vmscan: avoid multiplication overflow in shrink_zone()
[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/slab.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/notifier.h>
36 #include <linux/rwsem.h>
37 #include <linux/delay.h>
38 #include <linux/kthread.h>
39 #include <linux/freezer.h>
40 #include <linux/memcontrol.h>
41 #include <linux/delayacct.h>
42 #include <linux/sysctl.h>
43
44 #include <asm/tlbflush.h>
45 #include <asm/div64.h>
46
47 #include <linux/swapops.h>
48
49 #include "internal.h"
50
51 struct scan_control {
52         /* Incremented by the number of inactive pages that were scanned */
53         unsigned long nr_scanned;
54
55         /* Number of pages freed so far during a call to shrink_zones() */
56         unsigned long nr_reclaimed;
57
58         /* This context's GFP mask */
59         gfp_t gfp_mask;
60
61         int may_writepage;
62
63         /* Can mapped pages be reclaimed? */
64         int may_unmap;
65
66         /* Can pages be swapped as part of reclaim? */
67         int may_swap;
68
69         /* This context's SWAP_CLUSTER_MAX. If freeing memory for
70          * suspend, we effectively ignore SWAP_CLUSTER_MAX.
71          * In this context, it doesn't matter that we scan the
72          * whole list at once. */
73         int swap_cluster_max;
74
75         int swappiness;
76
77         int all_unreclaimable;
78
79         int order;
80
81         /* Which cgroup do we reclaim from */
82         struct mem_cgroup *mem_cgroup;
83
84         /*
85          * Nodemask of nodes allowed by the caller. If NULL, all nodes
86          * are scanned.
87          */
88         nodemask_t      *nodemask;
89
90         /* Pluggable isolate pages callback */
91         unsigned long (*isolate_pages)(unsigned long nr, struct list_head *dst,
92                         unsigned long *scanned, int order, int mode,
93                         struct zone *z, struct mem_cgroup *mem_cont,
94                         int active, int file);
95 };
96
97 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
98
99 #ifdef ARCH_HAS_PREFETCH
100 #define prefetch_prev_lru_page(_page, _base, _field)                    \
101         do {                                                            \
102                 if ((_page)->lru.prev != _base) {                       \
103                         struct page *prev;                              \
104                                                                         \
105                         prev = lru_to_page(&(_page->lru));              \
106                         prefetch(&prev->_field);                        \
107                 }                                                       \
108         } while (0)
109 #else
110 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
111 #endif
112
113 #ifdef ARCH_HAS_PREFETCHW
114 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
115         do {                                                            \
116                 if ((_page)->lru.prev != _base) {                       \
117                         struct page *prev;                              \
118                                                                         \
119                         prev = lru_to_page(&(_page->lru));              \
120                         prefetchw(&prev->_field);                       \
121                 }                                                       \
122         } while (0)
123 #else
124 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
125 #endif
126
127 /*
128  * From 0 .. 100.  Higher means more swappy.
129  */
130 int vm_swappiness = 60;
131 long vm_total_pages;    /* The total number of pages which the VM controls */
132
133 static LIST_HEAD(shrinker_list);
134 static DECLARE_RWSEM(shrinker_rwsem);
135
136 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
137 #define scanning_global_lru(sc) (!(sc)->mem_cgroup)
138 #else
139 #define scanning_global_lru(sc) (1)
140 #endif
141
142 static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
143                                                   struct scan_control *sc)
144 {
145         if (!scanning_global_lru(sc))
146                 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
147
148         return &zone->reclaim_stat;
149 }
150
151 static unsigned long zone_nr_pages(struct zone *zone, struct scan_control *sc,
152                                    enum lru_list lru)
153 {
154         if (!scanning_global_lru(sc))
155                 return mem_cgroup_zone_nr_pages(sc->mem_cgroup, zone, lru);
156
157         return zone_page_state(zone, NR_LRU_BASE + lru);
158 }
159
160
161 /*
162  * Add a shrinker callback to be called from the vm
163  */
164 void register_shrinker(struct shrinker *shrinker)
165 {
166         shrinker->nr = 0;
167         down_write(&shrinker_rwsem);
168         list_add_tail(&shrinker->list, &shrinker_list);
169         up_write(&shrinker_rwsem);
170 }
171 EXPORT_SYMBOL(register_shrinker);
172
173 /*
174  * Remove one
175  */
176 void unregister_shrinker(struct shrinker *shrinker)
177 {
178         down_write(&shrinker_rwsem);
179         list_del(&shrinker->list);
180         up_write(&shrinker_rwsem);
181 }
182 EXPORT_SYMBOL(unregister_shrinker);
183
184 #define SHRINK_BATCH 128
185 /*
186  * Call the shrink functions to age shrinkable caches
187  *
188  * Here we assume it costs one seek to replace a lru page and that it also
189  * takes a seek to recreate a cache object.  With this in mind we age equal
190  * percentages of the lru and ageable caches.  This should balance the seeks
191  * generated by these structures.
192  *
193  * If the vm encountered mapped pages on the LRU it increase the pressure on
194  * slab to avoid swapping.
195  *
196  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
197  *
198  * `lru_pages' represents the number of on-LRU pages in all the zones which
199  * are eligible for the caller's allocation attempt.  It is used for balancing
200  * slab reclaim versus page reclaim.
201  *
202  * Returns the number of slab objects which we shrunk.
203  */
204 unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
205                         unsigned long lru_pages)
206 {
207         struct shrinker *shrinker;
208         unsigned long ret = 0;
209
210         if (scanned == 0)
211                 scanned = SWAP_CLUSTER_MAX;
212
213         if (!down_read_trylock(&shrinker_rwsem))
214                 return 1;       /* Assume we'll be able to shrink next time */
215
216         list_for_each_entry(shrinker, &shrinker_list, list) {
217                 unsigned long long delta;
218                 unsigned long total_scan;
219                 unsigned long max_pass = (*shrinker->shrink)(0, gfp_mask);
220
221                 delta = (4 * scanned) / shrinker->seeks;
222                 delta *= max_pass;
223                 do_div(delta, lru_pages + 1);
224                 shrinker->nr += delta;
225                 if (shrinker->nr < 0) {
226                         printk(KERN_ERR "shrink_slab: %pF negative objects to "
227                                "delete nr=%ld\n",
228                                shrinker->shrink, shrinker->nr);
229                         shrinker->nr = max_pass;
230                 }
231
232                 /*
233                  * Avoid risking looping forever due to too large nr value:
234                  * never try to free more than twice the estimate number of
235                  * freeable entries.
236                  */
237                 if (shrinker->nr > max_pass * 2)
238                         shrinker->nr = max_pass * 2;
239
240                 total_scan = shrinker->nr;
241                 shrinker->nr = 0;
242
243                 while (total_scan >= SHRINK_BATCH) {
244                         long this_scan = SHRINK_BATCH;
245                         int shrink_ret;
246                         int nr_before;
247
248                         nr_before = (*shrinker->shrink)(0, gfp_mask);
249                         shrink_ret = (*shrinker->shrink)(this_scan, gfp_mask);
250                         if (shrink_ret == -1)
251                                 break;
252                         if (shrink_ret < nr_before)
253                                 ret += nr_before - shrink_ret;
254                         count_vm_events(SLABS_SCANNED, this_scan);
255                         total_scan -= this_scan;
256
257                         cond_resched();
258                 }
259
260                 shrinker->nr += total_scan;
261         }
262         up_read(&shrinker_rwsem);
263         return ret;
264 }
265
266 /* Called without lock on whether page is mapped, so answer is unstable */
267 static inline int page_mapping_inuse(struct page *page)
268 {
269         struct address_space *mapping;
270
271         /* Page is in somebody's page tables. */
272         if (page_mapped(page))
273                 return 1;
274
275         /* Be more reluctant to reclaim swapcache than pagecache */
276         if (PageSwapCache(page))
277                 return 1;
278
279         mapping = page_mapping(page);
280         if (!mapping)
281                 return 0;
282
283         /* File is mmap'd by somebody? */
284         return mapping_mapped(mapping);
285 }
286
287 static inline int is_page_cache_freeable(struct page *page)
288 {
289         return page_count(page) - !!page_has_private(page) == 2;
290 }
291
292 static int may_write_to_queue(struct backing_dev_info *bdi)
293 {
294         if (current->flags & PF_SWAPWRITE)
295                 return 1;
296         if (!bdi_write_congested(bdi))
297                 return 1;
298         if (bdi == current->backing_dev_info)
299                 return 1;
300         return 0;
301 }
302
303 /*
304  * We detected a synchronous write error writing a page out.  Probably
305  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
306  * fsync(), msync() or close().
307  *
308  * The tricky part is that after writepage we cannot touch the mapping: nothing
309  * prevents it from being freed up.  But we have a ref on the page and once
310  * that page is locked, the mapping is pinned.
311  *
312  * We're allowed to run sleeping lock_page() here because we know the caller has
313  * __GFP_FS.
314  */
315 static void handle_write_error(struct address_space *mapping,
316                                 struct page *page, int error)
317 {
318         lock_page(page);
319         if (page_mapping(page) == mapping)
320                 mapping_set_error(mapping, error);
321         unlock_page(page);
322 }
323
324 /* Request for sync pageout. */
325 enum pageout_io {
326         PAGEOUT_IO_ASYNC,
327         PAGEOUT_IO_SYNC,
328 };
329
330 /* possible outcome of pageout() */
331 typedef enum {
332         /* failed to write page out, page is locked */
333         PAGE_KEEP,
334         /* move page to the active list, page is locked */
335         PAGE_ACTIVATE,
336         /* page has been sent to the disk successfully, page is unlocked */
337         PAGE_SUCCESS,
338         /* page is clean and locked */
339         PAGE_CLEAN,
340 } pageout_t;
341
342 /*
343  * pageout is called by shrink_page_list() for each dirty page.
344  * Calls ->writepage().
345  */
346 static pageout_t pageout(struct page *page, struct address_space *mapping,
347                                                 enum pageout_io sync_writeback)
348 {
349         /*
350          * If the page is dirty, only perform writeback if that write
351          * will be non-blocking.  To prevent this allocation from being
352          * stalled by pagecache activity.  But note that there may be
353          * stalls if we need to run get_block().  We could test
354          * PagePrivate for that.
355          *
356          * If this process is currently in generic_file_write() against
357          * this page's queue, we can perform writeback even if that
358          * will block.
359          *
360          * If the page is swapcache, write it back even if that would
361          * block, for some throttling. This happens by accident, because
362          * swap_backing_dev_info is bust: it doesn't reflect the
363          * congestion state of the swapdevs.  Easy to fix, if needed.
364          * See swapfile.c:page_queue_congested().
365          */
366         if (!is_page_cache_freeable(page))
367                 return PAGE_KEEP;
368         if (!mapping) {
369                 /*
370                  * Some data journaling orphaned pages can have
371                  * page->mapping == NULL while being dirty with clean buffers.
372                  */
373                 if (page_has_private(page)) {
374                         if (try_to_free_buffers(page)) {
375                                 ClearPageDirty(page);
376                                 printk("%s: orphaned page\n", __func__);
377                                 return PAGE_CLEAN;
378                         }
379                 }
380                 return PAGE_KEEP;
381         }
382         if (mapping->a_ops->writepage == NULL)
383                 return PAGE_ACTIVATE;
384         if (!may_write_to_queue(mapping->backing_dev_info))
385                 return PAGE_KEEP;
386
387         if (clear_page_dirty_for_io(page)) {
388                 int res;
389                 struct writeback_control wbc = {
390                         .sync_mode = WB_SYNC_NONE,
391                         .nr_to_write = SWAP_CLUSTER_MAX,
392                         .range_start = 0,
393                         .range_end = LLONG_MAX,
394                         .nonblocking = 1,
395                         .for_reclaim = 1,
396                 };
397
398                 SetPageReclaim(page);
399                 res = mapping->a_ops->writepage(page, &wbc);
400                 if (res < 0)
401                         handle_write_error(mapping, page, res);
402                 if (res == AOP_WRITEPAGE_ACTIVATE) {
403                         ClearPageReclaim(page);
404                         return PAGE_ACTIVATE;
405                 }
406
407                 /*
408                  * Wait on writeback if requested to. This happens when
409                  * direct reclaiming a large contiguous area and the
410                  * first attempt to free a range of pages fails.
411                  */
412                 if (PageWriteback(page) && sync_writeback == PAGEOUT_IO_SYNC)
413                         wait_on_page_writeback(page);
414
415                 if (!PageWriteback(page)) {
416                         /* synchronous write or broken a_ops? */
417                         ClearPageReclaim(page);
418                 }
419                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
420                 return PAGE_SUCCESS;
421         }
422
423         return PAGE_CLEAN;
424 }
425
426 /*
427  * Same as remove_mapping, but if the page is removed from the mapping, it
428  * gets returned with a refcount of 0.
429  */
430 static int __remove_mapping(struct address_space *mapping, struct page *page)
431 {
432         BUG_ON(!PageLocked(page));
433         BUG_ON(mapping != page_mapping(page));
434
435         spin_lock_irq(&mapping->tree_lock);
436         /*
437          * The non racy check for a busy page.
438          *
439          * Must be careful with the order of the tests. When someone has
440          * a ref to the page, it may be possible that they dirty it then
441          * drop the reference. So if PageDirty is tested before page_count
442          * here, then the following race may occur:
443          *
444          * get_user_pages(&page);
445          * [user mapping goes away]
446          * write_to(page);
447          *                              !PageDirty(page)    [good]
448          * SetPageDirty(page);
449          * put_page(page);
450          *                              !page_count(page)   [good, discard it]
451          *
452          * [oops, our write_to data is lost]
453          *
454          * Reversing the order of the tests ensures such a situation cannot
455          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
456          * load is not satisfied before that of page->_count.
457          *
458          * Note that if SetPageDirty is always performed via set_page_dirty,
459          * and thus under tree_lock, then this ordering is not required.
460          */
461         if (!page_freeze_refs(page, 2))
462                 goto cannot_free;
463         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
464         if (unlikely(PageDirty(page))) {
465                 page_unfreeze_refs(page, 2);
466                 goto cannot_free;
467         }
468
469         if (PageSwapCache(page)) {
470                 swp_entry_t swap = { .val = page_private(page) };
471                 __delete_from_swap_cache(page);
472                 spin_unlock_irq(&mapping->tree_lock);
473                 swap_free(swap);
474         } else {
475                 __remove_from_page_cache(page);
476                 spin_unlock_irq(&mapping->tree_lock);
477         }
478
479         return 1;
480
481 cannot_free:
482         spin_unlock_irq(&mapping->tree_lock);
483         return 0;
484 }
485
486 /*
487  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
488  * someone else has a ref on the page, abort and return 0.  If it was
489  * successfully detached, return 1.  Assumes the caller has a single ref on
490  * this page.
491  */
492 int remove_mapping(struct address_space *mapping, struct page *page)
493 {
494         if (__remove_mapping(mapping, page)) {
495                 /*
496                  * Unfreezing the refcount with 1 rather than 2 effectively
497                  * drops the pagecache ref for us without requiring another
498                  * atomic operation.
499                  */
500                 page_unfreeze_refs(page, 1);
501                 return 1;
502         }
503         return 0;
504 }
505
506 /**
507  * putback_lru_page - put previously isolated page onto appropriate LRU list
508  * @page: page to be put back to appropriate lru list
509  *
510  * Add previously isolated @page to appropriate LRU list.
511  * Page may still be unevictable for other reasons.
512  *
513  * lru_lock must not be held, interrupts must be enabled.
514  */
515 #ifdef CONFIG_UNEVICTABLE_LRU
516 void putback_lru_page(struct page *page)
517 {
518         int lru;
519         int active = !!TestClearPageActive(page);
520         int was_unevictable = PageUnevictable(page);
521
522         VM_BUG_ON(PageLRU(page));
523
524 redo:
525         ClearPageUnevictable(page);
526
527         if (page_evictable(page, NULL)) {
528                 /*
529                  * For evictable pages, we can use the cache.
530                  * In event of a race, worst case is we end up with an
531                  * unevictable page on [in]active list.
532                  * We know how to handle that.
533                  */
534                 lru = active + page_is_file_cache(page);
535                 lru_cache_add_lru(page, lru);
536         } else {
537                 /*
538                  * Put unevictable pages directly on zone's unevictable
539                  * list.
540                  */
541                 lru = LRU_UNEVICTABLE;
542                 add_page_to_unevictable_list(page);
543         }
544
545         /*
546          * page's status can change while we move it among lru. If an evictable
547          * page is on unevictable list, it never be freed. To avoid that,
548          * check after we added it to the list, again.
549          */
550         if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
551                 if (!isolate_lru_page(page)) {
552                         put_page(page);
553                         goto redo;
554                 }
555                 /* This means someone else dropped this page from LRU
556                  * So, it will be freed or putback to LRU again. There is
557                  * nothing to do here.
558                  */
559         }
560
561         if (was_unevictable && lru != LRU_UNEVICTABLE)
562                 count_vm_event(UNEVICTABLE_PGRESCUED);
563         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
564                 count_vm_event(UNEVICTABLE_PGCULLED);
565
566         put_page(page);         /* drop ref from isolate */
567 }
568
569 #else /* CONFIG_UNEVICTABLE_LRU */
570
571 void putback_lru_page(struct page *page)
572 {
573         int lru;
574         VM_BUG_ON(PageLRU(page));
575
576         lru = !!TestClearPageActive(page) + page_is_file_cache(page);
577         lru_cache_add_lru(page, lru);
578         put_page(page);
579 }
580 #endif /* CONFIG_UNEVICTABLE_LRU */
581
582
583 /*
584  * shrink_page_list() returns the number of reclaimed pages
585  */
586 static unsigned long shrink_page_list(struct list_head *page_list,
587                                         struct scan_control *sc,
588                                         enum pageout_io sync_writeback)
589 {
590         LIST_HEAD(ret_pages);
591         struct pagevec freed_pvec;
592         int pgactivate = 0;
593         unsigned long nr_reclaimed = 0;
594
595         cond_resched();
596
597         pagevec_init(&freed_pvec, 1);
598         while (!list_empty(page_list)) {
599                 struct address_space *mapping;
600                 struct page *page;
601                 int may_enter_fs;
602                 int referenced;
603
604                 cond_resched();
605
606                 page = lru_to_page(page_list);
607                 list_del(&page->lru);
608
609                 if (!trylock_page(page))
610                         goto keep;
611
612                 VM_BUG_ON(PageActive(page));
613
614                 sc->nr_scanned++;
615
616                 if (unlikely(!page_evictable(page, NULL)))
617                         goto cull_mlocked;
618
619                 if (!sc->may_unmap && page_mapped(page))
620                         goto keep_locked;
621
622                 /* Double the slab pressure for mapped and swapcache pages */
623                 if (page_mapped(page) || PageSwapCache(page))
624                         sc->nr_scanned++;
625
626                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
627                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
628
629                 if (PageWriteback(page)) {
630                         /*
631                          * Synchronous reclaim is performed in two passes,
632                          * first an asynchronous pass over the list to
633                          * start parallel writeback, and a second synchronous
634                          * pass to wait for the IO to complete.  Wait here
635                          * for any page for which writeback has already
636                          * started.
637                          */
638                         if (sync_writeback == PAGEOUT_IO_SYNC && may_enter_fs)
639                                 wait_on_page_writeback(page);
640                         else
641                                 goto keep_locked;
642                 }
643
644                 referenced = page_referenced(page, 1, sc->mem_cgroup);
645                 /* In active use or really unfreeable?  Activate it. */
646                 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER &&
647                                         referenced && page_mapping_inuse(page))
648                         goto activate_locked;
649
650                 /*
651                  * Anonymous process memory has backing store?
652                  * Try to allocate it some swap space here.
653                  */
654                 if (PageAnon(page) && !PageSwapCache(page)) {
655                         if (!(sc->gfp_mask & __GFP_IO))
656                                 goto keep_locked;
657                         if (!add_to_swap(page))
658                                 goto activate_locked;
659                         may_enter_fs = 1;
660                 }
661
662                 mapping = page_mapping(page);
663
664                 /*
665                  * The page is mapped into the page tables of one or more
666                  * processes. Try to unmap it here.
667                  */
668                 if (page_mapped(page) && mapping) {
669                         switch (try_to_unmap(page, 0)) {
670                         case SWAP_FAIL:
671                                 goto activate_locked;
672                         case SWAP_AGAIN:
673                                 goto keep_locked;
674                         case SWAP_MLOCK:
675                                 goto cull_mlocked;
676                         case SWAP_SUCCESS:
677                                 ; /* try to free the page below */
678                         }
679                 }
680
681                 if (PageDirty(page)) {
682                         if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && referenced)
683                                 goto keep_locked;
684                         if (!may_enter_fs)
685                                 goto keep_locked;
686                         if (!sc->may_writepage)
687                                 goto keep_locked;
688
689                         /* Page is dirty, try to write it out here */
690                         switch (pageout(page, mapping, sync_writeback)) {
691                         case PAGE_KEEP:
692                                 goto keep_locked;
693                         case PAGE_ACTIVATE:
694                                 goto activate_locked;
695                         case PAGE_SUCCESS:
696                                 if (PageWriteback(page) || PageDirty(page))
697                                         goto keep;
698                                 /*
699                                  * A synchronous write - probably a ramdisk.  Go
700                                  * ahead and try to reclaim the page.
701                                  */
702                                 if (!trylock_page(page))
703                                         goto keep;
704                                 if (PageDirty(page) || PageWriteback(page))
705                                         goto keep_locked;
706                                 mapping = page_mapping(page);
707                         case PAGE_CLEAN:
708                                 ; /* try to free the page below */
709                         }
710                 }
711
712                 /*
713                  * If the page has buffers, try to free the buffer mappings
714                  * associated with this page. If we succeed we try to free
715                  * the page as well.
716                  *
717                  * We do this even if the page is PageDirty().
718                  * try_to_release_page() does not perform I/O, but it is
719                  * possible for a page to have PageDirty set, but it is actually
720                  * clean (all its buffers are clean).  This happens if the
721                  * buffers were written out directly, with submit_bh(). ext3
722                  * will do this, as well as the blockdev mapping.
723                  * try_to_release_page() will discover that cleanness and will
724                  * drop the buffers and mark the page clean - it can be freed.
725                  *
726                  * Rarely, pages can have buffers and no ->mapping.  These are
727                  * the pages which were not successfully invalidated in
728                  * truncate_complete_page().  We try to drop those buffers here
729                  * and if that worked, and the page is no longer mapped into
730                  * process address space (page_count == 1) it can be freed.
731                  * Otherwise, leave the page on the LRU so it is swappable.
732                  */
733                 if (page_has_private(page)) {
734                         if (!try_to_release_page(page, sc->gfp_mask))
735                                 goto activate_locked;
736                         if (!mapping && page_count(page) == 1) {
737                                 unlock_page(page);
738                                 if (put_page_testzero(page))
739                                         goto free_it;
740                                 else {
741                                         /*
742                                          * rare race with speculative reference.
743                                          * the speculative reference will free
744                                          * this page shortly, so we may
745                                          * increment nr_reclaimed here (and
746                                          * leave it off the LRU).
747                                          */
748                                         nr_reclaimed++;
749                                         continue;
750                                 }
751                         }
752                 }
753
754                 if (!mapping || !__remove_mapping(mapping, page))
755                         goto keep_locked;
756
757                 /*
758                  * At this point, we have no other references and there is
759                  * no way to pick any more up (removed from LRU, removed
760                  * from pagecache). Can use non-atomic bitops now (and
761                  * we obviously don't have to worry about waking up a process
762                  * waiting on the page lock, because there are no references.
763                  */
764                 __clear_page_locked(page);
765 free_it:
766                 nr_reclaimed++;
767                 if (!pagevec_add(&freed_pvec, page)) {
768                         __pagevec_free(&freed_pvec);
769                         pagevec_reinit(&freed_pvec);
770                 }
771                 continue;
772
773 cull_mlocked:
774                 if (PageSwapCache(page))
775                         try_to_free_swap(page);
776                 unlock_page(page);
777                 putback_lru_page(page);
778                 continue;
779
780 activate_locked:
781                 /* Not a candidate for swapping, so reclaim swap space. */
782                 if (PageSwapCache(page) && vm_swap_full())
783                         try_to_free_swap(page);
784                 VM_BUG_ON(PageActive(page));
785                 SetPageActive(page);
786                 pgactivate++;
787 keep_locked:
788                 unlock_page(page);
789 keep:
790                 list_add(&page->lru, &ret_pages);
791                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
792         }
793         list_splice(&ret_pages, page_list);
794         if (pagevec_count(&freed_pvec))
795                 __pagevec_free(&freed_pvec);
796         count_vm_events(PGACTIVATE, pgactivate);
797         return nr_reclaimed;
798 }
799
800 /* LRU Isolation modes. */
801 #define ISOLATE_INACTIVE 0      /* Isolate inactive pages. */
802 #define ISOLATE_ACTIVE 1        /* Isolate active pages. */
803 #define ISOLATE_BOTH 2          /* Isolate both active and inactive pages. */
804
805 /*
806  * Attempt to remove the specified page from its LRU.  Only take this page
807  * if it is of the appropriate PageActive status.  Pages which are being
808  * freed elsewhere are also ignored.
809  *
810  * page:        page to consider
811  * mode:        one of the LRU isolation modes defined above
812  *
813  * returns 0 on success, -ve errno on failure.
814  */
815 int __isolate_lru_page(struct page *page, int mode, int file)
816 {
817         int ret = -EINVAL;
818
819         /* Only take pages on the LRU. */
820         if (!PageLRU(page))
821                 return ret;
822
823         /*
824          * When checking the active state, we need to be sure we are
825          * dealing with comparible boolean values.  Take the logical not
826          * of each.
827          */
828         if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
829                 return ret;
830
831         if (mode != ISOLATE_BOTH && (!page_is_file_cache(page) != !file))
832                 return ret;
833
834         /*
835          * When this function is being called for lumpy reclaim, we
836          * initially look into all LRU pages, active, inactive and
837          * unevictable; only give shrink_page_list evictable pages.
838          */
839         if (PageUnevictable(page))
840                 return ret;
841
842         ret = -EBUSY;
843
844         if (likely(get_page_unless_zero(page))) {
845                 /*
846                  * Be careful not to clear PageLRU until after we're
847                  * sure the page is not being freed elsewhere -- the
848                  * page release code relies on it.
849                  */
850                 ClearPageLRU(page);
851                 ret = 0;
852                 mem_cgroup_del_lru(page);
853         }
854
855         return ret;
856 }
857
858 /*
859  * zone->lru_lock is heavily contended.  Some of the functions that
860  * shrink the lists perform better by taking out a batch of pages
861  * and working on them outside the LRU lock.
862  *
863  * For pagecache intensive workloads, this function is the hottest
864  * spot in the kernel (apart from copy_*_user functions).
865  *
866  * Appropriate locks must be held before calling this function.
867  *
868  * @nr_to_scan: The number of pages to look through on the list.
869  * @src:        The LRU list to pull pages off.
870  * @dst:        The temp list to put pages on to.
871  * @scanned:    The number of pages that were scanned.
872  * @order:      The caller's attempted allocation order
873  * @mode:       One of the LRU isolation modes
874  * @file:       True [1] if isolating file [!anon] pages
875  *
876  * returns how many pages were moved onto *@dst.
877  */
878 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
879                 struct list_head *src, struct list_head *dst,
880                 unsigned long *scanned, int order, int mode, int file)
881 {
882         unsigned long nr_taken = 0;
883         unsigned long scan;
884
885         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
886                 struct page *page;
887                 unsigned long pfn;
888                 unsigned long end_pfn;
889                 unsigned long page_pfn;
890                 int zone_id;
891
892                 page = lru_to_page(src);
893                 prefetchw_prev_lru_page(page, src, flags);
894
895                 VM_BUG_ON(!PageLRU(page));
896
897                 switch (__isolate_lru_page(page, mode, file)) {
898                 case 0:
899                         list_move(&page->lru, dst);
900                         nr_taken++;
901                         break;
902
903                 case -EBUSY:
904                         /* else it is being freed elsewhere */
905                         list_move(&page->lru, src);
906                         continue;
907
908                 default:
909                         BUG();
910                 }
911
912                 if (!order)
913                         continue;
914
915                 /*
916                  * Attempt to take all pages in the order aligned region
917                  * surrounding the tag page.  Only take those pages of
918                  * the same active state as that tag page.  We may safely
919                  * round the target page pfn down to the requested order
920                  * as the mem_map is guarenteed valid out to MAX_ORDER,
921                  * where that page is in a different zone we will detect
922                  * it from its zone id and abort this block scan.
923                  */
924                 zone_id = page_zone_id(page);
925                 page_pfn = page_to_pfn(page);
926                 pfn = page_pfn & ~((1 << order) - 1);
927                 end_pfn = pfn + (1 << order);
928                 for (; pfn < end_pfn; pfn++) {
929                         struct page *cursor_page;
930
931                         /* The target page is in the block, ignore it. */
932                         if (unlikely(pfn == page_pfn))
933                                 continue;
934
935                         /* Avoid holes within the zone. */
936                         if (unlikely(!pfn_valid_within(pfn)))
937                                 break;
938
939                         cursor_page = pfn_to_page(pfn);
940
941                         /* Check that we have not crossed a zone boundary. */
942                         if (unlikely(page_zone_id(cursor_page) != zone_id))
943                                 continue;
944                         switch (__isolate_lru_page(cursor_page, mode, file)) {
945                         case 0:
946                                 list_move(&cursor_page->lru, dst);
947                                 nr_taken++;
948                                 scan++;
949                                 break;
950
951                         case -EBUSY:
952                                 /* else it is being freed elsewhere */
953                                 list_move(&cursor_page->lru, src);
954                         default:
955                                 break;  /* ! on LRU or wrong list */
956                         }
957                 }
958         }
959
960         *scanned = scan;
961         return nr_taken;
962 }
963
964 static unsigned long isolate_pages_global(unsigned long nr,
965                                         struct list_head *dst,
966                                         unsigned long *scanned, int order,
967                                         int mode, struct zone *z,
968                                         struct mem_cgroup *mem_cont,
969                                         int active, int file)
970 {
971         int lru = LRU_BASE;
972         if (active)
973                 lru += LRU_ACTIVE;
974         if (file)
975                 lru += LRU_FILE;
976         return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
977                                                                 mode, !!file);
978 }
979
980 /*
981  * clear_active_flags() is a helper for shrink_active_list(), clearing
982  * any active bits from the pages in the list.
983  */
984 static unsigned long clear_active_flags(struct list_head *page_list,
985                                         unsigned int *count)
986 {
987         int nr_active = 0;
988         int lru;
989         struct page *page;
990
991         list_for_each_entry(page, page_list, lru) {
992                 lru = page_is_file_cache(page);
993                 if (PageActive(page)) {
994                         lru += LRU_ACTIVE;
995                         ClearPageActive(page);
996                         nr_active++;
997                 }
998                 count[lru]++;
999         }
1000
1001         return nr_active;
1002 }
1003
1004 /**
1005  * isolate_lru_page - tries to isolate a page from its LRU list
1006  * @page: page to isolate from its LRU list
1007  *
1008  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1009  * vmstat statistic corresponding to whatever LRU list the page was on.
1010  *
1011  * Returns 0 if the page was removed from an LRU list.
1012  * Returns -EBUSY if the page was not on an LRU list.
1013  *
1014  * The returned page will have PageLRU() cleared.  If it was found on
1015  * the active list, it will have PageActive set.  If it was found on
1016  * the unevictable list, it will have the PageUnevictable bit set. That flag
1017  * may need to be cleared by the caller before letting the page go.
1018  *
1019  * The vmstat statistic corresponding to the list on which the page was
1020  * found will be decremented.
1021  *
1022  * Restrictions:
1023  * (1) Must be called with an elevated refcount on the page. This is a
1024  *     fundamentnal difference from isolate_lru_pages (which is called
1025  *     without a stable reference).
1026  * (2) the lru_lock must not be held.
1027  * (3) interrupts must be enabled.
1028  */
1029 int isolate_lru_page(struct page *page)
1030 {
1031         int ret = -EBUSY;
1032
1033         if (PageLRU(page)) {
1034                 struct zone *zone = page_zone(page);
1035
1036                 spin_lock_irq(&zone->lru_lock);
1037                 if (PageLRU(page) && get_page_unless_zero(page)) {
1038                         int lru = page_lru(page);
1039                         ret = 0;
1040                         ClearPageLRU(page);
1041
1042                         del_page_from_lru_list(zone, page, lru);
1043                 }
1044                 spin_unlock_irq(&zone->lru_lock);
1045         }
1046         return ret;
1047 }
1048
1049 /*
1050  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1051  * of reclaimed pages
1052  */
1053 static unsigned long shrink_inactive_list(unsigned long max_scan,
1054                         struct zone *zone, struct scan_control *sc,
1055                         int priority, int file)
1056 {
1057         LIST_HEAD(page_list);
1058         struct pagevec pvec;
1059         unsigned long nr_scanned = 0;
1060         unsigned long nr_reclaimed = 0;
1061         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1062
1063         pagevec_init(&pvec, 1);
1064
1065         lru_add_drain();
1066         spin_lock_irq(&zone->lru_lock);
1067         do {
1068                 struct page *page;
1069                 unsigned long nr_taken;
1070                 unsigned long nr_scan;
1071                 unsigned long nr_freed;
1072                 unsigned long nr_active;
1073                 unsigned int count[NR_LRU_LISTS] = { 0, };
1074                 int mode = ISOLATE_INACTIVE;
1075
1076                 /*
1077                  * If we need a large contiguous chunk of memory, or have
1078                  * trouble getting a small set of contiguous pages, we
1079                  * will reclaim both active and inactive pages.
1080                  *
1081                  * We use the same threshold as pageout congestion_wait below.
1082                  */
1083                 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1084                         mode = ISOLATE_BOTH;
1085                 else if (sc->order && priority < DEF_PRIORITY - 2)
1086                         mode = ISOLATE_BOTH;
1087
1088                 nr_taken = sc->isolate_pages(sc->swap_cluster_max,
1089                              &page_list, &nr_scan, sc->order, mode,
1090                                 zone, sc->mem_cgroup, 0, file);
1091                 nr_active = clear_active_flags(&page_list, count);
1092                 __count_vm_events(PGDEACTIVATE, nr_active);
1093
1094                 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1095                                                 -count[LRU_ACTIVE_FILE]);
1096                 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1097                                                 -count[LRU_INACTIVE_FILE]);
1098                 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1099                                                 -count[LRU_ACTIVE_ANON]);
1100                 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1101                                                 -count[LRU_INACTIVE_ANON]);
1102
1103                 if (scanning_global_lru(sc))
1104                         zone->pages_scanned += nr_scan;
1105
1106                 reclaim_stat->recent_scanned[0] += count[LRU_INACTIVE_ANON];
1107                 reclaim_stat->recent_scanned[0] += count[LRU_ACTIVE_ANON];
1108                 reclaim_stat->recent_scanned[1] += count[LRU_INACTIVE_FILE];
1109                 reclaim_stat->recent_scanned[1] += count[LRU_ACTIVE_FILE];
1110
1111                 spin_unlock_irq(&zone->lru_lock);
1112
1113                 nr_scanned += nr_scan;
1114                 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
1115
1116                 /*
1117                  * If we are direct reclaiming for contiguous pages and we do
1118                  * not reclaim everything in the list, try again and wait
1119                  * for IO to complete. This will stall high-order allocations
1120                  * but that should be acceptable to the caller
1121                  */
1122                 if (nr_freed < nr_taken && !current_is_kswapd() &&
1123                                         sc->order > PAGE_ALLOC_COSTLY_ORDER) {
1124                         congestion_wait(WRITE, HZ/10);
1125
1126                         /*
1127                          * The attempt at page out may have made some
1128                          * of the pages active, mark them inactive again.
1129                          */
1130                         nr_active = clear_active_flags(&page_list, count);
1131                         count_vm_events(PGDEACTIVATE, nr_active);
1132
1133                         nr_freed += shrink_page_list(&page_list, sc,
1134                                                         PAGEOUT_IO_SYNC);
1135                 }
1136
1137                 nr_reclaimed += nr_freed;
1138                 local_irq_disable();
1139                 if (current_is_kswapd()) {
1140                         __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scan);
1141                         __count_vm_events(KSWAPD_STEAL, nr_freed);
1142                 } else if (scanning_global_lru(sc))
1143                         __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scan);
1144
1145                 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
1146
1147                 if (nr_taken == 0)
1148                         goto done;
1149
1150                 spin_lock(&zone->lru_lock);
1151                 /*
1152                  * Put back any unfreeable pages.
1153                  */
1154                 while (!list_empty(&page_list)) {
1155                         int lru;
1156                         page = lru_to_page(&page_list);
1157                         VM_BUG_ON(PageLRU(page));
1158                         list_del(&page->lru);
1159                         if (unlikely(!page_evictable(page, NULL))) {
1160                                 spin_unlock_irq(&zone->lru_lock);
1161                                 putback_lru_page(page);
1162                                 spin_lock_irq(&zone->lru_lock);
1163                                 continue;
1164                         }
1165                         SetPageLRU(page);
1166                         lru = page_lru(page);
1167                         add_page_to_lru_list(zone, page, lru);
1168                         if (PageActive(page)) {
1169                                 int file = !!page_is_file_cache(page);
1170                                 reclaim_stat->recent_rotated[file]++;
1171                         }
1172                         if (!pagevec_add(&pvec, page)) {
1173                                 spin_unlock_irq(&zone->lru_lock);
1174                                 __pagevec_release(&pvec);
1175                                 spin_lock_irq(&zone->lru_lock);
1176                         }
1177                 }
1178         } while (nr_scanned < max_scan);
1179         spin_unlock(&zone->lru_lock);
1180 done:
1181         local_irq_enable();
1182         pagevec_release(&pvec);
1183         return nr_reclaimed;
1184 }
1185
1186 /*
1187  * We are about to scan this zone at a certain priority level.  If that priority
1188  * level is smaller (ie: more urgent) than the previous priority, then note
1189  * that priority level within the zone.  This is done so that when the next
1190  * process comes in to scan this zone, it will immediately start out at this
1191  * priority level rather than having to build up its own scanning priority.
1192  * Here, this priority affects only the reclaim-mapped threshold.
1193  */
1194 static inline void note_zone_scanning_priority(struct zone *zone, int priority)
1195 {
1196         if (priority < zone->prev_priority)
1197                 zone->prev_priority = priority;
1198 }
1199
1200 /*
1201  * This moves pages from the active list to the inactive list.
1202  *
1203  * We move them the other way if the page is referenced by one or more
1204  * processes, from rmap.
1205  *
1206  * If the pages are mostly unmapped, the processing is fast and it is
1207  * appropriate to hold zone->lru_lock across the whole operation.  But if
1208  * the pages are mapped, the processing is slow (page_referenced()) so we
1209  * should drop zone->lru_lock around each page.  It's impossible to balance
1210  * this, so instead we remove the pages from the LRU while processing them.
1211  * It is safe to rely on PG_active against the non-LRU pages in here because
1212  * nobody will play with that bit on a non-LRU page.
1213  *
1214  * The downside is that we have to touch page->_count against each page.
1215  * But we had to alter page->flags anyway.
1216  */
1217
1218
1219 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1220                         struct scan_control *sc, int priority, int file)
1221 {
1222         unsigned long pgmoved;
1223         int pgdeactivate = 0;
1224         unsigned long pgscanned;
1225         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1226         LIST_HEAD(l_inactive);
1227         struct page *page;
1228         struct pagevec pvec;
1229         enum lru_list lru;
1230         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1231
1232         lru_add_drain();
1233         spin_lock_irq(&zone->lru_lock);
1234         pgmoved = sc->isolate_pages(nr_pages, &l_hold, &pgscanned, sc->order,
1235                                         ISOLATE_ACTIVE, zone,
1236                                         sc->mem_cgroup, 1, file);
1237         /*
1238          * zone->pages_scanned is used for detect zone's oom
1239          * mem_cgroup remembers nr_scan by itself.
1240          */
1241         if (scanning_global_lru(sc)) {
1242                 zone->pages_scanned += pgscanned;
1243         }
1244         reclaim_stat->recent_scanned[!!file] += pgmoved;
1245
1246         if (file)
1247                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -pgmoved);
1248         else
1249                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -pgmoved);
1250         spin_unlock_irq(&zone->lru_lock);
1251
1252         pgmoved = 0;
1253         while (!list_empty(&l_hold)) {
1254                 cond_resched();
1255                 page = lru_to_page(&l_hold);
1256                 list_del(&page->lru);
1257
1258                 if (unlikely(!page_evictable(page, NULL))) {
1259                         putback_lru_page(page);
1260                         continue;
1261                 }
1262
1263                 /* page_referenced clears PageReferenced */
1264                 if (page_mapping_inuse(page) &&
1265                     page_referenced(page, 0, sc->mem_cgroup))
1266                         pgmoved++;
1267
1268                 list_add(&page->lru, &l_inactive);
1269         }
1270
1271         /*
1272          * Move the pages to the [file or anon] inactive list.
1273          */
1274         pagevec_init(&pvec, 1);
1275         lru = LRU_BASE + file * LRU_FILE;
1276
1277         spin_lock_irq(&zone->lru_lock);
1278         /*
1279          * Count referenced pages from currently used mappings as
1280          * rotated, even though they are moved to the inactive list.
1281          * This helps balance scan pressure between file and anonymous
1282          * pages in get_scan_ratio.
1283          */
1284         reclaim_stat->recent_rotated[!!file] += pgmoved;
1285
1286         pgmoved = 0;
1287         while (!list_empty(&l_inactive)) {
1288                 page = lru_to_page(&l_inactive);
1289                 prefetchw_prev_lru_page(page, &l_inactive, flags);
1290                 VM_BUG_ON(PageLRU(page));
1291                 SetPageLRU(page);
1292                 VM_BUG_ON(!PageActive(page));
1293                 ClearPageActive(page);
1294
1295                 list_move(&page->lru, &zone->lru[lru].list);
1296                 mem_cgroup_add_lru_list(page, lru);
1297                 pgmoved++;
1298                 if (!pagevec_add(&pvec, page)) {
1299                         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1300                         spin_unlock_irq(&zone->lru_lock);
1301                         pgdeactivate += pgmoved;
1302                         pgmoved = 0;
1303                         if (buffer_heads_over_limit)
1304                                 pagevec_strip(&pvec);
1305                         __pagevec_release(&pvec);
1306                         spin_lock_irq(&zone->lru_lock);
1307                 }
1308         }
1309         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1310         pgdeactivate += pgmoved;
1311         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1312         __count_vm_events(PGDEACTIVATE, pgdeactivate);
1313         spin_unlock_irq(&zone->lru_lock);
1314         if (buffer_heads_over_limit)
1315                 pagevec_strip(&pvec);
1316         pagevec_release(&pvec);
1317 }
1318
1319 static int inactive_anon_is_low_global(struct zone *zone)
1320 {
1321         unsigned long active, inactive;
1322
1323         active = zone_page_state(zone, NR_ACTIVE_ANON);
1324         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1325
1326         if (inactive * zone->inactive_ratio < active)
1327                 return 1;
1328
1329         return 0;
1330 }
1331
1332 /**
1333  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1334  * @zone: zone to check
1335  * @sc:   scan control of this context
1336  *
1337  * Returns true if the zone does not have enough inactive anon pages,
1338  * meaning some active anon pages need to be deactivated.
1339  */
1340 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1341 {
1342         int low;
1343
1344         if (scanning_global_lru(sc))
1345                 low = inactive_anon_is_low_global(zone);
1346         else
1347                 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1348         return low;
1349 }
1350
1351 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1352         struct zone *zone, struct scan_control *sc, int priority)
1353 {
1354         int file = is_file_lru(lru);
1355
1356         if (lru == LRU_ACTIVE_FILE) {
1357                 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1358                 return 0;
1359         }
1360
1361         if (lru == LRU_ACTIVE_ANON && inactive_anon_is_low(zone, sc)) {
1362                 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1363                 return 0;
1364         }
1365         return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1366 }
1367
1368 /*
1369  * Determine how aggressively the anon and file LRU lists should be
1370  * scanned.  The relative value of each set of LRU lists is determined
1371  * by looking at the fraction of the pages scanned we did rotate back
1372  * onto the active list instead of evict.
1373  *
1374  * percent[0] specifies how much pressure to put on ram/swap backed
1375  * memory, while percent[1] determines pressure on the file LRUs.
1376  */
1377 static void get_scan_ratio(struct zone *zone, struct scan_control *sc,
1378                                         unsigned long *percent)
1379 {
1380         unsigned long anon, file, free;
1381         unsigned long anon_prio, file_prio;
1382         unsigned long ap, fp;
1383         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1384
1385         /* If we have no swap space, do not bother scanning anon pages. */
1386         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1387                 percent[0] = 0;
1388                 percent[1] = 100;
1389                 return;
1390         }
1391
1392         anon  = zone_nr_pages(zone, sc, LRU_ACTIVE_ANON) +
1393                 zone_nr_pages(zone, sc, LRU_INACTIVE_ANON);
1394         file  = zone_nr_pages(zone, sc, LRU_ACTIVE_FILE) +
1395                 zone_nr_pages(zone, sc, LRU_INACTIVE_FILE);
1396
1397         if (scanning_global_lru(sc)) {
1398                 free  = zone_page_state(zone, NR_FREE_PAGES);
1399                 /* If we have very few page cache pages,
1400                    force-scan anon pages. */
1401                 if (unlikely(file + free <= zone->pages_high)) {
1402                         percent[0] = 100;
1403                         percent[1] = 0;
1404                         return;
1405                 }
1406         }
1407
1408         /*
1409          * OK, so we have swap space and a fair amount of page cache
1410          * pages.  We use the recently rotated / recently scanned
1411          * ratios to determine how valuable each cache is.
1412          *
1413          * Because workloads change over time (and to avoid overflow)
1414          * we keep these statistics as a floating average, which ends
1415          * up weighing recent references more than old ones.
1416          *
1417          * anon in [0], file in [1]
1418          */
1419         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1420                 spin_lock_irq(&zone->lru_lock);
1421                 reclaim_stat->recent_scanned[0] /= 2;
1422                 reclaim_stat->recent_rotated[0] /= 2;
1423                 spin_unlock_irq(&zone->lru_lock);
1424         }
1425
1426         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1427                 spin_lock_irq(&zone->lru_lock);
1428                 reclaim_stat->recent_scanned[1] /= 2;
1429                 reclaim_stat->recent_rotated[1] /= 2;
1430                 spin_unlock_irq(&zone->lru_lock);
1431         }
1432
1433         /*
1434          * With swappiness at 100, anonymous and file have the same priority.
1435          * This scanning priority is essentially the inverse of IO cost.
1436          */
1437         anon_prio = sc->swappiness;
1438         file_prio = 200 - sc->swappiness;
1439
1440         /*
1441          * The amount of pressure on anon vs file pages is inversely
1442          * proportional to the fraction of recently scanned pages on
1443          * each list that were recently referenced and in active use.
1444          */
1445         ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1446         ap /= reclaim_stat->recent_rotated[0] + 1;
1447
1448         fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1449         fp /= reclaim_stat->recent_rotated[1] + 1;
1450
1451         /* Normalize to percentages */
1452         percent[0] = 100 * ap / (ap + fp + 1);
1453         percent[1] = 100 - percent[0];
1454 }
1455
1456
1457 /*
1458  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1459  */
1460 static void shrink_zone(int priority, struct zone *zone,
1461                                 struct scan_control *sc)
1462 {
1463         unsigned long nr[NR_LRU_LISTS];
1464         unsigned long nr_to_scan;
1465         unsigned long percent[2];       /* anon @ 0; file @ 1 */
1466         enum lru_list l;
1467         unsigned long nr_reclaimed = sc->nr_reclaimed;
1468         unsigned long swap_cluster_max = sc->swap_cluster_max;
1469
1470         get_scan_ratio(zone, sc, percent);
1471
1472         for_each_evictable_lru(l) {
1473                 int file = is_file_lru(l);
1474                 unsigned long scan;
1475
1476                 scan = zone_nr_pages(zone, sc, l);
1477                 if (priority) {
1478                         scan >>= priority;
1479                         scan = (scan * percent[file]) / 100;
1480                 }
1481                 if (scanning_global_lru(sc)) {
1482                         zone->lru[l].nr_scan += scan;
1483                         nr[l] = zone->lru[l].nr_scan;
1484                         if (nr[l] >= swap_cluster_max)
1485                                 zone->lru[l].nr_scan = 0;
1486                         else
1487                                 nr[l] = 0;
1488                 } else
1489                         nr[l] = scan;
1490         }
1491
1492         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1493                                         nr[LRU_INACTIVE_FILE]) {
1494                 for_each_evictable_lru(l) {
1495                         if (nr[l]) {
1496                                 nr_to_scan = min(nr[l], swap_cluster_max);
1497                                 nr[l] -= nr_to_scan;
1498
1499                                 nr_reclaimed += shrink_list(l, nr_to_scan,
1500                                                             zone, sc, priority);
1501                         }
1502                 }
1503                 /*
1504                  * On large memory systems, scan >> priority can become
1505                  * really large. This is fine for the starting priority;
1506                  * we want to put equal scanning pressure on each zone.
1507                  * However, if the VM has a harder time of freeing pages,
1508                  * with multiple processes reclaiming pages, the total
1509                  * freeing target can get unreasonably large.
1510                  */
1511                 if (nr_reclaimed > swap_cluster_max &&
1512                         priority < DEF_PRIORITY && !current_is_kswapd())
1513                         break;
1514         }
1515
1516         sc->nr_reclaimed = nr_reclaimed;
1517
1518         /*
1519          * Even if we did not try to evict anon pages at all, we want to
1520          * rebalance the anon lru active/inactive ratio.
1521          */
1522         if (inactive_anon_is_low(zone, sc))
1523                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1524
1525         throttle_vm_writeout(sc->gfp_mask);
1526 }
1527
1528 /*
1529  * This is the direct reclaim path, for page-allocating processes.  We only
1530  * try to reclaim pages from zones which will satisfy the caller's allocation
1531  * request.
1532  *
1533  * We reclaim from a zone even if that zone is over pages_high.  Because:
1534  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1535  *    allocation or
1536  * b) The zones may be over pages_high but they must go *over* pages_high to
1537  *    satisfy the `incremental min' zone defense algorithm.
1538  *
1539  * If a zone is deemed to be full of pinned pages then just give it a light
1540  * scan then give up on it.
1541  */
1542 static void shrink_zones(int priority, struct zonelist *zonelist,
1543                                         struct scan_control *sc)
1544 {
1545         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1546         struct zoneref *z;
1547         struct zone *zone;
1548
1549         sc->all_unreclaimable = 1;
1550         for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1551                                         sc->nodemask) {
1552                 if (!populated_zone(zone))
1553                         continue;
1554                 /*
1555                  * Take care memory controller reclaiming has small influence
1556                  * to global LRU.
1557                  */
1558                 if (scanning_global_lru(sc)) {
1559                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1560                                 continue;
1561                         note_zone_scanning_priority(zone, priority);
1562
1563                         if (zone_is_all_unreclaimable(zone) &&
1564                                                 priority != DEF_PRIORITY)
1565                                 continue;       /* Let kswapd poll it */
1566                         sc->all_unreclaimable = 0;
1567                 } else {
1568                         /*
1569                          * Ignore cpuset limitation here. We just want to reduce
1570                          * # of used pages by us regardless of memory shortage.
1571                          */
1572                         sc->all_unreclaimable = 0;
1573                         mem_cgroup_note_reclaim_priority(sc->mem_cgroup,
1574                                                         priority);
1575                 }
1576
1577                 shrink_zone(priority, zone, sc);
1578         }
1579 }
1580
1581 /*
1582  * This is the main entry point to direct page reclaim.
1583  *
1584  * If a full scan of the inactive list fails to free enough memory then we
1585  * are "out of memory" and something needs to be killed.
1586  *
1587  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1588  * high - the zone may be full of dirty or under-writeback pages, which this
1589  * caller can't do much about.  We kick pdflush and take explicit naps in the
1590  * hope that some of these pages can be written.  But if the allocating task
1591  * holds filesystem locks which prevent writeout this might not work, and the
1592  * allocation attempt will fail.
1593  *
1594  * returns:     0, if no pages reclaimed
1595  *              else, the number of pages reclaimed
1596  */
1597 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
1598                                         struct scan_control *sc)
1599 {
1600         int priority;
1601         unsigned long ret = 0;
1602         unsigned long total_scanned = 0;
1603         struct reclaim_state *reclaim_state = current->reclaim_state;
1604         unsigned long lru_pages = 0;
1605         struct zoneref *z;
1606         struct zone *zone;
1607         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1608
1609         delayacct_freepages_start();
1610
1611         if (scanning_global_lru(sc))
1612                 count_vm_event(ALLOCSTALL);
1613         /*
1614          * mem_cgroup will not do shrink_slab.
1615          */
1616         if (scanning_global_lru(sc)) {
1617                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1618
1619                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1620                                 continue;
1621
1622                         lru_pages += zone_lru_pages(zone);
1623                 }
1624         }
1625
1626         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1627                 sc->nr_scanned = 0;
1628                 if (!priority)
1629                         disable_swap_token();
1630                 shrink_zones(priority, zonelist, sc);
1631                 /*
1632                  * Don't shrink slabs when reclaiming memory from
1633                  * over limit cgroups
1634                  */
1635                 if (scanning_global_lru(sc)) {
1636                         shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
1637                         if (reclaim_state) {
1638                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
1639                                 reclaim_state->reclaimed_slab = 0;
1640                         }
1641                 }
1642                 total_scanned += sc->nr_scanned;
1643                 if (sc->nr_reclaimed >= sc->swap_cluster_max) {
1644                         ret = sc->nr_reclaimed;
1645                         goto out;
1646                 }
1647
1648                 /*
1649                  * Try to write back as many pages as we just scanned.  This
1650                  * tends to cause slow streaming writers to write data to the
1651                  * disk smoothly, at the dirtying rate, which is nice.   But
1652                  * that's undesirable in laptop mode, where we *want* lumpy
1653                  * writeout.  So in laptop mode, write out the whole world.
1654                  */
1655                 if (total_scanned > sc->swap_cluster_max +
1656                                         sc->swap_cluster_max / 2) {
1657                         wakeup_pdflush(laptop_mode ? 0 : total_scanned);
1658                         sc->may_writepage = 1;
1659                 }
1660
1661                 /* Take a nap, wait for some writeback to complete */
1662                 if (sc->nr_scanned && priority < DEF_PRIORITY - 2)
1663                         congestion_wait(WRITE, HZ/10);
1664         }
1665         /* top priority shrink_zones still had more to do? don't OOM, then */
1666         if (!sc->all_unreclaimable && scanning_global_lru(sc))
1667                 ret = sc->nr_reclaimed;
1668 out:
1669         /*
1670          * Now that we've scanned all the zones at this priority level, note
1671          * that level within the zone so that the next thread which performs
1672          * scanning of this zone will immediately start out at this priority
1673          * level.  This affects only the decision whether or not to bring
1674          * mapped pages onto the inactive list.
1675          */
1676         if (priority < 0)
1677                 priority = 0;
1678
1679         if (scanning_global_lru(sc)) {
1680                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1681
1682                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1683                                 continue;
1684
1685                         zone->prev_priority = priority;
1686                 }
1687         } else
1688                 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority);
1689
1690         delayacct_freepages_end();
1691
1692         return ret;
1693 }
1694
1695 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
1696                                 gfp_t gfp_mask, nodemask_t *nodemask)
1697 {
1698         struct scan_control sc = {
1699                 .gfp_mask = gfp_mask,
1700                 .may_writepage = !laptop_mode,
1701                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1702                 .may_unmap = 1,
1703                 .may_swap = 1,
1704                 .swappiness = vm_swappiness,
1705                 .order = order,
1706                 .mem_cgroup = NULL,
1707                 .isolate_pages = isolate_pages_global,
1708                 .nodemask = nodemask,
1709         };
1710
1711         return do_try_to_free_pages(zonelist, &sc);
1712 }
1713
1714 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
1715
1716 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
1717                                            gfp_t gfp_mask,
1718                                            bool noswap,
1719                                            unsigned int swappiness)
1720 {
1721         struct scan_control sc = {
1722                 .may_writepage = !laptop_mode,
1723                 .may_unmap = 1,
1724                 .may_swap = !noswap,
1725                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1726                 .swappiness = swappiness,
1727                 .order = 0,
1728                 .mem_cgroup = mem_cont,
1729                 .isolate_pages = mem_cgroup_isolate_pages,
1730                 .nodemask = NULL, /* we don't care the placement */
1731         };
1732         struct zonelist *zonelist;
1733
1734         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1735                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1736         zonelist = NODE_DATA(numa_node_id())->node_zonelists;
1737         return do_try_to_free_pages(zonelist, &sc);
1738 }
1739 #endif
1740
1741 /*
1742  * For kswapd, balance_pgdat() will work across all this node's zones until
1743  * they are all at pages_high.
1744  *
1745  * Returns the number of pages which were actually freed.
1746  *
1747  * There is special handling here for zones which are full of pinned pages.
1748  * This can happen if the pages are all mlocked, or if they are all used by
1749  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1750  * What we do is to detect the case where all pages in the zone have been
1751  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1752  * dead and from now on, only perform a short scan.  Basically we're polling
1753  * the zone for when the problem goes away.
1754  *
1755  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
1756  * zones which have free_pages > pages_high, but once a zone is found to have
1757  * free_pages <= pages_high, we scan that zone and the lower zones regardless
1758  * of the number of free pages in the lower zones.  This interoperates with
1759  * the page allocator fallback scheme to ensure that aging of pages is balanced
1760  * across the zones.
1761  */
1762 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
1763 {
1764         int all_zones_ok;
1765         int priority;
1766         int i;
1767         unsigned long total_scanned;
1768         struct reclaim_state *reclaim_state = current->reclaim_state;
1769         struct scan_control sc = {
1770                 .gfp_mask = GFP_KERNEL,
1771                 .may_unmap = 1,
1772                 .may_swap = 1,
1773                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1774                 .swappiness = vm_swappiness,
1775                 .order = order,
1776                 .mem_cgroup = NULL,
1777                 .isolate_pages = isolate_pages_global,
1778         };
1779         /*
1780          * temp_priority is used to remember the scanning priority at which
1781          * this zone was successfully refilled to free_pages == pages_high.
1782          */
1783         int temp_priority[MAX_NR_ZONES];
1784
1785 loop_again:
1786         total_scanned = 0;
1787         sc.nr_reclaimed = 0;
1788         sc.may_writepage = !laptop_mode;
1789         count_vm_event(PAGEOUTRUN);
1790
1791         for (i = 0; i < pgdat->nr_zones; i++)
1792                 temp_priority[i] = DEF_PRIORITY;
1793
1794         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1795                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1796                 unsigned long lru_pages = 0;
1797
1798                 /* The swap token gets in the way of swapout... */
1799                 if (!priority)
1800                         disable_swap_token();
1801
1802                 all_zones_ok = 1;
1803
1804                 /*
1805                  * Scan in the highmem->dma direction for the highest
1806                  * zone which needs scanning
1807                  */
1808                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1809                         struct zone *zone = pgdat->node_zones + i;
1810
1811                         if (!populated_zone(zone))
1812                                 continue;
1813
1814                         if (zone_is_all_unreclaimable(zone) &&
1815                             priority != DEF_PRIORITY)
1816                                 continue;
1817
1818                         /*
1819                          * Do some background aging of the anon list, to give
1820                          * pages a chance to be referenced before reclaiming.
1821                          */
1822                         if (inactive_anon_is_low(zone, &sc))
1823                                 shrink_active_list(SWAP_CLUSTER_MAX, zone,
1824                                                         &sc, priority, 0);
1825
1826                         if (!zone_watermark_ok(zone, order, zone->pages_high,
1827                                                0, 0)) {
1828                                 end_zone = i;
1829                                 break;
1830                         }
1831                 }
1832                 if (i < 0)
1833                         goto out;
1834
1835                 for (i = 0; i <= end_zone; i++) {
1836                         struct zone *zone = pgdat->node_zones + i;
1837
1838                         lru_pages += zone_lru_pages(zone);
1839                 }
1840
1841                 /*
1842                  * Now scan the zone in the dma->highmem direction, stopping
1843                  * at the last zone which needs scanning.
1844                  *
1845                  * We do this because the page allocator works in the opposite
1846                  * direction.  This prevents the page allocator from allocating
1847                  * pages behind kswapd's direction of progress, which would
1848                  * cause too much scanning of the lower zones.
1849                  */
1850                 for (i = 0; i <= end_zone; i++) {
1851                         struct zone *zone = pgdat->node_zones + i;
1852                         int nr_slab;
1853
1854                         if (!populated_zone(zone))
1855                                 continue;
1856
1857                         if (zone_is_all_unreclaimable(zone) &&
1858                                         priority != DEF_PRIORITY)
1859                                 continue;
1860
1861                         if (!zone_watermark_ok(zone, order, zone->pages_high,
1862                                                end_zone, 0))
1863                                 all_zones_ok = 0;
1864                         temp_priority[i] = priority;
1865                         sc.nr_scanned = 0;
1866                         note_zone_scanning_priority(zone, priority);
1867                         /*
1868                          * We put equal pressure on every zone, unless one
1869                          * zone has way too many pages free already.
1870                          */
1871                         if (!zone_watermark_ok(zone, order, 8*zone->pages_high,
1872                                                 end_zone, 0))
1873                                 shrink_zone(priority, zone, &sc);
1874                         reclaim_state->reclaimed_slab = 0;
1875                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1876                                                 lru_pages);
1877                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1878                         total_scanned += sc.nr_scanned;
1879                         if (zone_is_all_unreclaimable(zone))
1880                                 continue;
1881                         if (nr_slab == 0 && zone->pages_scanned >=
1882                                                 (zone_lru_pages(zone) * 6))
1883                                         zone_set_flag(zone,
1884                                                       ZONE_ALL_UNRECLAIMABLE);
1885                         /*
1886                          * If we've done a decent amount of scanning and
1887                          * the reclaim ratio is low, start doing writepage
1888                          * even in laptop mode
1889                          */
1890                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1891                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
1892                                 sc.may_writepage = 1;
1893                 }
1894                 if (all_zones_ok)
1895                         break;          /* kswapd: all done */
1896                 /*
1897                  * OK, kswapd is getting into trouble.  Take a nap, then take
1898                  * another pass across the zones.
1899                  */
1900                 if (total_scanned && priority < DEF_PRIORITY - 2)
1901                         congestion_wait(WRITE, HZ/10);
1902
1903                 /*
1904                  * We do this so kswapd doesn't build up large priorities for
1905                  * example when it is freeing in parallel with allocators. It
1906                  * matches the direct reclaim path behaviour in terms of impact
1907                  * on zone->*_priority.
1908                  */
1909                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
1910                         break;
1911         }
1912 out:
1913         /*
1914          * Note within each zone the priority level at which this zone was
1915          * brought into a happy state.  So that the next thread which scans this
1916          * zone will start out at that priority level.
1917          */
1918         for (i = 0; i < pgdat->nr_zones; i++) {
1919                 struct zone *zone = pgdat->node_zones + i;
1920
1921                 zone->prev_priority = temp_priority[i];
1922         }
1923         if (!all_zones_ok) {
1924                 cond_resched();
1925
1926                 try_to_freeze();
1927
1928                 /*
1929                  * Fragmentation may mean that the system cannot be
1930                  * rebalanced for high-order allocations in all zones.
1931                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
1932                  * it means the zones have been fully scanned and are still
1933                  * not balanced. For high-order allocations, there is
1934                  * little point trying all over again as kswapd may
1935                  * infinite loop.
1936                  *
1937                  * Instead, recheck all watermarks at order-0 as they
1938                  * are the most important. If watermarks are ok, kswapd will go
1939                  * back to sleep. High-order users can still perform direct
1940                  * reclaim if they wish.
1941                  */
1942                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
1943                         order = sc.order = 0;
1944
1945                 goto loop_again;
1946         }
1947
1948         return sc.nr_reclaimed;
1949 }
1950
1951 /*
1952  * The background pageout daemon, started as a kernel thread
1953  * from the init process.
1954  *
1955  * This basically trickles out pages so that we have _some_
1956  * free memory available even if there is no other activity
1957  * that frees anything up. This is needed for things like routing
1958  * etc, where we otherwise might have all activity going on in
1959  * asynchronous contexts that cannot page things out.
1960  *
1961  * If there are applications that are active memory-allocators
1962  * (most normal use), this basically shouldn't matter.
1963  */
1964 static int kswapd(void *p)
1965 {
1966         unsigned long order;
1967         pg_data_t *pgdat = (pg_data_t*)p;
1968         struct task_struct *tsk = current;
1969         DEFINE_WAIT(wait);
1970         struct reclaim_state reclaim_state = {
1971                 .reclaimed_slab = 0,
1972         };
1973         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1974
1975         lockdep_set_current_reclaim_state(GFP_KERNEL);
1976
1977         if (!cpumask_empty(cpumask))
1978                 set_cpus_allowed_ptr(tsk, cpumask);
1979         current->reclaim_state = &reclaim_state;
1980
1981         /*
1982          * Tell the memory management that we're a "memory allocator",
1983          * and that if we need more memory we should get access to it
1984          * regardless (see "__alloc_pages()"). "kswapd" should
1985          * never get caught in the normal page freeing logic.
1986          *
1987          * (Kswapd normally doesn't need memory anyway, but sometimes
1988          * you need a small amount of memory in order to be able to
1989          * page out something else, and this flag essentially protects
1990          * us from recursively trying to free more memory as we're
1991          * trying to free the first piece of memory in the first place).
1992          */
1993         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
1994         set_freezable();
1995
1996         order = 0;
1997         for ( ; ; ) {
1998                 unsigned long new_order;
1999
2000                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2001                 new_order = pgdat->kswapd_max_order;
2002                 pgdat->kswapd_max_order = 0;
2003                 if (order < new_order) {
2004                         /*
2005                          * Don't sleep if someone wants a larger 'order'
2006                          * allocation
2007                          */
2008                         order = new_order;
2009                 } else {
2010                         if (!freezing(current))
2011                                 schedule();
2012
2013                         order = pgdat->kswapd_max_order;
2014                 }
2015                 finish_wait(&pgdat->kswapd_wait, &wait);
2016
2017                 if (!try_to_freeze()) {
2018                         /* We can speed up thawing tasks if we don't call
2019                          * balance_pgdat after returning from the refrigerator
2020                          */
2021                         balance_pgdat(pgdat, order);
2022                 }
2023         }
2024         return 0;
2025 }
2026
2027 /*
2028  * A zone is low on free memory, so wake its kswapd task to service it.
2029  */
2030 void wakeup_kswapd(struct zone *zone, int order)
2031 {
2032         pg_data_t *pgdat;
2033
2034         if (!populated_zone(zone))
2035                 return;
2036
2037         pgdat = zone->zone_pgdat;
2038         if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0))
2039                 return;
2040         if (pgdat->kswapd_max_order < order)
2041                 pgdat->kswapd_max_order = order;
2042         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2043                 return;
2044         if (!waitqueue_active(&pgdat->kswapd_wait))
2045                 return;
2046         wake_up_interruptible(&pgdat->kswapd_wait);
2047 }
2048
2049 unsigned long global_lru_pages(void)
2050 {
2051         return global_page_state(NR_ACTIVE_ANON)
2052                 + global_page_state(NR_ACTIVE_FILE)
2053                 + global_page_state(NR_INACTIVE_ANON)
2054                 + global_page_state(NR_INACTIVE_FILE);
2055 }
2056
2057 #ifdef CONFIG_PM
2058 /*
2059  * Helper function for shrink_all_memory().  Tries to reclaim 'nr_pages' pages
2060  * from LRU lists system-wide, for given pass and priority.
2061  *
2062  * For pass > 3 we also try to shrink the LRU lists that contain a few pages
2063  */
2064 static void shrink_all_zones(unsigned long nr_pages, int prio,
2065                                       int pass, struct scan_control *sc)
2066 {
2067         struct zone *zone;
2068         unsigned long nr_reclaimed = 0;
2069
2070         for_each_populated_zone(zone) {
2071                 enum lru_list l;
2072
2073                 if (zone_is_all_unreclaimable(zone) && prio != DEF_PRIORITY)
2074                         continue;
2075
2076                 for_each_evictable_lru(l) {
2077                         enum zone_stat_item ls = NR_LRU_BASE + l;
2078                         unsigned long lru_pages = zone_page_state(zone, ls);
2079
2080                         /* For pass = 0, we don't shrink the active list */
2081                         if (pass == 0 && (l == LRU_ACTIVE_ANON ||
2082                                                 l == LRU_ACTIVE_FILE))
2083                                 continue;
2084
2085                         zone->lru[l].nr_scan += (lru_pages >> prio) + 1;
2086                         if (zone->lru[l].nr_scan >= nr_pages || pass > 3) {
2087                                 unsigned long nr_to_scan;
2088
2089                                 zone->lru[l].nr_scan = 0;
2090                                 nr_to_scan = min(nr_pages, lru_pages);
2091                                 nr_reclaimed += shrink_list(l, nr_to_scan, zone,
2092                                                                 sc, prio);
2093                                 if (nr_reclaimed >= nr_pages) {
2094                                         sc->nr_reclaimed += nr_reclaimed;
2095                                         return;
2096                                 }
2097                         }
2098                 }
2099         }
2100         sc->nr_reclaimed += nr_reclaimed;
2101 }
2102
2103 /*
2104  * Try to free `nr_pages' of memory, system-wide, and return the number of
2105  * freed pages.
2106  *
2107  * Rather than trying to age LRUs the aim is to preserve the overall
2108  * LRU order by reclaiming preferentially
2109  * inactive > active > active referenced > active mapped
2110  */
2111 unsigned long shrink_all_memory(unsigned long nr_pages)
2112 {
2113         unsigned long lru_pages, nr_slab;
2114         int pass;
2115         struct reclaim_state reclaim_state;
2116         struct scan_control sc = {
2117                 .gfp_mask = GFP_KERNEL,
2118                 .may_unmap = 0,
2119                 .may_writepage = 1,
2120                 .isolate_pages = isolate_pages_global,
2121                 .nr_reclaimed = 0,
2122         };
2123
2124         current->reclaim_state = &reclaim_state;
2125
2126         lru_pages = global_lru_pages();
2127         nr_slab = global_page_state(NR_SLAB_RECLAIMABLE);
2128         /* If slab caches are huge, it's better to hit them first */
2129         while (nr_slab >= lru_pages) {
2130                 reclaim_state.reclaimed_slab = 0;
2131                 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
2132                 if (!reclaim_state.reclaimed_slab)
2133                         break;
2134
2135                 sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2136                 if (sc.nr_reclaimed >= nr_pages)
2137                         goto out;
2138
2139                 nr_slab -= reclaim_state.reclaimed_slab;
2140         }
2141
2142         /*
2143          * We try to shrink LRUs in 5 passes:
2144          * 0 = Reclaim from inactive_list only
2145          * 1 = Reclaim from active list but don't reclaim mapped
2146          * 2 = 2nd pass of type 1
2147          * 3 = Reclaim mapped (normal reclaim)
2148          * 4 = 2nd pass of type 3
2149          */
2150         for (pass = 0; pass < 5; pass++) {
2151                 int prio;
2152
2153                 /* Force reclaiming mapped pages in the passes #3 and #4 */
2154                 if (pass > 2)
2155                         sc.may_unmap = 1;
2156
2157                 for (prio = DEF_PRIORITY; prio >= 0; prio--) {
2158                         unsigned long nr_to_scan = nr_pages - sc.nr_reclaimed;
2159
2160                         sc.nr_scanned = 0;
2161                         sc.swap_cluster_max = nr_to_scan;
2162                         shrink_all_zones(nr_to_scan, prio, pass, &sc);
2163                         if (sc.nr_reclaimed >= nr_pages)
2164                                 goto out;
2165
2166                         reclaim_state.reclaimed_slab = 0;
2167                         shrink_slab(sc.nr_scanned, sc.gfp_mask,
2168                                         global_lru_pages());
2169                         sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2170                         if (sc.nr_reclaimed >= nr_pages)
2171                                 goto out;
2172
2173                         if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
2174                                 congestion_wait(WRITE, HZ / 10);
2175                 }
2176         }
2177
2178         /*
2179          * If sc.nr_reclaimed = 0, we could not shrink LRUs, but there may be
2180          * something in slab caches
2181          */
2182         if (!sc.nr_reclaimed) {
2183                 do {
2184                         reclaim_state.reclaimed_slab = 0;
2185                         shrink_slab(nr_pages, sc.gfp_mask, global_lru_pages());
2186                         sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2187                 } while (sc.nr_reclaimed < nr_pages &&
2188                                 reclaim_state.reclaimed_slab > 0);
2189         }
2190
2191
2192 out:
2193         current->reclaim_state = NULL;
2194
2195         return sc.nr_reclaimed;
2196 }
2197 #endif
2198
2199 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2200    not required for correctness.  So if the last cpu in a node goes
2201    away, we get changed to run anywhere: as the first one comes back,
2202    restore their cpu bindings. */
2203 static int __devinit cpu_callback(struct notifier_block *nfb,
2204                                   unsigned long action, void *hcpu)
2205 {
2206         int nid;
2207
2208         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2209                 for_each_node_state(nid, N_HIGH_MEMORY) {
2210                         pg_data_t *pgdat = NODE_DATA(nid);
2211                         const struct cpumask *mask;
2212
2213                         mask = cpumask_of_node(pgdat->node_id);
2214
2215                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2216                                 /* One of our CPUs online: restore mask */
2217                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2218                 }
2219         }
2220         return NOTIFY_OK;
2221 }
2222
2223 /*
2224  * This kswapd start function will be called by init and node-hot-add.
2225  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2226  */
2227 int kswapd_run(int nid)
2228 {
2229         pg_data_t *pgdat = NODE_DATA(nid);
2230         int ret = 0;
2231
2232         if (pgdat->kswapd)
2233                 return 0;
2234
2235         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2236         if (IS_ERR(pgdat->kswapd)) {
2237                 /* failure at boot is fatal */
2238                 BUG_ON(system_state == SYSTEM_BOOTING);
2239                 printk("Failed to start kswapd on node %d\n",nid);
2240                 ret = -1;
2241         }
2242         return ret;
2243 }
2244
2245 static int __init kswapd_init(void)
2246 {
2247         int nid;
2248
2249         swap_setup();
2250         for_each_node_state(nid, N_HIGH_MEMORY)
2251                 kswapd_run(nid);
2252         hotcpu_notifier(cpu_callback, 0);
2253         return 0;
2254 }
2255
2256 module_init(kswapd_init)
2257
2258 #ifdef CONFIG_NUMA
2259 /*
2260  * Zone reclaim mode
2261  *
2262  * If non-zero call zone_reclaim when the number of free pages falls below
2263  * the watermarks.
2264  */
2265 int zone_reclaim_mode __read_mostly;
2266
2267 #define RECLAIM_OFF 0
2268 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
2269 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
2270 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
2271
2272 /*
2273  * Priority for ZONE_RECLAIM. This determines the fraction of pages
2274  * of a node considered for each zone_reclaim. 4 scans 1/16th of
2275  * a zone.
2276  */
2277 #define ZONE_RECLAIM_PRIORITY 4
2278
2279 /*
2280  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2281  * occur.
2282  */
2283 int sysctl_min_unmapped_ratio = 1;
2284
2285 /*
2286  * If the number of slab pages in a zone grows beyond this percentage then
2287  * slab reclaim needs to occur.
2288  */
2289 int sysctl_min_slab_ratio = 5;
2290
2291 /*
2292  * Try to free up some pages from this zone through reclaim.
2293  */
2294 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2295 {
2296         /* Minimum pages needed in order to stay on node */
2297         const unsigned long nr_pages = 1 << order;
2298         struct task_struct *p = current;
2299         struct reclaim_state reclaim_state;
2300         int priority;
2301         struct scan_control sc = {
2302                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2303                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2304                 .may_swap = 1,
2305                 .swap_cluster_max = max_t(unsigned long, nr_pages,
2306                                         SWAP_CLUSTER_MAX),
2307                 .gfp_mask = gfp_mask,
2308                 .swappiness = vm_swappiness,
2309                 .order = order,
2310                 .isolate_pages = isolate_pages_global,
2311         };
2312         unsigned long slab_reclaimable;
2313
2314         disable_swap_token();
2315         cond_resched();
2316         /*
2317          * We need to be able to allocate from the reserves for RECLAIM_SWAP
2318          * and we also need to be able to write out pages for RECLAIM_WRITE
2319          * and RECLAIM_SWAP.
2320          */
2321         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2322         reclaim_state.reclaimed_slab = 0;
2323         p->reclaim_state = &reclaim_state;
2324
2325         if (zone_page_state(zone, NR_FILE_PAGES) -
2326                 zone_page_state(zone, NR_FILE_MAPPED) >
2327                 zone->min_unmapped_pages) {
2328                 /*
2329                  * Free memory by calling shrink zone with increasing
2330                  * priorities until we have enough memory freed.
2331                  */
2332                 priority = ZONE_RECLAIM_PRIORITY;
2333                 do {
2334                         note_zone_scanning_priority(zone, priority);
2335                         shrink_zone(priority, zone, &sc);
2336                         priority--;
2337                 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
2338         }
2339
2340         slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2341         if (slab_reclaimable > zone->min_slab_pages) {
2342                 /*
2343                  * shrink_slab() does not currently allow us to determine how
2344                  * many pages were freed in this zone. So we take the current
2345                  * number of slab pages and shake the slab until it is reduced
2346                  * by the same nr_pages that we used for reclaiming unmapped
2347                  * pages.
2348                  *
2349                  * Note that shrink_slab will free memory on all zones and may
2350                  * take a long time.
2351                  */
2352                 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
2353                         zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
2354                                 slab_reclaimable - nr_pages)
2355                         ;
2356
2357                 /*
2358                  * Update nr_reclaimed by the number of slab pages we
2359                  * reclaimed from this zone.
2360                  */
2361                 sc.nr_reclaimed += slab_reclaimable -
2362                         zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2363         }
2364
2365         p->reclaim_state = NULL;
2366         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
2367         return sc.nr_reclaimed >= nr_pages;
2368 }
2369
2370 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2371 {
2372         int node_id;
2373         int ret;
2374
2375         /*
2376          * Zone reclaim reclaims unmapped file backed pages and
2377          * slab pages if we are over the defined limits.
2378          *
2379          * A small portion of unmapped file backed pages is needed for
2380          * file I/O otherwise pages read by file I/O will be immediately
2381          * thrown out if the zone is overallocated. So we do not reclaim
2382          * if less than a specified percentage of the zone is used by
2383          * unmapped file backed pages.
2384          */
2385         if (zone_page_state(zone, NR_FILE_PAGES) -
2386             zone_page_state(zone, NR_FILE_MAPPED) <= zone->min_unmapped_pages
2387             && zone_page_state(zone, NR_SLAB_RECLAIMABLE)
2388                         <= zone->min_slab_pages)
2389                 return 0;
2390
2391         if (zone_is_all_unreclaimable(zone))
2392                 return 0;
2393
2394         /*
2395          * Do not scan if the allocation should not be delayed.
2396          */
2397         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
2398                         return 0;
2399
2400         /*
2401          * Only run zone reclaim on the local zone or on zones that do not
2402          * have associated processors. This will favor the local processor
2403          * over remote processors and spread off node memory allocations
2404          * as wide as possible.
2405          */
2406         node_id = zone_to_nid(zone);
2407         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
2408                 return 0;
2409
2410         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2411                 return 0;
2412         ret = __zone_reclaim(zone, gfp_mask, order);
2413         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2414
2415         return ret;
2416 }
2417 #endif
2418
2419 #ifdef CONFIG_UNEVICTABLE_LRU
2420 /*
2421  * page_evictable - test whether a page is evictable
2422  * @page: the page to test
2423  * @vma: the VMA in which the page is or will be mapped, may be NULL
2424  *
2425  * Test whether page is evictable--i.e., should be placed on active/inactive
2426  * lists vs unevictable list.  The vma argument is !NULL when called from the
2427  * fault path to determine how to instantate a new page.
2428  *
2429  * Reasons page might not be evictable:
2430  * (1) page's mapping marked unevictable
2431  * (2) page is part of an mlocked VMA
2432  *
2433  */
2434 int page_evictable(struct page *page, struct vm_area_struct *vma)
2435 {
2436
2437         if (mapping_unevictable(page_mapping(page)))
2438                 return 0;
2439
2440         if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2441                 return 0;
2442
2443         return 1;
2444 }
2445
2446 /**
2447  * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2448  * @page: page to check evictability and move to appropriate lru list
2449  * @zone: zone page is in
2450  *
2451  * Checks a page for evictability and moves the page to the appropriate
2452  * zone lru list.
2453  *
2454  * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2455  * have PageUnevictable set.
2456  */
2457 static void check_move_unevictable_page(struct page *page, struct zone *zone)
2458 {
2459         VM_BUG_ON(PageActive(page));
2460
2461 retry:
2462         ClearPageUnevictable(page);
2463         if (page_evictable(page, NULL)) {
2464                 enum lru_list l = LRU_INACTIVE_ANON + page_is_file_cache(page);
2465
2466                 __dec_zone_state(zone, NR_UNEVICTABLE);
2467                 list_move(&page->lru, &zone->lru[l].list);
2468                 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
2469                 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
2470                 __count_vm_event(UNEVICTABLE_PGRESCUED);
2471         } else {
2472                 /*
2473                  * rotate unevictable list
2474                  */
2475                 SetPageUnevictable(page);
2476                 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
2477                 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
2478                 if (page_evictable(page, NULL))
2479                         goto retry;
2480         }
2481 }
2482
2483 /**
2484  * scan_mapping_unevictable_pages - scan an address space for evictable pages
2485  * @mapping: struct address_space to scan for evictable pages
2486  *
2487  * Scan all pages in mapping.  Check unevictable pages for
2488  * evictability and move them to the appropriate zone lru list.
2489  */
2490 void scan_mapping_unevictable_pages(struct address_space *mapping)
2491 {
2492         pgoff_t next = 0;
2493         pgoff_t end   = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
2494                          PAGE_CACHE_SHIFT;
2495         struct zone *zone;
2496         struct pagevec pvec;
2497
2498         if (mapping->nrpages == 0)
2499                 return;
2500
2501         pagevec_init(&pvec, 0);
2502         while (next < end &&
2503                 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
2504                 int i;
2505                 int pg_scanned = 0;
2506
2507                 zone = NULL;
2508
2509                 for (i = 0; i < pagevec_count(&pvec); i++) {
2510                         struct page *page = pvec.pages[i];
2511                         pgoff_t page_index = page->index;
2512                         struct zone *pagezone = page_zone(page);
2513
2514                         pg_scanned++;
2515                         if (page_index > next)
2516                                 next = page_index;
2517                         next++;
2518
2519                         if (pagezone != zone) {
2520                                 if (zone)
2521                                         spin_unlock_irq(&zone->lru_lock);
2522                                 zone = pagezone;
2523                                 spin_lock_irq(&zone->lru_lock);
2524                         }
2525
2526                         if (PageLRU(page) && PageUnevictable(page))
2527                                 check_move_unevictable_page(page, zone);
2528                 }
2529                 if (zone)
2530                         spin_unlock_irq(&zone->lru_lock);
2531                 pagevec_release(&pvec);
2532
2533                 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
2534         }
2535
2536 }
2537
2538 /**
2539  * scan_zone_unevictable_pages - check unevictable list for evictable pages
2540  * @zone - zone of which to scan the unevictable list
2541  *
2542  * Scan @zone's unevictable LRU lists to check for pages that have become
2543  * evictable.  Move those that have to @zone's inactive list where they
2544  * become candidates for reclaim, unless shrink_inactive_zone() decides
2545  * to reactivate them.  Pages that are still unevictable are rotated
2546  * back onto @zone's unevictable list.
2547  */
2548 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
2549 static void scan_zone_unevictable_pages(struct zone *zone)
2550 {
2551         struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
2552         unsigned long scan;
2553         unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
2554
2555         while (nr_to_scan > 0) {
2556                 unsigned long batch_size = min(nr_to_scan,
2557                                                 SCAN_UNEVICTABLE_BATCH_SIZE);
2558
2559                 spin_lock_irq(&zone->lru_lock);
2560                 for (scan = 0;  scan < batch_size; scan++) {
2561                         struct page *page = lru_to_page(l_unevictable);
2562
2563                         if (!trylock_page(page))
2564                                 continue;
2565
2566                         prefetchw_prev_lru_page(page, l_unevictable, flags);
2567
2568                         if (likely(PageLRU(page) && PageUnevictable(page)))
2569                                 check_move_unevictable_page(page, zone);
2570
2571                         unlock_page(page);
2572                 }
2573                 spin_unlock_irq(&zone->lru_lock);
2574
2575                 nr_to_scan -= batch_size;
2576         }
2577 }
2578
2579
2580 /**
2581  * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
2582  *
2583  * A really big hammer:  scan all zones' unevictable LRU lists to check for
2584  * pages that have become evictable.  Move those back to the zones'
2585  * inactive list where they become candidates for reclaim.
2586  * This occurs when, e.g., we have unswappable pages on the unevictable lists,
2587  * and we add swap to the system.  As such, it runs in the context of a task
2588  * that has possibly/probably made some previously unevictable pages
2589  * evictable.
2590  */
2591 static void scan_all_zones_unevictable_pages(void)
2592 {
2593         struct zone *zone;
2594
2595         for_each_zone(zone) {
2596                 scan_zone_unevictable_pages(zone);
2597         }
2598 }
2599
2600 /*
2601  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
2602  * all nodes' unevictable lists for evictable pages
2603  */
2604 unsigned long scan_unevictable_pages;
2605
2606 int scan_unevictable_handler(struct ctl_table *table, int write,
2607                            struct file *file, void __user *buffer,
2608                            size_t *length, loff_t *ppos)
2609 {
2610         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
2611
2612         if (write && *(unsigned long *)table->data)
2613                 scan_all_zones_unevictable_pages();
2614
2615         scan_unevictable_pages = 0;
2616         return 0;
2617 }
2618
2619 /*
2620  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
2621  * a specified node's per zone unevictable lists for evictable pages.
2622  */
2623
2624 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
2625                                           struct sysdev_attribute *attr,
2626                                           char *buf)
2627 {
2628         return sprintf(buf, "0\n");     /* always zero; should fit... */
2629 }
2630
2631 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
2632                                            struct sysdev_attribute *attr,
2633                                         const char *buf, size_t count)
2634 {
2635         struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
2636         struct zone *zone;
2637         unsigned long res;
2638         unsigned long req = strict_strtoul(buf, 10, &res);
2639
2640         if (!req)
2641                 return 1;       /* zero is no-op */
2642
2643         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2644                 if (!populated_zone(zone))
2645                         continue;
2646                 scan_zone_unevictable_pages(zone);
2647         }
2648         return 1;
2649 }
2650
2651
2652 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
2653                         read_scan_unevictable_node,
2654                         write_scan_unevictable_node);
2655
2656 int scan_unevictable_register_node(struct node *node)
2657 {
2658         return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
2659 }
2660
2661 void scan_unevictable_unregister_node(struct node *node)
2662 {
2663         sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
2664 }
2665
2666 #endif