vmscan: shrinker->nr updates race and go wrong
[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 swappiness;
99
100         int order;
101
102         /*
103          * Intend to reclaim enough continuous memory rather than reclaim
104          * enough amount of memory. i.e, mode for high order allocation.
105          */
106         reclaim_mode_t reclaim_mode;
107
108         /* Which cgroup do we reclaim from */
109         struct mem_cgroup *mem_cgroup;
110
111         /*
112          * Nodemask of nodes allowed by the caller. If NULL, all nodes
113          * are scanned.
114          */
115         nodemask_t      *nodemask;
116 };
117
118 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
119
120 #ifdef ARCH_HAS_PREFETCH
121 #define prefetch_prev_lru_page(_page, _base, _field)                    \
122         do {                                                            \
123                 if ((_page)->lru.prev != _base) {                       \
124                         struct page *prev;                              \
125                                                                         \
126                         prev = lru_to_page(&(_page->lru));              \
127                         prefetch(&prev->_field);                        \
128                 }                                                       \
129         } while (0)
130 #else
131 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
132 #endif
133
134 #ifdef ARCH_HAS_PREFETCHW
135 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
136         do {                                                            \
137                 if ((_page)->lru.prev != _base) {                       \
138                         struct page *prev;                              \
139                                                                         \
140                         prev = lru_to_page(&(_page->lru));              \
141                         prefetchw(&prev->_field);                       \
142                 }                                                       \
143         } while (0)
144 #else
145 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
146 #endif
147
148 /*
149  * From 0 .. 100.  Higher means more swappy.
150  */
151 int vm_swappiness = 60;
152 long vm_total_pages;    /* The total number of pages which the VM controls */
153
154 static LIST_HEAD(shrinker_list);
155 static DECLARE_RWSEM(shrinker_rwsem);
156
157 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
158 #define scanning_global_lru(sc) (!(sc)->mem_cgroup)
159 #else
160 #define scanning_global_lru(sc) (1)
161 #endif
162
163 static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
164                                                   struct scan_control *sc)
165 {
166         if (!scanning_global_lru(sc))
167                 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
168
169         return &zone->reclaim_stat;
170 }
171
172 static unsigned long zone_nr_lru_pages(struct zone *zone,
173                                 struct scan_control *sc, enum lru_list lru)
174 {
175         if (!scanning_global_lru(sc))
176                 return mem_cgroup_zone_nr_lru_pages(sc->mem_cgroup, zone, lru);
177
178         return zone_page_state(zone, NR_LRU_BASE + lru);
179 }
180
181
182 /*
183  * Add a shrinker callback to be called from the vm
184  */
185 void register_shrinker(struct shrinker *shrinker)
186 {
187         shrinker->nr = 0;
188         down_write(&shrinker_rwsem);
189         list_add_tail(&shrinker->list, &shrinker_list);
190         up_write(&shrinker_rwsem);
191 }
192 EXPORT_SYMBOL(register_shrinker);
193
194 /*
195  * Remove one
196  */
197 void unregister_shrinker(struct shrinker *shrinker)
198 {
199         down_write(&shrinker_rwsem);
200         list_del(&shrinker->list);
201         up_write(&shrinker_rwsem);
202 }
203 EXPORT_SYMBOL(unregister_shrinker);
204
205 static inline int do_shrinker_shrink(struct shrinker *shrinker,
206                                      struct shrink_control *sc,
207                                      unsigned long nr_to_scan)
208 {
209         sc->nr_to_scan = nr_to_scan;
210         return (*shrinker->shrink)(shrinker, sc);
211 }
212
213 #define SHRINK_BATCH 128
214 /*
215  * Call the shrink functions to age shrinkable caches
216  *
217  * Here we assume it costs one seek to replace a lru page and that it also
218  * takes a seek to recreate a cache object.  With this in mind we age equal
219  * percentages of the lru and ageable caches.  This should balance the seeks
220  * generated by these structures.
221  *
222  * If the vm encountered mapped pages on the LRU it increase the pressure on
223  * slab to avoid swapping.
224  *
225  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
226  *
227  * `lru_pages' represents the number of on-LRU pages in all the zones which
228  * are eligible for the caller's allocation attempt.  It is used for balancing
229  * slab reclaim versus page reclaim.
230  *
231  * Returns the number of slab objects which we shrunk.
232  */
233 unsigned long shrink_slab(struct shrink_control *shrink,
234                           unsigned long nr_pages_scanned,
235                           unsigned long lru_pages)
236 {
237         struct shrinker *shrinker;
238         unsigned long ret = 0;
239
240         if (nr_pages_scanned == 0)
241                 nr_pages_scanned = SWAP_CLUSTER_MAX;
242
243         if (!down_read_trylock(&shrinker_rwsem)) {
244                 /* Assume we'll be able to shrink next time */
245                 ret = 1;
246                 goto out;
247         }
248
249         list_for_each_entry(shrinker, &shrinker_list, list) {
250                 unsigned long long delta;
251                 unsigned long total_scan;
252                 unsigned long max_pass;
253                 int shrink_ret = 0;
254                 long nr;
255                 long new_nr;
256
257                 /*
258                  * copy the current shrinker scan count into a local variable
259                  * and zero it so that other concurrent shrinker invocations
260                  * don't also do this scanning work.
261                  */
262                 do {
263                         nr = shrinker->nr;
264                 } while (cmpxchg(&shrinker->nr, nr, 0) != nr);
265
266                 total_scan = nr;
267                 max_pass = do_shrinker_shrink(shrinker, shrink, 0);
268                 delta = (4 * nr_pages_scanned) / shrinker->seeks;
269                 delta *= max_pass;
270                 do_div(delta, lru_pages + 1);
271                 total_scan += delta;
272                 if (total_scan < 0) {
273                         printk(KERN_ERR "shrink_slab: %pF negative objects to "
274                                "delete nr=%ld\n",
275                                shrinker->shrink, total_scan);
276                         total_scan = max_pass;
277                 }
278
279                 /*
280                  * Avoid risking looping forever due to too large nr value:
281                  * never try to free more than twice the estimate number of
282                  * freeable entries.
283                  */
284                 if (total_scan > max_pass * 2)
285                         total_scan = max_pass * 2;
286
287                 trace_mm_shrink_slab_start(shrinker, shrink, nr,
288                                         nr_pages_scanned, lru_pages,
289                                         max_pass, delta, total_scan);
290
291                 while (total_scan >= SHRINK_BATCH) {
292                         long this_scan = SHRINK_BATCH;
293                         int nr_before;
294
295                         nr_before = do_shrinker_shrink(shrinker, shrink, 0);
296                         shrink_ret = do_shrinker_shrink(shrinker, shrink,
297                                                         this_scan);
298                         if (shrink_ret == -1)
299                                 break;
300                         if (shrink_ret < nr_before)
301                                 ret += nr_before - shrink_ret;
302                         count_vm_events(SLABS_SCANNED, this_scan);
303                         total_scan -= this_scan;
304
305                         cond_resched();
306                 }
307
308                 /*
309                  * move the unused scan count back into the shrinker in a
310                  * manner that handles concurrent updates. If we exhausted the
311                  * scan, there is no need to do an update.
312                  */
313                 do {
314                         nr = shrinker->nr;
315                         new_nr = total_scan + nr;
316                         if (total_scan <= 0)
317                                 break;
318                 } while (cmpxchg(&shrinker->nr, nr, new_nr) != nr);
319
320                 trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr);
321         }
322         up_read(&shrinker_rwsem);
323 out:
324         cond_resched();
325         return ret;
326 }
327
328 static void set_reclaim_mode(int priority, struct scan_control *sc,
329                                    bool sync)
330 {
331         reclaim_mode_t syncmode = sync ? RECLAIM_MODE_SYNC : RECLAIM_MODE_ASYNC;
332
333         /*
334          * Initially assume we are entering either lumpy reclaim or
335          * reclaim/compaction.Depending on the order, we will either set the
336          * sync mode or just reclaim order-0 pages later.
337          */
338         if (COMPACTION_BUILD)
339                 sc->reclaim_mode = RECLAIM_MODE_COMPACTION;
340         else
341                 sc->reclaim_mode = RECLAIM_MODE_LUMPYRECLAIM;
342
343         /*
344          * Avoid using lumpy reclaim or reclaim/compaction if possible by
345          * restricting when its set to either costly allocations or when
346          * under memory pressure
347          */
348         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
349                 sc->reclaim_mode |= syncmode;
350         else if (sc->order && priority < DEF_PRIORITY - 2)
351                 sc->reclaim_mode |= syncmode;
352         else
353                 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
354 }
355
356 static void reset_reclaim_mode(struct scan_control *sc)
357 {
358         sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
359 }
360
361 static inline int is_page_cache_freeable(struct page *page)
362 {
363         /*
364          * A freeable page cache page is referenced only by the caller
365          * that isolated the page, the page cache radix tree and
366          * optional buffer heads at page->private.
367          */
368         return page_count(page) - page_has_private(page) == 2;
369 }
370
371 static int may_write_to_queue(struct backing_dev_info *bdi,
372                               struct scan_control *sc)
373 {
374         if (current->flags & PF_SWAPWRITE)
375                 return 1;
376         if (!bdi_write_congested(bdi))
377                 return 1;
378         if (bdi == current->backing_dev_info)
379                 return 1;
380
381         /* lumpy reclaim for hugepage often need a lot of write */
382         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
383                 return 1;
384         return 0;
385 }
386
387 /*
388  * We detected a synchronous write error writing a page out.  Probably
389  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
390  * fsync(), msync() or close().
391  *
392  * The tricky part is that after writepage we cannot touch the mapping: nothing
393  * prevents it from being freed up.  But we have a ref on the page and once
394  * that page is locked, the mapping is pinned.
395  *
396  * We're allowed to run sleeping lock_page() here because we know the caller has
397  * __GFP_FS.
398  */
399 static void handle_write_error(struct address_space *mapping,
400                                 struct page *page, int error)
401 {
402         lock_page(page);
403         if (page_mapping(page) == mapping)
404                 mapping_set_error(mapping, error);
405         unlock_page(page);
406 }
407
408 /* possible outcome of pageout() */
409 typedef enum {
410         /* failed to write page out, page is locked */
411         PAGE_KEEP,
412         /* move page to the active list, page is locked */
413         PAGE_ACTIVATE,
414         /* page has been sent to the disk successfully, page is unlocked */
415         PAGE_SUCCESS,
416         /* page is clean and locked */
417         PAGE_CLEAN,
418 } pageout_t;
419
420 /*
421  * pageout is called by shrink_page_list() for each dirty page.
422  * Calls ->writepage().
423  */
424 static pageout_t pageout(struct page *page, struct address_space *mapping,
425                          struct scan_control *sc)
426 {
427         /*
428          * If the page is dirty, only perform writeback if that write
429          * will be non-blocking.  To prevent this allocation from being
430          * stalled by pagecache activity.  But note that there may be
431          * stalls if we need to run get_block().  We could test
432          * PagePrivate for that.
433          *
434          * If this process is currently in __generic_file_aio_write() against
435          * this page's queue, we can perform writeback even if that
436          * will block.
437          *
438          * If the page is swapcache, write it back even if that would
439          * block, for some throttling. This happens by accident, because
440          * swap_backing_dev_info is bust: it doesn't reflect the
441          * congestion state of the swapdevs.  Easy to fix, if needed.
442          */
443         if (!is_page_cache_freeable(page))
444                 return PAGE_KEEP;
445         if (!mapping) {
446                 /*
447                  * Some data journaling orphaned pages can have
448                  * page->mapping == NULL while being dirty with clean buffers.
449                  */
450                 if (page_has_private(page)) {
451                         if (try_to_free_buffers(page)) {
452                                 ClearPageDirty(page);
453                                 printk("%s: orphaned page\n", __func__);
454                                 return PAGE_CLEAN;
455                         }
456                 }
457                 return PAGE_KEEP;
458         }
459         if (mapping->a_ops->writepage == NULL)
460                 return PAGE_ACTIVATE;
461         if (!may_write_to_queue(mapping->backing_dev_info, sc))
462                 return PAGE_KEEP;
463
464         if (clear_page_dirty_for_io(page)) {
465                 int res;
466                 struct writeback_control wbc = {
467                         .sync_mode = WB_SYNC_NONE,
468                         .nr_to_write = SWAP_CLUSTER_MAX,
469                         .range_start = 0,
470                         .range_end = LLONG_MAX,
471                         .for_reclaim = 1,
472                 };
473
474                 SetPageReclaim(page);
475                 res = mapping->a_ops->writepage(page, &wbc);
476                 if (res < 0)
477                         handle_write_error(mapping, page, res);
478                 if (res == AOP_WRITEPAGE_ACTIVATE) {
479                         ClearPageReclaim(page);
480                         return PAGE_ACTIVATE;
481                 }
482
483                 /*
484                  * Wait on writeback if requested to. This happens when
485                  * direct reclaiming a large contiguous area and the
486                  * first attempt to free a range of pages fails.
487                  */
488                 if (PageWriteback(page) &&
489                     (sc->reclaim_mode & RECLAIM_MODE_SYNC))
490                         wait_on_page_writeback(page);
491
492                 if (!PageWriteback(page)) {
493                         /* synchronous write or broken a_ops? */
494                         ClearPageReclaim(page);
495                 }
496                 trace_mm_vmscan_writepage(page,
497                         trace_reclaim_flags(page, sc->reclaim_mode));
498                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
499                 return PAGE_SUCCESS;
500         }
501
502         return PAGE_CLEAN;
503 }
504
505 /*
506  * Same as remove_mapping, but if the page is removed from the mapping, it
507  * gets returned with a refcount of 0.
508  */
509 static int __remove_mapping(struct address_space *mapping, struct page *page)
510 {
511         BUG_ON(!PageLocked(page));
512         BUG_ON(mapping != page_mapping(page));
513
514         spin_lock_irq(&mapping->tree_lock);
515         /*
516          * The non racy check for a busy page.
517          *
518          * Must be careful with the order of the tests. When someone has
519          * a ref to the page, it may be possible that they dirty it then
520          * drop the reference. So if PageDirty is tested before page_count
521          * here, then the following race may occur:
522          *
523          * get_user_pages(&page);
524          * [user mapping goes away]
525          * write_to(page);
526          *                              !PageDirty(page)    [good]
527          * SetPageDirty(page);
528          * put_page(page);
529          *                              !page_count(page)   [good, discard it]
530          *
531          * [oops, our write_to data is lost]
532          *
533          * Reversing the order of the tests ensures such a situation cannot
534          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
535          * load is not satisfied before that of page->_count.
536          *
537          * Note that if SetPageDirty is always performed via set_page_dirty,
538          * and thus under tree_lock, then this ordering is not required.
539          */
540         if (!page_freeze_refs(page, 2))
541                 goto cannot_free;
542         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
543         if (unlikely(PageDirty(page))) {
544                 page_unfreeze_refs(page, 2);
545                 goto cannot_free;
546         }
547
548         if (PageSwapCache(page)) {
549                 swp_entry_t swap = { .val = page_private(page) };
550                 __delete_from_swap_cache(page);
551                 spin_unlock_irq(&mapping->tree_lock);
552                 swapcache_free(swap, page);
553         } else {
554                 void (*freepage)(struct page *);
555
556                 freepage = mapping->a_ops->freepage;
557
558                 __delete_from_page_cache(page);
559                 spin_unlock_irq(&mapping->tree_lock);
560                 mem_cgroup_uncharge_cache_page(page);
561
562                 if (freepage != NULL)
563                         freepage(page);
564         }
565
566         return 1;
567
568 cannot_free:
569         spin_unlock_irq(&mapping->tree_lock);
570         return 0;
571 }
572
573 /*
574  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
575  * someone else has a ref on the page, abort and return 0.  If it was
576  * successfully detached, return 1.  Assumes the caller has a single ref on
577  * this page.
578  */
579 int remove_mapping(struct address_space *mapping, struct page *page)
580 {
581         if (__remove_mapping(mapping, page)) {
582                 /*
583                  * Unfreezing the refcount with 1 rather than 2 effectively
584                  * drops the pagecache ref for us without requiring another
585                  * atomic operation.
586                  */
587                 page_unfreeze_refs(page, 1);
588                 return 1;
589         }
590         return 0;
591 }
592
593 /**
594  * putback_lru_page - put previously isolated page onto appropriate LRU list
595  * @page: page to be put back to appropriate lru list
596  *
597  * Add previously isolated @page to appropriate LRU list.
598  * Page may still be unevictable for other reasons.
599  *
600  * lru_lock must not be held, interrupts must be enabled.
601  */
602 void putback_lru_page(struct page *page)
603 {
604         int lru;
605         int active = !!TestClearPageActive(page);
606         int was_unevictable = PageUnevictable(page);
607
608         VM_BUG_ON(PageLRU(page));
609
610 redo:
611         ClearPageUnevictable(page);
612
613         if (page_evictable(page, NULL)) {
614                 /*
615                  * For evictable pages, we can use the cache.
616                  * In event of a race, worst case is we end up with an
617                  * unevictable page on [in]active list.
618                  * We know how to handle that.
619                  */
620                 lru = active + page_lru_base_type(page);
621                 lru_cache_add_lru(page, lru);
622         } else {
623                 /*
624                  * Put unevictable pages directly on zone's unevictable
625                  * list.
626                  */
627                 lru = LRU_UNEVICTABLE;
628                 add_page_to_unevictable_list(page);
629                 /*
630                  * When racing with an mlock clearing (page is
631                  * unlocked), make sure that if the other thread does
632                  * not observe our setting of PG_lru and fails
633                  * isolation, we see PG_mlocked cleared below and move
634                  * the page back to the evictable list.
635                  *
636                  * The other side is TestClearPageMlocked().
637                  */
638                 smp_mb();
639         }
640
641         /*
642          * page's status can change while we move it among lru. If an evictable
643          * page is on unevictable list, it never be freed. To avoid that,
644          * check after we added it to the list, again.
645          */
646         if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
647                 if (!isolate_lru_page(page)) {
648                         put_page(page);
649                         goto redo;
650                 }
651                 /* This means someone else dropped this page from LRU
652                  * So, it will be freed or putback to LRU again. There is
653                  * nothing to do here.
654                  */
655         }
656
657         if (was_unevictable && lru != LRU_UNEVICTABLE)
658                 count_vm_event(UNEVICTABLE_PGRESCUED);
659         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
660                 count_vm_event(UNEVICTABLE_PGCULLED);
661
662         put_page(page);         /* drop ref from isolate */
663 }
664
665 enum page_references {
666         PAGEREF_RECLAIM,
667         PAGEREF_RECLAIM_CLEAN,
668         PAGEREF_KEEP,
669         PAGEREF_ACTIVATE,
670 };
671
672 static enum page_references page_check_references(struct page *page,
673                                                   struct scan_control *sc)
674 {
675         int referenced_ptes, referenced_page;
676         unsigned long vm_flags;
677
678         referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
679         referenced_page = TestClearPageReferenced(page);
680
681         /* Lumpy reclaim - ignore references */
682         if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
683                 return PAGEREF_RECLAIM;
684
685         /*
686          * Mlock lost the isolation race with us.  Let try_to_unmap()
687          * move the page to the unevictable list.
688          */
689         if (vm_flags & VM_LOCKED)
690                 return PAGEREF_RECLAIM;
691
692         if (referenced_ptes) {
693                 if (PageAnon(page))
694                         return PAGEREF_ACTIVATE;
695                 /*
696                  * All mapped pages start out with page table
697                  * references from the instantiating fault, so we need
698                  * to look twice if a mapped file page is used more
699                  * than once.
700                  *
701                  * Mark it and spare it for another trip around the
702                  * inactive list.  Another page table reference will
703                  * lead to its activation.
704                  *
705                  * Note: the mark is set for activated pages as well
706                  * so that recently deactivated but used pages are
707                  * quickly recovered.
708                  */
709                 SetPageReferenced(page);
710
711                 if (referenced_page)
712                         return PAGEREF_ACTIVATE;
713
714                 return PAGEREF_KEEP;
715         }
716
717         /* Reclaim if clean, defer dirty pages to writeback */
718         if (referenced_page && !PageSwapBacked(page))
719                 return PAGEREF_RECLAIM_CLEAN;
720
721         return PAGEREF_RECLAIM;
722 }
723
724 static noinline_for_stack void free_page_list(struct list_head *free_pages)
725 {
726         struct pagevec freed_pvec;
727         struct page *page, *tmp;
728
729         pagevec_init(&freed_pvec, 1);
730
731         list_for_each_entry_safe(page, tmp, free_pages, lru) {
732                 list_del(&page->lru);
733                 if (!pagevec_add(&freed_pvec, page)) {
734                         __pagevec_free(&freed_pvec);
735                         pagevec_reinit(&freed_pvec);
736                 }
737         }
738
739         pagevec_free(&freed_pvec);
740 }
741
742 /*
743  * shrink_page_list() returns the number of reclaimed pages
744  */
745 static unsigned long shrink_page_list(struct list_head *page_list,
746                                       struct zone *zone,
747                                       struct scan_control *sc)
748 {
749         LIST_HEAD(ret_pages);
750         LIST_HEAD(free_pages);
751         int pgactivate = 0;
752         unsigned long nr_dirty = 0;
753         unsigned long nr_congested = 0;
754         unsigned long nr_reclaimed = 0;
755
756         cond_resched();
757
758         while (!list_empty(page_list)) {
759                 enum page_references references;
760                 struct address_space *mapping;
761                 struct page *page;
762                 int may_enter_fs;
763
764                 cond_resched();
765
766                 page = lru_to_page(page_list);
767                 list_del(&page->lru);
768
769                 if (!trylock_page(page))
770                         goto keep;
771
772                 VM_BUG_ON(PageActive(page));
773                 VM_BUG_ON(page_zone(page) != zone);
774
775                 sc->nr_scanned++;
776
777                 if (unlikely(!page_evictable(page, NULL)))
778                         goto cull_mlocked;
779
780                 if (!sc->may_unmap && page_mapped(page))
781                         goto keep_locked;
782
783                 /* Double the slab pressure for mapped and swapcache pages */
784                 if (page_mapped(page) || PageSwapCache(page))
785                         sc->nr_scanned++;
786
787                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
788                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
789
790                 if (PageWriteback(page)) {
791                         /*
792                          * Synchronous reclaim is performed in two passes,
793                          * first an asynchronous pass over the list to
794                          * start parallel writeback, and a second synchronous
795                          * pass to wait for the IO to complete.  Wait here
796                          * for any page for which writeback has already
797                          * started.
798                          */
799                         if ((sc->reclaim_mode & RECLAIM_MODE_SYNC) &&
800                             may_enter_fs)
801                                 wait_on_page_writeback(page);
802                         else {
803                                 unlock_page(page);
804                                 goto keep_lumpy;
805                         }
806                 }
807
808                 references = page_check_references(page, sc);
809                 switch (references) {
810                 case PAGEREF_ACTIVATE:
811                         goto activate_locked;
812                 case PAGEREF_KEEP:
813                         goto keep_locked;
814                 case PAGEREF_RECLAIM:
815                 case PAGEREF_RECLAIM_CLEAN:
816                         ; /* try to reclaim the page below */
817                 }
818
819                 /*
820                  * Anonymous process memory has backing store?
821                  * Try to allocate it some swap space here.
822                  */
823                 if (PageAnon(page) && !PageSwapCache(page)) {
824                         if (!(sc->gfp_mask & __GFP_IO))
825                                 goto keep_locked;
826                         if (!add_to_swap(page))
827                                 goto activate_locked;
828                         may_enter_fs = 1;
829                 }
830
831                 mapping = page_mapping(page);
832
833                 /*
834                  * The page is mapped into the page tables of one or more
835                  * processes. Try to unmap it here.
836                  */
837                 if (page_mapped(page) && mapping) {
838                         switch (try_to_unmap(page, TTU_UNMAP)) {
839                         case SWAP_FAIL:
840                                 goto activate_locked;
841                         case SWAP_AGAIN:
842                                 goto keep_locked;
843                         case SWAP_MLOCK:
844                                 goto cull_mlocked;
845                         case SWAP_SUCCESS:
846                                 ; /* try to free the page below */
847                         }
848                 }
849
850                 if (PageDirty(page)) {
851                         nr_dirty++;
852
853                         if (references == PAGEREF_RECLAIM_CLEAN)
854                                 goto keep_locked;
855                         if (!may_enter_fs)
856                                 goto keep_locked;
857                         if (!sc->may_writepage)
858                                 goto keep_locked;
859
860                         /* Page is dirty, try to write it out here */
861                         switch (pageout(page, mapping, sc)) {
862                         case PAGE_KEEP:
863                                 nr_congested++;
864                                 goto keep_locked;
865                         case PAGE_ACTIVATE:
866                                 goto activate_locked;
867                         case PAGE_SUCCESS:
868                                 if (PageWriteback(page))
869                                         goto keep_lumpy;
870                                 if (PageDirty(page))
871                                         goto keep;
872
873                                 /*
874                                  * A synchronous write - probably a ramdisk.  Go
875                                  * ahead and try to reclaim the page.
876                                  */
877                                 if (!trylock_page(page))
878                                         goto keep;
879                                 if (PageDirty(page) || PageWriteback(page))
880                                         goto keep_locked;
881                                 mapping = page_mapping(page);
882                         case PAGE_CLEAN:
883                                 ; /* try to free the page below */
884                         }
885                 }
886
887                 /*
888                  * If the page has buffers, try to free the buffer mappings
889                  * associated with this page. If we succeed we try to free
890                  * the page as well.
891                  *
892                  * We do this even if the page is PageDirty().
893                  * try_to_release_page() does not perform I/O, but it is
894                  * possible for a page to have PageDirty set, but it is actually
895                  * clean (all its buffers are clean).  This happens if the
896                  * buffers were written out directly, with submit_bh(). ext3
897                  * will do this, as well as the blockdev mapping.
898                  * try_to_release_page() will discover that cleanness and will
899                  * drop the buffers and mark the page clean - it can be freed.
900                  *
901                  * Rarely, pages can have buffers and no ->mapping.  These are
902                  * the pages which were not successfully invalidated in
903                  * truncate_complete_page().  We try to drop those buffers here
904                  * and if that worked, and the page is no longer mapped into
905                  * process address space (page_count == 1) it can be freed.
906                  * Otherwise, leave the page on the LRU so it is swappable.
907                  */
908                 if (page_has_private(page)) {
909                         if (!try_to_release_page(page, sc->gfp_mask))
910                                 goto activate_locked;
911                         if (!mapping && page_count(page) == 1) {
912                                 unlock_page(page);
913                                 if (put_page_testzero(page))
914                                         goto free_it;
915                                 else {
916                                         /*
917                                          * rare race with speculative reference.
918                                          * the speculative reference will free
919                                          * this page shortly, so we may
920                                          * increment nr_reclaimed here (and
921                                          * leave it off the LRU).
922                                          */
923                                         nr_reclaimed++;
924                                         continue;
925                                 }
926                         }
927                 }
928
929                 if (!mapping || !__remove_mapping(mapping, page))
930                         goto keep_locked;
931
932                 /*
933                  * At this point, we have no other references and there is
934                  * no way to pick any more up (removed from LRU, removed
935                  * from pagecache). Can use non-atomic bitops now (and
936                  * we obviously don't have to worry about waking up a process
937                  * waiting on the page lock, because there are no references.
938                  */
939                 __clear_page_locked(page);
940 free_it:
941                 nr_reclaimed++;
942
943                 /*
944                  * Is there need to periodically free_page_list? It would
945                  * appear not as the counts should be low
946                  */
947                 list_add(&page->lru, &free_pages);
948                 continue;
949
950 cull_mlocked:
951                 if (PageSwapCache(page))
952                         try_to_free_swap(page);
953                 unlock_page(page);
954                 putback_lru_page(page);
955                 reset_reclaim_mode(sc);
956                 continue;
957
958 activate_locked:
959                 /* Not a candidate for swapping, so reclaim swap space. */
960                 if (PageSwapCache(page) && vm_swap_full())
961                         try_to_free_swap(page);
962                 VM_BUG_ON(PageActive(page));
963                 SetPageActive(page);
964                 pgactivate++;
965 keep_locked:
966                 unlock_page(page);
967 keep:
968                 reset_reclaim_mode(sc);
969 keep_lumpy:
970                 list_add(&page->lru, &ret_pages);
971                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
972         }
973
974         /*
975          * Tag a zone as congested if all the dirty pages encountered were
976          * backed by a congested BDI. In this case, reclaimers should just
977          * back off and wait for congestion to clear because further reclaim
978          * will encounter the same problem
979          */
980         if (nr_dirty && nr_dirty == nr_congested && scanning_global_lru(sc))
981                 zone_set_flag(zone, ZONE_CONGESTED);
982
983         free_page_list(&free_pages);
984
985         list_splice(&ret_pages, page_list);
986         count_vm_events(PGACTIVATE, pgactivate);
987         return nr_reclaimed;
988 }
989
990 /*
991  * Attempt to remove the specified page from its LRU.  Only take this page
992  * if it is of the appropriate PageActive status.  Pages which are being
993  * freed elsewhere are also ignored.
994  *
995  * page:        page to consider
996  * mode:        one of the LRU isolation modes defined above
997  *
998  * returns 0 on success, -ve errno on failure.
999  */
1000 int __isolate_lru_page(struct page *page, int mode, int file)
1001 {
1002         int ret = -EINVAL;
1003
1004         /* Only take pages on the LRU. */
1005         if (!PageLRU(page))
1006                 return ret;
1007
1008         /*
1009          * When checking the active state, we need to be sure we are
1010          * dealing with comparible boolean values.  Take the logical not
1011          * of each.
1012          */
1013         if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
1014                 return ret;
1015
1016         if (mode != ISOLATE_BOTH && page_is_file_cache(page) != file)
1017                 return ret;
1018
1019         /*
1020          * When this function is being called for lumpy reclaim, we
1021          * initially look into all LRU pages, active, inactive and
1022          * unevictable; only give shrink_page_list evictable pages.
1023          */
1024         if (PageUnevictable(page))
1025                 return ret;
1026
1027         ret = -EBUSY;
1028
1029         if (likely(get_page_unless_zero(page))) {
1030                 /*
1031                  * Be careful not to clear PageLRU until after we're
1032                  * sure the page is not being freed elsewhere -- the
1033                  * page release code relies on it.
1034                  */
1035                 ClearPageLRU(page);
1036                 ret = 0;
1037         }
1038
1039         return ret;
1040 }
1041
1042 /*
1043  * zone->lru_lock is heavily contended.  Some of the functions that
1044  * shrink the lists perform better by taking out a batch of pages
1045  * and working on them outside the LRU lock.
1046  *
1047  * For pagecache intensive workloads, this function is the hottest
1048  * spot in the kernel (apart from copy_*_user functions).
1049  *
1050  * Appropriate locks must be held before calling this function.
1051  *
1052  * @nr_to_scan: The number of pages to look through on the list.
1053  * @src:        The LRU list to pull pages off.
1054  * @dst:        The temp list to put pages on to.
1055  * @scanned:    The number of pages that were scanned.
1056  * @order:      The caller's attempted allocation order
1057  * @mode:       One of the LRU isolation modes
1058  * @file:       True [1] if isolating file [!anon] pages
1059  *
1060  * returns how many pages were moved onto *@dst.
1061  */
1062 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1063                 struct list_head *src, struct list_head *dst,
1064                 unsigned long *scanned, int order, int mode, int file)
1065 {
1066         unsigned long nr_taken = 0;
1067         unsigned long nr_lumpy_taken = 0;
1068         unsigned long nr_lumpy_dirty = 0;
1069         unsigned long nr_lumpy_failed = 0;
1070         unsigned long scan;
1071
1072         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1073                 struct page *page;
1074                 unsigned long pfn;
1075                 unsigned long end_pfn;
1076                 unsigned long page_pfn;
1077                 int zone_id;
1078
1079                 page = lru_to_page(src);
1080                 prefetchw_prev_lru_page(page, src, flags);
1081
1082                 VM_BUG_ON(!PageLRU(page));
1083
1084                 switch (__isolate_lru_page(page, mode, file)) {
1085                 case 0:
1086                         list_move(&page->lru, dst);
1087                         mem_cgroup_del_lru(page);
1088                         nr_taken += hpage_nr_pages(page);
1089                         break;
1090
1091                 case -EBUSY:
1092                         /* else it is being freed elsewhere */
1093                         list_move(&page->lru, src);
1094                         mem_cgroup_rotate_lru_list(page, page_lru(page));
1095                         continue;
1096
1097                 default:
1098                         BUG();
1099                 }
1100
1101                 if (!order)
1102                         continue;
1103
1104                 /*
1105                  * Attempt to take all pages in the order aligned region
1106                  * surrounding the tag page.  Only take those pages of
1107                  * the same active state as that tag page.  We may safely
1108                  * round the target page pfn down to the requested order
1109                  * as the mem_map is guaranteed valid out to MAX_ORDER,
1110                  * where that page is in a different zone we will detect
1111                  * it from its zone id and abort this block scan.
1112                  */
1113                 zone_id = page_zone_id(page);
1114                 page_pfn = page_to_pfn(page);
1115                 pfn = page_pfn & ~((1 << order) - 1);
1116                 end_pfn = pfn + (1 << order);
1117                 for (; pfn < end_pfn; pfn++) {
1118                         struct page *cursor_page;
1119
1120                         /* The target page is in the block, ignore it. */
1121                         if (unlikely(pfn == page_pfn))
1122                                 continue;
1123
1124                         /* Avoid holes within the zone. */
1125                         if (unlikely(!pfn_valid_within(pfn)))
1126                                 break;
1127
1128                         cursor_page = pfn_to_page(pfn);
1129
1130                         /* Check that we have not crossed a zone boundary. */
1131                         if (unlikely(page_zone_id(cursor_page) != zone_id))
1132                                 break;
1133
1134                         /*
1135                          * If we don't have enough swap space, reclaiming of
1136                          * anon page which don't already have a swap slot is
1137                          * pointless.
1138                          */
1139                         if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
1140                             !PageSwapCache(cursor_page))
1141                                 break;
1142
1143                         if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1144                                 list_move(&cursor_page->lru, dst);
1145                                 mem_cgroup_del_lru(cursor_page);
1146                                 nr_taken += hpage_nr_pages(page);
1147                                 nr_lumpy_taken++;
1148                                 if (PageDirty(cursor_page))
1149                                         nr_lumpy_dirty++;
1150                                 scan++;
1151                         } else {
1152                                 /*
1153                                  * Check if the page is freed already.
1154                                  *
1155                                  * We can't use page_count() as that
1156                                  * requires compound_head and we don't
1157                                  * have a pin on the page here. If a
1158                                  * page is tail, we may or may not
1159                                  * have isolated the head, so assume
1160                                  * it's not free, it'd be tricky to
1161                                  * track the head status without a
1162                                  * page pin.
1163                                  */
1164                                 if (!PageTail(cursor_page) &&
1165                                     !atomic_read(&cursor_page->_count))
1166                                         continue;
1167                                 break;
1168                         }
1169                 }
1170
1171                 /* If we break out of the loop above, lumpy reclaim failed */
1172                 if (pfn < end_pfn)
1173                         nr_lumpy_failed++;
1174         }
1175
1176         *scanned = scan;
1177
1178         trace_mm_vmscan_lru_isolate(order,
1179                         nr_to_scan, scan,
1180                         nr_taken,
1181                         nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1182                         mode);
1183         return nr_taken;
1184 }
1185
1186 static unsigned long isolate_pages_global(unsigned long nr,
1187                                         struct list_head *dst,
1188                                         unsigned long *scanned, int order,
1189                                         int mode, struct zone *z,
1190                                         int active, int file)
1191 {
1192         int lru = LRU_BASE;
1193         if (active)
1194                 lru += LRU_ACTIVE;
1195         if (file)
1196                 lru += LRU_FILE;
1197         return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1198                                                                 mode, file);
1199 }
1200
1201 /*
1202  * clear_active_flags() is a helper for shrink_active_list(), clearing
1203  * any active bits from the pages in the list.
1204  */
1205 static unsigned long clear_active_flags(struct list_head *page_list,
1206                                         unsigned int *count)
1207 {
1208         int nr_active = 0;
1209         int lru;
1210         struct page *page;
1211
1212         list_for_each_entry(page, page_list, lru) {
1213                 int numpages = hpage_nr_pages(page);
1214                 lru = page_lru_base_type(page);
1215                 if (PageActive(page)) {
1216                         lru += LRU_ACTIVE;
1217                         ClearPageActive(page);
1218                         nr_active += numpages;
1219                 }
1220                 if (count)
1221                         count[lru] += numpages;
1222         }
1223
1224         return nr_active;
1225 }
1226
1227 /**
1228  * isolate_lru_page - tries to isolate a page from its LRU list
1229  * @page: page to isolate from its LRU list
1230  *
1231  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1232  * vmstat statistic corresponding to whatever LRU list the page was on.
1233  *
1234  * Returns 0 if the page was removed from an LRU list.
1235  * Returns -EBUSY if the page was not on an LRU list.
1236  *
1237  * The returned page will have PageLRU() cleared.  If it was found on
1238  * the active list, it will have PageActive set.  If it was found on
1239  * the unevictable list, it will have the PageUnevictable bit set. That flag
1240  * may need to be cleared by the caller before letting the page go.
1241  *
1242  * The vmstat statistic corresponding to the list on which the page was
1243  * found will be decremented.
1244  *
1245  * Restrictions:
1246  * (1) Must be called with an elevated refcount on the page. This is a
1247  *     fundamentnal difference from isolate_lru_pages (which is called
1248  *     without a stable reference).
1249  * (2) the lru_lock must not be held.
1250  * (3) interrupts must be enabled.
1251  */
1252 int isolate_lru_page(struct page *page)
1253 {
1254         int ret = -EBUSY;
1255
1256         VM_BUG_ON(!page_count(page));
1257
1258         if (PageLRU(page)) {
1259                 struct zone *zone = page_zone(page);
1260
1261                 spin_lock_irq(&zone->lru_lock);
1262                 if (PageLRU(page)) {
1263                         int lru = page_lru(page);
1264                         ret = 0;
1265                         get_page(page);
1266                         ClearPageLRU(page);
1267
1268                         del_page_from_lru_list(zone, page, lru);
1269                 }
1270                 spin_unlock_irq(&zone->lru_lock);
1271         }
1272         return ret;
1273 }
1274
1275 /*
1276  * Are there way too many processes in the direct reclaim path already?
1277  */
1278 static int too_many_isolated(struct zone *zone, int file,
1279                 struct scan_control *sc)
1280 {
1281         unsigned long inactive, isolated;
1282
1283         if (current_is_kswapd())
1284                 return 0;
1285
1286         if (!scanning_global_lru(sc))
1287                 return 0;
1288
1289         if (file) {
1290                 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1291                 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1292         } else {
1293                 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1294                 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1295         }
1296
1297         return isolated > inactive;
1298 }
1299
1300 /*
1301  * TODO: Try merging with migrations version of putback_lru_pages
1302  */
1303 static noinline_for_stack void
1304 putback_lru_pages(struct zone *zone, struct scan_control *sc,
1305                                 unsigned long nr_anon, unsigned long nr_file,
1306                                 struct list_head *page_list)
1307 {
1308         struct page *page;
1309         struct pagevec pvec;
1310         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1311
1312         pagevec_init(&pvec, 1);
1313
1314         /*
1315          * Put back any unfreeable pages.
1316          */
1317         spin_lock(&zone->lru_lock);
1318         while (!list_empty(page_list)) {
1319                 int lru;
1320                 page = lru_to_page(page_list);
1321                 VM_BUG_ON(PageLRU(page));
1322                 list_del(&page->lru);
1323                 if (unlikely(!page_evictable(page, NULL))) {
1324                         spin_unlock_irq(&zone->lru_lock);
1325                         putback_lru_page(page);
1326                         spin_lock_irq(&zone->lru_lock);
1327                         continue;
1328                 }
1329                 SetPageLRU(page);
1330                 lru = page_lru(page);
1331                 add_page_to_lru_list(zone, page, lru);
1332                 if (is_active_lru(lru)) {
1333                         int file = is_file_lru(lru);
1334                         int numpages = hpage_nr_pages(page);
1335                         reclaim_stat->recent_rotated[file] += numpages;
1336                 }
1337                 if (!pagevec_add(&pvec, page)) {
1338                         spin_unlock_irq(&zone->lru_lock);
1339                         __pagevec_release(&pvec);
1340                         spin_lock_irq(&zone->lru_lock);
1341                 }
1342         }
1343         __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1344         __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1345
1346         spin_unlock_irq(&zone->lru_lock);
1347         pagevec_release(&pvec);
1348 }
1349
1350 static noinline_for_stack void update_isolated_counts(struct zone *zone,
1351                                         struct scan_control *sc,
1352                                         unsigned long *nr_anon,
1353                                         unsigned long *nr_file,
1354                                         struct list_head *isolated_list)
1355 {
1356         unsigned long nr_active;
1357         unsigned int count[NR_LRU_LISTS] = { 0, };
1358         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1359
1360         nr_active = clear_active_flags(isolated_list, count);
1361         __count_vm_events(PGDEACTIVATE, nr_active);
1362
1363         __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1364                               -count[LRU_ACTIVE_FILE]);
1365         __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1366                               -count[LRU_INACTIVE_FILE]);
1367         __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1368                               -count[LRU_ACTIVE_ANON]);
1369         __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1370                               -count[LRU_INACTIVE_ANON]);
1371
1372         *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1373         *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1374         __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1375         __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1376
1377         reclaim_stat->recent_scanned[0] += *nr_anon;
1378         reclaim_stat->recent_scanned[1] += *nr_file;
1379 }
1380
1381 /*
1382  * Returns true if the caller should wait to clean dirty/writeback pages.
1383  *
1384  * If we are direct reclaiming for contiguous pages and we do not reclaim
1385  * everything in the list, try again and wait for writeback IO to complete.
1386  * This will stall high-order allocations noticeably. Only do that when really
1387  * need to free the pages under high memory pressure.
1388  */
1389 static inline bool should_reclaim_stall(unsigned long nr_taken,
1390                                         unsigned long nr_freed,
1391                                         int priority,
1392                                         struct scan_control *sc)
1393 {
1394         int lumpy_stall_priority;
1395
1396         /* kswapd should not stall on sync IO */
1397         if (current_is_kswapd())
1398                 return false;
1399
1400         /* Only stall on lumpy reclaim */
1401         if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
1402                 return false;
1403
1404         /* If we have relaimed everything on the isolated list, no stall */
1405         if (nr_freed == nr_taken)
1406                 return false;
1407
1408         /*
1409          * For high-order allocations, there are two stall thresholds.
1410          * High-cost allocations stall immediately where as lower
1411          * order allocations such as stacks require the scanning
1412          * priority to be much higher before stalling.
1413          */
1414         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1415                 lumpy_stall_priority = DEF_PRIORITY;
1416         else
1417                 lumpy_stall_priority = DEF_PRIORITY / 3;
1418
1419         return priority <= lumpy_stall_priority;
1420 }
1421
1422 /*
1423  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1424  * of reclaimed pages
1425  */
1426 static noinline_for_stack unsigned long
1427 shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1428                         struct scan_control *sc, int priority, int file)
1429 {
1430         LIST_HEAD(page_list);
1431         unsigned long nr_scanned;
1432         unsigned long nr_reclaimed = 0;
1433         unsigned long nr_taken;
1434         unsigned long nr_anon;
1435         unsigned long nr_file;
1436
1437         while (unlikely(too_many_isolated(zone, file, sc))) {
1438                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1439
1440                 /* We are about to die and free our memory. Return now. */
1441                 if (fatal_signal_pending(current))
1442                         return SWAP_CLUSTER_MAX;
1443         }
1444
1445         set_reclaim_mode(priority, sc, false);
1446         lru_add_drain();
1447         spin_lock_irq(&zone->lru_lock);
1448
1449         if (scanning_global_lru(sc)) {
1450                 nr_taken = isolate_pages_global(nr_to_scan,
1451                         &page_list, &nr_scanned, sc->order,
1452                         sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM ?
1453                                         ISOLATE_BOTH : ISOLATE_INACTIVE,
1454                         zone, 0, file);
1455                 zone->pages_scanned += nr_scanned;
1456                 if (current_is_kswapd())
1457                         __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1458                                                nr_scanned);
1459                 else
1460                         __count_zone_vm_events(PGSCAN_DIRECT, zone,
1461                                                nr_scanned);
1462         } else {
1463                 nr_taken = mem_cgroup_isolate_pages(nr_to_scan,
1464                         &page_list, &nr_scanned, sc->order,
1465                         sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM ?
1466                                         ISOLATE_BOTH : ISOLATE_INACTIVE,
1467                         zone, sc->mem_cgroup,
1468                         0, file);
1469                 /*
1470                  * mem_cgroup_isolate_pages() keeps track of
1471                  * scanned pages on its own.
1472                  */
1473         }
1474
1475         if (nr_taken == 0) {
1476                 spin_unlock_irq(&zone->lru_lock);
1477                 return 0;
1478         }
1479
1480         update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
1481
1482         spin_unlock_irq(&zone->lru_lock);
1483
1484         nr_reclaimed = shrink_page_list(&page_list, zone, sc);
1485
1486         /* Check if we should syncronously wait for writeback */
1487         if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1488                 set_reclaim_mode(priority, sc, true);
1489                 nr_reclaimed += shrink_page_list(&page_list, zone, sc);
1490         }
1491
1492         local_irq_disable();
1493         if (current_is_kswapd())
1494                 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1495         __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1496
1497         putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
1498
1499         trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1500                 zone_idx(zone),
1501                 nr_scanned, nr_reclaimed,
1502                 priority,
1503                 trace_shrink_flags(file, sc->reclaim_mode));
1504         return nr_reclaimed;
1505 }
1506
1507 /*
1508  * This moves pages from the active list to the inactive list.
1509  *
1510  * We move them the other way if the page is referenced by one or more
1511  * processes, from rmap.
1512  *
1513  * If the pages are mostly unmapped, the processing is fast and it is
1514  * appropriate to hold zone->lru_lock across the whole operation.  But if
1515  * the pages are mapped, the processing is slow (page_referenced()) so we
1516  * should drop zone->lru_lock around each page.  It's impossible to balance
1517  * this, so instead we remove the pages from the LRU while processing them.
1518  * It is safe to rely on PG_active against the non-LRU pages in here because
1519  * nobody will play with that bit on a non-LRU page.
1520  *
1521  * The downside is that we have to touch page->_count against each page.
1522  * But we had to alter page->flags anyway.
1523  */
1524
1525 static void move_active_pages_to_lru(struct zone *zone,
1526                                      struct list_head *list,
1527                                      enum lru_list lru)
1528 {
1529         unsigned long pgmoved = 0;
1530         struct pagevec pvec;
1531         struct page *page;
1532
1533         pagevec_init(&pvec, 1);
1534
1535         while (!list_empty(list)) {
1536                 page = lru_to_page(list);
1537
1538                 VM_BUG_ON(PageLRU(page));
1539                 SetPageLRU(page);
1540
1541                 list_move(&page->lru, &zone->lru[lru].list);
1542                 mem_cgroup_add_lru_list(page, lru);
1543                 pgmoved += hpage_nr_pages(page);
1544
1545                 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1546                         spin_unlock_irq(&zone->lru_lock);
1547                         if (buffer_heads_over_limit)
1548                                 pagevec_strip(&pvec);
1549                         __pagevec_release(&pvec);
1550                         spin_lock_irq(&zone->lru_lock);
1551                 }
1552         }
1553         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1554         if (!is_active_lru(lru))
1555                 __count_vm_events(PGDEACTIVATE, pgmoved);
1556 }
1557
1558 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1559                         struct scan_control *sc, int priority, int file)
1560 {
1561         unsigned long nr_taken;
1562         unsigned long pgscanned;
1563         unsigned long vm_flags;
1564         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1565         LIST_HEAD(l_active);
1566         LIST_HEAD(l_inactive);
1567         struct page *page;
1568         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1569         unsigned long nr_rotated = 0;
1570
1571         lru_add_drain();
1572         spin_lock_irq(&zone->lru_lock);
1573         if (scanning_global_lru(sc)) {
1574                 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1575                                                 &pgscanned, sc->order,
1576                                                 ISOLATE_ACTIVE, zone,
1577                                                 1, file);
1578                 zone->pages_scanned += pgscanned;
1579         } else {
1580                 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1581                                                 &pgscanned, sc->order,
1582                                                 ISOLATE_ACTIVE, zone,
1583                                                 sc->mem_cgroup, 1, file);
1584                 /*
1585                  * mem_cgroup_isolate_pages() keeps track of
1586                  * scanned pages on its own.
1587                  */
1588         }
1589
1590         reclaim_stat->recent_scanned[file] += nr_taken;
1591
1592         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1593         if (file)
1594                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1595         else
1596                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1597         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1598         spin_unlock_irq(&zone->lru_lock);
1599
1600         while (!list_empty(&l_hold)) {
1601                 cond_resched();
1602                 page = lru_to_page(&l_hold);
1603                 list_del(&page->lru);
1604
1605                 if (unlikely(!page_evictable(page, NULL))) {
1606                         putback_lru_page(page);
1607                         continue;
1608                 }
1609
1610                 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1611                         nr_rotated += hpage_nr_pages(page);
1612                         /*
1613                          * Identify referenced, file-backed active pages and
1614                          * give them one more trip around the active list. So
1615                          * that executable code get better chances to stay in
1616                          * memory under moderate memory pressure.  Anon pages
1617                          * are not likely to be evicted by use-once streaming
1618                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1619                          * so we ignore them here.
1620                          */
1621                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1622                                 list_add(&page->lru, &l_active);
1623                                 continue;
1624                         }
1625                 }
1626
1627                 ClearPageActive(page);  /* we are de-activating */
1628                 list_add(&page->lru, &l_inactive);
1629         }
1630
1631         /*
1632          * Move pages back to the lru list.
1633          */
1634         spin_lock_irq(&zone->lru_lock);
1635         /*
1636          * Count referenced pages from currently used mappings as rotated,
1637          * even though only some of them are actually re-activated.  This
1638          * helps balance scan pressure between file and anonymous pages in
1639          * get_scan_ratio.
1640          */
1641         reclaim_stat->recent_rotated[file] += nr_rotated;
1642
1643         move_active_pages_to_lru(zone, &l_active,
1644                                                 LRU_ACTIVE + file * LRU_FILE);
1645         move_active_pages_to_lru(zone, &l_inactive,
1646                                                 LRU_BASE   + file * LRU_FILE);
1647         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1648         spin_unlock_irq(&zone->lru_lock);
1649 }
1650
1651 #ifdef CONFIG_SWAP
1652 static int inactive_anon_is_low_global(struct zone *zone)
1653 {
1654         unsigned long active, inactive;
1655
1656         active = zone_page_state(zone, NR_ACTIVE_ANON);
1657         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1658
1659         if (inactive * zone->inactive_ratio < active)
1660                 return 1;
1661
1662         return 0;
1663 }
1664
1665 /**
1666  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1667  * @zone: zone to check
1668  * @sc:   scan control of this context
1669  *
1670  * Returns true if the zone does not have enough inactive anon pages,
1671  * meaning some active anon pages need to be deactivated.
1672  */
1673 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1674 {
1675         int low;
1676
1677         /*
1678          * If we don't have swap space, anonymous page deactivation
1679          * is pointless.
1680          */
1681         if (!total_swap_pages)
1682                 return 0;
1683
1684         if (scanning_global_lru(sc))
1685                 low = inactive_anon_is_low_global(zone);
1686         else
1687                 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1688         return low;
1689 }
1690 #else
1691 static inline int inactive_anon_is_low(struct zone *zone,
1692                                         struct scan_control *sc)
1693 {
1694         return 0;
1695 }
1696 #endif
1697
1698 static int inactive_file_is_low_global(struct zone *zone)
1699 {
1700         unsigned long active, inactive;
1701
1702         active = zone_page_state(zone, NR_ACTIVE_FILE);
1703         inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1704
1705         return (active > inactive);
1706 }
1707
1708 /**
1709  * inactive_file_is_low - check if file pages need to be deactivated
1710  * @zone: zone to check
1711  * @sc:   scan control of this context
1712  *
1713  * When the system is doing streaming IO, memory pressure here
1714  * ensures that active file pages get deactivated, until more
1715  * than half of the file pages are on the inactive list.
1716  *
1717  * Once we get to that situation, protect the system's working
1718  * set from being evicted by disabling active file page aging.
1719  *
1720  * This uses a different ratio than the anonymous pages, because
1721  * the page cache uses a use-once replacement algorithm.
1722  */
1723 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1724 {
1725         int low;
1726
1727         if (scanning_global_lru(sc))
1728                 low = inactive_file_is_low_global(zone);
1729         else
1730                 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1731         return low;
1732 }
1733
1734 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1735                                 int file)
1736 {
1737         if (file)
1738                 return inactive_file_is_low(zone, sc);
1739         else
1740                 return inactive_anon_is_low(zone, sc);
1741 }
1742
1743 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1744         struct zone *zone, struct scan_control *sc, int priority)
1745 {
1746         int file = is_file_lru(lru);
1747
1748         if (is_active_lru(lru)) {
1749                 if (inactive_list_is_low(zone, sc, file))
1750                     shrink_active_list(nr_to_scan, zone, sc, priority, file);
1751                 return 0;
1752         }
1753
1754         return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1755 }
1756
1757 /*
1758  * Determine how aggressively the anon and file LRU lists should be
1759  * scanned.  The relative value of each set of LRU lists is determined
1760  * by looking at the fraction of the pages scanned we did rotate back
1761  * onto the active list instead of evict.
1762  *
1763  * nr[0] = anon pages to scan; nr[1] = file pages to scan
1764  */
1765 static void get_scan_count(struct zone *zone, struct scan_control *sc,
1766                                         unsigned long *nr, int priority)
1767 {
1768         unsigned long anon, file, free;
1769         unsigned long anon_prio, file_prio;
1770         unsigned long ap, fp;
1771         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1772         u64 fraction[2], denominator;
1773         enum lru_list l;
1774         int noswap = 0;
1775         int force_scan = 0;
1776
1777
1778         anon  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1779                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1780         file  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1781                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1782
1783         if (((anon + file) >> priority) < SWAP_CLUSTER_MAX) {
1784                 /* kswapd does zone balancing and need to scan this zone */
1785                 if (scanning_global_lru(sc) && current_is_kswapd())
1786                         force_scan = 1;
1787                 /* memcg may have small limit and need to avoid priority drop */
1788                 if (!scanning_global_lru(sc))
1789                         force_scan = 1;
1790         }
1791
1792         /* If we have no swap space, do not bother scanning anon pages. */
1793         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1794                 noswap = 1;
1795                 fraction[0] = 0;
1796                 fraction[1] = 1;
1797                 denominator = 1;
1798                 goto out;
1799         }
1800
1801         if (scanning_global_lru(sc)) {
1802                 free  = zone_page_state(zone, NR_FREE_PAGES);
1803                 /* If we have very few page cache pages,
1804                    force-scan anon pages. */
1805                 if (unlikely(file + free <= high_wmark_pages(zone))) {
1806                         fraction[0] = 1;
1807                         fraction[1] = 0;
1808                         denominator = 1;
1809                         goto out;
1810                 }
1811         }
1812
1813         /*
1814          * With swappiness at 100, anonymous and file have the same priority.
1815          * This scanning priority is essentially the inverse of IO cost.
1816          */
1817         anon_prio = sc->swappiness;
1818         file_prio = 200 - sc->swappiness;
1819
1820         /*
1821          * OK, so we have swap space and a fair amount of page cache
1822          * pages.  We use the recently rotated / recently scanned
1823          * ratios to determine how valuable each cache is.
1824          *
1825          * Because workloads change over time (and to avoid overflow)
1826          * we keep these statistics as a floating average, which ends
1827          * up weighing recent references more than old ones.
1828          *
1829          * anon in [0], file in [1]
1830          */
1831         spin_lock_irq(&zone->lru_lock);
1832         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1833                 reclaim_stat->recent_scanned[0] /= 2;
1834                 reclaim_stat->recent_rotated[0] /= 2;
1835         }
1836
1837         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1838                 reclaim_stat->recent_scanned[1] /= 2;
1839                 reclaim_stat->recent_rotated[1] /= 2;
1840         }
1841
1842         /*
1843          * The amount of pressure on anon vs file pages is inversely
1844          * proportional to the fraction of recently scanned pages on
1845          * each list that were recently referenced and in active use.
1846          */
1847         ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1848         ap /= reclaim_stat->recent_rotated[0] + 1;
1849
1850         fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1851         fp /= reclaim_stat->recent_rotated[1] + 1;
1852         spin_unlock_irq(&zone->lru_lock);
1853
1854         fraction[0] = ap;
1855         fraction[1] = fp;
1856         denominator = ap + fp + 1;
1857 out:
1858         for_each_evictable_lru(l) {
1859                 int file = is_file_lru(l);
1860                 unsigned long scan;
1861
1862                 scan = zone_nr_lru_pages(zone, sc, l);
1863                 if (priority || noswap) {
1864                         scan >>= priority;
1865                         scan = div64_u64(scan * fraction[file], denominator);
1866                 }
1867
1868                 /*
1869                  * If zone is small or memcg is small, nr[l] can be 0.
1870                  * This results no-scan on this priority and priority drop down.
1871                  * For global direct reclaim, it can visit next zone and tend
1872                  * not to have problems. For global kswapd, it's for zone
1873                  * balancing and it need to scan a small amounts. When using
1874                  * memcg, priority drop can cause big latency. So, it's better
1875                  * to scan small amount. See may_noscan above.
1876                  */
1877                 if (!scan && force_scan) {
1878                         if (file)
1879                                 scan = SWAP_CLUSTER_MAX;
1880                         else if (!noswap)
1881                                 scan = SWAP_CLUSTER_MAX;
1882                 }
1883                 nr[l] = scan;
1884         }
1885 }
1886
1887 /*
1888  * Reclaim/compaction depends on a number of pages being freed. To avoid
1889  * disruption to the system, a small number of order-0 pages continue to be
1890  * rotated and reclaimed in the normal fashion. However, by the time we get
1891  * back to the allocator and call try_to_compact_zone(), we ensure that
1892  * there are enough free pages for it to be likely successful
1893  */
1894 static inline bool should_continue_reclaim(struct zone *zone,
1895                                         unsigned long nr_reclaimed,
1896                                         unsigned long nr_scanned,
1897                                         struct scan_control *sc)
1898 {
1899         unsigned long pages_for_compaction;
1900         unsigned long inactive_lru_pages;
1901
1902         /* If not in reclaim/compaction mode, stop */
1903         if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
1904                 return false;
1905
1906         /* Consider stopping depending on scan and reclaim activity */
1907         if (sc->gfp_mask & __GFP_REPEAT) {
1908                 /*
1909                  * For __GFP_REPEAT allocations, stop reclaiming if the
1910                  * full LRU list has been scanned and we are still failing
1911                  * to reclaim pages. This full LRU scan is potentially
1912                  * expensive but a __GFP_REPEAT caller really wants to succeed
1913                  */
1914                 if (!nr_reclaimed && !nr_scanned)
1915                         return false;
1916         } else {
1917                 /*
1918                  * For non-__GFP_REPEAT allocations which can presumably
1919                  * fail without consequence, stop if we failed to reclaim
1920                  * any pages from the last SWAP_CLUSTER_MAX number of
1921                  * pages that were scanned. This will return to the
1922                  * caller faster at the risk reclaim/compaction and
1923                  * the resulting allocation attempt fails
1924                  */
1925                 if (!nr_reclaimed)
1926                         return false;
1927         }
1928
1929         /*
1930          * If we have not reclaimed enough pages for compaction and the
1931          * inactive lists are large enough, continue reclaiming
1932          */
1933         pages_for_compaction = (2UL << sc->order);
1934         inactive_lru_pages = zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON) +
1935                                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1936         if (sc->nr_reclaimed < pages_for_compaction &&
1937                         inactive_lru_pages > pages_for_compaction)
1938                 return true;
1939
1940         /* If compaction would go ahead or the allocation would succeed, stop */
1941         switch (compaction_suitable(zone, sc->order)) {
1942         case COMPACT_PARTIAL:
1943         case COMPACT_CONTINUE:
1944                 return false;
1945         default:
1946                 return true;
1947         }
1948 }
1949
1950 /*
1951  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1952  */
1953 static void shrink_zone(int priority, struct zone *zone,
1954                                 struct scan_control *sc)
1955 {
1956         unsigned long nr[NR_LRU_LISTS];
1957         unsigned long nr_to_scan;
1958         enum lru_list l;
1959         unsigned long nr_reclaimed, nr_scanned;
1960         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1961
1962 restart:
1963         nr_reclaimed = 0;
1964         nr_scanned = sc->nr_scanned;
1965         get_scan_count(zone, sc, nr, priority);
1966
1967         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1968                                         nr[LRU_INACTIVE_FILE]) {
1969                 for_each_evictable_lru(l) {
1970                         if (nr[l]) {
1971                                 nr_to_scan = min_t(unsigned long,
1972                                                    nr[l], SWAP_CLUSTER_MAX);
1973                                 nr[l] -= nr_to_scan;
1974
1975                                 nr_reclaimed += shrink_list(l, nr_to_scan,
1976                                                             zone, sc, priority);
1977                         }
1978                 }
1979                 /*
1980                  * On large memory systems, scan >> priority can become
1981                  * really large. This is fine for the starting priority;
1982                  * we want to put equal scanning pressure on each zone.
1983                  * However, if the VM has a harder time of freeing pages,
1984                  * with multiple processes reclaiming pages, the total
1985                  * freeing target can get unreasonably large.
1986                  */
1987                 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
1988                         break;
1989         }
1990         sc->nr_reclaimed += nr_reclaimed;
1991
1992         /*
1993          * Even if we did not try to evict anon pages at all, we want to
1994          * rebalance the anon lru active/inactive ratio.
1995          */
1996         if (inactive_anon_is_low(zone, sc))
1997                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1998
1999         /* reclaim/compaction might need reclaim to continue */
2000         if (should_continue_reclaim(zone, nr_reclaimed,
2001                                         sc->nr_scanned - nr_scanned, sc))
2002                 goto restart;
2003
2004         throttle_vm_writeout(sc->gfp_mask);
2005 }
2006
2007 /*
2008  * This is the direct reclaim path, for page-allocating processes.  We only
2009  * try to reclaim pages from zones which will satisfy the caller's allocation
2010  * request.
2011  *
2012  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2013  * Because:
2014  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2015  *    allocation or
2016  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2017  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2018  *    zone defense algorithm.
2019  *
2020  * If a zone is deemed to be full of pinned pages then just give it a light
2021  * scan then give up on it.
2022  */
2023 static void shrink_zones(int priority, struct zonelist *zonelist,
2024                                         struct scan_control *sc)
2025 {
2026         struct zoneref *z;
2027         struct zone *zone;
2028         unsigned long nr_soft_reclaimed;
2029         unsigned long nr_soft_scanned;
2030
2031         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2032                                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2033                 if (!populated_zone(zone))
2034                         continue;
2035                 /*
2036                  * Take care memory controller reclaiming has small influence
2037                  * to global LRU.
2038                  */
2039                 if (scanning_global_lru(sc)) {
2040                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2041                                 continue;
2042                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2043                                 continue;       /* Let kswapd poll it */
2044                         /*
2045                          * This steals pages from memory cgroups over softlimit
2046                          * and returns the number of reclaimed pages and
2047                          * scanned pages. This works for global memory pressure
2048                          * and balancing, not for a memcg's limit.
2049                          */
2050                         nr_soft_scanned = 0;
2051                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2052                                                 sc->order, sc->gfp_mask,
2053                                                 &nr_soft_scanned);
2054                         sc->nr_reclaimed += nr_soft_reclaimed;
2055                         sc->nr_scanned += nr_soft_scanned;
2056                         /* need some check for avoid more shrink_zone() */
2057                 }
2058
2059                 shrink_zone(priority, zone, sc);
2060         }
2061 }
2062
2063 static bool zone_reclaimable(struct zone *zone)
2064 {
2065         return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2066 }
2067
2068 /* All zones in zonelist are unreclaimable? */
2069 static bool all_unreclaimable(struct zonelist *zonelist,
2070                 struct scan_control *sc)
2071 {
2072         struct zoneref *z;
2073         struct zone *zone;
2074
2075         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2076                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2077                 if (!populated_zone(zone))
2078                         continue;
2079                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2080                         continue;
2081                 if (!zone->all_unreclaimable)
2082                         return false;
2083         }
2084
2085         return true;
2086 }
2087
2088 /*
2089  * This is the main entry point to direct page reclaim.
2090  *
2091  * If a full scan of the inactive list fails to free enough memory then we
2092  * are "out of memory" and something needs to be killed.
2093  *
2094  * If the caller is !__GFP_FS then the probability of a failure is reasonably
2095  * high - the zone may be full of dirty or under-writeback pages, which this
2096  * caller can't do much about.  We kick the writeback threads and take explicit
2097  * naps in the hope that some of these pages can be written.  But if the
2098  * allocating task holds filesystem locks which prevent writeout this might not
2099  * work, and the allocation attempt will fail.
2100  *
2101  * returns:     0, if no pages reclaimed
2102  *              else, the number of pages reclaimed
2103  */
2104 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2105                                         struct scan_control *sc,
2106                                         struct shrink_control *shrink)
2107 {
2108         int priority;
2109         unsigned long total_scanned = 0;
2110         struct reclaim_state *reclaim_state = current->reclaim_state;
2111         struct zoneref *z;
2112         struct zone *zone;
2113         unsigned long writeback_threshold;
2114
2115         get_mems_allowed();
2116         delayacct_freepages_start();
2117
2118         if (scanning_global_lru(sc))
2119                 count_vm_event(ALLOCSTALL);
2120
2121         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2122                 sc->nr_scanned = 0;
2123                 if (!priority)
2124                         disable_swap_token(sc->mem_cgroup);
2125                 shrink_zones(priority, zonelist, sc);
2126                 /*
2127                  * Don't shrink slabs when reclaiming memory from
2128                  * over limit cgroups
2129                  */
2130                 if (scanning_global_lru(sc)) {
2131                         unsigned long lru_pages = 0;
2132                         for_each_zone_zonelist(zone, z, zonelist,
2133                                         gfp_zone(sc->gfp_mask)) {
2134                                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2135                                         continue;
2136
2137                                 lru_pages += zone_reclaimable_pages(zone);
2138                         }
2139
2140                         shrink_slab(shrink, sc->nr_scanned, lru_pages);
2141                         if (reclaim_state) {
2142                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2143                                 reclaim_state->reclaimed_slab = 0;
2144                         }
2145                 }
2146                 total_scanned += sc->nr_scanned;
2147                 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2148                         goto out;
2149
2150                 /*
2151                  * Try to write back as many pages as we just scanned.  This
2152                  * tends to cause slow streaming writers to write data to the
2153                  * disk smoothly, at the dirtying rate, which is nice.   But
2154                  * that's undesirable in laptop mode, where we *want* lumpy
2155                  * writeout.  So in laptop mode, write out the whole world.
2156                  */
2157                 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2158                 if (total_scanned > writeback_threshold) {
2159                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
2160                         sc->may_writepage = 1;
2161                 }
2162
2163                 /* Take a nap, wait for some writeback to complete */
2164                 if (!sc->hibernation_mode && sc->nr_scanned &&
2165                     priority < DEF_PRIORITY - 2) {
2166                         struct zone *preferred_zone;
2167
2168                         first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2169                                                 &cpuset_current_mems_allowed,
2170                                                 &preferred_zone);
2171                         wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2172                 }
2173         }
2174
2175 out:
2176         delayacct_freepages_end();
2177         put_mems_allowed();
2178
2179         if (sc->nr_reclaimed)
2180                 return sc->nr_reclaimed;
2181
2182         /*
2183          * As hibernation is going on, kswapd is freezed so that it can't mark
2184          * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2185          * check.
2186          */
2187         if (oom_killer_disabled)
2188                 return 0;
2189
2190         /* top priority shrink_zones still had more to do? don't OOM, then */
2191         if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
2192                 return 1;
2193
2194         return 0;
2195 }
2196
2197 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2198                                 gfp_t gfp_mask, nodemask_t *nodemask)
2199 {
2200         unsigned long nr_reclaimed;
2201         struct scan_control sc = {
2202                 .gfp_mask = gfp_mask,
2203                 .may_writepage = !laptop_mode,
2204                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2205                 .may_unmap = 1,
2206                 .may_swap = 1,
2207                 .swappiness = vm_swappiness,
2208                 .order = order,
2209                 .mem_cgroup = NULL,
2210                 .nodemask = nodemask,
2211         };
2212         struct shrink_control shrink = {
2213                 .gfp_mask = sc.gfp_mask,
2214         };
2215
2216         trace_mm_vmscan_direct_reclaim_begin(order,
2217                                 sc.may_writepage,
2218                                 gfp_mask);
2219
2220         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2221
2222         trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2223
2224         return nr_reclaimed;
2225 }
2226
2227 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2228
2229 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2230                                                 gfp_t gfp_mask, bool noswap,
2231                                                 unsigned int swappiness,
2232                                                 struct zone *zone,
2233                                                 unsigned long *nr_scanned)
2234 {
2235         struct scan_control sc = {
2236                 .nr_scanned = 0,
2237                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2238                 .may_writepage = !laptop_mode,
2239                 .may_unmap = 1,
2240                 .may_swap = !noswap,
2241                 .swappiness = swappiness,
2242                 .order = 0,
2243                 .mem_cgroup = mem,
2244         };
2245
2246         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2247                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2248
2249         trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2250                                                       sc.may_writepage,
2251                                                       sc.gfp_mask);
2252
2253         /*
2254          * NOTE: Although we can get the priority field, using it
2255          * here is not a good idea, since it limits the pages we can scan.
2256          * if we don't reclaim here, the shrink_zone from balance_pgdat
2257          * will pick up pages from other mem cgroup's as well. We hack
2258          * the priority and make it zero.
2259          */
2260         shrink_zone(0, zone, &sc);
2261
2262         trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2263
2264         *nr_scanned = sc.nr_scanned;
2265         return sc.nr_reclaimed;
2266 }
2267
2268 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
2269                                            gfp_t gfp_mask,
2270                                            bool noswap,
2271                                            unsigned int swappiness)
2272 {
2273         struct zonelist *zonelist;
2274         unsigned long nr_reclaimed;
2275         int nid;
2276         struct scan_control sc = {
2277                 .may_writepage = !laptop_mode,
2278                 .may_unmap = 1,
2279                 .may_swap = !noswap,
2280                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2281                 .swappiness = swappiness,
2282                 .order = 0,
2283                 .mem_cgroup = mem_cont,
2284                 .nodemask = NULL, /* we don't care the placement */
2285                 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2286                                 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2287         };
2288         struct shrink_control shrink = {
2289                 .gfp_mask = sc.gfp_mask,
2290         };
2291
2292         /*
2293          * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2294          * take care of from where we get pages. So the node where we start the
2295          * scan does not need to be the current node.
2296          */
2297         nid = mem_cgroup_select_victim_node(mem_cont);
2298
2299         zonelist = NODE_DATA(nid)->node_zonelists;
2300
2301         trace_mm_vmscan_memcg_reclaim_begin(0,
2302                                             sc.may_writepage,
2303                                             sc.gfp_mask);
2304
2305         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2306
2307         trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2308
2309         return nr_reclaimed;
2310 }
2311 #endif
2312
2313 /*
2314  * pgdat_balanced is used when checking if a node is balanced for high-order
2315  * allocations. Only zones that meet watermarks and are in a zone allowed
2316  * by the callers classzone_idx are added to balanced_pages. The total of
2317  * balanced pages must be at least 25% of the zones allowed by classzone_idx
2318  * for the node to be considered balanced. Forcing all zones to be balanced
2319  * for high orders can cause excessive reclaim when there are imbalanced zones.
2320  * The choice of 25% is due to
2321  *   o a 16M DMA zone that is balanced will not balance a zone on any
2322  *     reasonable sized machine
2323  *   o On all other machines, the top zone must be at least a reasonable
2324  *     percentage of the middle zones. For example, on 32-bit x86, highmem
2325  *     would need to be at least 256M for it to be balance a whole node.
2326  *     Similarly, on x86-64 the Normal zone would need to be at least 1G
2327  *     to balance a node on its own. These seemed like reasonable ratios.
2328  */
2329 static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2330                                                 int classzone_idx)
2331 {
2332         unsigned long present_pages = 0;
2333         int i;
2334
2335         for (i = 0; i <= classzone_idx; i++)
2336                 present_pages += pgdat->node_zones[i].present_pages;
2337
2338         /* A special case here: if zone has no page, we think it's balanced */
2339         return balanced_pages >= (present_pages >> 2);
2340 }
2341
2342 /* is kswapd sleeping prematurely? */
2343 static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2344                                         int classzone_idx)
2345 {
2346         int i;
2347         unsigned long balanced = 0;
2348         bool all_zones_ok = true;
2349
2350         /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2351         if (remaining)
2352                 return true;
2353
2354         /* Check the watermark levels */
2355         for (i = 0; i <= classzone_idx; i++) {
2356                 struct zone *zone = pgdat->node_zones + i;
2357
2358                 if (!populated_zone(zone))
2359                         continue;
2360
2361                 /*
2362                  * balance_pgdat() skips over all_unreclaimable after
2363                  * DEF_PRIORITY. Effectively, it considers them balanced so
2364                  * they must be considered balanced here as well if kswapd
2365                  * is to sleep
2366                  */
2367                 if (zone->all_unreclaimable) {
2368                         balanced += zone->present_pages;
2369                         continue;
2370                 }
2371
2372                 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2373                                                         i, 0))
2374                         all_zones_ok = false;
2375                 else
2376                         balanced += zone->present_pages;
2377         }
2378
2379         /*
2380          * For high-order requests, the balanced zones must contain at least
2381          * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2382          * must be balanced
2383          */
2384         if (order)
2385                 return !pgdat_balanced(pgdat, balanced, classzone_idx);
2386         else
2387                 return !all_zones_ok;
2388 }
2389
2390 /*
2391  * For kswapd, balance_pgdat() will work across all this node's zones until
2392  * they are all at high_wmark_pages(zone).
2393  *
2394  * Returns the final order kswapd was reclaiming at
2395  *
2396  * There is special handling here for zones which are full of pinned pages.
2397  * This can happen if the pages are all mlocked, or if they are all used by
2398  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
2399  * What we do is to detect the case where all pages in the zone have been
2400  * scanned twice and there has been zero successful reclaim.  Mark the zone as
2401  * dead and from now on, only perform a short scan.  Basically we're polling
2402  * the zone for when the problem goes away.
2403  *
2404  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
2405  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2406  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2407  * lower zones regardless of the number of free pages in the lower zones. This
2408  * interoperates with the page allocator fallback scheme to ensure that aging
2409  * of pages is balanced across the zones.
2410  */
2411 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2412                                                         int *classzone_idx)
2413 {
2414         int all_zones_ok;
2415         unsigned long balanced;
2416         int priority;
2417         int i;
2418         int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
2419         unsigned long total_scanned;
2420         struct reclaim_state *reclaim_state = current->reclaim_state;
2421         unsigned long nr_soft_reclaimed;
2422         unsigned long nr_soft_scanned;
2423         struct scan_control sc = {
2424                 .gfp_mask = GFP_KERNEL,
2425                 .may_unmap = 1,
2426                 .may_swap = 1,
2427                 /*
2428                  * kswapd doesn't want to be bailed out while reclaim. because
2429                  * we want to put equal scanning pressure on each zone.
2430                  */
2431                 .nr_to_reclaim = ULONG_MAX,
2432                 .swappiness = vm_swappiness,
2433                 .order = order,
2434                 .mem_cgroup = NULL,
2435         };
2436         struct shrink_control shrink = {
2437                 .gfp_mask = sc.gfp_mask,
2438         };
2439 loop_again:
2440         total_scanned = 0;
2441         sc.nr_reclaimed = 0;
2442         sc.may_writepage = !laptop_mode;
2443         count_vm_event(PAGEOUTRUN);
2444
2445         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2446                 unsigned long lru_pages = 0;
2447                 int has_under_min_watermark_zone = 0;
2448
2449                 /* The swap token gets in the way of swapout... */
2450                 if (!priority)
2451                         disable_swap_token(NULL);
2452
2453                 all_zones_ok = 1;
2454                 balanced = 0;
2455
2456                 /*
2457                  * Scan in the highmem->dma direction for the highest
2458                  * zone which needs scanning
2459                  */
2460                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2461                         struct zone *zone = pgdat->node_zones + i;
2462
2463                         if (!populated_zone(zone))
2464                                 continue;
2465
2466                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2467                                 continue;
2468
2469                         /*
2470                          * Do some background aging of the anon list, to give
2471                          * pages a chance to be referenced before reclaiming.
2472                          */
2473                         if (inactive_anon_is_low(zone, &sc))
2474                                 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2475                                                         &sc, priority, 0);
2476
2477                         if (!zone_watermark_ok_safe(zone, order,
2478                                         high_wmark_pages(zone), 0, 0)) {
2479                                 end_zone = i;
2480                                 break;
2481                         }
2482                 }
2483                 if (i < 0)
2484                         goto out;
2485
2486                 for (i = 0; i <= end_zone; i++) {
2487                         struct zone *zone = pgdat->node_zones + i;
2488
2489                         lru_pages += zone_reclaimable_pages(zone);
2490                 }
2491
2492                 /*
2493                  * Now scan the zone in the dma->highmem direction, stopping
2494                  * at the last zone which needs scanning.
2495                  *
2496                  * We do this because the page allocator works in the opposite
2497                  * direction.  This prevents the page allocator from allocating
2498                  * pages behind kswapd's direction of progress, which would
2499                  * cause too much scanning of the lower zones.
2500                  */
2501                 for (i = 0; i <= end_zone; i++) {
2502                         struct zone *zone = pgdat->node_zones + i;
2503                         int nr_slab;
2504                         unsigned long balance_gap;
2505
2506                         if (!populated_zone(zone))
2507                                 continue;
2508
2509                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2510                                 continue;
2511
2512                         sc.nr_scanned = 0;
2513
2514                         nr_soft_scanned = 0;
2515                         /*
2516                          * Call soft limit reclaim before calling shrink_zone.
2517                          */
2518                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2519                                                         order, sc.gfp_mask,
2520                                                         &nr_soft_scanned);
2521                         sc.nr_reclaimed += nr_soft_reclaimed;
2522                         total_scanned += nr_soft_scanned;
2523
2524                         /*
2525                          * We put equal pressure on every zone, unless
2526                          * one zone has way too many pages free
2527                          * already. The "too many pages" is defined
2528                          * as the high wmark plus a "gap" where the
2529                          * gap is either the low watermark or 1%
2530                          * of the zone, whichever is smaller.
2531                          */
2532                         balance_gap = min(low_wmark_pages(zone),
2533                                 (zone->present_pages +
2534                                         KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2535                                 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2536                         if (!zone_watermark_ok_safe(zone, order,
2537                                         high_wmark_pages(zone) + balance_gap,
2538                                         end_zone, 0)) {
2539                                 shrink_zone(priority, zone, &sc);
2540
2541                                 reclaim_state->reclaimed_slab = 0;
2542                                 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2543                                 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2544                                 total_scanned += sc.nr_scanned;
2545
2546                                 if (nr_slab == 0 && !zone_reclaimable(zone))
2547                                         zone->all_unreclaimable = 1;
2548                         }
2549
2550                         /*
2551                          * If we've done a decent amount of scanning and
2552                          * the reclaim ratio is low, start doing writepage
2553                          * even in laptop mode
2554                          */
2555                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2556                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2557                                 sc.may_writepage = 1;
2558
2559                         if (zone->all_unreclaimable) {
2560                                 if (end_zone && end_zone == i)
2561                                         end_zone--;
2562                                 continue;
2563                         }
2564
2565                         if (!zone_watermark_ok_safe(zone, order,
2566                                         high_wmark_pages(zone), end_zone, 0)) {
2567                                 all_zones_ok = 0;
2568                                 /*
2569                                  * We are still under min water mark.  This
2570                                  * means that we have a GFP_ATOMIC allocation
2571                                  * failure risk. Hurry up!
2572                                  */
2573                                 if (!zone_watermark_ok_safe(zone, order,
2574                                             min_wmark_pages(zone), end_zone, 0))
2575                                         has_under_min_watermark_zone = 1;
2576                         } else {
2577                                 /*
2578                                  * If a zone reaches its high watermark,
2579                                  * consider it to be no longer congested. It's
2580                                  * possible there are dirty pages backed by
2581                                  * congested BDIs but as pressure is relieved,
2582                                  * spectulatively avoid congestion waits
2583                                  */
2584                                 zone_clear_flag(zone, ZONE_CONGESTED);
2585                                 if (i <= *classzone_idx)
2586                                         balanced += zone->present_pages;
2587                         }
2588
2589                 }
2590                 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2591                         break;          /* kswapd: all done */
2592                 /*
2593                  * OK, kswapd is getting into trouble.  Take a nap, then take
2594                  * another pass across the zones.
2595                  */
2596                 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2597                         if (has_under_min_watermark_zone)
2598                                 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2599                         else
2600                                 congestion_wait(BLK_RW_ASYNC, HZ/10);
2601                 }
2602
2603                 /*
2604                  * We do this so kswapd doesn't build up large priorities for
2605                  * example when it is freeing in parallel with allocators. It
2606                  * matches the direct reclaim path behaviour in terms of impact
2607                  * on zone->*_priority.
2608                  */
2609                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2610                         break;
2611         }
2612 out:
2613
2614         /*
2615          * order-0: All zones must meet high watermark for a balanced node
2616          * high-order: Balanced zones must make up at least 25% of the node
2617          *             for the node to be balanced
2618          */
2619         if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2620                 cond_resched();
2621
2622                 try_to_freeze();
2623
2624                 /*
2625                  * Fragmentation may mean that the system cannot be
2626                  * rebalanced for high-order allocations in all zones.
2627                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2628                  * it means the zones have been fully scanned and are still
2629                  * not balanced. For high-order allocations, there is
2630                  * little point trying all over again as kswapd may
2631                  * infinite loop.
2632                  *
2633                  * Instead, recheck all watermarks at order-0 as they
2634                  * are the most important. If watermarks are ok, kswapd will go
2635                  * back to sleep. High-order users can still perform direct
2636                  * reclaim if they wish.
2637                  */
2638                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2639                         order = sc.order = 0;
2640
2641                 goto loop_again;
2642         }
2643
2644         /*
2645          * If kswapd was reclaiming at a higher order, it has the option of
2646          * sleeping without all zones being balanced. Before it does, it must
2647          * ensure that the watermarks for order-0 on *all* zones are met and
2648          * that the congestion flags are cleared. The congestion flag must
2649          * be cleared as kswapd is the only mechanism that clears the flag
2650          * and it is potentially going to sleep here.
2651          */
2652         if (order) {
2653                 for (i = 0; i <= end_zone; i++) {
2654                         struct zone *zone = pgdat->node_zones + i;
2655
2656                         if (!populated_zone(zone))
2657                                 continue;
2658
2659                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2660                                 continue;
2661
2662                         /* Confirm the zone is balanced for order-0 */
2663                         if (!zone_watermark_ok(zone, 0,
2664                                         high_wmark_pages(zone), 0, 0)) {
2665                                 order = sc.order = 0;
2666                                 goto loop_again;
2667                         }
2668
2669                         /* If balanced, clear the congested flag */
2670                         zone_clear_flag(zone, ZONE_CONGESTED);
2671                 }
2672         }
2673
2674         /*
2675          * Return the order we were reclaiming at so sleeping_prematurely()
2676          * makes a decision on the order we were last reclaiming at. However,
2677          * if another caller entered the allocator slow path while kswapd
2678          * was awake, order will remain at the higher level
2679          */
2680         *classzone_idx = end_zone;
2681         return order;
2682 }
2683
2684 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2685 {
2686         long remaining = 0;
2687         DEFINE_WAIT(wait);
2688
2689         if (freezing(current) || kthread_should_stop())
2690                 return;
2691
2692         prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2693
2694         /* Try to sleep for a short interval */
2695         if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2696                 remaining = schedule_timeout(HZ/10);
2697                 finish_wait(&pgdat->kswapd_wait, &wait);
2698                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2699         }
2700
2701         /*
2702          * After a short sleep, check if it was a premature sleep. If not, then
2703          * go fully to sleep until explicitly woken up.
2704          */
2705         if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2706                 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2707
2708                 /*
2709                  * vmstat counters are not perfectly accurate and the estimated
2710                  * value for counters such as NR_FREE_PAGES can deviate from the
2711                  * true value by nr_online_cpus * threshold. To avoid the zone
2712                  * watermarks being breached while under pressure, we reduce the
2713                  * per-cpu vmstat threshold while kswapd is awake and restore
2714                  * them before going back to sleep.
2715                  */
2716                 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2717                 schedule();
2718                 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2719         } else {
2720                 if (remaining)
2721                         count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2722                 else
2723                         count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2724         }
2725         finish_wait(&pgdat->kswapd_wait, &wait);
2726 }
2727
2728 /*
2729  * The background pageout daemon, started as a kernel thread
2730  * from the init process.
2731  *
2732  * This basically trickles out pages so that we have _some_
2733  * free memory available even if there is no other activity
2734  * that frees anything up. This is needed for things like routing
2735  * etc, where we otherwise might have all activity going on in
2736  * asynchronous contexts that cannot page things out.
2737  *
2738  * If there are applications that are active memory-allocators
2739  * (most normal use), this basically shouldn't matter.
2740  */
2741 static int kswapd(void *p)
2742 {
2743         unsigned long order, new_order;
2744         int classzone_idx, new_classzone_idx;
2745         pg_data_t *pgdat = (pg_data_t*)p;
2746         struct task_struct *tsk = current;
2747
2748         struct reclaim_state reclaim_state = {
2749                 .reclaimed_slab = 0,
2750         };
2751         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2752
2753         lockdep_set_current_reclaim_state(GFP_KERNEL);
2754
2755         if (!cpumask_empty(cpumask))
2756                 set_cpus_allowed_ptr(tsk, cpumask);
2757         current->reclaim_state = &reclaim_state;
2758
2759         /*
2760          * Tell the memory management that we're a "memory allocator",
2761          * and that if we need more memory we should get access to it
2762          * regardless (see "__alloc_pages()"). "kswapd" should
2763          * never get caught in the normal page freeing logic.
2764          *
2765          * (Kswapd normally doesn't need memory anyway, but sometimes
2766          * you need a small amount of memory in order to be able to
2767          * page out something else, and this flag essentially protects
2768          * us from recursively trying to free more memory as we're
2769          * trying to free the first piece of memory in the first place).
2770          */
2771         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2772         set_freezable();
2773
2774         order = new_order = 0;
2775         classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
2776         for ( ; ; ) {
2777                 int ret;
2778
2779                 /*
2780                  * If the last balance_pgdat was unsuccessful it's unlikely a
2781                  * new request of a similar or harder type will succeed soon
2782                  * so consider going to sleep on the basis we reclaimed at
2783                  */
2784                 if (classzone_idx >= new_classzone_idx && order == new_order) {
2785                         new_order = pgdat->kswapd_max_order;
2786                         new_classzone_idx = pgdat->classzone_idx;
2787                         pgdat->kswapd_max_order =  0;
2788                         pgdat->classzone_idx = pgdat->nr_zones - 1;
2789                 }
2790
2791                 if (order < new_order || classzone_idx > new_classzone_idx) {
2792                         /*
2793                          * Don't sleep if someone wants a larger 'order'
2794                          * allocation or has tigher zone constraints
2795                          */
2796                         order = new_order;
2797                         classzone_idx = new_classzone_idx;
2798                 } else {
2799                         kswapd_try_to_sleep(pgdat, order, classzone_idx);
2800                         order = pgdat->kswapd_max_order;
2801                         classzone_idx = pgdat->classzone_idx;
2802                         pgdat->kswapd_max_order = 0;
2803                         pgdat->classzone_idx = pgdat->nr_zones - 1;
2804                 }
2805
2806                 ret = try_to_freeze();
2807                 if (kthread_should_stop())
2808                         break;
2809
2810                 /*
2811                  * We can speed up thawing tasks if we don't call balance_pgdat
2812                  * after returning from the refrigerator
2813                  */
2814                 if (!ret) {
2815                         trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2816                         order = balance_pgdat(pgdat, order, &classzone_idx);
2817                 }
2818         }
2819         return 0;
2820 }
2821
2822 /*
2823  * A zone is low on free memory, so wake its kswapd task to service it.
2824  */
2825 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
2826 {
2827         pg_data_t *pgdat;
2828
2829         if (!populated_zone(zone))
2830                 return;
2831
2832         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2833                 return;
2834         pgdat = zone->zone_pgdat;
2835         if (pgdat->kswapd_max_order < order) {
2836                 pgdat->kswapd_max_order = order;
2837                 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
2838         }
2839         if (!waitqueue_active(&pgdat->kswapd_wait))
2840                 return;
2841         if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
2842                 return;
2843
2844         trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
2845         wake_up_interruptible(&pgdat->kswapd_wait);
2846 }
2847
2848 /*
2849  * The reclaimable count would be mostly accurate.
2850  * The less reclaimable pages may be
2851  * - mlocked pages, which will be moved to unevictable list when encountered
2852  * - mapped pages, which may require several travels to be reclaimed
2853  * - dirty pages, which is not "instantly" reclaimable
2854  */
2855 unsigned long global_reclaimable_pages(void)
2856 {
2857         int nr;
2858
2859         nr = global_page_state(NR_ACTIVE_FILE) +
2860              global_page_state(NR_INACTIVE_FILE);
2861
2862         if (nr_swap_pages > 0)
2863                 nr += global_page_state(NR_ACTIVE_ANON) +
2864                       global_page_state(NR_INACTIVE_ANON);
2865
2866         return nr;
2867 }
2868
2869 unsigned long zone_reclaimable_pages(struct zone *zone)
2870 {
2871         int nr;
2872
2873         nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2874              zone_page_state(zone, NR_INACTIVE_FILE);
2875
2876         if (nr_swap_pages > 0)
2877                 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2878                       zone_page_state(zone, NR_INACTIVE_ANON);
2879
2880         return nr;
2881 }
2882
2883 #ifdef CONFIG_HIBERNATION
2884 /*
2885  * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2886  * freed pages.
2887  *
2888  * Rather than trying to age LRUs the aim is to preserve the overall
2889  * LRU order by reclaiming preferentially
2890  * inactive > active > active referenced > active mapped
2891  */
2892 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
2893 {
2894         struct reclaim_state reclaim_state;
2895         struct scan_control sc = {
2896                 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2897                 .may_swap = 1,
2898                 .may_unmap = 1,
2899                 .may_writepage = 1,
2900                 .nr_to_reclaim = nr_to_reclaim,
2901                 .hibernation_mode = 1,
2902                 .swappiness = vm_swappiness,
2903                 .order = 0,
2904         };
2905         struct shrink_control shrink = {
2906                 .gfp_mask = sc.gfp_mask,
2907         };
2908         struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2909         struct task_struct *p = current;
2910         unsigned long nr_reclaimed;
2911
2912         p->flags |= PF_MEMALLOC;
2913         lockdep_set_current_reclaim_state(sc.gfp_mask);
2914         reclaim_state.reclaimed_slab = 0;
2915         p->reclaim_state = &reclaim_state;
2916
2917         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2918
2919         p->reclaim_state = NULL;
2920         lockdep_clear_current_reclaim_state();
2921         p->flags &= ~PF_MEMALLOC;
2922
2923         return nr_reclaimed;
2924 }
2925 #endif /* CONFIG_HIBERNATION */
2926
2927 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2928    not required for correctness.  So if the last cpu in a node goes
2929    away, we get changed to run anywhere: as the first one comes back,
2930    restore their cpu bindings. */
2931 static int __devinit cpu_callback(struct notifier_block *nfb,
2932                                   unsigned long action, void *hcpu)
2933 {
2934         int nid;
2935
2936         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2937                 for_each_node_state(nid, N_HIGH_MEMORY) {
2938                         pg_data_t *pgdat = NODE_DATA(nid);
2939                         const struct cpumask *mask;
2940
2941                         mask = cpumask_of_node(pgdat->node_id);
2942
2943                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2944                                 /* One of our CPUs online: restore mask */
2945                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2946                 }
2947         }
2948         return NOTIFY_OK;
2949 }
2950
2951 /*
2952  * This kswapd start function will be called by init and node-hot-add.
2953  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2954  */
2955 int kswapd_run(int nid)
2956 {
2957         pg_data_t *pgdat = NODE_DATA(nid);
2958         int ret = 0;
2959
2960         if (pgdat->kswapd)
2961                 return 0;
2962
2963         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2964         if (IS_ERR(pgdat->kswapd)) {
2965                 /* failure at boot is fatal */
2966                 BUG_ON(system_state == SYSTEM_BOOTING);
2967                 printk("Failed to start kswapd on node %d\n",nid);
2968                 ret = -1;
2969         }
2970         return ret;
2971 }
2972
2973 /*
2974  * Called by memory hotplug when all memory in a node is offlined.
2975  */
2976 void kswapd_stop(int nid)
2977 {
2978         struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
2979
2980         if (kswapd)
2981                 kthread_stop(kswapd);
2982 }
2983
2984 static int __init kswapd_init(void)
2985 {
2986         int nid;
2987
2988         swap_setup();
2989         for_each_node_state(nid, N_HIGH_MEMORY)
2990                 kswapd_run(nid);
2991         hotcpu_notifier(cpu_callback, 0);
2992         return 0;
2993 }
2994
2995 module_init(kswapd_init)
2996
2997 #ifdef CONFIG_NUMA
2998 /*
2999  * Zone reclaim mode
3000  *
3001  * If non-zero call zone_reclaim when the number of free pages falls below
3002  * the watermarks.
3003  */
3004 int zone_reclaim_mode __read_mostly;
3005
3006 #define RECLAIM_OFF 0
3007 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
3008 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
3009 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
3010
3011 /*
3012  * Priority for ZONE_RECLAIM. This determines the fraction of pages
3013  * of a node considered for each zone_reclaim. 4 scans 1/16th of
3014  * a zone.
3015  */
3016 #define ZONE_RECLAIM_PRIORITY 4
3017
3018 /*
3019  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3020  * occur.
3021  */
3022 int sysctl_min_unmapped_ratio = 1;
3023
3024 /*
3025  * If the number of slab pages in a zone grows beyond this percentage then
3026  * slab reclaim needs to occur.
3027  */
3028 int sysctl_min_slab_ratio = 5;
3029
3030 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3031 {
3032         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3033         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3034                 zone_page_state(zone, NR_ACTIVE_FILE);
3035
3036         /*
3037          * It's possible for there to be more file mapped pages than
3038          * accounted for by the pages on the file LRU lists because
3039          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3040          */
3041         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3042 }
3043
3044 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3045 static long zone_pagecache_reclaimable(struct zone *zone)
3046 {
3047         long nr_pagecache_reclaimable;
3048         long delta = 0;
3049
3050         /*
3051          * If RECLAIM_SWAP is set, then all file pages are considered
3052          * potentially reclaimable. Otherwise, we have to worry about
3053          * pages like swapcache and zone_unmapped_file_pages() provides
3054          * a better estimate
3055          */
3056         if (zone_reclaim_mode & RECLAIM_SWAP)
3057                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3058         else
3059                 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3060
3061         /* If we can't clean pages, remove dirty pages from consideration */
3062         if (!(zone_reclaim_mode & RECLAIM_WRITE))
3063                 delta += zone_page_state(zone, NR_FILE_DIRTY);
3064
3065         /* Watch for any possible underflows due to delta */
3066         if (unlikely(delta > nr_pagecache_reclaimable))
3067                 delta = nr_pagecache_reclaimable;
3068
3069         return nr_pagecache_reclaimable - delta;
3070 }
3071
3072 /*
3073  * Try to free up some pages from this zone through reclaim.
3074  */
3075 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3076 {
3077         /* Minimum pages needed in order to stay on node */
3078         const unsigned long nr_pages = 1 << order;
3079         struct task_struct *p = current;
3080         struct reclaim_state reclaim_state;
3081         int priority;
3082         struct scan_control sc = {
3083                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3084                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3085                 .may_swap = 1,
3086                 .nr_to_reclaim = max_t(unsigned long, nr_pages,
3087                                        SWAP_CLUSTER_MAX),
3088                 .gfp_mask = gfp_mask,
3089                 .swappiness = vm_swappiness,
3090                 .order = order,
3091         };
3092         struct shrink_control shrink = {
3093                 .gfp_mask = sc.gfp_mask,
3094         };
3095         unsigned long nr_slab_pages0, nr_slab_pages1;
3096
3097         cond_resched();
3098         /*
3099          * We need to be able to allocate from the reserves for RECLAIM_SWAP
3100          * and we also need to be able to write out pages for RECLAIM_WRITE
3101          * and RECLAIM_SWAP.
3102          */
3103         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3104         lockdep_set_current_reclaim_state(gfp_mask);
3105         reclaim_state.reclaimed_slab = 0;
3106         p->reclaim_state = &reclaim_state;
3107
3108         if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3109                 /*
3110                  * Free memory by calling shrink zone with increasing
3111                  * priorities until we have enough memory freed.
3112                  */
3113                 priority = ZONE_RECLAIM_PRIORITY;
3114                 do {
3115                         shrink_zone(priority, zone, &sc);
3116                         priority--;
3117                 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
3118         }
3119
3120         nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3121         if (nr_slab_pages0 > zone->min_slab_pages) {
3122                 /*
3123                  * shrink_slab() does not currently allow us to determine how
3124                  * many pages were freed in this zone. So we take the current
3125                  * number of slab pages and shake the slab until it is reduced
3126                  * by the same nr_pages that we used for reclaiming unmapped
3127                  * pages.
3128                  *
3129                  * Note that shrink_slab will free memory on all zones and may
3130                  * take a long time.
3131                  */
3132                 for (;;) {
3133                         unsigned long lru_pages = zone_reclaimable_pages(zone);
3134
3135                         /* No reclaimable slab or very low memory pressure */
3136                         if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3137                                 break;
3138
3139                         /* Freed enough memory */
3140                         nr_slab_pages1 = zone_page_state(zone,
3141                                                         NR_SLAB_RECLAIMABLE);
3142                         if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3143                                 break;
3144                 }
3145
3146                 /*
3147                  * Update nr_reclaimed by the number of slab pages we
3148                  * reclaimed from this zone.
3149                  */
3150                 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3151                 if (nr_slab_pages1 < nr_slab_pages0)
3152                         sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3153         }
3154
3155         p->reclaim_state = NULL;
3156         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3157         lockdep_clear_current_reclaim_state();
3158         return sc.nr_reclaimed >= nr_pages;
3159 }
3160
3161 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3162 {
3163         int node_id;
3164         int ret;
3165
3166         /*
3167          * Zone reclaim reclaims unmapped file backed pages and
3168          * slab pages if we are over the defined limits.
3169          *
3170          * A small portion of unmapped file backed pages is needed for
3171          * file I/O otherwise pages read by file I/O will be immediately
3172          * thrown out if the zone is overallocated. So we do not reclaim
3173          * if less than a specified percentage of the zone is used by
3174          * unmapped file backed pages.
3175          */
3176         if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3177             zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3178                 return ZONE_RECLAIM_FULL;
3179
3180         if (zone->all_unreclaimable)
3181                 return ZONE_RECLAIM_FULL;
3182
3183         /*
3184          * Do not scan if the allocation should not be delayed.
3185          */
3186         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3187                 return ZONE_RECLAIM_NOSCAN;
3188
3189         /*
3190          * Only run zone reclaim on the local zone or on zones that do not
3191          * have associated processors. This will favor the local processor
3192          * over remote processors and spread off node memory allocations
3193          * as wide as possible.
3194          */
3195         node_id = zone_to_nid(zone);
3196         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3197                 return ZONE_RECLAIM_NOSCAN;
3198
3199         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3200                 return ZONE_RECLAIM_NOSCAN;
3201
3202         ret = __zone_reclaim(zone, gfp_mask, order);
3203         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3204
3205         if (!ret)
3206                 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3207
3208         return ret;
3209 }
3210 #endif
3211
3212 /*
3213  * page_evictable - test whether a page is evictable
3214  * @page: the page to test
3215  * @vma: the VMA in which the page is or will be mapped, may be NULL
3216  *
3217  * Test whether page is evictable--i.e., should be placed on active/inactive
3218  * lists vs unevictable list.  The vma argument is !NULL when called from the
3219  * fault path to determine how to instantate a new page.
3220  *
3221  * Reasons page might not be evictable:
3222  * (1) page's mapping marked unevictable
3223  * (2) page is part of an mlocked VMA
3224  *
3225  */
3226 int page_evictable(struct page *page, struct vm_area_struct *vma)
3227 {
3228
3229         if (mapping_unevictable(page_mapping(page)))
3230                 return 0;
3231
3232         if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
3233                 return 0;
3234
3235         return 1;
3236 }
3237
3238 /**
3239  * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
3240  * @page: page to check evictability and move to appropriate lru list
3241  * @zone: zone page is in
3242  *
3243  * Checks a page for evictability and moves the page to the appropriate
3244  * zone lru list.
3245  *
3246  * Restrictions: zone->lru_lock must be held, page must be on LRU and must
3247  * have PageUnevictable set.
3248  */
3249 static void check_move_unevictable_page(struct page *page, struct zone *zone)
3250 {
3251         VM_BUG_ON(PageActive(page));
3252
3253 retry:
3254         ClearPageUnevictable(page);
3255         if (page_evictable(page, NULL)) {
3256                 enum lru_list l = page_lru_base_type(page);
3257
3258                 __dec_zone_state(zone, NR_UNEVICTABLE);
3259                 list_move(&page->lru, &zone->lru[l].list);
3260                 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
3261                 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
3262                 __count_vm_event(UNEVICTABLE_PGRESCUED);
3263         } else {
3264                 /*
3265                  * rotate unevictable list
3266                  */
3267                 SetPageUnevictable(page);
3268                 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
3269                 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
3270                 if (page_evictable(page, NULL))
3271                         goto retry;
3272         }
3273 }
3274
3275 /**
3276  * scan_mapping_unevictable_pages - scan an address space for evictable pages
3277  * @mapping: struct address_space to scan for evictable pages
3278  *
3279  * Scan all pages in mapping.  Check unevictable pages for
3280  * evictability and move them to the appropriate zone lru list.
3281  */
3282 void scan_mapping_unevictable_pages(struct address_space *mapping)
3283 {
3284         pgoff_t next = 0;
3285         pgoff_t end   = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
3286                          PAGE_CACHE_SHIFT;
3287         struct zone *zone;
3288         struct pagevec pvec;
3289
3290         if (mapping->nrpages == 0)
3291                 return;
3292
3293         pagevec_init(&pvec, 0);
3294         while (next < end &&
3295                 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
3296                 int i;
3297                 int pg_scanned = 0;
3298
3299                 zone = NULL;
3300
3301                 for (i = 0; i < pagevec_count(&pvec); i++) {
3302                         struct page *page = pvec.pages[i];
3303                         pgoff_t page_index = page->index;
3304                         struct zone *pagezone = page_zone(page);
3305
3306                         pg_scanned++;
3307                         if (page_index > next)
3308                                 next = page_index;
3309                         next++;
3310
3311                         if (pagezone != zone) {
3312                                 if (zone)
3313                                         spin_unlock_irq(&zone->lru_lock);
3314                                 zone = pagezone;
3315                                 spin_lock_irq(&zone->lru_lock);
3316                         }
3317
3318                         if (PageLRU(page) && PageUnevictable(page))
3319                                 check_move_unevictable_page(page, zone);
3320                 }
3321                 if (zone)
3322                         spin_unlock_irq(&zone->lru_lock);
3323                 pagevec_release(&pvec);
3324
3325                 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
3326         }
3327
3328 }
3329
3330 /**
3331  * scan_zone_unevictable_pages - check unevictable list for evictable pages
3332  * @zone - zone of which to scan the unevictable list
3333  *
3334  * Scan @zone's unevictable LRU lists to check for pages that have become
3335  * evictable.  Move those that have to @zone's inactive list where they
3336  * become candidates for reclaim, unless shrink_inactive_zone() decides
3337  * to reactivate them.  Pages that are still unevictable are rotated
3338  * back onto @zone's unevictable list.
3339  */
3340 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
3341 static void scan_zone_unevictable_pages(struct zone *zone)
3342 {
3343         struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
3344         unsigned long scan;
3345         unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
3346
3347         while (nr_to_scan > 0) {
3348                 unsigned long batch_size = min(nr_to_scan,
3349                                                 SCAN_UNEVICTABLE_BATCH_SIZE);
3350
3351                 spin_lock_irq(&zone->lru_lock);
3352                 for (scan = 0;  scan < batch_size; scan++) {
3353                         struct page *page = lru_to_page(l_unevictable);
3354
3355                         if (!trylock_page(page))
3356                                 continue;
3357
3358                         prefetchw_prev_lru_page(page, l_unevictable, flags);
3359
3360                         if (likely(PageLRU(page) && PageUnevictable(page)))
3361                                 check_move_unevictable_page(page, zone);
3362
3363                         unlock_page(page);
3364                 }
3365                 spin_unlock_irq(&zone->lru_lock);
3366
3367                 nr_to_scan -= batch_size;
3368         }
3369 }
3370
3371
3372 /**
3373  * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
3374  *
3375  * A really big hammer:  scan all zones' unevictable LRU lists to check for
3376  * pages that have become evictable.  Move those back to the zones'
3377  * inactive list where they become candidates for reclaim.
3378  * This occurs when, e.g., we have unswappable pages on the unevictable lists,
3379  * and we add swap to the system.  As such, it runs in the context of a task
3380  * that has possibly/probably made some previously unevictable pages
3381  * evictable.
3382  */
3383 static void scan_all_zones_unevictable_pages(void)
3384 {
3385         struct zone *zone;
3386
3387         for_each_zone(zone) {
3388                 scan_zone_unevictable_pages(zone);
3389         }
3390 }
3391
3392 /*
3393  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
3394  * all nodes' unevictable lists for evictable pages
3395  */
3396 unsigned long scan_unevictable_pages;
3397
3398 int scan_unevictable_handler(struct ctl_table *table, int write,
3399                            void __user *buffer,
3400                            size_t *length, loff_t *ppos)
3401 {
3402         proc_doulongvec_minmax(table, write, buffer, length, ppos);
3403
3404         if (write && *(unsigned long *)table->data)
3405                 scan_all_zones_unevictable_pages();
3406
3407         scan_unevictable_pages = 0;
3408         return 0;
3409 }
3410
3411 #ifdef CONFIG_NUMA
3412 /*
3413  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
3414  * a specified node's per zone unevictable lists for evictable pages.
3415  */
3416
3417 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
3418                                           struct sysdev_attribute *attr,
3419                                           char *buf)
3420 {
3421         return sprintf(buf, "0\n");     /* always zero; should fit... */
3422 }
3423
3424 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
3425                                            struct sysdev_attribute *attr,
3426                                         const char *buf, size_t count)
3427 {
3428         struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
3429         struct zone *zone;
3430         unsigned long res;
3431         unsigned long req = strict_strtoul(buf, 10, &res);
3432
3433         if (!req)
3434                 return 1;       /* zero is no-op */
3435
3436         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
3437                 if (!populated_zone(zone))
3438                         continue;
3439                 scan_zone_unevictable_pages(zone);
3440         }
3441         return 1;
3442 }
3443
3444
3445 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3446                         read_scan_unevictable_node,
3447                         write_scan_unevictable_node);
3448
3449 int scan_unevictable_register_node(struct node *node)
3450 {
3451         return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
3452 }
3453
3454 void scan_unevictable_unregister_node(struct node *node)
3455 {
3456         sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
3457 }
3458 #endif