                                         batch_size);
313                         if (shrink_ret == -1)
314                                 break;
315                         if (shrink_ret < nr_before)
316                                 ret += nr_before - shrink_ret;
317                         count_vm_events(SLABS_SCANNED, batch_size);
318                         total_scan -= batch_size;
319
320                         cond_resched();
321                 }
322
323                 /*
324                  * move the unused scan count back into the shrinker in a
325                  * manner that handles concurrent updates. If we exhausted the
326                  * scan, there is no need to do an update.
327                  */
328                 do {
329                         nr = shrinker->nr;
330                         new_nr = total_scan + nr;
331                         if (total_scan <= 0)
332                                 break;
333                 } while (cmpxchg(&shrinker->nr, nr, new_nr) != nr);
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 clearing (page is
637                  * unlocked), make sure that if the other thread does
638                  * not observe our setting of PG_lru and fails
639                  * isolation, we see PG_mlocked cleared below and move
640                  * the page back to the evictable list.
641                  *
642                  * The other side is TestClearPageMlocked().
643                  */
644                 smp_mb();
645         }
646
647         /*
648          * page's status can change while we move it among lru. If an evictable
649          * page is on unevictable list, it never be freed. To avoid that,
650          * check after we added it to the list, again.
651          */
652         if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
653                 if (!isolate_lru_page(page)) {
654                         put_page(page);
655                         goto redo;
656                 }
657                 /* This means someone else dropped this page from LRU
658                  * So, it will be freed or putback to LRU again. There is
659                  * nothing to do here.
660                  */
661         }
662
663         if (was_unevictable && lru != LRU_UNEVICTABLE)
664                 count_vm_event(UNEVICTABLE_PGRESCUED);
665         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
666                 count_vm_event(UNEVICTABLE_PGCULLED);
667
668         put_page(page);         /* drop ref from isolate */
669 }
670
671 enum page_references {
672         PAGEREF_RECLAIM,
673         PAGEREF_RECLAIM_CLEAN,
674         PAGEREF_KEEP,
675         PAGEREF_ACTIVATE,
676 };
677
678 static enum page_references page_check_references(struct page *page,
679                                                   struct scan_control *sc)
680 {
681         int referenced_ptes, referenced_page;
682         unsigned long vm_flags;
683
684         referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
685         referenced_page = TestClearPageReferenced(page);
686
687         /* Lumpy reclaim - ignore references */
688         if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
689                 return PAGEREF_RECLAIM;
690
691         /*
692          * Mlock lost the isolation race with us.  Let try_to_unmap()
693          * move the page to the unevictable list.
694          */
695         if (vm_flags & VM_LOCKED)
696                 return PAGEREF_RECLAIM;
697
698         if (referenced_ptes) {
699                 if (PageAnon(page))
700                         return PAGEREF_ACTIVATE;
701                 /*
702                  * All mapped pages start out with page table
703                  * references from the instantiating fault, so we need
704                  * to look twice if a mapped file page is used more
705                  * than once.
706                  *
707                  * Mark it and spare it for another trip around the
708                  * inactive list.  Another page table reference will
709                  * lead to its activation.
710                  *
711                  * Note: the mark is set for activated pages as well
712                  * so that recently deactivated but used pages are
713                  * quickly recovered.
714                  */
715                 SetPageReferenced(page);
716
717                 if (referenced_page)
718                         return PAGEREF_ACTIVATE;
719
720                 return PAGEREF_KEEP;
721         }
722
723         /* Reclaim if clean, defer dirty pages to writeback */
724         if (referenced_page && !PageSwapBacked(page))
725                 return PAGEREF_RECLAIM_CLEAN;
726
727         return PAGEREF_RECLAIM;
728 }
729
730 static noinline_for_stack void free_page_list(struct list_head *free_pages)
731 {
732         struct pagevec freed_pvec;
733         struct page *page, *tmp;
734
735         pagevec_init(&freed_pvec, 1);
736
737         list_for_each_entry_safe(page, tmp, free_pages, lru) {
738                 list_del(&page->lru);
739                 if (!pagevec_add(&freed_pvec, page)) {
740                         __pagevec_free(&freed_pvec);
741                         pagevec_reinit(&freed_pvec);
742                 }
743         }
744
745         pagevec_free(&freed_pvec);
746 }
747
748 /*
749  * shrink_page_list() returns the number of reclaimed pages
750  */
751 static unsigned long shrink_page_list(struct list_head *page_list,
752                                       struct zone *zone,
753                                       struct scan_control *sc,
754                                       int priority,
755                                       unsigned long *ret_nr_dirty,
756                                       unsigned long *ret_nr_writeback)
757 {
758         LIST_HEAD(ret_pages);
759         LIST_HEAD(free_pages);
760         int pgactivate = 0;
761         unsigned long nr_dirty = 0;
762         unsigned long nr_congested = 0;
763         unsigned long nr_reclaimed = 0;
764         unsigned long nr_writeback = 0;
765
766         cond_resched();
767
768         while (!list_empty(page_list)) {
769                 enum page_references references;
770                 struct address_space *mapping;
771                 struct page *page;
772                 int may_enter_fs;
773
774                 cond_resched();
775
776                 page = lru_to_page(page_list);
777                 list_del(&page->lru);
778
779                 if (!trylock_page(page))
780                         goto keep;
781
782                 VM_BUG_ON(PageActive(page));
783                 VM_BUG_ON(page_zone(page) != zone);
784
785                 sc->nr_scanned++;
786
787                 if (unlikely(!page_evictable(page, NULL)))
788                         goto cull_mlocked;
789
790                 if (!sc->may_unmap && page_mapped(page))
791                         goto keep_locked;
792
793                 /* Double the slab pressure for mapped and swapcache pages */
794                 if (page_mapped(page) || PageSwapCache(page))
795                         sc->nr_scanned++;
796
797                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
798                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
799
800                 if (PageWriteback(page)) {
801                         nr_writeback++;
802                         /*
803                          * Synchronous reclaim cannot queue pages for
804                          * writeback due to the possibility of stack overflow
805                          * but if it encounters a page under writeback, wait
806                          * for the IO to complete.
807                          */
808                         if ((sc->reclaim_mode & RECLAIM_MODE_SYNC) &&
809                             may_enter_fs)
810                                 wait_on_page_writeback(page);
811                         else {
812                                 unlock_page(page);
813                                 goto keep_lumpy;
814                         }
815                 }
816
817                 references = page_check_references(page, sc);
818                 switch (references) {
819                 case PAGEREF_ACTIVATE:
820                         goto activate_locked;
821                 case PAGEREF_KEEP:
822                         goto keep_locked;
823                 case PAGEREF_RECLAIM:
824                 case PAGEREF_RECLAIM_CLEAN:
825                         ; /* try to reclaim the page below */
826                 }
827
828                 /*
829                  * Anonymous process memory has backing store?
830                  * Try to allocate it some swap space here.
831                  */
832                 if (PageAnon(page) && !PageSwapCache(page)) {
833                         if (!(sc->gfp_mask & __GFP_IO))
834                                 goto keep_locked;
835                         if (!add_to_swap(page))
836                                 goto activate_locked;
837                         may_enter_fs = 1;
838                 }
839
840                 mapping = page_mapping(page);
841
842                 /*
843                  * The page is mapped into the page tables of one or more
844                  * processes. Try to unmap it here.
845                  */
846                 if (page_mapped(page) && mapping) {
847                         switch (try_to_unmap(page, TTU_UNMAP)) {
848                         case SWAP_FAIL:
849                                 goto activate_locked;
850                         case SWAP_AGAIN:
851                                 goto keep_locked;
852                         case SWAP_MLOCK:
853                                 goto cull_mlocked;
854                         case SWAP_SUCCESS:
855                                 ; /* try to free the page below */
856                         }
857                 }
858
859                 if (PageDirty(page)) {
860                         nr_dirty++;
861
862                         /*
863                          * Only kswapd can writeback filesystem pages to
864                          * avoid risk of stack overflow but do not writeback
865                          * unless under significant pressure.
866                          */
867                         if (page_is_file_cache(page) &&
868                                         (!current_is_kswapd() || priority >= DEF_PRIORITY - 2)) {
869                                 /*
870                                  * Immediately reclaim when written back.
871                                  * Similar in principal to deactivate_page()
872                                  * except we already have the page isolated
873                                  * and know it's dirty
874                                  */
875                                 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
876                                 SetPageReclaim(page);
877
878                                 goto keep_locked;
879                         }
880
881                         if (references == PAGEREF_RECLAIM_CLEAN)
882                                 goto keep_locked;
883                         if (!may_enter_fs)
884                                 goto keep_locked;
885                         if (!sc->may_writepage)
886                                 goto keep_locked;
887
888                         /* Page is dirty, try to write it out here */
889                         switch (pageout(page, mapping, sc)) {
890                         case PAGE_KEEP:
891                                 nr_congested++;
892                                 goto keep_locked;
893                         case PAGE_ACTIVATE:
894                                 goto activate_locked;
895                         case PAGE_SUCCESS:
896                                 if (PageWriteback(page))
897                                         goto keep_lumpy;
898                                 if (PageDirty(page))
899                                         goto keep;
900
901                                 /*
902                                  * A synchronous write - probably a ramdisk.  Go
903                                  * ahead and try to reclaim the page.
904                                  */
905                                 if (!trylock_page(page))
906                                         goto keep;
907                                 if (PageDirty(page) || PageWriteback(page))
908                                         goto keep_locked;
909                                 mapping = page_mapping(page);
910                         case PAGE_CLEAN:
911                                 ; /* try to free the page below */
912                         }
913                 }
914
915                 /*
916                  * If the page has buffers, try to free the buffer mappings
917                  * associated with this page. If we succeed we try to free
918                  * the page as well.
919                  *
920                  * We do this even if the page is PageDirty().
921                  * try_to_release_page() does not perform I/O, but it is
922                  * possible for a page to have PageDirty set, but it is actually
923                  * clean (all its buffers are clean).  This happens if the
924                  * buffers were written out directly, with submit_bh(). ext3
925                  * will do this, as well as the blockdev mapping.
926                  * try_to_release_page() will discover that cleanness and will
927                  * drop the buffers and mark the page clean - it can be freed.
928                  *
929                  * Rarely, pages can have buffers and no ->mapping.  These are
930                  * the pages which were not successfully invalidated in
931                  * truncate_complete_page().  We try to drop those buffers here
932                  * and if that worked, and the page is no longer mapped into
933                  * process address space (page_count == 1) it can be freed.
934                  * Otherwise, leave the page on the LRU so it is swappable.
935                  */
936                 if (page_has_private(page)) {
937                         if (!try_to_release_page(page, sc->gfp_mask))
938                                 goto activate_locked;
939                         if (!mapping && page_count(page) == 1) {
940                                 unlock_page(page);
941                                 if (put_page_testzero(page))
942                                         goto free_it;
943                                 else {
944                                         /*
945                                          * rare race with speculative reference.
946                                          * the speculative reference will free
947                                          * this page shortly, so we may
948                                          * increment nr_reclaimed here (and
949                                          * leave it off the LRU).
950                                          */
951                                         nr_reclaimed++;
952                                         continue;
953                                 }
954                         }
955                 }
956
957                 if (!mapping || !__remove_mapping(mapping, page))
958                         goto keep_locked;
959
960                 /*
961                  * At this point, we have no other references and there is
962                  * no way to pick any more up (removed from LRU, removed
963                  * from pagecache). Can use non-atomic bitops now (and
964                  * we obviously don't have to worry about waking up a process
965                  * waiting on the page lock, because there are no references.
966                  */
967                 __clear_page_locked(page);
968 free_it:
969                 nr_reclaimed++;
970
971                 /*
972                  * Is there need to periodically free_page_list? It would
973                  * appear not as the counts should be low
974                  */
975                 list_add(&page->lru, &free_pages);
976                 continue;
977
978 cull_mlocked:
979                 if (PageSwapCache(page))
980                         try_to_free_swap(page);
981                 unlock_page(page);
982                 putback_lru_page(page);
983                 reset_reclaim_mode(sc);
984                 continue;
985
986 activate_locked:
987                 /* Not a candidate for swapping, so reclaim swap space. */
988                 if (PageSwapCache(page) && vm_swap_full())
989                         try_to_free_swap(page);
990                 VM_BUG_ON(PageActive(page));
991                 SetPageActive(page);
992                 pgactivate++;
993 keep_locked:
994                 unlock_page(page);
995 keep:
996                 reset_reclaim_mode(sc);
997 keep_lumpy:
998                 list_add(&page->lru, &ret_pages);
999                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
1000         }
1001
1002         /*
1003          * Tag a zone as congested if all the dirty pages encountered were
1004          * backed by a congested BDI. In this case, reclaimers should just
1005          * back off and wait for congestion to clear because further reclaim
1006          * will encounter the same problem
1007          */
1008         if (nr_dirty && nr_dirty == nr_congested && scanning_global_lru(sc))
1009                 zone_set_flag(zone, ZONE_CONGESTED);
1010
1011         free_page_list(&free_pages);
1012
1013         list_splice(&ret_pages, page_list);
1014         count_vm_events(PGACTIVATE, pgactivate);
1015         *ret_nr_dirty += nr_dirty;
1016         *ret_nr_writeback += nr_writeback;
1017         return nr_reclaimed;
1018 }
1019
1020 /*
1021  * Attempt to remove the specified page from its LRU.  Only take this page
1022  * if it is of the appropriate PageActive status.  Pages which are being
1023  * freed elsewhere are also ignored.
1024  *
1025  * page:        page to consider
1026  * mode:        one of the LRU isolation modes defined above
1027  *
1028  * returns 0 on success, -ve errno on failure.
1029  */
1030 int __isolate_lru_page(struct page *page, isolate_mode_t mode, int file)
1031 {
1032         bool all_lru_mode;
1033         int ret = -EINVAL;
1034
1035         /* Only take pages on the LRU. */
1036         if (!PageLRU(page))
1037                 return ret;
1038
1039         all_lru_mode = (mode & (ISOLATE_ACTIVE|ISOLATE_INACTIVE)) ==
1040                 (ISOLATE_ACTIVE|ISOLATE_INACTIVE);
1041
1042         /*
1043          * When checking the active state, we need to be sure we are
1044          * dealing with comparible boolean values.  Take the logical not
1045          * of each.
1046          */
1047         if (!all_lru_mode && !PageActive(page) != !(mode & ISOLATE_ACTIVE))
1048                 return ret;
1049
1050         if (!all_lru_mode && !!page_is_file_cache(page) != file)
1051                 return ret;
1052
1053         /*
1054          * When this function is being called for lumpy reclaim, we
1055          * initially look into all LRU pages, active, inactive and
1056          * unevictable; only give shrink_page_list evictable pages.
1057          */
1058         if (PageUnevictable(page))
1059                 return ret;
1060
1061         ret = -EBUSY;
1062
1063         if ((mode & ISOLATE_CLEAN) && (PageDirty(page) || PageWriteback(page)))
1064                 return ret;
1065
1066         if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1067                 return ret;
1068
1069         if (likely(get_page_unless_zero(page))) {
1070                 /*
1071                  * Be careful not to clear PageLRU until after we're
1072                  * sure the page is not being freed elsewhere -- the
1073                  * page release code relies on it.
1074                  */
1075                 ClearPageLRU(page);
1076                 ret = 0;
1077         }
1078
1079         return ret;
1080 }
1081
1082 /*
1083  * zone->lru_lock is heavily contended.  Some of the functions that
1084  * shrink the lists perform better by taking out a batch of pages
1085  * and working on them outside the LRU lock.
1086  *
1087  * For pagecache intensive workloads, this function is the hottest
1088  * spot in the kernel (apart from copy_*_user functions).
1089  *
1090  * Appropriate locks must be held before calling this function.
1091  *
1092  * @nr_to_scan: The number of pages to look through on the list.
1093  * @src:        The LRU list to pull pages off.
1094  * @dst:        The temp list to put pages on to.
1095  * @scanned:    The number of pages that were scanned.
1096  * @order:      The caller's attempted allocation order
1097  * @mode:       One of the LRU isolation modes
1098  * @file:       True [1] if isolating file [!anon] pages
1099  *
1100  * returns how many pages were moved onto *@dst.
1101  */
1102 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1103                 struct list_head *src, struct list_head *dst,
1104                 unsigned long *scanned, int order, isolate_mode_t mode,
1105                 int file)
1106 {
1107         unsigned long nr_taken = 0;
1108         unsigned long nr_lumpy_taken = 0;
1109         unsigned long nr_lumpy_dirty = 0;
1110         unsigned long nr_lumpy_failed = 0;
1111         unsigned long scan;
1112
1113         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1114                 struct page *page;
1115                 unsigned long pfn;
1116                 unsigned long end_pfn;
1117                 unsigned long page_pfn;
1118                 int zone_id;
1119
1120                 page = lru_to_page(src);
1121                 prefetchw_prev_lru_page(page, src, flags);
1122
1123                 VM_BUG_ON(!PageLRU(page));
1124
1125                 switch (__isolate_lru_page(page, mode, file)) {
1126                 case 0:
1127                         list_move(&page->lru, dst);
1128                         mem_cgroup_del_lru(page);
1129                         nr_taken += hpage_nr_pages(page);
1130                         break;
1131
1132                 case -EBUSY:
1133                         /* else it is being freed elsewhere */
1134                         list_move(&page->lru, src);
1135                         mem_cgroup_rotate_lru_list(page, page_lru(page));
1136                         continue;
1137
1138                 default:
1139                         BUG();
1140                 }
1141
1142                 if (!order)
1143                         continue;
1144
1145                 /*
1146                  * Attempt to take all pages in the order aligned region
1147                  * surrounding the tag page.  Only take those pages of
1148                  * the same active state as that tag page.  We may safely
1149                  * round the target page pfn down to the requested order
1150                  * as the mem_map is guaranteed valid out to MAX_ORDER,
1151                  * where that page is in a different zone we will detect
1152                  * it from its zone id and abort this block scan.
1153                  */
1154                 zone_id = page_zone_id(page);
1155                 page_pfn = page_to_pfn(page);
1156                 pfn = page_pfn & ~((1 << order) - 1);
1157                 end_pfn = pfn + (1 << order);
1158                 for (; pfn < end_pfn; pfn++) {
1159                         struct page *cursor_page;
1160
1161                         /* The target page is in the block, ignore it. */
1162                         if (unlikely(pfn == page_pfn))
1163                                 continue;
1164
1165                         /* Avoid holes within the zone. */
1166                         if (unlikely(!pfn_valid_within(pfn)))
1167                                 break;
1168
1169                         cursor_page = pfn_to_page(pfn);
1170
1171                         /* Check that we have not crossed a zone boundary. */
1172                         if (unlikely(page_zone_id(cursor_page) != zone_id))
1173                                 break;
1174
1175                         /*
1176                          * If we don't have enough swap space, reclaiming of
1177                          * anon page which don't already have a swap slot is
1178                          * pointless.
1179                          */
1180                         if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
1181                             !PageSwapCache(cursor_page))
1182                                 break;
1183
1184                         if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1185                                 list_move(&cursor_page->lru, dst);
1186                                 mem_cgroup_del_lru(cursor_page);
1187                                 nr_taken += hpage_nr_pages(page);
1188                                 nr_lumpy_taken++;
1189                                 if (PageDirty(cursor_page))
1190                                         nr_lumpy_dirty++;
1191                                 scan++;
1192                         } else {
1193                                 /*
1194                                  * Check if the page is freed already.
1195                                  *
1196                                  * We can't use page_count() as that
1197                                  * requires compound_head and we don't
1198                                  * have a pin on the page here. If a
1199                                  * page is tail, we may or may not
1200                                  * have isolated the head, so assume
1201                                  * it's not free, it'd be tricky to
1202                                  * track the head status without a
1203                                  * page pin.
1204                                  */
1205                                 if (!PageTail(cursor_page) &&
1206                                     !atomic_read(&cursor_page->_count))
1207                                         continue;
1208                                 break;
1209                         }
1210                 }
1211
1212                 /* If we break out of the loop above, lumpy reclaim failed */
1213                 if (pfn < end_pfn)
1214                         nr_lumpy_failed++;
1215         }
1216
1217         *scanned = scan;
1218
1219         trace_mm_vmscan_lru_isolate(order,
1220                         nr_to_scan, scan,
1221                         nr_taken,
1222                         nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1223                         mode);
1224         return nr_taken;
1225 }
1226
1227 static unsigned long isolate_pages_global(unsigned long nr,
1228                                         struct list_head *dst,
1229                                         unsigned long *scanned, int order,
1230                                         isolate_mode_t mode,
1231                                         struct zone *z, int active, int file)
1232 {
1233         int lru = LRU_BASE;
1234         if (active)
1235                 lru += LRU_ACTIVE;
1236         if (file)
1237                 lru += LRU_FILE;
1238         return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1239                                                                 mode, file);
1240 }
1241
1242 /*
1243  * clear_active_flags() is a helper for shrink_active_list(), clearing
1244  * any active bits from the pages in the list.
1245  */
1246 static unsigned long clear_active_flags(struct list_head *page_list,
1247                                         unsigned int *count)
1248 {
1249         int nr_active = 0;
1250         int lru;
1251         struct page *page;
1252
1253         list_for_each_entry(page, page_list, lru) {
1254                 int numpages = hpage_nr_pages(page);
1255                 lru = page_lru_base_type(page);
1256                 if (PageActive(page)) {
1257                         lru += LRU_ACTIVE;
1258                         ClearPageActive(page);
1259                         nr_active += numpages;
1260                 }
1261                 if (count)
1262                         count[lru] += numpages;
1263         }
1264
1265         return nr_active;
1266 }
1267
1268 /**
1269  * isolate_lru_page - tries to isolate a page from its LRU list
1270  * @page: page to isolate from its LRU list
1271  *
1272  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1273  * vmstat statistic corresponding to whatever LRU list the page was on.
1274  *
1275  * Returns 0 if the page was removed from an LRU list.
1276  * Returns -EBUSY if the page was not on an LRU list.
1277  *
1278  * The returned page will have PageLRU() cleared.  If it was found on
1279  * the active list, it will have PageActive set.  If it was found on
1280  * the unevictable list, it will have the PageUnevictable bit set. That flag
1281  * may need to be cleared by the caller before letting the page go.
1282  *
1283  * The vmstat statistic corresponding to the list on which the page was
1284  * found will be decremented.
1285  *
1286  * Restrictions:
1287  * (1) Must be called with an elevated refcount on the page. This is a
1288  *     fundamentnal difference from isolate_lru_pages (which is called
1289  *     without a stable reference).
1290  * (2) the lru_lock must not be held.
1291  * (3) interrupts must be enabled.
1292  */
1293 int isolate_lru_page(struct page *page)
1294 {
1295         int ret = -EBUSY;
1296
1297         VM_BUG_ON(!page_count(page));
1298
1299         if (PageLRU(page)) {
1300                 struct zone *zone = page_zone(page);
1301
1302                 spin_lock_irq(&zone->lru_lock);
1303                 if (PageLRU(page)) {
1304                         int lru = page_lru(page);
1305                         ret = 0;
1306                         get_page(page);
1307                         ClearPageLRU(page);
1308
1309                         del_page_from_lru_list(zone, page, lru);
1310                 }
1311                 spin_unlock_irq(&zone->lru_lock);
1312         }
1313         return ret;
1314 }
1315
1316 /*
1317  * Are there way too many processes in the direct reclaim path already?
1318  */
1319 static int too_many_isolated(struct zone *zone, int file,
1320                 struct scan_control *sc)
1321 {
1322         unsigned long inactive, isolated;
1323
1324         if (current_is_kswapd())
1325                 return 0;
1326
1327         if (!scanning_global_lru(sc))
1328                 return 0;
1329
1330         if (file) {
1331                 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1332                 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1333         } else {
1334                 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1335                 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1336         }
1337
1338         return isolated > inactive;
1339 }
1340
1341 /*
1342  * TODO: Try merging with migrations version of putback_lru_pages
1343  */
1344 static noinline_for_stack void
1345 putback_lru_pages(struct zone *zone, struct scan_control *sc,
1346                                 unsigned long nr_anon, unsigned long nr_file,
1347                                 struct list_head *page_list)
1348 {
1349         struct page *page;
1350         struct pagevec pvec;
1351         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1352
1353         pagevec_init(&pvec, 1);
1354
1355         /*
1356          * Put back any unfreeable pages.
1357          */
1358         spin_lock(&zone->lru_lock);
1359         while (!list_empty(page_list)) {
1360                 int lru;
1361                 page = lru_to_page(page_list);
1362                 VM_BUG_ON(PageLRU(page));
1363                 list_del(&page->lru);
1364                 if (unlikely(!page_evictable(page, NULL))) {
1365                         spin_unlock_irq(&zone->lru_lock);
1366                         putback_lru_page(page);
1367                         spin_lock_irq(&zone->lru_lock);
1368                         continue;
1369                 }
1370                 SetPageLRU(page);
1371                 lru = page_lru(page);
1372                 add_page_to_lru_list(zone, page, lru);
1373                 if (is_active_lru(lru)) {
1374                         int file = is_file_lru(lru);
1375                         int numpages = hpage_nr_pages(page);
1376                         reclaim_stat->recent_rotated[file] += numpages;
1377                 }
1378                 if (!pagevec_add(&pvec, page)) {
1379                         spin_unlock_irq(&zone->lru_lock);
1380                         __pagevec_release(&pvec);
1381                         spin_lock_irq(&zone->lru_lock);
1382                 }
1383         }
1384         __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1385         __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1386
1387         spin_unlock_irq(&zone->lru_lock);
1388         pagevec_release(&pvec);
1389 }
1390
1391 static noinline_for_stack void update_isolated_counts(struct zone *zone,
1392                                         struct scan_control *sc,
1393                                         unsigned long *nr_anon,
1394                                         unsigned long *nr_file,
1395                                         struct list_head *isolated_list)
1396 {
1397         unsigned long nr_active;
1398         unsigned int count[NR_LRU_LISTS] = { 0, };
1399         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1400
1401         nr_active = clear_active_flags(isolated_list, count);
1402         __count_vm_events(PGDEACTIVATE, nr_active);
1403
1404         __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1405                               -count[LRU_ACTIVE_FILE]);
1406         __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1407                               -count[LRU_INACTIVE_FILE]);
1408         __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1409                               -count[LRU_ACTIVE_ANON]);
1410         __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1411                               -count[LRU_INACTIVE_ANON]);
1412
1413         *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1414         *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1415         __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1416         __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1417
1418         reclaim_stat->recent_scanned[0] += *nr_anon;
1419         reclaim_stat->recent_scanned[1] += *nr_file;
1420 }
1421
1422 /*
1423  * Returns true if a direct reclaim should wait on pages under writeback.
1424  *
1425  * If we are direct reclaiming for contiguous pages and we do not reclaim
1426  * everything in the list, try again and wait for writeback IO to complete.
1427  * This will stall high-order allocations noticeably. Only do that when really
1428  * need to free the pages under high memory pressure.
1429  */
1430 static inline bool should_reclaim_stall(unsigned long nr_taken,
1431                                         unsigned long nr_freed,
1432                                         int priority,
1433                                         struct scan_control *sc)
1434 {
1435         int lumpy_stall_priority;
1436
1437         /* kswapd should not stall on sync IO */
1438         if (current_is_kswapd())
1439                 return false;
1440
1441         /* Only stall on lumpy reclaim */
1442         if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
1443                 return false;
1444
1445         /* If we have reclaimed everything on the isolated list, no stall */
1446         if (nr_freed == nr_taken)
1447                 return false;
1448
1449         /*
1450          * For high-order allocations, there are two stall thresholds.
1451          * High-cost allocations stall immediately where as lower
1452          * order allocations such as stacks require the scanning
1453          * priority to be much higher before stalling.
1454          */
1455         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1456                 lumpy_stall_priority = DEF_PRIORITY;
1457         else
1458                 lumpy_stall_priority = DEF_PRIORITY / 3;
1459
1460         return priority <= lumpy_stall_priority;
1461 }
1462
1463 /*
1464  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1465  * of reclaimed pages
1466  */
1467 static noinline_for_stack unsigned long
1468 shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1469                         struct scan_control *sc, int priority, int file)
1470 {
1471         LIST_HEAD(page_list);
1472         unsigned long nr_scanned;
1473         unsigned long nr_reclaimed = 0;
1474         unsigned long nr_taken;
1475         unsigned long nr_anon;
1476         unsigned long nr_file;
1477         unsigned long nr_dirty = 0;
1478         unsigned long nr_writeback = 0;
1479         isolate_mode_t reclaim_mode = ISOLATE_INACTIVE;
1480
1481         while (unlikely(too_many_isolated(zone, file, sc))) {
1482                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1483
1484                 /* We are about to die and free our memory. Return now. */
1485                 if (fatal_signal_pending(current))
1486                         return SWAP_CLUSTER_MAX;
1487         }
1488
1489         set_reclaim_mode(priority, sc, false);
1490         if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
1491                 reclaim_mode |= ISOLATE_ACTIVE;
1492
1493         lru_add_drain();
1494
1495         if (!sc->may_unmap)
1496                 reclaim_mode |= ISOLATE_UNMAPPED;
1497         if (!sc->may_writepage)
1498                 reclaim_mode |= ISOLATE_CLEAN;
1499
1500         spin_lock_irq(&zone->lru_lock);
1501
1502         if (scanning_global_lru(sc)) {
1503                 nr_taken = isolate_pages_global(nr_to_scan, &page_list,
1504                         &nr_scanned, sc->order, reclaim_mode, zone, 0, file);
1505                 zone->pages_scanned += nr_scanned;
1506                 if (current_is_kswapd())
1507                         __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1508                                                nr_scanned);
1509                 else
1510                         __count_zone_vm_events(PGSCAN_DIRECT, zone,
1511                                                nr_scanned);
1512         } else {
1513                 nr_taken = mem_cgroup_isolate_pages(nr_to_scan, &page_list,
1514                         &nr_scanned, sc->order, reclaim_mode, zone,
1515                         sc->mem_cgroup, 0, file);
1516                 /*
1517                  * mem_cgroup_isolate_pages() keeps track of
1518                  * scanned pages on its own.
1519                  */
1520         }
1521
1522         if (nr_taken == 0) {
1523                 spin_unlock_irq(&zone->lru_lock);
1524                 return 0;
1525         }
1526
1527         update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
1528
1529         spin_unlock_irq(&zone->lru_lock);
1530
1531         nr_reclaimed = shrink_page_list(&page_list, zone, sc, priority,
1532                                                 &nr_dirty, &nr_writeback);
1533
1534         /* Check if we should syncronously wait for writeback */
1535         if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1536                 set_reclaim_mode(priority, sc, true);
1537                 nr_reclaimed += shrink_page_list(&page_list, zone, sc,
1538                                         priority, &nr_dirty, &nr_writeback);
1539         }
1540
1541         local_irq_disable();
1542         if (current_is_kswapd())
1543                 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1544         __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1545
1546         putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
1547
1548         /*
1549          * If reclaim is isolating dirty pages under writeback, it implies
1550          * that the long-lived page allocation rate is exceeding the page
1551          * laundering rate. Either the global limits are not being effective
1552          * at throttling processes due to the page distribution throughout
1553          * zones or there is heavy usage of a slow backing device. The
1554          * only option is to throttle from reclaim context which is not ideal
1555          * as there is no guarantee the dirtying process is throttled in the
1556          * same way balance_dirty_pages() manages.
1557          *
1558          * This scales the number of dirty pages that must be under writeback
1559          * before throttling depending on priority. It is a simple backoff
1560          * function that has the most effect in the range DEF_PRIORITY to
1561          * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1562          * in trouble and reclaim is considered to be in trouble.
1563          *
1564          * DEF_PRIORITY   100% isolated pages must be PageWriteback to throttle
1565          * DEF_PRIORITY-1  50% must be PageWriteback
1566          * DEF_PRIORITY-2  25% must be PageWriteback, kswapd in trouble
1567          * ...
1568          * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1569          *                     isolated page is PageWriteback
1570          */
1571         if (nr_writeback && nr_writeback >= (nr_taken >> (DEF_PRIORITY-priority)))
1572                 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1573
1574         trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1575                 zone_idx(zone),
1576                 nr_scanned, nr_reclaimed,
1577                 priority,
1578                 trace_shrink_flags(file, sc->reclaim_mode));
1579         return nr_reclaimed;
1580 }
1581
1582 /*
1583  * This moves pages from the active list to the inactive list.
1584  *
1585  * We move them the other way if the page is referenced by one or more
1586  * processes, from rmap.
1587  *
1588  * If the pages are mostly unmapped, the processing is fast and it is
1589  * appropriate to hold zone->lru_lock across the whole operation.  But if
1590  * the pages are mapped, the processing is slow (page_referenced()) so we
1591  * should drop zone->lru_lock around each page.  It's impossible to balance
1592  * this, so instead we remove the pages from the LRU while processing them.
1593  * It is safe to rely on PG_active against the non-LRU pages in here because
1594  * nobody will play with that bit on a non-LRU page.
1595  *
1596  * The downside is that we have to touch page->_count against each page.
1597  * But we had to alter page->flags anyway.
1598  */
1599
1600 static void move_active_pages_to_lru(struct zone *zone,
1601                                      struct list_head *list,
1602                                      enum lru_list lru)
1603 {
1604         unsigned long pgmoved = 0;
1605         struct pagevec pvec;
1606         struct page *page;
1607
1608         pagevec_init(&pvec, 1);
1609
1610         while (!list_empty(list)) {
1611                 page = lru_to_page(list);
1612
1613                 VM_BUG_ON(PageLRU(page));
1614                 SetPageLRU(page);
1615
1616                 list_move(&page->lru, &zone->lru[lru].list);
1617                 mem_cgroup_add_lru_list(page, lru);
1618                 pgmoved += hpage_nr_pages(page);
1619
1620                 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1621                         spin_unlock_irq(&zone->lru_lock);
1622                         if (buffer_heads_over_limit)
1623                                 pagevec_strip(&pvec);
1624                         __pagevec_release(&pvec);
1625                         spin_lock_irq(&zone->lru_lock);
1626                 }
1627         }
1628         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1629         if (!is_active_lru(lru))
1630                 __count_vm_events(PGDEACTIVATE, pgmoved);
1631 }
1632
1633 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1634                         struct scan_control *sc, int priority, int file)
1635 {
1636         unsigned long nr_taken;
1637         unsigned long pgscanned;
1638         unsigned long vm_flags;
1639         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1640         LIST_HEAD(l_active);
1641         LIST_HEAD(l_inactive);
1642         struct page *page;
1643         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1644         unsigned long nr_rotated = 0;
1645         isolate_mode_t reclaim_mode = ISOLATE_ACTIVE;
1646
1647         lru_add_drain();
1648
1649         if (!sc->may_unmap)
1650                 reclaim_mode |= ISOLATE_UNMAPPED;
1651         if (!sc->may_writepage)
1652                 reclaim_mode |= ISOLATE_CLEAN;
1653
1654         spin_lock_irq(&zone->lru_lock);
1655         if (scanning_global_lru(sc)) {
1656                 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1657                                                 &pgscanned, sc->order,
1658                                                 reclaim_mode, zone,
1659                                                 1, file);
1660                 zone->pages_scanned += pgscanned;
1661         } else {
1662                 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1663                                                 &pgscanned, sc->order,
1664                                                 reclaim_mode, zone,
1665                                                 sc->mem_cgroup, 1, file);
1666                 /*
1667                  * mem_cgroup_isolate_pages() keeps track of
1668                  * scanned pages on its own.
1669                  */
1670         }
1671
1672         reclaim_stat->recent_scanned[file] += nr_taken;
1673
1674         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1675         if (file)
1676                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1677         else
1678                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1679         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1680         spin_unlock_irq(&zone->lru_lock);
1681
1682         while (!list_empty(&l_hold)) {
1683                 cond_resched();
1684                 page = lru_to_page(&l_hold);
1685                 list_del(&page->lru);
1686
1687                 if (unlikely(!page_evictable(page, NULL))) {
1688                         putback_lru_page(page);
1689                         continue;
1690                 }
1691
1692                 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1693                         nr_rotated += hpage_nr_pages(page);
1694                         /*
1695                          * Identify referenced, file-backed active pages and
1696                          * give them one more trip around the active list. So
1697                          * that executable code get better chances to stay in
1698                          * memory under moderate memory pressure.  Anon pages
1699                          * are not likely to be evicted by use-once streaming
1700                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1701                          * so we ignore them here.
1702                          */
1703                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1704                                 list_add(&page->lru, &l_active);
1705                                 continue;
1706                         }
1707                 }
1708
1709                 ClearPageActive(page);  /* we are de-activating */
1710                 list_add(&page->lru, &l_inactive);
1711         }
1712
1713         /*
1714          * Move pages back to the lru list.
1715          */
1716         spin_lock_irq(&zone->lru_lock);
1717         /*
1718          * Count referenced pages from currently used mappings as rotated,
1719          * even though only some of them are actually re-activated.  This
1720          * helps balance scan pressure between file and anonymous pages in
1721          * get_scan_ratio.
1722          */
1723         reclaim_stat->recent_rotated[file] += nr_rotated;
1724
1725         move_active_pages_to_lru(zone, &l_active,
1726                                                 LRU_ACTIVE + file * LRU_FILE);
1727         move_active_pages_to_lru(zone, &l_inactive,
1728                                                 LRU_BASE   + file * LRU_FILE);
1729         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1730         spin_unlock_irq(&zone->lru_lock);
1731 }
1732
1733 #ifdef CONFIG_SWAP
1734 static int inactive_anon_is_low_global(struct zone *zone)
1735 {
1736         unsigned long active, inactive;
1737
1738         active = zone_page_state(zone, NR_ACTIVE_ANON);
1739         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1740
1741         if (inactive * zone->inactive_ratio < active)
1742                 return 1;
1743
1744         return 0;
1745 }
1746
1747 /**
1748  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1749  * @zone: zone to check
1750  * @sc:   scan control of this context
1751  *
1752  * Returns true if the zone does not have enough inactive anon pages,
1753  * meaning some active anon pages need to be deactivated.
1754  */
1755 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1756 {
1757         int low;
1758
1759         /*
1760          * If we don't have swap space, anonymous page deactivation
1761          * is pointless.
1762          */
1763         if (!total_swap_pages)
1764                 return 0;
1765
1766         if (scanning_global_lru(sc))
1767                 low = inactive_anon_is_low_global(zone);
1768         else
1769                 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1770         return low;
1771 }
1772 #else
1773 static inline int inactive_anon_is_low(struct zone *zone,
1774                                         struct scan_control *sc)
1775 {
1776         return 0;
1777 }
1778 #endif
1779
1780 static int inactive_file_is_low_global(struct zone *zone)
1781 {
1782         unsigned long active, inactive;
1783
1784         active = zone_page_state(zone, NR_ACTIVE_FILE);
1785         inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1786
1787         return (active > inactive);
1788 }
1789
1790 /**
1791  * inactive_file_is_low - check if file pages need to be deactivated
1792  * @zone: zone to check
1793  * @sc:   scan control of this context
1794  *
1795  * When the system is doing streaming IO, memory pressure here
1796  * ensures that active file pages get deactivated, until more
1797  * than half of the file pages are on the inactive list.
1798  *
1799  * Once we get to that situation, protect the system's working
1800  * set from being evicted by disabling active file page aging.
1801  *
1802  * This uses a different ratio than the anonymous pages, because
1803  * the page cache uses a use-once replacement algorithm.
1804  */
1805 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1806 {
1807         int low;
1808
1809         if (scanning_global_lru(sc))
1810                 low = inactive_file_is_low_global(zone);
1811         else
1812                 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1813         return low;
1814 }
1815
1816 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1817                                 int file)
1818 {
1819         if (file)
1820                 return inactive_file_is_low(zone, sc);
1821         else
1822                 return inactive_anon_is_low(zone, sc);
1823 }
1824
1825 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1826         struct zone *zone, struct scan_control *sc, int priority)
1827 {
1828         int file = is_file_lru(lru);
1829
1830         if (is_active_lru(lru)) {
1831                 if (inactive_list_is_low(zone, sc, file))
1832                     shrink_active_list(nr_to_scan, zone, sc, priority, file);
1833                 return 0;
1834         }
1835
1836         return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1837 }
1838
1839 static int vmscan_swappiness(struct scan_control *sc)
1840 {
1841         if (scanning_global_lru(sc))
1842                 return vm_swappiness;
1843         return mem_cgroup_swappiness(sc->mem_cgroup);
1844 }
1845
1846 /*
1847  * Determine how aggressively the anon and file LRU lists should be
1848  * scanned.  The relative value of each set of LRU lists is determined
1849  * by looking at the fraction of the pages scanned we did rotate back
1850  * onto the active list instead of evict.
1851  *
1852  * nr[0] = anon pages to scan; nr[1] = file pages to scan
1853  */
1854 static void get_scan_count(struct zone *zone, struct scan_control *sc,
1855                                         unsigned long *nr, int priority)
1856 {
1857         unsigned long anon, file, free;
1858         unsigned long anon_prio, file_prio;
1859         unsigned long ap, fp;
1860         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1861         u64 fraction[2], denominator;
1862         enum lru_list l;
1863         int noswap = 0;
1864         bool force_scan = false;
1865
1866         /*
1867          * If the zone or memcg is small, nr[l] can be 0.  This
1868          * results in no scanning on this priority and a potential
1869          * priority drop.  Global direct reclaim can go to the next
1870          * zone and tends to have no problems. Global kswapd is for
1871          * zone balancing and it needs to scan a minimum amount. When
1872          * reclaiming for a memcg, a priority drop can cause high
1873          * latencies, so it's better to scan a minimum amount there as
1874          * well.
1875          */
1876         if (scanning_global_lru(sc) && current_is_kswapd())
1877                 force_scan = true;
1878         if (!scanning_global_lru(sc))
1879                 force_scan = true;
1880
1881         /* If we have no swap space, do not bother scanning anon pages. */
1882         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1883                 noswap = 1;
1884                 fraction[0] = 0;
1885                 fraction[1] = 1;
1886                 denominator = 1;
1887                 goto out;
1888         }
1889
1890         anon  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1891                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1892         file  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1893                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1894
1895         if (scanning_global_lru(sc)) {
1896                 free  = zone_page_state(zone, NR_FREE_PAGES);
1897                 /* If we have very few page cache pages,
1898                    force-scan anon pages. */
1899                 if (unlikely(file + free <= high_wmark_pages(zone))) {
1900                         fraction[0] = 1;
1901                         fraction[1] = 0;
1902                         denominator = 1;
1903                         goto out;
1904                 }
1905         }
1906
1907         /*
1908          * With swappiness at 100, anonymous and file have the same priority.
1909          * This scanning priority is essentially the inverse of IO cost.
1910          */
1911         anon_prio = vmscan_swappiness(sc);
1912         file_prio = 200 - vmscan_swappiness(sc);
1913
1914         /*
1915          * OK, so we have swap space and a fair amount of page cache
1916          * pages.  We use the recently rotated / recently scanned
1917          * ratios to determine how valuable each cache is.
1918          *
1919          * Because workloads change over time (and to avoid overflow)
1920          * we keep these statistics as a floating average, which ends
1921          * up weighing recent references more than old ones.
1922          *
1923          * anon in [0], file in [1]
1924          */
1925         spin_lock_irq(&zone->lru_lock);
1926         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1927                 reclaim_stat->recent_scanned[0] /= 2;
1928                 reclaim_stat->recent_rotated[0] /= 2;
1929         }
1930
1931         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1932                 reclaim_stat->recent_scanned[1] /= 2;
1933                 reclaim_stat->recent_rotated[1] /= 2;
1934         }
1935
1936         /*
1937          * The amount of pressure on anon vs file pages is inversely
1938          * proportional to the fraction of recently scanned pages on
1939          * each list that were recently referenced and in active use.
1940          */
1941         ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1942         ap /= reclaim_stat->recent_rotated[0] + 1;
1943
1944         fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1945         fp /= reclaim_stat->recent_rotated[1] + 1;
1946         spin_unlock_irq(&zone->lru_lock);
1947
1948         fraction[0] = ap;
1949         fraction[1] = fp;
1950         denominator = ap + fp + 1;
1951 out:
1952         for_each_evictable_lru(l) {
1953                 int file = is_file_lru(l);
1954                 unsigned long scan;
1955
1956                 scan = zone_nr_lru_pages(zone, sc, l);
1957                 if (priority || noswap) {
1958                         scan >>= priority;
1959                         if (!scan && force_scan)
1960                                 scan = SWAP_CLUSTER_MAX;
1961                         scan = div64_u64(scan * fraction[file], denominator);
1962                 }
1963                 nr[l] = scan;
1964         }
1965 }
1966
1967 /*
1968  * Reclaim/compaction depends on a number of pages being freed. To avoid
1969  * disruption to the system, a small number of order-0 pages continue to be
1970  * rotated and reclaimed in the normal fashion. However, by the time we get
1971  * back to the allocator and call try_to_compact_zone(), we ensure that
1972  * there are enough free pages for it to be likely successful
1973  */
1974 static inline bool should_continue_reclaim(struct zone *zone,
1975                                         unsigned long nr_reclaimed,
1976                                         unsigned long nr_scanned,
1977                                         struct scan_control *sc)
1978 {
1979         unsigned long pages_for_compaction;
1980         unsigned long inactive_lru_pages;
1981
1982         /* If not in reclaim/compaction mode, stop */
1983         if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
1984                 return false;
1985
1986         /* Consider stopping depending on scan and reclaim activity */
1987         if (sc->gfp_mask & __GFP_REPEAT) {
1988                 /*
1989                  * For __GFP_REPEAT allocations, stop reclaiming if the
1990                  * full LRU list has been scanned and we are still failing
1991                  * to reclaim pages. This full LRU scan is potentially
1992                  * expensive but a __GFP_REPEAT caller really wants to succeed
1993                  */
1994                 if (!nr_reclaimed && !nr_scanned)
1995                         return false;
1996         } else {
1997                 /*
1998                  * For non-__GFP_REPEAT allocations which can presumably
1999                  * fail without consequence, stop if we failed to reclaim
2000                  * any pages from the last SWAP_CLUSTER_MAX number of
2001                  * pages that were scanned. This will return to the
2002                  * caller faster at the risk reclaim/compaction and
2003                  * the resulting allocation attempt fails
2004                  */
2005                 if (!nr_reclaimed)
2006                         return false;
2007         }
2008
2009         /*
2010          * If we have not reclaimed enough pages for compaction and the
2011          * inactive lists are large enough, continue reclaiming
2012          */
2013         pages_for_compaction = (2UL << sc->order);
2014         inactive_lru_pages = zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON) +
2015                                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
2016         if (sc->nr_reclaimed < pages_for_compaction &&
2017                         inactive_lru_pages > pages_for_compaction)
2018                 return true;
2019
2020         /* If compaction would go ahead or the allocation would succeed, stop */
2021         switch (compaction_suitable(zone, sc->order)) {
2022         case COMPACT_PARTIAL:
2023         case COMPACT_CONTINUE:
2024                 return false;
2025         default:
2026                 return true;
2027         }
2028 }
2029
2030 /*
2031  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
2032  */
2033 static void shrink_zone(int priority, struct zone *zone,
2034                                 struct scan_control *sc)
2035 {
2036         unsigned long nr[NR_LRU_LISTS];
2037         unsigned long nr_to_scan;
2038         enum lru_list l;
2039         unsigned long nr_reclaimed, nr_scanned;
2040         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2041         struct blk_plug plug;
2042
2043 restart:
2044         nr_reclaimed = 0;
2045         nr_scanned = sc->nr_scanned;
2046         get_scan_count(zone, sc, nr, priority);
2047
2048         blk_start_plug(&plug);
2049         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2050                                         nr[LRU_INACTIVE_FILE]) {
2051                 for_each_evictable_lru(l) {
2052                         if (nr[l]) {
2053                                 nr_to_scan = min_t(unsigned long,
2054                                                    nr[l], SWAP_CLUSTER_MAX);
2055                                 nr[l] -= nr_to_scan;
2056
2057                                 nr_reclaimed += shrink_list(l, nr_to_scan,
2058                                                             zone, sc, priority);
2059                         }
2060                 }
2061                 /*
2062                  * On large memory systems, scan >> priority can become
2063                  * really large. This is fine for the starting priority;
2064                  * we want to put equal scanning pressure on each zone.
2065                  * However, if the VM has a harder time of freeing pages,
2066                  * with multiple processes reclaiming pages, the total
2067                  * freeing target can get unreasonably large.
2068                  */
2069                 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
2070                         break;
2071         }
2072         blk_finish_plug(&plug);
2073         sc->nr_reclaimed += nr_reclaimed;
2074
2075         /*
2076          * Even if we did not try to evict anon pages at all, we want to
2077          * rebalance the anon lru active/inactive ratio.
2078          */
2079         if (inactive_anon_is_low(zone, sc))
2080                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
2081
2082         /* reclaim/compaction might need reclaim to continue */
2083         if (should_continue_reclaim(zone, nr_reclaimed,
2084                                         sc->nr_scanned - nr_scanned, sc))
2085                 goto restart;
2086
2087         throttle_vm_writeout(sc->gfp_mask);
2088 }
2089
2090 /*
2091  * This is the direct reclaim path, for page-allocating processes.  We only
2092  * try to reclaim pages from zones which will satisfy the caller's allocation
2093  * request.
2094  *
2095  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2096  * Because:
2097  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2098  *    allocation or
2099  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2100  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2101  *    zone defense algorithm.
2102  *
2103  * If a zone is deemed to be full of pinned pages then just give it a light
2104  * scan then give up on it.
2105  */
2106 static void shrink_zones(int priority, struct zonelist *zonelist,
2107                                         struct scan_control *sc)
2108 {
2109         struct zoneref *z;
2110         struct zone *zone;
2111         unsigned long nr_soft_reclaimed;
2112         unsigned long nr_soft_scanned;
2113
2114         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2115                                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2116                 if (!populated_zone(zone))
2117                         continue;
2118                 /*
2119                  * Take care memory controller reclaiming has small influence
2120                  * to global LRU.
2121                  */
2122                 if (scanning_global_lru(sc)) {
2123                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2124                                 continue;
2125                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2126                                 continue;       /* Let kswapd poll it */
2127                         /*
2128                          * This steals pages from memory cgroups over softlimit
2129                          * and returns the number of reclaimed pages and
2130                          * scanned pages. This works for global memory pressure
2131                          * and balancing, not for a memcg's limit.
2132                          */
2133                         nr_soft_scanned = 0;
2134                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2135                                                 sc->order, sc->gfp_mask,
2136                                                 &nr_soft_scanned);
2137                         sc->nr_reclaimed += nr_soft_reclaimed;
2138                         sc->nr_scanned += nr_soft_scanned;
2139                         /* need some check for avoid more shrink_zone() */
2140                 }
2141
2142                 shrink_zone(priority, zone, sc);
2143         }
2144 }
2145
2146 static bool zone_reclaimable(struct zone *zone)
2147 {
2148         return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2149 }
2150
2151 /* All zones in zonelist are unreclaimable? */
2152 static bool all_unreclaimable(struct zonelist *zonelist,
2153                 struct scan_control *sc)
2154 {
2155         struct zoneref *z;
2156         struct zone *zone;
2157
2158         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2159                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2160                 if (!populated_zone(zone))
2161                         continue;
2162                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2163                         continue;
2164                 if (!zone->all_unreclaimable)
2165                         return false;
2166         }
2167
2168         return true;
2169 }
2170
2171 /*
2172  * This is the main entry point to direct page reclaim.
2173  *
2174  * If a full scan of the inactive list fails to free enough memory then we
2175  * are "out of memory" and something needs to be killed.
2176  *
2177  * If the caller is !__GFP_FS then the probability of a failure is reasonably
2178  * high - the zone may be full of dirty or under-writeback pages, which this
2179  * caller can't do much about.  We kick the writeback threads and take explicit
2180  * naps in the hope that some of these pages can be written.  But if the
2181  * allocating task holds filesystem locks which prevent writeout this might not
2182  * work, and the allocation attempt will fail.
2183  *
2184  * returns:     0, if no pages reclaimed
2185  *              else, the number of pages reclaimed
2186  */
2187 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2188                                         struct scan_control *sc,
2189                                         struct shrink_control *shrink)
2190 {
2191         int priority;
2192         unsigned long total_scanned = 0;
2193         struct reclaim_state *reclaim_state = current->reclaim_state;
2194         struct zoneref *z;
2195         struct zone *zone;
2196         unsigned long writeback_threshold;
2197
2198         get_mems_allowed();
2199         delayacct_freepages_start();
2200
2201         if (scanning_global_lru(sc))
2202                 count_vm_event(ALLOCSTALL);
2203
2204         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2205                 sc->nr_scanned = 0;
2206                 if (!priority)
2207                         disable_swap_token(sc->mem_cgroup);
2208                 shrink_zones(priority, zonelist, sc);
2209                 /*
2210                  * Don't shrink slabs when reclaiming memory from
2211                  * over limit cgroups
2212                  */
2213                 if (scanning_global_lru(sc)) {
2214                         unsigned long lru_pages = 0;
2215                         for_each_zone_zonelist(zone, z, zonelist,
2216                                         gfp_zone(sc->gfp_mask)) {
2217                                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2218                                         continue;
2219
2220                                 lru_pages += zone_reclaimable_pages(zone);
2221                         }
2222
2223                         shrink_slab(shrink, sc->nr_scanned, lru_pages);
2224                         if (reclaim_state) {
2225                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2226                                 reclaim_state->reclaimed_slab = 0;
2227                         }
2228                 }
2229                 total_scanned += sc->nr_scanned;
2230                 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2231                         goto out;
2232
2233                 /*
2234                  * Try to write back as many pages as we just scanned.  This
2235                  * tends to cause slow streaming writers to write data to the
2236                  * disk smoothly, at the dirtying rate, which is nice.   But
2237                  * that's undesirable in laptop mode, where we *want* lumpy
2238                  * writeout.  So in laptop mode, write out the whole world.
2239                  */
2240                 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2241                 if (total_scanned > writeback_threshold) {
2242                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
2243                         sc->may_writepage = 1;
2244                 }
2245
2246                 /* Take a nap, wait for some writeback to complete */
2247                 if (!sc->hibernation_mode && sc->nr_scanned &&
2248                     priority < DEF_PRIORITY - 2) {
2249                         struct zone *preferred_zone;
2250
2251                         first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2252                                                 &cpuset_current_mems_allowed,
2253                                                 &preferred_zone);
2254                         wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2255                 }
2256         }
2257
2258 out:
2259         delayacct_freepages_end();
2260         put_mems_allowed();
2261
2262         if (sc->nr_reclaimed)
2263                 return sc->nr_reclaimed;
2264
2265         /*
2266          * As hibernation is going on, kswapd is freezed so that it can't mark
2267          * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2268          * check.
2269          */
2270         if (oom_killer_disabled)
2271                 return 0;
2272
2273         /* top priority shrink_zones still had more to do? don't OOM, then */
2274         if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
2275                 return 1;
2276
2277         return 0;
2278 }
2279
2280 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2281                                 gfp_t gfp_mask, nodemask_t *nodemask)
2282 {
2283         unsigned long nr_reclaimed;
2284         struct scan_control sc = {
2285                 .gfp_mask = gfp_mask,
2286                 .may_writepage = !laptop_mode,
2287                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2288                 .may_unmap = 1,
2289                 .may_swap = 1,
2290                 .order = order,
2291                 .mem_cgroup = NULL,
2292                 .nodemask = nodemask,
2293         };
2294         struct shrink_control shrink = {
2295                 .gfp_mask = sc.gfp_mask,
2296         };
2297
2298         trace_mm_vmscan_direct_reclaim_begin(order,
2299                                 sc.may_writepage,
2300                                 gfp_mask);
2301
2302         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2303
2304         trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2305
2306         return nr_reclaimed;
2307 }
2308
2309 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2310
2311 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2312                                                 gfp_t gfp_mask, bool noswap,
2313                                                 struct zone *zone,
2314                                                 unsigned long *nr_scanned)
2315 {
2316         struct scan_control sc = {
2317                 .nr_scanned = 0,
2318                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2319                 .may_writepage = !laptop_mode,
2320                 .may_unmap = 1,
2321                 .may_swap = !noswap,
2322                 .order = 0,
2323                 .mem_cgroup = mem,
2324         };
2325
2326         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2327                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2328
2329         trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2330                                                       sc.may_writepage,
2331                                                       sc.gfp_mask);
2332
2333         /*
2334          * NOTE: Although we can get the priority field, using it
2335          * here is not a good idea, since it limits the pages we can scan.
2336          * if we don't reclaim here, the shrink_zone from balance_pgdat
2337          * will pick up pages from other mem cgroup's as well. We hack
2338          * the priority and make it zero.
2339          */
2340         shrink_zone(0, zone, &sc);
2341
2342         trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2343
2344         *nr_scanned = sc.nr_scanned;
2345         return sc.nr_reclaimed;
2346 }
2347
2348 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
2349                                            gfp_t gfp_mask,
2350                                            bool noswap)
2351 {
2352         struct zonelist *zonelist;
2353         unsigned long nr_reclaimed;
2354         int nid;
2355         struct scan_control sc = {
2356                 .may_writepage = !laptop_mode,
2357                 .may_unmap = 1,
2358                 .may_swap = !noswap,
2359                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2360                 .order = 0,
2361                 .mem_cgroup = mem_cont,
2362                 .nodemask = NULL, /* we don't care the placement */
2363                 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2364                                 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2365         };
2366         struct shrink_control shrink = {
2367                 .gfp_mask = sc.gfp_mask,
2368         };
2369
2370         /*
2371          * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2372          * take care of from where we get pages. So the node where we start the
2373          * scan does not need to be the current node.
2374          */
2375         nid = mem_cgroup_select_victim_node(mem_cont);
2376
2377         zonelist = NODE_DATA(nid)->node_zonelists;
2378
2379         trace_mm_vmscan_memcg_reclaim_begin(0,
2380                                             sc.may_writepage,
2381                                             sc.gfp_mask);
2382
2383         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2384
2385         trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2386
2387         return nr_reclaimed;
2388 }
2389 #endif
2390
2391 /*
2392  * pgdat_balanced is used when checking if a node is balanced for high-order
2393  * allocations. Only zones that meet watermarks and are in a zone allowed
2394  * by the callers classzone_idx are added to balanced_pages. The total of
2395  * balanced pages must be at least 25% of the zones allowed by classzone_idx
2396  * for the node to be considered balanced. Forcing all zones to be balanced
2397  * for high orders can cause excessive reclaim when there are imbalanced zones.
2398  * The choice of 25% is due to
2399  *   o a 16M DMA zone that is balanced will not balance a zone on any
2400  *     reasonable sized machine
2401  *   o On all other machines, the top zone must be at least a reasonable
2402  *     percentage of the middle zones. For example, on 32-bit x86, highmem
2403  *     would need to be at least 256M for it to be balance a whole node.
2404  *     Similarly, on x86-64 the Normal zone would need to be at least 1G
2405  *     to balance a node on its own. These seemed like reasonable ratios.
2406  */
2407 static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2408                                                 int classzone_idx)
2409 {
2410         unsigned long present_pages = 0;
2411         int i;
2412
2413         for (i = 0; i <= classzone_idx; i++)
2414                 present_pages += pgdat->node_zones[i].present_pages;
2415
2416         /* A special case here: if zone has no page, we think it's balanced */
2417         return balanced_pages >= (present_pages >> 2);
2418 }
2419
2420 /* is kswapd sleeping prematurely? */
2421 static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2422                                         int classzone_idx)
2423 {
2424         int i;
2425         unsigned long balanced = 0;
2426         bool all_zones_ok = true;
2427
2428         /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2429         if (remaining)
2430                 return true;
2431
2432         /* Check the watermark levels */
2433         for (i = 0; i <= classzone_idx; i++) {
2434                 struct zone *zone = pgdat->node_zones + i;
2435
2436                 if (!populated_zone(zone))
2437                         continue;
2438
2439                 /*
2440                  * balance_pgdat() skips over all_unreclaimable after
2441                  * DEF_PRIORITY. Effectively, it considers them balanced so
2442                  * they must be considered balanced here as well if kswapd
2443                  * is to sleep
2444                  */
2445                 if (zone->all_unreclaimable) {
2446                         balanced += zone->present_pages;
2447                         continue;
2448                 }
2449
2450                 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2451                                                         i, 0))
2452                         all_zones_ok = false;
2453                 else
2454                         balanced += zone->present_pages;
2455         }
2456
2457         /*
2458          * For high-order requests, the balanced zones must contain at least
2459          * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2460          * must be balanced
2461          */
2462         if (order)
2463                 return !pgdat_balanced(pgdat, balanced, classzone_idx);
2464         else
2465                 return !all_zones_ok;
2466 }
2467
2468 /*
2469  * For kswapd, balance_pgdat() will work across all this node's zones until
2470  * they are all at high_wmark_pages(zone).
2471  *
2472  * Returns the final order kswapd was reclaiming at
2473  *
2474  * There is special handling here for zones which are full of pinned pages.
2475  * This can happen if the pages are all mlocked, or if they are all used by
2476  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
2477  * What we do is to detect the case where all pages in the zone have been
2478  * scanned twice and there has been zero successful reclaim.  Mark the zone as
2479  * dead and from now on, only perform a short scan.  Basically we're polling
2480  * the zone for when the problem goes away.
2481  *
2482  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
2483  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2484  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2485  * lower zones regardless of the number of free pages in the lower zones. This
2486  * interoperates with the page allocator fallback scheme to ensure that aging
2487  * of pages is balanced across the zones.
2488  */
2489 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2490                                                         int *classzone_idx)
2491 {
2492         int all_zones_ok;
2493         unsigned long balanced;
2494         int priority;
2495         int i;
2496         int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
2497         unsigned long total_scanned;
2498         struct reclaim_state *reclaim_state = current->reclaim_state;
2499         unsigned long nr_soft_reclaimed;
2500         unsigned long nr_soft_scanned;
2501         struct scan_control sc = {
2502                 .gfp_mask = GFP_KERNEL,
2503                 .may_unmap = 1,
2504                 .may_swap = 1,
2505                 /*
2506                  * kswapd doesn't want to be bailed out while reclaim. because
2507                  * we want to put equal scanning pressure on each zone.
2508                  */
2509                 .nr_to_reclaim = ULONG_MAX,
2510                 .order = order,
2511                 .mem_cgroup = NULL,
2512         };
2513         struct shrink_control shrink = {
2514                 .gfp_mask = sc.gfp_mask,
2515         };
2516 loop_again:
2517         total_scanned = 0;
2518         sc.nr_reclaimed = 0;
2519         sc.may_writepage = !laptop_mode;
2520         count_vm_event(PAGEOUTRUN);
2521
2522         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2523                 unsigned long lru_pages = 0;
2524                 int has_under_min_watermark_zone = 0;
2525
2526                 /* The swap token gets in the way of swapout... */
2527                 if (!priority)
2528                         disable_swap_token(NULL);
2529
2530                 all_zones_ok = 1;
2531                 balanced = 0;
2532
2533                 /*
2534                  * Scan in the highmem->dma direction for the highest
2535                  * zone which needs scanning
2536                  */
2537                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2538                         struct zone *zone = pgdat->node_zones + i;
2539
2540                         if (!populated_zone(zone))
2541                                 continue;
2542
2543                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2544                                 continue;
2545
2546                         /*
2547                          * Do some background aging of the anon list, to give
2548                          * pages a chance to be referenced before reclaiming.
2549                          */
2550                         if (inactive_anon_is_low(zone, &sc))
2551                                 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2552                                                         &sc, priority, 0);
2553
2554                         if (!zone_watermark_ok_safe(zone, order,
2555                                         high_wmark_pages(zone), 0, 0)) {
2556                                 end_zone = i;
2557                                 break;
2558                         } else {
2559                                 /* If balanced, clear the congested flag */
2560                                 zone_clear_flag(zone, ZONE_CONGESTED);
2561                         }
2562                 }
2563                 if (i < 0)
2564                         goto out;
2565
2566                 for (i = 0; i <= end_zone; i++) {
2567                         struct zone *zone = pgdat->node_zones + i;
2568
2569                         lru_pages += zone_reclaimable_pages(zone);
2570                 }
2571
2572                 /*
2573                  * Now scan the zone in the dma->highmem direction, stopping
2574                  * at the last zone which needs scanning.
2575                  *
2576                  * We do this because the page allocator works in the opposite
2577                  * direction.  This prevents the page allocator from allocating
2578                  * pages behind kswapd's direction of progress, which would
2579                  * cause too much scanning of the lower zones.
2580                  */
2581                 for (i = 0; i <= end_zone; i++) {
2582                         struct zone *zone = pgdat->node_zones + i;
2583                         int nr_slab;
2584                         unsigned long balance_gap;
2585
2586                         if (!populated_zone(zone))
2587                                 continue;
2588
2589                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2590                                 continue;
2591
2592                         sc.nr_scanned = 0;
2593
2594                         nr_soft_scanned = 0;
2595                         /*
2596                          * Call soft limit reclaim before calling shrink_zone.
2597                          */
2598                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2599                                                         order, sc.gfp_mask,
2600                                                         &nr_soft_scanned);
2601                         sc.nr_reclaimed += nr_soft_reclaimed;
2602                         total_scanned += nr_soft_scanned;
2603
2604                         /*
2605                          * We put equal pressure on every zone, unless
2606                          * one zone has way too many pages free
2607                          * already. The "too many pages" is defined
2608                          * as the high wmark plus a "gap" where the
2609                          * gap is either the low watermark or 1%
2610                          * of the zone, whichever is smaller.
2611                          */
2612                         balance_gap = min(low_wmark_pages(zone),
2613                                 (zone->present_pages +
2614                                         KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2615                                 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2616                         if (!zone_watermark_ok_safe(zone, order,
2617                                         high_wmark_pages(zone) + balance_gap,
2618                                         end_zone, 0)) {
2619                                 shrink_zone(priority, zone, &sc);
2620
2621                                 reclaim_state->reclaimed_slab = 0;
2622                                 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2623                                 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2624                                 total_scanned += sc.nr_scanned;
2625
2626                                 if (nr_slab == 0 && !zone_reclaimable(zone))
2627                                         zone->all_unreclaimable = 1;
2628                         }
2629
2630                         /*
2631                          * If we've done a decent amount of scanning and
2632                          * the reclaim ratio is low, start doing writepage
2633                          * even in laptop mode
2634                          */
2635                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2636                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2637                                 sc.may_writepage = 1;
2638
2639                         if (zone->all_unreclaimable) {
2640                                 if (end_zone && end_zone == i)
2641                                         end_zone--;
2642                                 continue;
2643                         }
2644
2645                         if (!zone_watermark_ok_safe(zone, order,
2646                                         high_wmark_pages(zone), end_zone, 0)) {
2647                                 all_zones_ok = 0;
2648                                 /*
2649                                  * We are still under min water mark.  This
2650                                  * means that we have a GFP_ATOMIC allocation
2651                                  * failure risk. Hurry up!
2652                                  */
2653                                 if (!zone_watermark_ok_safe(zone, order,
2654                                             min_wmark_pages(zone), end_zone, 0))
2655                                         has_under_min_watermark_zone = 1;
2656                         } else {
2657                                 /*
2658                                  * If a zone reaches its high watermark,
2659                                  * consider it to be no longer congested. It's
2660                                  * possible there are dirty pages backed by
2661                                  * congested BDIs but as pressure is relieved,
2662                                  * spectulatively avoid congestion waits
2663                                  */
2664                                 zone_clear_flag(zone, ZONE_CONGESTED);
2665                                 if (i <= *classzone_idx)
2666                                         balanced += zone->present_pages;
2667                         }
2668
2669                 }
2670                 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2671                         break;          /* kswapd: all done */
2672                 /*
2673                  * OK, kswapd is getting into trouble.  Take a nap, then take
2674                  * another pass across the zones.
2675                  */
2676                 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2677                         if (has_under_min_watermark_zone)
2678                                 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2679                         else
2680                                 congestion_wait(BLK_RW_ASYNC, HZ/10);
2681                 }
2682
2683                 /*
2684                  * We do this so kswapd doesn't build up large priorities for
2685                  * example when it is freeing in parallel with allocators. It
2686                  * matches the direct reclaim path behaviour in terms of impact
2687                  * on zone->*_priority.
2688                  */
2689                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2690                         break;
2691         }
2692 out:
2693
2694         /*
2695          * order-0: All zones must meet high watermark for a balanced node
2696          * high-order: Balanced zones must make up at least 25% of the node
2697          *             for the node to be balanced
2698          */
2699         if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2700                 cond_resched();
2701
2702                 try_to_freeze();
2703
2704                 /*
2705                  * Fragmentation may mean that the system cannot be
2706                  * rebalanced for high-order allocations in all zones.
2707                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2708                  * it means the zones have been fully scanned and are still
2709                  * not balanced. For high-order allocations, there is
2710                  * little point trying all over again as kswapd may
2711                  * infinite loop.
2712                  *
2713                  * Instead, recheck all watermarks at order-0 as they
2714                  * are the most important. If watermarks are ok, kswapd will go
2715                  * back to sleep. High-order users can still perform direct
2716                  * reclaim if they wish.
2717                  */
2718                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2719                         order = sc.order = 0;
2720
2721                 goto loop_again;
2722         }
2723
2724         /*
2725          * If kswapd was reclaiming at a higher order, it has the option of
2726          * sleeping without all zones being balanced. Before it does, it must
2727          * ensure that the watermarks for order-0 on *all* zones are met and
2728          * that the congestion flags are cleared. The congestion flag must
2729          * be cleared as kswapd is the only mechanism that clears the flag
2730          * and it is potentially going to sleep here.
2731          */
2732         if (order) {
2733                 for (i = 0; i <= end_zone; i++) {
2734                         struct zone *zone = pgdat->node_zones + i;
2735
2736                         if (!populated_zone(zone))
2737                                 continue;
2738
2739                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2740                                 continue;
2741
2742                         /* Confirm the zone is balanced for order-0 */
2743                         if (!zone_watermark_ok(zone, 0,
2744                                         high_wmark_pages(zone), 0, 0)) {
2745                                 order = sc.order = 0;
2746                                 goto loop_again;
2747                         }
2748
2749                         /* If balanced, clear the congested flag */
2750                         zone_clear_flag(zone, ZONE_CONGESTED);
2751                         if (i <= *classzone_idx)
2752                                 balanced += zone->present_pages;
2753                 }
2754         }
2755
2756         /*
2757          * Return the order we were reclaiming at so sleeping_prematurely()
2758          * makes a decision on the order we were last reclaiming at. However,
2759          * if another caller entered the allocator slow path while kswapd
2760          * was awake, order will remain at the higher level
2761          */
2762         *classzone_idx = end_zone;
2763         return order;
2764 }
2765
2766 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2767 {
2768         long remaining = 0;
2769         DEFINE_WAIT(wait);
2770
2771         if (freezing(current) || kthread_should_stop())
2772                 return;
2773
2774         prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2775
2776         /* Try to sleep for a short interval */
2777         if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2778                 remaining = schedule_timeout(HZ/10);
2779                 finish_wait(&pgdat->kswapd_wait, &wait);
2780                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2781         }
2782
2783         /*
2784          * After a short sleep, check if it was a premature sleep. If not, then
2785          * go fully to sleep until explicitly woken up.
2786          */
2787         if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2788                 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2789
2790                 /*
2791                  * vmstat counters are not perfectly accurate and the estimated
2792                  * value for counters such as NR_FREE_PAGES can deviate from the
2793                  * true value by nr_online_cpus * threshold. To avoid the zone
2794                  * watermarks being breached while under pressure, we reduce the
2795                  * per-cpu vmstat threshold while kswapd is awake and restore
2796                  * them before going back to sleep.
2797                  */
2798                 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2799                 schedule();
2800                 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2801         } else {
2802                 if (remaining)
2803                         count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2804                 else
2805                         count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2806         }
2807         finish_wait(&pgdat->kswapd_wait, &wait);
2808 }
2809
2810 /*
2811  * The background pageout daemon, started as a kernel thread
2812  * from the init process.
2813  *
2814  * This basically trickles out pages so that we have _some_
2815  * free memory available even if there is no other activity
2816  * that frees anything up. This is needed for things like routing
2817  * etc, where we otherwise might have all activity going on in
2818  * asynchronous contexts that cannot page things out.
2819  *
2820  * If there are applications that are active memory-allocators
2821  * (most normal use), this basically shouldn't matter.
2822  */
2823 static int kswapd(void *p)
2824 {
2825         unsigned long order, new_order;
2826         int classzone_idx, new_classzone_idx;
2827         pg_data_t *pgdat = (pg_data_t*)p;
2828         struct task_struct *tsk = current;
2829
2830         struct reclaim_state reclaim_state = {
2831                 .reclaimed_slab = 0,
2832         };
2833         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2834
2835         lockdep_set_current_reclaim_state(GFP_KERNEL);
2836
2837         if (!cpumask_empty(cpumask))
2838                 set_cpus_allowed_ptr(tsk, cpumask);
2839         current->reclaim_state = &reclaim_state;
2840
2841         /*
2842          * Tell the memory management that we're a "memory allocator",
2843          * and that if we need more memory we should get access to it
2844          * regardless (see "__alloc_pages()"). "kswapd" should
2845          * never get caught in the normal page freeing logic.
2846          *
2847          * (Kswapd normally doesn't need memory anyway, but sometimes
2848          * you need a small amount of memory in order to be able to
2849          * page out something else, and this flag essentially protects
2850          * us from recursively trying to free more memory as we're
2851          * trying to free the first piece of memory in the first place).
2852          */
2853         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2854         set_freezable();
2855
2856         order = new_order = 0;
2857         classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
2858         for ( ; ; ) {
2859                 int ret;
2860
2861                 /*
2862                  * If the last balance_pgdat was unsuccessful it's unlikely a
2863                  * new request of a similar or harder type will succeed soon
2864                  * so consider going to sleep on the basis we reclaimed at
2865                  */
2866                 if (classzone_idx >= new_classzone_idx && order == new_order) {
2867                         new_order = pgdat->kswapd_max_order;
2868                         new_classzone_idx = pgdat->classzone_idx;
2869                         pgdat->kswapd_max_order =  0;
2870                         pgdat->classzone_idx = pgdat->nr_zones - 1;
2871                 }
2872
2873                 if (order < new_order || classzone_idx > new_classzone_idx) {
2874                         /*
2875                          * Don't sleep if someone wants a larger 'order'
2876                          * allocation or has tigher zone constraints
2877                          */
2878                         order = new_order;
2879                         classzone_idx = new_classzone_idx;
2880                 } else {
2881                         kswapd_try_to_sleep(pgdat, order, classzone_idx);
2882                         order = pgdat->kswapd_max_order;
2883                         classzone_idx = pgdat->classzone_idx;
2884                         pgdat->kswapd_max_order = 0;
2885                         pgdat->classzone_idx = pgdat->nr_zones - 1;
2886                 }
2887
2888                 ret = try_to_freeze();
2889                 if (kthread_should_stop())
2890                         break;
2891
2892                 /*
2893                  * We can speed up thawing tasks if we don't call balance_pgdat
2894                  * after returning from the refrigerator
2895                  */
2896                 if (!ret) {
2897                         trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2898                         order = balance_pgdat(pgdat, order, &classzone_idx);
2899                 }
2900         }
2901         return 0;
2902 }
2903
2904 /*
2905  * A zone is low on free memory, so wake its kswapd task to service it.
2906  */
2907 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
2908 {
2909         pg_data_t *pgdat;
2910
2911         if (!populated_zone(zone))
2912                 return;
2913
2914         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2915                 return;
2916         pgdat = zone->zone_pgdat;
2917         if (pgdat->kswapd_max_order < order) {
2918                 pgdat->kswapd_max_order = order;
2919                 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
2920         }
2921         if (!waitqueue_active(&pgdat->kswapd_wait))
2922                 return;
2923         if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
2924                 return;
2925
2926         trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
2927         wake_up_interruptible(&pgdat->kswapd_wait);
2928 }
2929
2930 /*
2931  * The reclaimable count would be mostly accurate.
2932  * The less reclaimable pages may be
2933  * - mlocked pages, which will be moved to unevictable list when encountered
2934  * - mapped pages, which may require several travels to be reclaimed
2935  * - dirty pages, which is not "instantly" reclaimable
2936  */
2937 unsigned long global_reclaimable_pages(void)
2938 {
2939         int nr;
2940
2941         nr = global_page_state(NR_ACTIVE_FILE) +
2942              global_page_state(NR_INACTIVE_FILE);
2943
2944         if (nr_swap_pages > 0)
2945                 nr += global_page_state(NR_ACTIVE_ANON) +
2946                       global_page_state(NR_INACTIVE_ANON);
2947
2948         return nr;
2949 }
2950
2951 unsigned long zone_reclaimable_pages(struct zone *zone)
2952 {
2953         int nr;
2954
2955         nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2956              zone_page_state(zone, NR_INACTIVE_FILE);
2957
2958         if (nr_swap_pages > 0)
2959                 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2960                       zone_page_state(zone, NR_INACTIVE_ANON);
2961
2962         return nr;
2963 }
2964
2965 #ifdef CONFIG_HIBERNATION
2966 /*
2967  * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2968  * freed pages.
2969  *
2970  * Rather than trying to age LRUs the aim is to preserve the overall
2971  * LRU order by reclaiming preferentially
2972  * inactive > active > active referenced > active mapped
2973  */
2974 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
2975 {
2976         struct reclaim_state reclaim_state;
2977         struct scan_control sc = {
2978                 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2979                 .may_swap = 1,
2980                 .may_unmap = 1,
2981                 .may_writepage = 1,
2982                 .nr_to_reclaim = nr_to_reclaim,
2983                 .hibernation_mode = 1,
2984                 .order = 0,
2985         };
2986         struct shrink_control shrink = {
2987                 .gfp_mask = sc.gfp_mask,
2988         };
2989         struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2990         struct task_struct *p = current;
2991         unsigned long nr_reclaimed;
2992
2993         p->flags |= PF_MEMALLOC;
2994         lockdep_set_current_reclaim_state(sc.gfp_mask);
2995         reclaim_state.reclaimed_slab = 0;
2996         p->reclaim_state = &reclaim_state;
2997
2998         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2999
3000         p->reclaim_state = NULL;
3001         lockdep_clear_current_reclaim_state();
3002         p->flags &= ~PF_MEMALLOC;
3003
3004         return nr_reclaimed;
3005 }
3006 #endif /* CONFIG_HIBERNATION */
3007
3008 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3009    not required for correctness.  So if the last cpu in a node goes
3010    away, we get changed to run anywhere: as the first one comes back,
3011    restore their cpu bindings. */
3012 static int __devinit cpu_callback(struct notifier_block *nfb,
3013                                   unsigned long action, void *hcpu)
3014 {
3015         int nid;
3016
3017         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3018                 for_each_node_state(nid, N_HIGH_MEMORY) {
3019                         pg_data_t *pgdat = NODE_DATA(nid);
3020                         const struct cpumask *mask;
3021
3022                         mask = cpumask_of_node(pgdat->node_id);
3023
3024                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3025                                 /* One of our CPUs online: restore mask */
3026                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3027                 }
3028         }
3029         return NOTIFY_OK;
3030 }
3031
3032 /*
3033  * This kswapd start function will be called by init and node-hot-add.
3034  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3035  */
3036 int kswapd_run(int nid)
3037 {
3038         pg_data_t *pgdat = NODE_DATA(nid);
3039         int ret = 0;
3040
3041         if (pgdat->kswapd)
3042                 return 0;
3043
3044         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3045         if (IS_ERR(pgdat->kswapd)) {
3046                 /* failure at boot is fatal */
3047                 BUG_ON(system_state == SYSTEM_BOOTING);
3048                 printk("Failed to start kswapd on node %d\n",nid);
3049                 ret = -1;
3050         }
3051         return ret;
3052 }
3053
3054 /*
3055  * Called by memory hotplug when all memory in a node is offlined.
3056  */
3057 void kswapd_stop(int nid)
3058 {
3059         struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3060
3061         if (kswapd)
3062                 kthread_stop(kswapd);
3063 }
3064
3065 static int __init kswapd_init(void)
3066 {
3067         int nid;
3068
3069         swap_setup();
3070         for_each_node_state(nid, N_HIGH_MEMORY)
3071                 kswapd_run(nid);
3072         hotcpu_notifier(cpu_callback, 0);
3073         return 0;
3074 }
3075
3076 module_init(kswapd_init)
3077
3078 #ifdef CONFIG_NUMA
3079 /*
3080  * Zone reclaim mode
3081  *
3082  * If non-zero call zone_reclaim when the number of free pages falls below
3083  * the watermarks.
3084  */
3085 int zone_reclaim_mode __read_mostly;
3086
3087 #define RECLAIM_OFF 0
3088 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
3089 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
3090 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
3091
3092 /*
3093  * Priority for ZONE_RECLAIM. This determines the fraction of pages
3094  * of a node considered for each zone_reclaim. 4 scans 1/16th of
3095  * a zone.
3096  */
3097 #define ZONE_RECLAIM_PRIORITY 4
3098
3099 /*
3100  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3101  * occur.
3102  */
3103 int sysctl_min_unmapped_ratio = 1;
3104
3105 /*
3106  * If the number of slab pages in a zone grows beyond this percentage then
3107  * slab reclaim needs to occur.
3108  */
3109 int sysctl_min_slab_ratio = 5;
3110
3111 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3112 {
3113         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3114         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3115                 zone_page_state(zone, NR_ACTIVE_FILE);
3116
3117         /*
3118          * It's possible for there to be more file mapped pages than
3119          * accounted for by the pages on the file LRU lists because
3120          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3121          */
3122         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3123 }
3124
3125 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3126 static long zone_pagecache_reclaimable(struct zone *zone)
3127 {
3128         long nr_pagecache_reclaimable;
3129         long delta = 0;
3130
3131         /*
3132          * If RECLAIM_SWAP is set, then all file pages are considered
3133          * potentially reclaimable. Otherwise, we have to worry about
3134          * pages like swapcache and zone_unmapped_file_pages() provides
3135          * a better estimate
3136          */
3137         if (zone_reclaim_mode & RECLAIM_SWAP)
3138                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3139         else
3140                 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3141
3142         /* If we can't clean pages, remove dirty pages from consideration */
3143         if (!(zone_reclaim_mode & RECLAIM_WRITE))
3144                 delta += zone_page_state(zone, NR_FILE_DIRTY);
3145
3146         /* Watch for any possible underflows due to delta */
3147         if (unlikely(delta > nr_pagecache_reclaimable))
3148                 delta = nr_pagecache_reclaimable;
3149
3150         return nr_pagecache_reclaimable - delta;
3151 }
3152
3153 /*
3154  * Try to free up some pages from this zone through reclaim.
3155  */
3156 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3157 {
3158         /* Minimum pages needed in order to stay on node */
3159         const unsigned long nr_pages = 1 << order;
3160         struct task_struct *p = current;
3161         struct reclaim_state reclaim_state;
3162         int priority;
3163         struct scan_control sc = {
3164                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3165                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3166                 .may_swap = 1,
3167                 .nr_to_reclaim = max_t(unsigned long, nr_pages,
3168                                        SWAP_CLUSTER_MAX),
3169                 .gfp_mask = gfp_mask,
3170                 .order = order,
3171         };
3172         struct shrink_control shrink = {
3173                 .gfp_mask = sc.gfp_mask,
3174         };
3175         unsigned long nr_slab_pages0, nr_slab_pages1;
3176
3177         cond_resched();
3178         /*
3179          * We need to be able to allocate from the reserves for RECLAIM_SWAP
3180          * and we also need to be able to write out pages for RECLAIM_WRITE
3181          * and RECLAIM_SWAP.
3182          */
3183         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3184         lockdep_set_current_reclaim_state(gfp_mask);
3185         reclaim_state.reclaimed_slab = 0;
3186         p->reclaim_state = &reclaim_state;
3187
3188         if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3189                 /*
3190                  * Free memory by calling shrink zone with increasing
3191                  * priorities until we have enough memory freed.
3192                  */
3193                 priority = ZONE_RECLAIM_PRIORITY;
3194                 do {
3195                         shrink_zone(priority, zone, &sc);
3196                         priority--;
3197                 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
3198         }
3199
3200         nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3201         if (nr_slab_pages0 > zone->min_slab_pages) {
3202                 /*
3203                  * shrink_slab() does not currently allow us to determine how
3204                  * many pages were freed in this zone. So we take the current
3205                  * number of slab pages and shake the slab until it is reduced
3206                  * by the same nr_pages that we used for reclaiming unmapped
3207                  * pages.
3208                  *
3209                  * Note that shrink_slab will free memory on all zones and may
3210                  * take a long time.
3211                  */
3212                 for (;;) {
3213                         unsigned long lru_pages = zone_reclaimable_pages(zone);
3214
3215                         /* No reclaimable slab or very low memory pressure */
3216                         if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3217                                 break;
3218
3219                         /* Freed enough memory */
3220                         nr_slab_pages1 = zone_page_state(zone,
3221                                                         NR_SLAB_RECLAIMABLE);
3222                         if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3223                                 break;
3224                 }
3225
3226                 /*
3227                  * Update nr_reclaimed by the number of slab pages we
3228                  * reclaimed from this zone.
3229                  */
3230                 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3231                 if (nr_slab_pages1 < nr_slab_pages0)
3232                         sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3233         }
3234
3235         p->reclaim_state = NULL;
3236         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3237         lockdep_clear_current_reclaim_state();
3238         return sc.nr_reclaimed >= nr_pages;
3239 }
3240
3241 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3242 {
3243         int node_id;
3244         int ret;
3245
3246         /*
3247          * Zone reclaim reclaims unmapped file backed pages and
3248          * slab pages if we are over the defined limits.
3249          *
3250          * A small portion of unmapped file backed pages is needed for
3251          * file I/O otherwise pages read by file I/O will be immediately
3252          * thrown out if the zone is overallocated. So we do not reclaim
3253          * if less than a specified percentage of the zone is used by
3254          * unmapped file backed pages.
3255          */
3256         if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3257             zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3258                 return ZONE_RECLAIM_FULL;
3259
3260         if (zone->all_unreclaimable)
3261                 return ZONE_RECLAIM_FULL;
3262
3263         /*
3264          * Do not scan if the allocation should not be delayed.
3265          */
3266         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3267                 return ZONE_RECLAIM_NOSCAN;
3268
3269         /*
3270          * Only run zone reclaim on the local zone or on zones that do not
3271          * have associated processors. This will favor the local processor
3272          * over remote processors and spread off node memory allocations
3273          * as wide as possible.
3274          */
3275         node_id = zone_to_nid(zone);
3276         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3277                 return ZONE_RECLAIM_NOSCAN;
3278
3279         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3280                 return ZONE_RECLAIM_NOSCAN;
3281
3282         ret = __zone_reclaim(zone, gfp_mask, order);
3283         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3284
3285         if (!ret)
3286                 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3287
3288         return ret;
3289 }
3290 #endif
3291
3292 /*
3293  * page_evictable - test whether a page is evictable
3294  * @page: the page to test
3295  * @vma: the VMA in which the page is or will be mapped, may be NULL
3296  *
3297  * Test whether page is evictable--i.e., should be placed on active/inactive
3298  * lists vs unevictable list.  The vma argument is !NULL when called from the
3299  * fault path to determine how to instantate a new page.
3300  *
3301  * Reasons page might not be evictable:
3302  * (1) page's mapping marked unevictable
3303  * (2) page is part of an mlocked VMA
3304  *
3305  */
3306 int page_evictable(struct page *page, struct vm_area_struct *vma)
3307 {
3308
3309         if (mapping_unevictable(page_mapping(page)))
3310                 return 0;
3311
3312         if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
3313                 return 0;
3314
3315         return 1;
3316 }
3317
3318 /**
3319  * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
3320  * @page: page to check evictability and move to appropriate lru list
3321  * @zone: zone page is in
3322  *
3323  * Checks a page for evictability and moves the page to the appropriate
3324  * zone lru list.
3325  *
3326  * Restrictions: zone->lru_lock must be held, page must be on LRU and must
3327  * have PageUnevictable set.
3328  */
3329 static void check_move_unevictable_page(struct page *page, struct zone *zone)
3330 {
3331         VM_BUG_ON(PageActive(page));
3332
3333 retry:
3334         ClearPageUnevictable(page);
3335         if (page_evictable(page, NULL)) {
3336                 enum lru_list l = page_lru_base_type(page);
3337
3338                 __dec_zone_state(zone, NR_UNEVICTABLE);
3339                 list_move(&page->lru, &zone->lru[l].list);
3340                 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
3341                 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
3342                 __count_vm_event(UNEVICTABLE_PGRESCUED);
3343         } else {
3344                 /*
3345                  * rotate unevictable list
3346                  */
3347                 SetPageUnevictable(page);
3348                 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
3349                 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
3350                 if (page_evictable(page, NULL))
3351                         goto retry;
3352         }
3353 }
3354
3355 /**
3356  * scan_mapping_unevictable_pages - scan an address space for evictable pages
3357  * @mapping: struct address_space to scan for evictable pages
3358  *
3359  * Scan all pages in mapping.  Check unevictable pages for
3360  * evictability and move them to the appropriate zone lru list.
3361  */
3362 void scan_mapping_unevictable_pages(struct address_space *mapping)
3363 {
3364         pgoff_t next = 0;
3365         pgoff_t end   = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
3366                          PAGE_CACHE_SHIFT;
3367         struct zone *zone;
3368         struct pagevec pvec;
3369
3370         if (mapping->nrpages == 0)
3371                 return;
3372
3373         pagevec_init(&pvec, 0);
3374         while (next < end &&
3375                 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
3376                 int i;
3377                 int pg_scanned = 0;
3378
3379                 zone = NULL;
3380
3381                 for (i = 0; i < pagevec_count(&pvec); i++) {
3382                         struct page *page = pvec.pages[i];
3383                         pgoff_t page_index = page->index;
3384                         struct zone *pagezone = page_zone(page);
3385
3386                         pg_scanned++;
3387                         if (page_index > next)
3388                                 next = page_index;
3389                         next++;
3390
3391                         if (pagezone != zone) {
3392                                 if (zone)
3393                                         spin_unlock_irq(&zone->lru_lock);
3394                                 zone = pagezone;
3395                                 spin_lock_irq(&zone->lru_lock);
3396                         }
3397
3398                         if (PageLRU(page) && PageUnevictable(page))
3399                                 check_move_unevictable_page(page, zone);
3400                 }
3401                 if (zone)
3402                         spin_unlock_irq(&zone->lru_lock);
3403                 pagevec_release(&pvec);
3404
3405                 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
3406         }
3407
3408 }
3409
3410 /**
3411  * scan_zone_unevictable_pages - check unevictable list for evictable pages
3412  * @zone - zone of which to scan the unevictable list
3413  *
3414  * Scan @zone's unevictable LRU lists to check for pages that have become
3415  * evictable.  Move those that have to @zone's inactive list where they
3416  * become candidates for reclaim, unless shrink_inactive_zone() decides
3417  * to reactivate them.  Pages that are still unevictable are rotated
3418  * back onto @zone's unevictable list.
3419  */
3420 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
3421 static void scan_zone_unevictable_pages(struct zone *zone)
3422 {
3423         struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
3424         unsigned long scan;
3425         unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
3426
3427         while (nr_to_scan > 0) {
3428                 unsigned long batch_size = min(nr_to_scan,
3429                                                 SCAN_UNEVICTABLE_BATCH_SIZE);
3430
3431                 spin_lock_irq(&zone->lru_lock);
3432                 for (scan = 0;  scan < batch_size; scan++) {
3433                         struct page *page = lru_to_page(l_unevictable);
3434
3435                         if (!trylock_page(page))
3436                                 continue;
3437
3438                         prefetchw_prev_lru_page(page, l_unevictable, flags);
3439
3440                         if (likely(PageLRU(page) && PageUnevictable(page)))
3441                                 check_move_unevictable_page(page, zone);
3442
3443                         unlock_page(page);
3444                 }
3445                 spin_unlock_irq(&zone->lru_lock);
3446
3447                 nr_to_scan -= batch_size;
3448         }
3449 }
3450
3451
3452 /**
3453  * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
3454  *
3455  * A really big hammer:  scan all zones' unevictable LRU lists to check for
3456  * pages that have become evictable.  Move those back to the zones'
3457  * inactive list where they become candidates for reclaim.
3458  * This occurs when, e.g., we have unswappable pages on the unevictable lists,
3459  * and we add swap to the system.  As such, it runs in the context of a task
3460  * that has possibly/probably made some previously unevictable pages
3461  * evictable.
3462  */
3463 static void scan_all_zones_unevictable_pages(void)
3464 {
3465         struct zone *zone;
3466
3467         for_each_zone(zone) {
3468                 scan_zone_unevictable_pages(zone);
3469         }
3470 }
3471
3472 /*
3473  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
3474  * all nodes' unevictable lists for evictable pages
3475  */
3476 unsigned long scan_unevictable_pages;
3477
3478 int scan_unevictable_handler(struct ctl_table *table, int write,
3479                            void __user *buffer,
3480                            size_t *length, loff_t *ppos)
3481 {
3482         proc_doulongvec_minmax(table, write, buffer, length, ppos);
3483
3484         if (write && *(unsigned long *)table->data)
3485                 scan_all_zones_unevictable_pages();
3486
3487         scan_unevictable_pages = 0;
3488         return 0;
3489 }
3490
3491 #ifdef CONFIG_NUMA
3492 /*
3493  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
3494  * a specified node's per zone unevictable lists for evictable pages.
3495  */
3496
3497 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
3498                                           struct sysdev_attribute *attr,
3499                                           char *buf)
3500 {
3501         return sprintf(buf, "0\n");     /* always zero; should fit... */
3502 }
3503
3504 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
3505                                            struct sysdev_attribute *attr,
3506                                         const char *buf, size_t count)
3507 {
3508         struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
3509         struct zone *zone;
3510         unsigned long res;
3511         unsigned long req = strict_strtoul(buf, 10, &res);
3512
3513         if (!req)
3514                 return 1;       /* zero is no-op */
3515
3516         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
3517                 if (!populated_zone(zone))
3518                         continue;
3519                 scan_zone_unevictable_pages(zone);
3520         }
3521         return 1;
3522 }
3523
3524
3525 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3526                         read_scan_unevictable_node,
3527                         write_scan_unevictable_node);
3528
3529 int scan_unevictable_register_node(struct node *node)
3530 {
3531         return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
3532 }
3533
3534 void scan_unevictable_unregister_node(struct node *node)
3535 {
3536         sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
3537 }
3538 #endif