mm: define PAGE_MAPPING_FLAGS
[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_lru_pages(struct zone *zone,
152                                 struct scan_control *sc, 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         /*
290          * A freeable page cache page is referenced only by the caller
291          * that isolated the page, the page cache radix tree and
292          * optional buffer heads at page->private.
293          */
294         return page_count(page) - page_has_private(page) == 2;
295 }
296
297 static int may_write_to_queue(struct backing_dev_info *bdi)
298 {
299         if (current->flags & PF_SWAPWRITE)
300                 return 1;
301         if (!bdi_write_congested(bdi))
302                 return 1;
303         if (bdi == current->backing_dev_info)
304                 return 1;
305         return 0;
306 }
307
308 /*
309  * We detected a synchronous write error writing a page out.  Probably
310  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
311  * fsync(), msync() or close().
312  *
313  * The tricky part is that after writepage we cannot touch the mapping: nothing
314  * prevents it from being freed up.  But we have a ref on the page and once
315  * that page is locked, the mapping is pinned.
316  *
317  * We're allowed to run sleeping lock_page() here because we know the caller has
318  * __GFP_FS.
319  */
320 static void handle_write_error(struct address_space *mapping,
321                                 struct page *page, int error)
322 {
323         lock_page(page);
324         if (page_mapping(page) == mapping)
325                 mapping_set_error(mapping, error);
326         unlock_page(page);
327 }
328
329 /* Request for sync pageout. */
330 enum pageout_io {
331         PAGEOUT_IO_ASYNC,
332         PAGEOUT_IO_SYNC,
333 };
334
335 /* possible outcome of pageout() */
336 typedef enum {
337         /* failed to write page out, page is locked */
338         PAGE_KEEP,
339         /* move page to the active list, page is locked */
340         PAGE_ACTIVATE,
341         /* page has been sent to the disk successfully, page is unlocked */
342         PAGE_SUCCESS,
343         /* page is clean and locked */
344         PAGE_CLEAN,
345 } pageout_t;
346
347 /*
348  * pageout is called by shrink_page_list() for each dirty page.
349  * Calls ->writepage().
350  */
351 static pageout_t pageout(struct page *page, struct address_space *mapping,
352                                                 enum pageout_io sync_writeback)
353 {
354         /*
355          * If the page is dirty, only perform writeback if that write
356          * will be non-blocking.  To prevent this allocation from being
357          * stalled by pagecache activity.  But note that there may be
358          * stalls if we need to run get_block().  We could test
359          * PagePrivate for that.
360          *
361          * If this process is currently in __generic_file_aio_write() against
362          * this page's queue, we can perform writeback even if that
363          * will block.
364          *
365          * If the page is swapcache, write it back even if that would
366          * block, for some throttling. This happens by accident, because
367          * swap_backing_dev_info is bust: it doesn't reflect the
368          * congestion state of the swapdevs.  Easy to fix, if needed.
369          */
370         if (!is_page_cache_freeable(page))
371                 return PAGE_KEEP;
372         if (!mapping) {
373                 /*
374                  * Some data journaling orphaned pages can have
375                  * page->mapping == NULL while being dirty with clean buffers.
376                  */
377                 if (page_has_private(page)) {
378                         if (try_to_free_buffers(page)) {
379                                 ClearPageDirty(page);
380                                 printk("%s: orphaned page\n", __func__);
381                                 return PAGE_CLEAN;
382                         }
383                 }
384                 return PAGE_KEEP;
385         }
386         if (mapping->a_ops->writepage == NULL)
387                 return PAGE_ACTIVATE;
388         if (!may_write_to_queue(mapping->backing_dev_info))
389                 return PAGE_KEEP;
390
391         if (clear_page_dirty_for_io(page)) {
392                 int res;
393                 struct writeback_control wbc = {
394                         .sync_mode = WB_SYNC_NONE,
395                         .nr_to_write = SWAP_CLUSTER_MAX,
396                         .range_start = 0,
397                         .range_end = LLONG_MAX,
398                         .nonblocking = 1,
399                         .for_reclaim = 1,
400                 };
401
402                 SetPageReclaim(page);
403                 res = mapping->a_ops->writepage(page, &wbc);
404                 if (res < 0)
405                         handle_write_error(mapping, page, res);
406                 if (res == AOP_WRITEPAGE_ACTIVATE) {
407                         ClearPageReclaim(page);
408                         return PAGE_ACTIVATE;
409                 }
410
411                 /*
412                  * Wait on writeback if requested to. This happens when
413                  * direct reclaiming a large contiguous area and the
414                  * first attempt to free a range of pages fails.
415                  */
416                 if (PageWriteback(page) && sync_writeback == PAGEOUT_IO_SYNC)
417                         wait_on_page_writeback(page);
418
419                 if (!PageWriteback(page)) {
420                         /* synchronous write or broken a_ops? */
421                         ClearPageReclaim(page);
422                 }
423                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
424                 return PAGE_SUCCESS;
425         }
426
427         return PAGE_CLEAN;
428 }
429
430 /*
431  * Same as remove_mapping, but if the page is removed from the mapping, it
432  * gets returned with a refcount of 0.
433  */
434 static int __remove_mapping(struct address_space *mapping, struct page *page)
435 {
436         BUG_ON(!PageLocked(page));
437         BUG_ON(mapping != page_mapping(page));
438
439         spin_lock_irq(&mapping->tree_lock);
440         /*
441          * The non racy check for a busy page.
442          *
443          * Must be careful with the order of the tests. When someone has
444          * a ref to the page, it may be possible that they dirty it then
445          * drop the reference. So if PageDirty is tested before page_count
446          * here, then the following race may occur:
447          *
448          * get_user_pages(&page);
449          * [user mapping goes away]
450          * write_to(page);
451          *                              !PageDirty(page)    [good]
452          * SetPageDirty(page);
453          * put_page(page);
454          *                              !page_count(page)   [good, discard it]
455          *
456          * [oops, our write_to data is lost]
457          *
458          * Reversing the order of the tests ensures such a situation cannot
459          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
460          * load is not satisfied before that of page->_count.
461          *
462          * Note that if SetPageDirty is always performed via set_page_dirty,
463          * and thus under tree_lock, then this ordering is not required.
464          */
465         if (!page_freeze_refs(page, 2))
466                 goto cannot_free;
467         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
468         if (unlikely(PageDirty(page))) {
469                 page_unfreeze_refs(page, 2);
470                 goto cannot_free;
471         }
472
473         if (PageSwapCache(page)) {
474                 swp_entry_t swap = { .val = page_private(page) };
475                 __delete_from_swap_cache(page);
476                 spin_unlock_irq(&mapping->tree_lock);
477                 swapcache_free(swap, page);
478         } else {
479                 __remove_from_page_cache(page);
480                 spin_unlock_irq(&mapping->tree_lock);
481                 mem_cgroup_uncharge_cache_page(page);
482         }
483
484         return 1;
485
486 cannot_free:
487         spin_unlock_irq(&mapping->tree_lock);
488         return 0;
489 }
490
491 /*
492  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
493  * someone else has a ref on the page, abort and return 0.  If it was
494  * successfully detached, return 1.  Assumes the caller has a single ref on
495  * this page.
496  */
497 int remove_mapping(struct address_space *mapping, struct page *page)
498 {
499         if (__remove_mapping(mapping, page)) {
500                 /*
501                  * Unfreezing the refcount with 1 rather than 2 effectively
502                  * drops the pagecache ref for us without requiring another
503                  * atomic operation.
504                  */
505                 page_unfreeze_refs(page, 1);
506                 return 1;
507         }
508         return 0;
509 }
510
511 /**
512  * putback_lru_page - put previously isolated page onto appropriate LRU list
513  * @page: page to be put back to appropriate lru list
514  *
515  * Add previously isolated @page to appropriate LRU list.
516  * Page may still be unevictable for other reasons.
517  *
518  * lru_lock must not be held, interrupts must be enabled.
519  */
520 void putback_lru_page(struct page *page)
521 {
522         int lru;
523         int active = !!TestClearPageActive(page);
524         int was_unevictable = PageUnevictable(page);
525
526         VM_BUG_ON(PageLRU(page));
527
528 redo:
529         ClearPageUnevictable(page);
530
531         if (page_evictable(page, NULL)) {
532                 /*
533                  * For evictable pages, we can use the cache.
534                  * In event of a race, worst case is we end up with an
535                  * unevictable page on [in]active list.
536                  * We know how to handle that.
537                  */
538                 lru = active + page_lru_base_type(page);
539                 lru_cache_add_lru(page, lru);
540         } else {
541                 /*
542                  * Put unevictable pages directly on zone's unevictable
543                  * list.
544                  */
545                 lru = LRU_UNEVICTABLE;
546                 add_page_to_unevictable_list(page);
547                 /*
548                  * When racing with an mlock clearing (page is
549                  * unlocked), make sure that if the other thread does
550                  * not observe our setting of PG_lru and fails
551                  * isolation, we see PG_mlocked cleared below and move
552                  * the page back to the evictable list.
553                  *
554                  * The other side is TestClearPageMlocked().
555                  */
556                 smp_mb();
557         }
558
559         /*
560          * page's status can change while we move it among lru. If an evictable
561          * page is on unevictable list, it never be freed. To avoid that,
562          * check after we added it to the list, again.
563          */
564         if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
565                 if (!isolate_lru_page(page)) {
566                         put_page(page);
567                         goto redo;
568                 }
569                 /* This means someone else dropped this page from LRU
570                  * So, it will be freed or putback to LRU again. There is
571                  * nothing to do here.
572                  */
573         }
574
575         if (was_unevictable && lru != LRU_UNEVICTABLE)
576                 count_vm_event(UNEVICTABLE_PGRESCUED);
577         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
578                 count_vm_event(UNEVICTABLE_PGCULLED);
579
580         put_page(page);         /* drop ref from isolate */
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         unsigned long vm_flags;
595
596         cond_resched();
597
598         pagevec_init(&freed_pvec, 1);
599         while (!list_empty(page_list)) {
600                 struct address_space *mapping;
601                 struct page *page;
602                 int may_enter_fs;
603                 int referenced;
604
605                 cond_resched();
606
607                 page = lru_to_page(page_list);
608                 list_del(&page->lru);
609
610                 if (!trylock_page(page))
611                         goto keep;
612
613                 VM_BUG_ON(PageActive(page));
614
615                 sc->nr_scanned++;
616
617                 if (unlikely(!page_evictable(page, NULL)))
618                         goto cull_mlocked;
619
620                 if (!sc->may_unmap && page_mapped(page))
621                         goto keep_locked;
622
623                 /* Double the slab pressure for mapped and swapcache pages */
624                 if (page_mapped(page) || PageSwapCache(page))
625                         sc->nr_scanned++;
626
627                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
628                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
629
630                 if (PageWriteback(page)) {
631                         /*
632                          * Synchronous reclaim is performed in two passes,
633                          * first an asynchronous pass over the list to
634                          * start parallel writeback, and a second synchronous
635                          * pass to wait for the IO to complete.  Wait here
636                          * for any page for which writeback has already
637                          * started.
638                          */
639                         if (sync_writeback == PAGEOUT_IO_SYNC && may_enter_fs)
640                                 wait_on_page_writeback(page);
641                         else
642                                 goto keep_locked;
643                 }
644
645                 referenced = page_referenced(page, 1,
646                                                 sc->mem_cgroup, &vm_flags);
647                 /*
648                  * In active use or really unfreeable?  Activate it.
649                  * If page which have PG_mlocked lost isoltation race,
650                  * try_to_unmap moves it to unevictable list
651                  */
652                 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER &&
653                                         referenced && page_mapping_inuse(page)
654                                         && !(vm_flags & VM_LOCKED))
655                         goto activate_locked;
656
657                 /*
658                  * Anonymous process memory has backing store?
659                  * Try to allocate it some swap space here.
660                  */
661                 if (PageAnon(page) && !PageSwapCache(page)) {
662                         if (!(sc->gfp_mask & __GFP_IO))
663                                 goto keep_locked;
664                         if (!add_to_swap(page))
665                                 goto activate_locked;
666                         may_enter_fs = 1;
667                 }
668
669                 mapping = page_mapping(page);
670
671                 /*
672                  * The page is mapped into the page tables of one or more
673                  * processes. Try to unmap it here.
674                  */
675                 if (page_mapped(page) && mapping) {
676                         switch (try_to_unmap(page, TTU_UNMAP)) {
677                         case SWAP_FAIL:
678                                 goto activate_locked;
679                         case SWAP_AGAIN:
680                                 goto keep_locked;
681                         case SWAP_MLOCK:
682                                 goto cull_mlocked;
683                         case SWAP_SUCCESS:
684                                 ; /* try to free the page below */
685                         }
686                 }
687
688                 if (PageDirty(page)) {
689                         if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && referenced)
690                                 goto keep_locked;
691                         if (!may_enter_fs)
692                                 goto keep_locked;
693                         if (!sc->may_writepage)
694                                 goto keep_locked;
695
696                         /* Page is dirty, try to write it out here */
697                         switch (pageout(page, mapping, sync_writeback)) {
698                         case PAGE_KEEP:
699                                 goto keep_locked;
700                         case PAGE_ACTIVATE:
701                                 goto activate_locked;
702                         case PAGE_SUCCESS:
703                                 if (PageWriteback(page) || PageDirty(page))
704                                         goto keep;
705                                 /*
706                                  * A synchronous write - probably a ramdisk.  Go
707                                  * ahead and try to reclaim the page.
708                                  */
709                                 if (!trylock_page(page))
710                                         goto keep;
711                                 if (PageDirty(page) || PageWriteback(page))
712                                         goto keep_locked;
713                                 mapping = page_mapping(page);
714                         case PAGE_CLEAN:
715                                 ; /* try to free the page below */
716                         }
717                 }
718
719                 /*
720                  * If the page has buffers, try to free the buffer mappings
721                  * associated with this page. If we succeed we try to free
722                  * the page as well.
723                  *
724                  * We do this even if the page is PageDirty().
725                  * try_to_release_page() does not perform I/O, but it is
726                  * possible for a page to have PageDirty set, but it is actually
727                  * clean (all its buffers are clean).  This happens if the
728                  * buffers were written out directly, with submit_bh(). ext3
729                  * will do this, as well as the blockdev mapping.
730                  * try_to_release_page() will discover that cleanness and will
731                  * drop the buffers and mark the page clean - it can be freed.
732                  *
733                  * Rarely, pages can have buffers and no ->mapping.  These are
734                  * the pages which were not successfully invalidated in
735                  * truncate_complete_page().  We try to drop those buffers here
736                  * and if that worked, and the page is no longer mapped into
737                  * process address space (page_count == 1) it can be freed.
738                  * Otherwise, leave the page on the LRU so it is swappable.
739                  */
740                 if (page_has_private(page)) {
741                         if (!try_to_release_page(page, sc->gfp_mask))
742                                 goto activate_locked;
743                         if (!mapping && page_count(page) == 1) {
744                                 unlock_page(page);
745                                 if (put_page_testzero(page))
746                                         goto free_it;
747                                 else {
748                                         /*
749                                          * rare race with speculative reference.
750                                          * the speculative reference will free
751                                          * this page shortly, so we may
752                                          * increment nr_reclaimed here (and
753                                          * leave it off the LRU).
754                                          */
755                                         nr_reclaimed++;
756                                         continue;
757                                 }
758                         }
759                 }
760
761                 if (!mapping || !__remove_mapping(mapping, page))
762                         goto keep_locked;
763
764                 /*
765                  * At this point, we have no other references and there is
766                  * no way to pick any more up (removed from LRU, removed
767                  * from pagecache). Can use non-atomic bitops now (and
768                  * we obviously don't have to worry about waking up a process
769                  * waiting on the page lock, because there are no references.
770                  */
771                 __clear_page_locked(page);
772 free_it:
773                 nr_reclaimed++;
774                 if (!pagevec_add(&freed_pvec, page)) {
775                         __pagevec_free(&freed_pvec);
776                         pagevec_reinit(&freed_pvec);
777                 }
778                 continue;
779
780 cull_mlocked:
781                 if (PageSwapCache(page))
782                         try_to_free_swap(page);
783                 unlock_page(page);
784                 putback_lru_page(page);
785                 continue;
786
787 activate_locked:
788                 /* Not a candidate for swapping, so reclaim swap space. */
789                 if (PageSwapCache(page) && vm_swap_full())
790                         try_to_free_swap(page);
791                 VM_BUG_ON(PageActive(page));
792                 SetPageActive(page);
793                 pgactivate++;
794 keep_locked:
795                 unlock_page(page);
796 keep:
797                 list_add(&page->lru, &ret_pages);
798                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
799         }
800         list_splice(&ret_pages, page_list);
801         if (pagevec_count(&freed_pvec))
802                 __pagevec_free(&freed_pvec);
803         count_vm_events(PGACTIVATE, pgactivate);
804         return nr_reclaimed;
805 }
806
807 /* LRU Isolation modes. */
808 #define ISOLATE_INACTIVE 0      /* Isolate inactive pages. */
809 #define ISOLATE_ACTIVE 1        /* Isolate active pages. */
810 #define ISOLATE_BOTH 2          /* Isolate both active and inactive pages. */
811
812 /*
813  * Attempt to remove the specified page from its LRU.  Only take this page
814  * if it is of the appropriate PageActive status.  Pages which are being
815  * freed elsewhere are also ignored.
816  *
817  * page:        page to consider
818  * mode:        one of the LRU isolation modes defined above
819  *
820  * returns 0 on success, -ve errno on failure.
821  */
822 int __isolate_lru_page(struct page *page, int mode, int file)
823 {
824         int ret = -EINVAL;
825
826         /* Only take pages on the LRU. */
827         if (!PageLRU(page))
828                 return ret;
829
830         /*
831          * When checking the active state, we need to be sure we are
832          * dealing with comparible boolean values.  Take the logical not
833          * of each.
834          */
835         if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
836                 return ret;
837
838         if (mode != ISOLATE_BOTH && page_is_file_cache(page) != file)
839                 return ret;
840
841         /*
842          * When this function is being called for lumpy reclaim, we
843          * initially look into all LRU pages, active, inactive and
844          * unevictable; only give shrink_page_list evictable pages.
845          */
846         if (PageUnevictable(page))
847                 return ret;
848
849         ret = -EBUSY;
850
851         if (likely(get_page_unless_zero(page))) {
852                 /*
853                  * Be careful not to clear PageLRU until after we're
854                  * sure the page is not being freed elsewhere -- the
855                  * page release code relies on it.
856                  */
857                 ClearPageLRU(page);
858                 ret = 0;
859         }
860
861         return ret;
862 }
863
864 /*
865  * zone->lru_lock is heavily contended.  Some of the functions that
866  * shrink the lists perform better by taking out a batch of pages
867  * and working on them outside the LRU lock.
868  *
869  * For pagecache intensive workloads, this function is the hottest
870  * spot in the kernel (apart from copy_*_user functions).
871  *
872  * Appropriate locks must be held before calling this function.
873  *
874  * @nr_to_scan: The number of pages to look through on the list.
875  * @src:        The LRU list to pull pages off.
876  * @dst:        The temp list to put pages on to.
877  * @scanned:    The number of pages that were scanned.
878  * @order:      The caller's attempted allocation order
879  * @mode:       One of the LRU isolation modes
880  * @file:       True [1] if isolating file [!anon] pages
881  *
882  * returns how many pages were moved onto *@dst.
883  */
884 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
885                 struct list_head *src, struct list_head *dst,
886                 unsigned long *scanned, int order, int mode, int file)
887 {
888         unsigned long nr_taken = 0;
889         unsigned long scan;
890
891         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
892                 struct page *page;
893                 unsigned long pfn;
894                 unsigned long end_pfn;
895                 unsigned long page_pfn;
896                 int zone_id;
897
898                 page = lru_to_page(src);
899                 prefetchw_prev_lru_page(page, src, flags);
900
901                 VM_BUG_ON(!PageLRU(page));
902
903                 switch (__isolate_lru_page(page, mode, file)) {
904                 case 0:
905                         list_move(&page->lru, dst);
906                         mem_cgroup_del_lru(page);
907                         nr_taken++;
908                         break;
909
910                 case -EBUSY:
911                         /* else it is being freed elsewhere */
912                         list_move(&page->lru, src);
913                         mem_cgroup_rotate_lru_list(page, page_lru(page));
914                         continue;
915
916                 default:
917                         BUG();
918                 }
919
920                 if (!order)
921                         continue;
922
923                 /*
924                  * Attempt to take all pages in the order aligned region
925                  * surrounding the tag page.  Only take those pages of
926                  * the same active state as that tag page.  We may safely
927                  * round the target page pfn down to the requested order
928                  * as the mem_map is guarenteed valid out to MAX_ORDER,
929                  * where that page is in a different zone we will detect
930                  * it from its zone id and abort this block scan.
931                  */
932                 zone_id = page_zone_id(page);
933                 page_pfn = page_to_pfn(page);
934                 pfn = page_pfn & ~((1 << order) - 1);
935                 end_pfn = pfn + (1 << order);
936                 for (; pfn < end_pfn; pfn++) {
937                         struct page *cursor_page;
938
939                         /* The target page is in the block, ignore it. */
940                         if (unlikely(pfn == page_pfn))
941                                 continue;
942
943                         /* Avoid holes within the zone. */
944                         if (unlikely(!pfn_valid_within(pfn)))
945                                 break;
946
947                         cursor_page = pfn_to_page(pfn);
948
949                         /* Check that we have not crossed a zone boundary. */
950                         if (unlikely(page_zone_id(cursor_page) != zone_id))
951                                 continue;
952
953                         /*
954                          * If we don't have enough swap space, reclaiming of
955                          * anon page which don't already have a swap slot is
956                          * pointless.
957                          */
958                         if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
959                                         !PageSwapCache(cursor_page))
960                                 continue;
961
962                         if (__isolate_lru_page(cursor_page, mode, file) == 0) {
963                                 list_move(&cursor_page->lru, dst);
964                                 mem_cgroup_del_lru(cursor_page);
965                                 nr_taken++;
966                                 scan++;
967                         }
968                 }
969         }
970
971         *scanned = scan;
972         return nr_taken;
973 }
974
975 static unsigned long isolate_pages_global(unsigned long nr,
976                                         struct list_head *dst,
977                                         unsigned long *scanned, int order,
978                                         int mode, struct zone *z,
979                                         struct mem_cgroup *mem_cont,
980                                         int active, int file)
981 {
982         int lru = LRU_BASE;
983         if (active)
984                 lru += LRU_ACTIVE;
985         if (file)
986                 lru += LRU_FILE;
987         return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
988                                                                 mode, file);
989 }
990
991 /*
992  * clear_active_flags() is a helper for shrink_active_list(), clearing
993  * any active bits from the pages in the list.
994  */
995 static unsigned long clear_active_flags(struct list_head *page_list,
996                                         unsigned int *count)
997 {
998         int nr_active = 0;
999         int lru;
1000         struct page *page;
1001
1002         list_for_each_entry(page, page_list, lru) {
1003                 lru = page_lru_base_type(page);
1004                 if (PageActive(page)) {
1005                         lru += LRU_ACTIVE;
1006                         ClearPageActive(page);
1007                         nr_active++;
1008                 }
1009                 count[lru]++;
1010         }
1011
1012         return nr_active;
1013 }
1014
1015 /**
1016  * isolate_lru_page - tries to isolate a page from its LRU list
1017  * @page: page to isolate from its LRU list
1018  *
1019  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1020  * vmstat statistic corresponding to whatever LRU list the page was on.
1021  *
1022  * Returns 0 if the page was removed from an LRU list.
1023  * Returns -EBUSY if the page was not on an LRU list.
1024  *
1025  * The returned page will have PageLRU() cleared.  If it was found on
1026  * the active list, it will have PageActive set.  If it was found on
1027  * the unevictable list, it will have the PageUnevictable bit set. That flag
1028  * may need to be cleared by the caller before letting the page go.
1029  *
1030  * The vmstat statistic corresponding to the list on which the page was
1031  * found will be decremented.
1032  *
1033  * Restrictions:
1034  * (1) Must be called with an elevated refcount on the page. This is a
1035  *     fundamentnal difference from isolate_lru_pages (which is called
1036  *     without a stable reference).
1037  * (2) the lru_lock must not be held.
1038  * (3) interrupts must be enabled.
1039  */
1040 int isolate_lru_page(struct page *page)
1041 {
1042         int ret = -EBUSY;
1043
1044         if (PageLRU(page)) {
1045                 struct zone *zone = page_zone(page);
1046
1047                 spin_lock_irq(&zone->lru_lock);
1048                 if (PageLRU(page) && get_page_unless_zero(page)) {
1049                         int lru = page_lru(page);
1050                         ret = 0;
1051                         ClearPageLRU(page);
1052
1053                         del_page_from_lru_list(zone, page, lru);
1054                 }
1055                 spin_unlock_irq(&zone->lru_lock);
1056         }
1057         return ret;
1058 }
1059
1060 /*
1061  * Are there way too many processes in the direct reclaim path already?
1062  */
1063 static int too_many_isolated(struct zone *zone, int file,
1064                 struct scan_control *sc)
1065 {
1066         unsigned long inactive, isolated;
1067
1068         if (current_is_kswapd())
1069                 return 0;
1070
1071         if (!scanning_global_lru(sc))
1072                 return 0;
1073
1074         if (file) {
1075                 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1076                 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1077         } else {
1078                 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1079                 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1080         }
1081
1082         return isolated > inactive;
1083 }
1084
1085 /*
1086  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1087  * of reclaimed pages
1088  */
1089 static unsigned long shrink_inactive_list(unsigned long max_scan,
1090                         struct zone *zone, struct scan_control *sc,
1091                         int priority, int file)
1092 {
1093         LIST_HEAD(page_list);
1094         struct pagevec pvec;
1095         unsigned long nr_scanned = 0;
1096         unsigned long nr_reclaimed = 0;
1097         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1098         int lumpy_reclaim = 0;
1099
1100         while (unlikely(too_many_isolated(zone, file, sc))) {
1101                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1102
1103                 /* We are about to die and free our memory. Return now. */
1104                 if (fatal_signal_pending(current))
1105                         return SWAP_CLUSTER_MAX;
1106         }
1107
1108         /*
1109          * If we need a large contiguous chunk of memory, or have
1110          * trouble getting a small set of contiguous pages, we
1111          * will reclaim both active and inactive pages.
1112          *
1113          * We use the same threshold as pageout congestion_wait below.
1114          */
1115         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1116                 lumpy_reclaim = 1;
1117         else if (sc->order && priority < DEF_PRIORITY - 2)
1118                 lumpy_reclaim = 1;
1119
1120         pagevec_init(&pvec, 1);
1121
1122         lru_add_drain();
1123         spin_lock_irq(&zone->lru_lock);
1124         do {
1125                 struct page *page;
1126                 unsigned long nr_taken;
1127                 unsigned long nr_scan;
1128                 unsigned long nr_freed;
1129                 unsigned long nr_active;
1130                 unsigned int count[NR_LRU_LISTS] = { 0, };
1131                 int mode = lumpy_reclaim ? ISOLATE_BOTH : ISOLATE_INACTIVE;
1132                 unsigned long nr_anon;
1133                 unsigned long nr_file;
1134
1135                 nr_taken = sc->isolate_pages(sc->swap_cluster_max,
1136                              &page_list, &nr_scan, sc->order, mode,
1137                                 zone, sc->mem_cgroup, 0, file);
1138
1139                 if (scanning_global_lru(sc)) {
1140                         zone->pages_scanned += nr_scan;
1141                         if (current_is_kswapd())
1142                                 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1143                                                        nr_scan);
1144                         else
1145                                 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1146                                                        nr_scan);
1147                 }
1148
1149                 if (nr_taken == 0)
1150                         goto done;
1151
1152                 nr_active = clear_active_flags(&page_list, count);
1153                 __count_vm_events(PGDEACTIVATE, nr_active);
1154
1155                 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1156                                                 -count[LRU_ACTIVE_FILE]);
1157                 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1158                                                 -count[LRU_INACTIVE_FILE]);
1159                 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1160                                                 -count[LRU_ACTIVE_ANON]);
1161                 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1162                                                 -count[LRU_INACTIVE_ANON]);
1163
1164                 nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1165                 nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1166                 __mod_zone_page_state(zone, NR_ISOLATED_ANON, nr_anon);
1167                 __mod_zone_page_state(zone, NR_ISOLATED_FILE, nr_file);
1168
1169                 reclaim_stat->recent_scanned[0] += count[LRU_INACTIVE_ANON];
1170                 reclaim_stat->recent_scanned[0] += count[LRU_ACTIVE_ANON];
1171                 reclaim_stat->recent_scanned[1] += count[LRU_INACTIVE_FILE];
1172                 reclaim_stat->recent_scanned[1] += count[LRU_ACTIVE_FILE];
1173
1174                 spin_unlock_irq(&zone->lru_lock);
1175
1176                 nr_scanned += nr_scan;
1177                 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
1178
1179                 /*
1180                  * If we are direct reclaiming for contiguous pages and we do
1181                  * not reclaim everything in the list, try again and wait
1182                  * for IO to complete. This will stall high-order allocations
1183                  * but that should be acceptable to the caller
1184                  */
1185                 if (nr_freed < nr_taken && !current_is_kswapd() &&
1186                     lumpy_reclaim) {
1187                         congestion_wait(BLK_RW_ASYNC, HZ/10);
1188
1189                         /*
1190                          * The attempt at page out may have made some
1191                          * of the pages active, mark them inactive again.
1192                          */
1193                         nr_active = clear_active_flags(&page_list, count);
1194                         count_vm_events(PGDEACTIVATE, nr_active);
1195
1196                         nr_freed += shrink_page_list(&page_list, sc,
1197                                                         PAGEOUT_IO_SYNC);
1198                 }
1199
1200                 nr_reclaimed += nr_freed;
1201
1202                 local_irq_disable();
1203                 if (current_is_kswapd())
1204                         __count_vm_events(KSWAPD_STEAL, nr_freed);
1205                 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
1206
1207                 spin_lock(&zone->lru_lock);
1208                 /*
1209                  * Put back any unfreeable pages.
1210                  */
1211                 while (!list_empty(&page_list)) {
1212                         int lru;
1213                         page = lru_to_page(&page_list);
1214                         VM_BUG_ON(PageLRU(page));
1215                         list_del(&page->lru);
1216                         if (unlikely(!page_evictable(page, NULL))) {
1217                                 spin_unlock_irq(&zone->lru_lock);
1218                                 putback_lru_page(page);
1219                                 spin_lock_irq(&zone->lru_lock);
1220                                 continue;
1221                         }
1222                         SetPageLRU(page);
1223                         lru = page_lru(page);
1224                         add_page_to_lru_list(zone, page, lru);
1225                         if (is_active_lru(lru)) {
1226                                 int file = is_file_lru(lru);
1227                                 reclaim_stat->recent_rotated[file]++;
1228                         }
1229                         if (!pagevec_add(&pvec, page)) {
1230                                 spin_unlock_irq(&zone->lru_lock);
1231                                 __pagevec_release(&pvec);
1232                                 spin_lock_irq(&zone->lru_lock);
1233                         }
1234                 }
1235                 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1236                 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1237
1238         } while (nr_scanned < max_scan);
1239
1240 done:
1241         spin_unlock_irq(&zone->lru_lock);
1242         pagevec_release(&pvec);
1243         return nr_reclaimed;
1244 }
1245
1246 /*
1247  * We are about to scan this zone at a certain priority level.  If that priority
1248  * level is smaller (ie: more urgent) than the previous priority, then note
1249  * that priority level within the zone.  This is done so that when the next
1250  * process comes in to scan this zone, it will immediately start out at this
1251  * priority level rather than having to build up its own scanning priority.
1252  * Here, this priority affects only the reclaim-mapped threshold.
1253  */
1254 static inline void note_zone_scanning_priority(struct zone *zone, int priority)
1255 {
1256         if (priority < zone->prev_priority)
1257                 zone->prev_priority = priority;
1258 }
1259
1260 /*
1261  * This moves pages from the active list to the inactive list.
1262  *
1263  * We move them the other way if the page is referenced by one or more
1264  * processes, from rmap.
1265  *
1266  * If the pages are mostly unmapped, the processing is fast and it is
1267  * appropriate to hold zone->lru_lock across the whole operation.  But if
1268  * the pages are mapped, the processing is slow (page_referenced()) so we
1269  * should drop zone->lru_lock around each page.  It's impossible to balance
1270  * this, so instead we remove the pages from the LRU while processing them.
1271  * It is safe to rely on PG_active against the non-LRU pages in here because
1272  * nobody will play with that bit on a non-LRU page.
1273  *
1274  * The downside is that we have to touch page->_count against each page.
1275  * But we had to alter page->flags anyway.
1276  */
1277
1278 static void move_active_pages_to_lru(struct zone *zone,
1279                                      struct list_head *list,
1280                                      enum lru_list lru)
1281 {
1282         unsigned long pgmoved = 0;
1283         struct pagevec pvec;
1284         struct page *page;
1285
1286         pagevec_init(&pvec, 1);
1287
1288         while (!list_empty(list)) {
1289                 page = lru_to_page(list);
1290
1291                 VM_BUG_ON(PageLRU(page));
1292                 SetPageLRU(page);
1293
1294                 list_move(&page->lru, &zone->lru[lru].list);
1295                 mem_cgroup_add_lru_list(page, lru);
1296                 pgmoved++;
1297
1298                 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1299                         spin_unlock_irq(&zone->lru_lock);
1300                         if (buffer_heads_over_limit)
1301                                 pagevec_strip(&pvec);
1302                         __pagevec_release(&pvec);
1303                         spin_lock_irq(&zone->lru_lock);
1304                 }
1305         }
1306         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1307         if (!is_active_lru(lru))
1308                 __count_vm_events(PGDEACTIVATE, pgmoved);
1309 }
1310
1311 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1312                         struct scan_control *sc, int priority, int file)
1313 {
1314         unsigned long nr_taken;
1315         unsigned long pgscanned;
1316         unsigned long vm_flags;
1317         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1318         LIST_HEAD(l_active);
1319         LIST_HEAD(l_inactive);
1320         struct page *page;
1321         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1322         unsigned long nr_rotated = 0;
1323
1324         lru_add_drain();
1325         spin_lock_irq(&zone->lru_lock);
1326         nr_taken = sc->isolate_pages(nr_pages, &l_hold, &pgscanned, sc->order,
1327                                         ISOLATE_ACTIVE, zone,
1328                                         sc->mem_cgroup, 1, file);
1329         /*
1330          * zone->pages_scanned is used for detect zone's oom
1331          * mem_cgroup remembers nr_scan by itself.
1332          */
1333         if (scanning_global_lru(sc)) {
1334                 zone->pages_scanned += pgscanned;
1335         }
1336         reclaim_stat->recent_scanned[file] += nr_taken;
1337
1338         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1339         if (file)
1340                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1341         else
1342                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1343         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1344         spin_unlock_irq(&zone->lru_lock);
1345
1346         while (!list_empty(&l_hold)) {
1347                 cond_resched();
1348                 page = lru_to_page(&l_hold);
1349                 list_del(&page->lru);
1350
1351                 if (unlikely(!page_evictable(page, NULL))) {
1352                         putback_lru_page(page);
1353                         continue;
1354                 }
1355
1356                 /* page_referenced clears PageReferenced */
1357                 if (page_mapping_inuse(page) &&
1358                     page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1359                         nr_rotated++;
1360                         /*
1361                          * Identify referenced, file-backed active pages and
1362                          * give them one more trip around the active list. So
1363                          * that executable code get better chances to stay in
1364                          * memory under moderate memory pressure.  Anon pages
1365                          * are not likely to be evicted by use-once streaming
1366                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1367                          * so we ignore them here.
1368                          */
1369                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1370                                 list_add(&page->lru, &l_active);
1371                                 continue;
1372                         }
1373                 }
1374
1375                 ClearPageActive(page);  /* we are de-activating */
1376                 list_add(&page->lru, &l_inactive);
1377         }
1378
1379         /*
1380          * Move pages back to the lru list.
1381          */
1382         spin_lock_irq(&zone->lru_lock);
1383         /*
1384          * Count referenced pages from currently used mappings as rotated,
1385          * even though only some of them are actually re-activated.  This
1386          * helps balance scan pressure between file and anonymous pages in
1387          * get_scan_ratio.
1388          */
1389         reclaim_stat->recent_rotated[file] += nr_rotated;
1390
1391         move_active_pages_to_lru(zone, &l_active,
1392                                                 LRU_ACTIVE + file * LRU_FILE);
1393         move_active_pages_to_lru(zone, &l_inactive,
1394                                                 LRU_BASE   + file * LRU_FILE);
1395         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1396         spin_unlock_irq(&zone->lru_lock);
1397 }
1398
1399 static int inactive_anon_is_low_global(struct zone *zone)
1400 {
1401         unsigned long active, inactive;
1402
1403         active = zone_page_state(zone, NR_ACTIVE_ANON);
1404         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1405
1406         if (inactive * zone->inactive_ratio < active)
1407                 return 1;
1408
1409         return 0;
1410 }
1411
1412 /**
1413  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1414  * @zone: zone to check
1415  * @sc:   scan control of this context
1416  *
1417  * Returns true if the zone does not have enough inactive anon pages,
1418  * meaning some active anon pages need to be deactivated.
1419  */
1420 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1421 {
1422         int low;
1423
1424         if (scanning_global_lru(sc))
1425                 low = inactive_anon_is_low_global(zone);
1426         else
1427                 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1428         return low;
1429 }
1430
1431 static int inactive_file_is_low_global(struct zone *zone)
1432 {
1433         unsigned long active, inactive;
1434
1435         active = zone_page_state(zone, NR_ACTIVE_FILE);
1436         inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1437
1438         return (active > inactive);
1439 }
1440
1441 /**
1442  * inactive_file_is_low - check if file pages need to be deactivated
1443  * @zone: zone to check
1444  * @sc:   scan control of this context
1445  *
1446  * When the system is doing streaming IO, memory pressure here
1447  * ensures that active file pages get deactivated, until more
1448  * than half of the file pages are on the inactive list.
1449  *
1450  * Once we get to that situation, protect the system's working
1451  * set from being evicted by disabling active file page aging.
1452  *
1453  * This uses a different ratio than the anonymous pages, because
1454  * the page cache uses a use-once replacement algorithm.
1455  */
1456 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1457 {
1458         int low;
1459
1460         if (scanning_global_lru(sc))
1461                 low = inactive_file_is_low_global(zone);
1462         else
1463                 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1464         return low;
1465 }
1466
1467 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1468         struct zone *zone, struct scan_control *sc, int priority)
1469 {
1470         int file = is_file_lru(lru);
1471
1472         if (lru == LRU_ACTIVE_FILE && inactive_file_is_low(zone, sc)) {
1473                 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1474                 return 0;
1475         }
1476
1477         if (lru == LRU_ACTIVE_ANON && inactive_anon_is_low(zone, sc)) {
1478                 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1479                 return 0;
1480         }
1481         return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1482 }
1483
1484 /*
1485  * Determine how aggressively the anon and file LRU lists should be
1486  * scanned.  The relative value of each set of LRU lists is determined
1487  * by looking at the fraction of the pages scanned we did rotate back
1488  * onto the active list instead of evict.
1489  *
1490  * percent[0] specifies how much pressure to put on ram/swap backed
1491  * memory, while percent[1] determines pressure on the file LRUs.
1492  */
1493 static void get_scan_ratio(struct zone *zone, struct scan_control *sc,
1494                                         unsigned long *percent)
1495 {
1496         unsigned long anon, file, free;
1497         unsigned long anon_prio, file_prio;
1498         unsigned long ap, fp;
1499         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1500
1501         anon  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1502                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1503         file  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1504                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1505
1506         if (scanning_global_lru(sc)) {
1507                 free  = zone_page_state(zone, NR_FREE_PAGES);
1508                 /* If we have very few page cache pages,
1509                    force-scan anon pages. */
1510                 if (unlikely(file + free <= high_wmark_pages(zone))) {
1511                         percent[0] = 100;
1512                         percent[1] = 0;
1513                         return;
1514                 }
1515         }
1516
1517         /*
1518          * OK, so we have swap space and a fair amount of page cache
1519          * pages.  We use the recently rotated / recently scanned
1520          * ratios to determine how valuable each cache is.
1521          *
1522          * Because workloads change over time (and to avoid overflow)
1523          * we keep these statistics as a floating average, which ends
1524          * up weighing recent references more than old ones.
1525          *
1526          * anon in [0], file in [1]
1527          */
1528         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1529                 spin_lock_irq(&zone->lru_lock);
1530                 reclaim_stat->recent_scanned[0] /= 2;
1531                 reclaim_stat->recent_rotated[0] /= 2;
1532                 spin_unlock_irq(&zone->lru_lock);
1533         }
1534
1535         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1536                 spin_lock_irq(&zone->lru_lock);
1537                 reclaim_stat->recent_scanned[1] /= 2;
1538                 reclaim_stat->recent_rotated[1] /= 2;
1539                 spin_unlock_irq(&zone->lru_lock);
1540         }
1541
1542         /*
1543          * With swappiness at 100, anonymous and file have the same priority.
1544          * This scanning priority is essentially the inverse of IO cost.
1545          */
1546         anon_prio = sc->swappiness;
1547         file_prio = 200 - sc->swappiness;
1548
1549         /*
1550          * The amount of pressure on anon vs file pages is inversely
1551          * proportional to the fraction of recently scanned pages on
1552          * each list that were recently referenced and in active use.
1553          */
1554         ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1555         ap /= reclaim_stat->recent_rotated[0] + 1;
1556
1557         fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1558         fp /= reclaim_stat->recent_rotated[1] + 1;
1559
1560         /* Normalize to percentages */
1561         percent[0] = 100 * ap / (ap + fp + 1);
1562         percent[1] = 100 - percent[0];
1563 }
1564
1565 /*
1566  * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1567  * until we collected @swap_cluster_max pages to scan.
1568  */
1569 static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
1570                                        unsigned long *nr_saved_scan,
1571                                        unsigned long swap_cluster_max)
1572 {
1573         unsigned long nr;
1574
1575         *nr_saved_scan += nr_to_scan;
1576         nr = *nr_saved_scan;
1577
1578         if (nr >= swap_cluster_max)
1579                 *nr_saved_scan = 0;
1580         else
1581                 nr = 0;
1582
1583         return nr;
1584 }
1585
1586 /*
1587  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1588  */
1589 static void shrink_zone(int priority, struct zone *zone,
1590                                 struct scan_control *sc)
1591 {
1592         unsigned long nr[NR_LRU_LISTS];
1593         unsigned long nr_to_scan;
1594         unsigned long percent[2];       /* anon @ 0; file @ 1 */
1595         enum lru_list l;
1596         unsigned long nr_reclaimed = sc->nr_reclaimed;
1597         unsigned long swap_cluster_max = sc->swap_cluster_max;
1598         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1599         int noswap = 0;
1600
1601         /* If we have no swap space, do not bother scanning anon pages. */
1602         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1603                 noswap = 1;
1604                 percent[0] = 0;
1605                 percent[1] = 100;
1606         } else
1607                 get_scan_ratio(zone, sc, percent);
1608
1609         for_each_evictable_lru(l) {
1610                 int file = is_file_lru(l);
1611                 unsigned long scan;
1612
1613                 scan = zone_nr_lru_pages(zone, sc, l);
1614                 if (priority || noswap) {
1615                         scan >>= priority;
1616                         scan = (scan * percent[file]) / 100;
1617                 }
1618                 nr[l] = nr_scan_try_batch(scan,
1619                                           &reclaim_stat->nr_saved_scan[l],
1620                                           swap_cluster_max);
1621         }
1622
1623         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1624                                         nr[LRU_INACTIVE_FILE]) {
1625                 for_each_evictable_lru(l) {
1626                         if (nr[l]) {
1627                                 nr_to_scan = min(nr[l], swap_cluster_max);
1628                                 nr[l] -= nr_to_scan;
1629
1630                                 nr_reclaimed += shrink_list(l, nr_to_scan,
1631                                                             zone, sc, priority);
1632                         }
1633                 }
1634                 /*
1635                  * On large memory systems, scan >> priority can become
1636                  * really large. This is fine for the starting priority;
1637                  * we want to put equal scanning pressure on each zone.
1638                  * However, if the VM has a harder time of freeing pages,
1639                  * with multiple processes reclaiming pages, the total
1640                  * freeing target can get unreasonably large.
1641                  */
1642                 if (nr_reclaimed > swap_cluster_max &&
1643                         priority < DEF_PRIORITY && !current_is_kswapd())
1644                         break;
1645         }
1646
1647         sc->nr_reclaimed = nr_reclaimed;
1648
1649         /*
1650          * Even if we did not try to evict anon pages at all, we want to
1651          * rebalance the anon lru active/inactive ratio.
1652          */
1653         if (inactive_anon_is_low(zone, sc) && nr_swap_pages > 0)
1654                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1655
1656         throttle_vm_writeout(sc->gfp_mask);
1657 }
1658
1659 /*
1660  * This is the direct reclaim path, for page-allocating processes.  We only
1661  * try to reclaim pages from zones which will satisfy the caller's allocation
1662  * request.
1663  *
1664  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1665  * Because:
1666  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1667  *    allocation or
1668  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1669  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1670  *    zone defense algorithm.
1671  *
1672  * If a zone is deemed to be full of pinned pages then just give it a light
1673  * scan then give up on it.
1674  */
1675 static void shrink_zones(int priority, struct zonelist *zonelist,
1676                                         struct scan_control *sc)
1677 {
1678         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1679         struct zoneref *z;
1680         struct zone *zone;
1681
1682         sc->all_unreclaimable = 1;
1683         for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1684                                         sc->nodemask) {
1685                 if (!populated_zone(zone))
1686                         continue;
1687                 /*
1688                  * Take care memory controller reclaiming has small influence
1689                  * to global LRU.
1690                  */
1691                 if (scanning_global_lru(sc)) {
1692                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1693                                 continue;
1694                         note_zone_scanning_priority(zone, priority);
1695
1696                         if (zone_is_all_unreclaimable(zone) &&
1697                                                 priority != DEF_PRIORITY)
1698                                 continue;       /* Let kswapd poll it */
1699                         sc->all_unreclaimable = 0;
1700                 } else {
1701                         /*
1702                          * Ignore cpuset limitation here. We just want to reduce
1703                          * # of used pages by us regardless of memory shortage.
1704                          */
1705                         sc->all_unreclaimable = 0;
1706                         mem_cgroup_note_reclaim_priority(sc->mem_cgroup,
1707                                                         priority);
1708                 }
1709
1710                 shrink_zone(priority, zone, sc);
1711         }
1712 }
1713
1714 /*
1715  * This is the main entry point to direct page reclaim.
1716  *
1717  * If a full scan of the inactive list fails to free enough memory then we
1718  * are "out of memory" and something needs to be killed.
1719  *
1720  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1721  * high - the zone may be full of dirty or under-writeback pages, which this
1722  * caller can't do much about.  We kick the writeback threads and take explicit
1723  * naps in the hope that some of these pages can be written.  But if the
1724  * allocating task holds filesystem locks which prevent writeout this might not
1725  * work, and the allocation attempt will fail.
1726  *
1727  * returns:     0, if no pages reclaimed
1728  *              else, the number of pages reclaimed
1729  */
1730 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
1731                                         struct scan_control *sc)
1732 {
1733         int priority;
1734         unsigned long ret = 0;
1735         unsigned long total_scanned = 0;
1736         struct reclaim_state *reclaim_state = current->reclaim_state;
1737         unsigned long lru_pages = 0;
1738         struct zoneref *z;
1739         struct zone *zone;
1740         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1741
1742         delayacct_freepages_start();
1743
1744         if (scanning_global_lru(sc))
1745                 count_vm_event(ALLOCSTALL);
1746         /*
1747          * mem_cgroup will not do shrink_slab.
1748          */
1749         if (scanning_global_lru(sc)) {
1750                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1751
1752                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1753                                 continue;
1754
1755                         lru_pages += zone_reclaimable_pages(zone);
1756                 }
1757         }
1758
1759         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1760                 sc->nr_scanned = 0;
1761                 if (!priority)
1762                         disable_swap_token();
1763                 shrink_zones(priority, zonelist, sc);
1764                 /*
1765                  * Don't shrink slabs when reclaiming memory from
1766                  * over limit cgroups
1767                  */
1768                 if (scanning_global_lru(sc)) {
1769                         shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
1770                         if (reclaim_state) {
1771                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
1772                                 reclaim_state->reclaimed_slab = 0;
1773                         }
1774                 }
1775                 total_scanned += sc->nr_scanned;
1776                 if (sc->nr_reclaimed >= sc->swap_cluster_max) {
1777                         ret = sc->nr_reclaimed;
1778                         goto out;
1779                 }
1780
1781                 /*
1782                  * Try to write back as many pages as we just scanned.  This
1783                  * tends to cause slow streaming writers to write data to the
1784                  * disk smoothly, at the dirtying rate, which is nice.   But
1785                  * that's undesirable in laptop mode, where we *want* lumpy
1786                  * writeout.  So in laptop mode, write out the whole world.
1787                  */
1788                 if (total_scanned > sc->swap_cluster_max +
1789                                         sc->swap_cluster_max / 2) {
1790                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
1791                         sc->may_writepage = 1;
1792                 }
1793
1794                 /* Take a nap, wait for some writeback to complete */
1795                 if (sc->nr_scanned && priority < DEF_PRIORITY - 2)
1796                         congestion_wait(BLK_RW_ASYNC, HZ/10);
1797         }
1798         /* top priority shrink_zones still had more to do? don't OOM, then */
1799         if (!sc->all_unreclaimable && scanning_global_lru(sc))
1800                 ret = sc->nr_reclaimed;
1801 out:
1802         /*
1803          * Now that we've scanned all the zones at this priority level, note
1804          * that level within the zone so that the next thread which performs
1805          * scanning of this zone will immediately start out at this priority
1806          * level.  This affects only the decision whether or not to bring
1807          * mapped pages onto the inactive list.
1808          */
1809         if (priority < 0)
1810                 priority = 0;
1811
1812         if (scanning_global_lru(sc)) {
1813                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1814
1815                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1816                                 continue;
1817
1818                         zone->prev_priority = priority;
1819                 }
1820         } else
1821                 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority);
1822
1823         delayacct_freepages_end();
1824
1825         return ret;
1826 }
1827
1828 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
1829                                 gfp_t gfp_mask, nodemask_t *nodemask)
1830 {
1831         struct scan_control sc = {
1832                 .gfp_mask = gfp_mask,
1833                 .may_writepage = !laptop_mode,
1834                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1835                 .may_unmap = 1,
1836                 .may_swap = 1,
1837                 .swappiness = vm_swappiness,
1838                 .order = order,
1839                 .mem_cgroup = NULL,
1840                 .isolate_pages = isolate_pages_global,
1841                 .nodemask = nodemask,
1842         };
1843
1844         return do_try_to_free_pages(zonelist, &sc);
1845 }
1846
1847 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
1848
1849 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
1850                                                 gfp_t gfp_mask, bool noswap,
1851                                                 unsigned int swappiness,
1852                                                 struct zone *zone, int nid)
1853 {
1854         struct scan_control sc = {
1855                 .may_writepage = !laptop_mode,
1856                 .may_unmap = 1,
1857                 .may_swap = !noswap,
1858                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1859                 .swappiness = swappiness,
1860                 .order = 0,
1861                 .mem_cgroup = mem,
1862                 .isolate_pages = mem_cgroup_isolate_pages,
1863         };
1864         nodemask_t nm  = nodemask_of_node(nid);
1865
1866         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1867                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1868         sc.nodemask = &nm;
1869         sc.nr_reclaimed = 0;
1870         sc.nr_scanned = 0;
1871         /*
1872          * NOTE: Although we can get the priority field, using it
1873          * here is not a good idea, since it limits the pages we can scan.
1874          * if we don't reclaim here, the shrink_zone from balance_pgdat
1875          * will pick up pages from other mem cgroup's as well. We hack
1876          * the priority and make it zero.
1877          */
1878         shrink_zone(0, zone, &sc);
1879         return sc.nr_reclaimed;
1880 }
1881
1882 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
1883                                            gfp_t gfp_mask,
1884                                            bool noswap,
1885                                            unsigned int swappiness)
1886 {
1887         struct zonelist *zonelist;
1888         struct scan_control sc = {
1889                 .may_writepage = !laptop_mode,
1890                 .may_unmap = 1,
1891                 .may_swap = !noswap,
1892                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1893                 .swappiness = swappiness,
1894                 .order = 0,
1895                 .mem_cgroup = mem_cont,
1896                 .isolate_pages = mem_cgroup_isolate_pages,
1897                 .nodemask = NULL, /* we don't care the placement */
1898         };
1899
1900         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1901                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1902         zonelist = NODE_DATA(numa_node_id())->node_zonelists;
1903         return do_try_to_free_pages(zonelist, &sc);
1904 }
1905 #endif
1906
1907 /* is kswapd sleeping prematurely? */
1908 static int sleeping_prematurely(pg_data_t *pgdat, int order, long remaining)
1909 {
1910         int i;
1911
1912         /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
1913         if (remaining)
1914                 return 1;
1915
1916         /* If after HZ/10, a zone is below the high mark, it's premature */
1917         for (i = 0; i < pgdat->nr_zones; i++) {
1918                 struct zone *zone = pgdat->node_zones + i;
1919
1920                 if (!populated_zone(zone))
1921                         continue;
1922
1923                 if (!zone_watermark_ok(zone, order, high_wmark_pages(zone),
1924                                                                 0, 0))
1925                         return 1;
1926         }
1927
1928         return 0;
1929 }
1930
1931 /*
1932  * For kswapd, balance_pgdat() will work across all this node's zones until
1933  * they are all at high_wmark_pages(zone).
1934  *
1935  * Returns the number of pages which were actually freed.
1936  *
1937  * There is special handling here for zones which are full of pinned pages.
1938  * This can happen if the pages are all mlocked, or if they are all used by
1939  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1940  * What we do is to detect the case where all pages in the zone have been
1941  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1942  * dead and from now on, only perform a short scan.  Basically we're polling
1943  * the zone for when the problem goes away.
1944  *
1945  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
1946  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
1947  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
1948  * lower zones regardless of the number of free pages in the lower zones. This
1949  * interoperates with the page allocator fallback scheme to ensure that aging
1950  * of pages is balanced across the zones.
1951  */
1952 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
1953 {
1954         int all_zones_ok;
1955         int priority;
1956         int i;
1957         unsigned long total_scanned;
1958         struct reclaim_state *reclaim_state = current->reclaim_state;
1959         struct scan_control sc = {
1960                 .gfp_mask = GFP_KERNEL,
1961                 .may_unmap = 1,
1962                 .may_swap = 1,
1963                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1964                 .swappiness = vm_swappiness,
1965                 .order = order,
1966                 .mem_cgroup = NULL,
1967                 .isolate_pages = isolate_pages_global,
1968         };
1969         /*
1970          * temp_priority is used to remember the scanning priority at which
1971          * this zone was successfully refilled to
1972          * free_pages == high_wmark_pages(zone).
1973          */
1974         int temp_priority[MAX_NR_ZONES];
1975
1976 loop_again:
1977         total_scanned = 0;
1978         sc.nr_reclaimed = 0;
1979         sc.may_writepage = !laptop_mode;
1980         count_vm_event(PAGEOUTRUN);
1981
1982         for (i = 0; i < pgdat->nr_zones; i++)
1983                 temp_priority[i] = DEF_PRIORITY;
1984
1985         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1986                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1987                 unsigned long lru_pages = 0;
1988                 int has_under_min_watermark_zone = 0;
1989
1990                 /* The swap token gets in the way of swapout... */
1991                 if (!priority)
1992                         disable_swap_token();
1993
1994                 all_zones_ok = 1;
1995
1996                 /*
1997                  * Scan in the highmem->dma direction for the highest
1998                  * zone which needs scanning
1999                  */
2000                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2001                         struct zone *zone = pgdat->node_zones + i;
2002
2003                         if (!populated_zone(zone))
2004                                 continue;
2005
2006                         if (zone_is_all_unreclaimable(zone) &&
2007                             priority != DEF_PRIORITY)
2008                                 continue;
2009
2010                         /*
2011                          * Do some background aging of the anon list, to give
2012                          * pages a chance to be referenced before reclaiming.
2013                          */
2014                         if (inactive_anon_is_low(zone, &sc))
2015                                 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2016                                                         &sc, priority, 0);
2017
2018                         if (!zone_watermark_ok(zone, order,
2019                                         high_wmark_pages(zone), 0, 0)) {
2020                                 end_zone = i;
2021                                 break;
2022                         }
2023                 }
2024                 if (i < 0)
2025                         goto out;
2026
2027                 for (i = 0; i <= end_zone; i++) {
2028                         struct zone *zone = pgdat->node_zones + i;
2029
2030                         lru_pages += zone_reclaimable_pages(zone);
2031                 }
2032
2033                 /*
2034                  * Now scan the zone in the dma->highmem direction, stopping
2035                  * at the last zone which needs scanning.
2036                  *
2037                  * We do this because the page allocator works in the opposite
2038                  * direction.  This prevents the page allocator from allocating
2039                  * pages behind kswapd's direction of progress, which would
2040                  * cause too much scanning of the lower zones.
2041                  */
2042                 for (i = 0; i <= end_zone; i++) {
2043                         struct zone *zone = pgdat->node_zones + i;
2044                         int nr_slab;
2045                         int nid, zid;
2046
2047                         if (!populated_zone(zone))
2048                                 continue;
2049
2050                         if (zone_is_all_unreclaimable(zone) &&
2051                                         priority != DEF_PRIORITY)
2052                                 continue;
2053
2054                         if (!zone_watermark_ok(zone, order,
2055                                         high_wmark_pages(zone), end_zone, 0))
2056                                 all_zones_ok = 0;
2057                         temp_priority[i] = priority;
2058                         sc.nr_scanned = 0;
2059                         note_zone_scanning_priority(zone, priority);
2060
2061                         nid = pgdat->node_id;
2062                         zid = zone_idx(zone);
2063                         /*
2064                          * Call soft limit reclaim before calling shrink_zone.
2065                          * For now we ignore the return value
2066                          */
2067                         mem_cgroup_soft_limit_reclaim(zone, order, sc.gfp_mask,
2068                                                         nid, zid);
2069                         /*
2070                          * We put equal pressure on every zone, unless one
2071                          * zone has way too many pages free already.
2072                          */
2073                         if (!zone_watermark_ok(zone, order,
2074                                         8*high_wmark_pages(zone), end_zone, 0))
2075                                 shrink_zone(priority, zone, &sc);
2076                         reclaim_state->reclaimed_slab = 0;
2077                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
2078                                                 lru_pages);
2079                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2080                         total_scanned += sc.nr_scanned;
2081                         if (zone_is_all_unreclaimable(zone))
2082                                 continue;
2083                         if (nr_slab == 0 && zone->pages_scanned >=
2084                                         (zone_reclaimable_pages(zone) * 6))
2085                                         zone_set_flag(zone,
2086                                                       ZONE_ALL_UNRECLAIMABLE);
2087                         /*
2088                          * If we've done a decent amount of scanning and
2089                          * the reclaim ratio is low, start doing writepage
2090                          * even in laptop mode
2091                          */
2092                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2093                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2094                                 sc.may_writepage = 1;
2095
2096                         /*
2097                          * We are still under min water mark. it mean we have
2098                          * GFP_ATOMIC allocation failure risk. Hurry up!
2099                          */
2100                         if (!zone_watermark_ok(zone, order, min_wmark_pages(zone),
2101                                               end_zone, 0))
2102                                 has_under_min_watermark_zone = 1;
2103
2104                 }
2105                 if (all_zones_ok)
2106                         break;          /* kswapd: all done */
2107                 /*
2108                  * OK, kswapd is getting into trouble.  Take a nap, then take
2109                  * another pass across the zones.
2110                  */
2111                 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2112                         if (has_under_min_watermark_zone)
2113                                 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2114                         else
2115                                 congestion_wait(BLK_RW_ASYNC, HZ/10);
2116                 }
2117
2118                 /*
2119                  * We do this so kswapd doesn't build up large priorities for
2120                  * example when it is freeing in parallel with allocators. It
2121                  * matches the direct reclaim path behaviour in terms of impact
2122                  * on zone->*_priority.
2123                  */
2124                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2125                         break;
2126         }
2127 out:
2128         /*
2129          * Note within each zone the priority level at which this zone was
2130          * brought into a happy state.  So that the next thread which scans this
2131          * zone will start out at that priority level.
2132          */
2133         for (i = 0; i < pgdat->nr_zones; i++) {
2134                 struct zone *zone = pgdat->node_zones + i;
2135
2136                 zone->prev_priority = temp_priority[i];
2137         }
2138         if (!all_zones_ok) {
2139                 cond_resched();
2140
2141                 try_to_freeze();
2142
2143                 /*
2144                  * Fragmentation may mean that the system cannot be
2145                  * rebalanced for high-order allocations in all zones.
2146                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2147                  * it means the zones have been fully scanned and are still
2148                  * not balanced. For high-order allocations, there is
2149                  * little point trying all over again as kswapd may
2150                  * infinite loop.
2151                  *
2152                  * Instead, recheck all watermarks at order-0 as they
2153                  * are the most important. If watermarks are ok, kswapd will go
2154                  * back to sleep. High-order users can still perform direct
2155                  * reclaim if they wish.
2156                  */
2157                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2158                         order = sc.order = 0;
2159
2160                 goto loop_again;
2161         }
2162
2163         return sc.nr_reclaimed;
2164 }
2165
2166 /*
2167  * The background pageout daemon, started as a kernel thread
2168  * from the init process.
2169  *
2170  * This basically trickles out pages so that we have _some_
2171  * free memory available even if there is no other activity
2172  * that frees anything up. This is needed for things like routing
2173  * etc, where we otherwise might have all activity going on in
2174  * asynchronous contexts that cannot page things out.
2175  *
2176  * If there are applications that are active memory-allocators
2177  * (most normal use), this basically shouldn't matter.
2178  */
2179 static int kswapd(void *p)
2180 {
2181         unsigned long order;
2182         pg_data_t *pgdat = (pg_data_t*)p;
2183         struct task_struct *tsk = current;
2184         DEFINE_WAIT(wait);
2185         struct reclaim_state reclaim_state = {
2186                 .reclaimed_slab = 0,
2187         };
2188         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2189
2190         lockdep_set_current_reclaim_state(GFP_KERNEL);
2191
2192         if (!cpumask_empty(cpumask))
2193                 set_cpus_allowed_ptr(tsk, cpumask);
2194         current->reclaim_state = &reclaim_state;
2195
2196         /*
2197          * Tell the memory management that we're a "memory allocator",
2198          * and that if we need more memory we should get access to it
2199          * regardless (see "__alloc_pages()"). "kswapd" should
2200          * never get caught in the normal page freeing logic.
2201          *
2202          * (Kswapd normally doesn't need memory anyway, but sometimes
2203          * you need a small amount of memory in order to be able to
2204          * page out something else, and this flag essentially protects
2205          * us from recursively trying to free more memory as we're
2206          * trying to free the first piece of memory in the first place).
2207          */
2208         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2209         set_freezable();
2210
2211         order = 0;
2212         for ( ; ; ) {
2213                 unsigned long new_order;
2214                 int ret;
2215
2216                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2217                 new_order = pgdat->kswapd_max_order;
2218                 pgdat->kswapd_max_order = 0;
2219                 if (order < new_order) {
2220                         /*
2221                          * Don't sleep if someone wants a larger 'order'
2222                          * allocation
2223                          */
2224                         order = new_order;
2225                 } else {
2226                         if (!freezing(current) && !kthread_should_stop()) {
2227                                 long remaining = 0;
2228
2229                                 /* Try to sleep for a short interval */
2230                                 if (!sleeping_prematurely(pgdat, order, remaining)) {
2231                                         remaining = schedule_timeout(HZ/10);
2232                                         finish_wait(&pgdat->kswapd_wait, &wait);
2233                                         prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2234                                 }
2235
2236                                 /*
2237                                  * After a short sleep, check if it was a
2238                                  * premature sleep. If not, then go fully
2239                                  * to sleep until explicitly woken up
2240                                  */
2241                                 if (!sleeping_prematurely(pgdat, order, remaining))
2242                                         schedule();
2243                                 else {
2244                                         if (remaining)
2245                                                 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2246                                         else
2247                                                 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2248                                 }
2249                         }
2250
2251                         order = pgdat->kswapd_max_order;
2252                 }
2253                 finish_wait(&pgdat->kswapd_wait, &wait);
2254
2255                 ret = try_to_freeze();
2256                 if (kthread_should_stop())
2257                         break;
2258
2259                 /*
2260                  * We can speed up thawing tasks if we don't call balance_pgdat
2261                  * after returning from the refrigerator
2262                  */
2263                 if (!ret)
2264                         balance_pgdat(pgdat, order);
2265         }
2266         return 0;
2267 }
2268
2269 /*
2270  * A zone is low on free memory, so wake its kswapd task to service it.
2271  */
2272 void wakeup_kswapd(struct zone *zone, int order)
2273 {
2274         pg_data_t *pgdat;
2275
2276         if (!populated_zone(zone))
2277                 return;
2278
2279         pgdat = zone->zone_pgdat;
2280         if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
2281                 return;
2282         if (pgdat->kswapd_max_order < order)
2283                 pgdat->kswapd_max_order = order;
2284         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2285                 return;
2286         if (!waitqueue_active(&pgdat->kswapd_wait))
2287                 return;
2288         wake_up_interruptible(&pgdat->kswapd_wait);
2289 }
2290
2291 /*
2292  * The reclaimable count would be mostly accurate.
2293  * The less reclaimable pages may be
2294  * - mlocked pages, which will be moved to unevictable list when encountered
2295  * - mapped pages, which may require several travels to be reclaimed
2296  * - dirty pages, which is not "instantly" reclaimable
2297  */
2298 unsigned long global_reclaimable_pages(void)
2299 {
2300         int nr;
2301
2302         nr = global_page_state(NR_ACTIVE_FILE) +
2303              global_page_state(NR_INACTIVE_FILE);
2304
2305         if (nr_swap_pages > 0)
2306                 nr += global_page_state(NR_ACTIVE_ANON) +
2307                       global_page_state(NR_INACTIVE_ANON);
2308
2309         return nr;
2310 }
2311
2312 unsigned long zone_reclaimable_pages(struct zone *zone)
2313 {
2314         int nr;
2315
2316         nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2317              zone_page_state(zone, NR_INACTIVE_FILE);
2318
2319         if (nr_swap_pages > 0)
2320                 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2321                       zone_page_state(zone, NR_INACTIVE_ANON);
2322
2323         return nr;
2324 }
2325
2326 #ifdef CONFIG_HIBERNATION
2327 /*
2328  * Helper function for shrink_all_memory().  Tries to reclaim 'nr_pages' pages
2329  * from LRU lists system-wide, for given pass and priority.
2330  *
2331  * For pass > 3 we also try to shrink the LRU lists that contain a few pages
2332  */
2333 static void shrink_all_zones(unsigned long nr_pages, int prio,
2334                                       int pass, struct scan_control *sc)
2335 {
2336         struct zone *zone;
2337         unsigned long nr_reclaimed = 0;
2338         struct zone_reclaim_stat *reclaim_stat;
2339
2340         for_each_populated_zone(zone) {
2341                 enum lru_list l;
2342
2343                 if (zone_is_all_unreclaimable(zone) && prio != DEF_PRIORITY)
2344                         continue;
2345
2346                 for_each_evictable_lru(l) {
2347                         enum zone_stat_item ls = NR_LRU_BASE + l;
2348                         unsigned long lru_pages = zone_page_state(zone, ls);
2349
2350                         /* For pass = 0, we don't shrink the active list */
2351                         if (pass == 0 && (l == LRU_ACTIVE_ANON ||
2352                                                 l == LRU_ACTIVE_FILE))
2353                                 continue;
2354
2355                         reclaim_stat = get_reclaim_stat(zone, sc);
2356                         reclaim_stat->nr_saved_scan[l] +=
2357                                                 (lru_pages >> prio) + 1;
2358                         if (reclaim_stat->nr_saved_scan[l]
2359                                                 >= nr_pages || pass > 3) {
2360                                 unsigned long nr_to_scan;
2361
2362                                 reclaim_stat->nr_saved_scan[l] = 0;
2363                                 nr_to_scan = min(nr_pages, lru_pages);
2364                                 nr_reclaimed += shrink_list(l, nr_to_scan, zone,
2365                                                                 sc, prio);
2366                                 if (nr_reclaimed >= nr_pages) {
2367                                         sc->nr_reclaimed += nr_reclaimed;
2368                                         return;
2369                                 }
2370                         }
2371                 }
2372         }
2373         sc->nr_reclaimed += nr_reclaimed;
2374 }
2375
2376 /*
2377  * Try to free `nr_pages' of memory, system-wide, and return the number of
2378  * freed pages.
2379  *
2380  * Rather than trying to age LRUs the aim is to preserve the overall
2381  * LRU order by reclaiming preferentially
2382  * inactive > active > active referenced > active mapped
2383  */
2384 unsigned long shrink_all_memory(unsigned long nr_pages)
2385 {
2386         unsigned long lru_pages, nr_slab;
2387         int pass;
2388         struct reclaim_state reclaim_state;
2389         struct scan_control sc = {
2390                 .gfp_mask = GFP_KERNEL,
2391                 .may_unmap = 0,
2392                 .may_writepage = 1,
2393                 .isolate_pages = isolate_pages_global,
2394                 .nr_reclaimed = 0,
2395         };
2396
2397         current->reclaim_state = &reclaim_state;
2398
2399         lru_pages = global_reclaimable_pages();
2400         nr_slab = global_page_state(NR_SLAB_RECLAIMABLE);
2401         /* If slab caches are huge, it's better to hit them first */
2402         while (nr_slab >= lru_pages) {
2403                 reclaim_state.reclaimed_slab = 0;
2404                 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
2405                 if (!reclaim_state.reclaimed_slab)
2406                         break;
2407
2408                 sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2409                 if (sc.nr_reclaimed >= nr_pages)
2410                         goto out;
2411
2412                 nr_slab -= reclaim_state.reclaimed_slab;
2413         }
2414
2415         /*
2416          * We try to shrink LRUs in 5 passes:
2417          * 0 = Reclaim from inactive_list only
2418          * 1 = Reclaim from active list but don't reclaim mapped
2419          * 2 = 2nd pass of type 1
2420          * 3 = Reclaim mapped (normal reclaim)
2421          * 4 = 2nd pass of type 3
2422          */
2423         for (pass = 0; pass < 5; pass++) {
2424                 int prio;
2425
2426                 /* Force reclaiming mapped pages in the passes #3 and #4 */
2427                 if (pass > 2)
2428                         sc.may_unmap = 1;
2429
2430                 for (prio = DEF_PRIORITY; prio >= 0; prio--) {
2431                         unsigned long nr_to_scan = nr_pages - sc.nr_reclaimed;
2432
2433                         sc.nr_scanned = 0;
2434                         sc.swap_cluster_max = nr_to_scan;
2435                         shrink_all_zones(nr_to_scan, prio, pass, &sc);
2436                         if (sc.nr_reclaimed >= nr_pages)
2437                                 goto out;
2438
2439                         reclaim_state.reclaimed_slab = 0;
2440                         shrink_slab(sc.nr_scanned, sc.gfp_mask,
2441                                     global_reclaimable_pages());
2442                         sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2443                         if (sc.nr_reclaimed >= nr_pages)
2444                                 goto out;
2445
2446                         if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
2447                                 congestion_wait(BLK_RW_ASYNC, HZ / 10);
2448                 }
2449         }
2450
2451         /*
2452          * If sc.nr_reclaimed = 0, we could not shrink LRUs, but there may be
2453          * something in slab caches
2454          */
2455         if (!sc.nr_reclaimed) {
2456                 do {
2457                         reclaim_state.reclaimed_slab = 0;
2458                         shrink_slab(nr_pages, sc.gfp_mask,
2459                                     global_reclaimable_pages());
2460                         sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2461                 } while (sc.nr_reclaimed < nr_pages &&
2462                                 reclaim_state.reclaimed_slab > 0);
2463         }
2464
2465
2466 out:
2467         current->reclaim_state = NULL;
2468
2469         return sc.nr_reclaimed;
2470 }
2471 #endif /* CONFIG_HIBERNATION */
2472
2473 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2474    not required for correctness.  So if the last cpu in a node goes
2475    away, we get changed to run anywhere: as the first one comes back,
2476    restore their cpu bindings. */
2477 static int __devinit cpu_callback(struct notifier_block *nfb,
2478                                   unsigned long action, void *hcpu)
2479 {
2480         int nid;
2481
2482         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2483                 for_each_node_state(nid, N_HIGH_MEMORY) {
2484                         pg_data_t *pgdat = NODE_DATA(nid);
2485                         const struct cpumask *mask;
2486
2487                         mask = cpumask_of_node(pgdat->node_id);
2488
2489                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2490                                 /* One of our CPUs online: restore mask */
2491                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2492                 }
2493         }
2494         return NOTIFY_OK;
2495 }
2496
2497 /*
2498  * This kswapd start function will be called by init and node-hot-add.
2499  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2500  */
2501 int kswapd_run(int nid)
2502 {
2503         pg_data_t *pgdat = NODE_DATA(nid);
2504         int ret = 0;
2505
2506         if (pgdat->kswapd)
2507                 return 0;
2508
2509         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2510         if (IS_ERR(pgdat->kswapd)) {
2511                 /* failure at boot is fatal */
2512                 BUG_ON(system_state == SYSTEM_BOOTING);
2513                 printk("Failed to start kswapd on node %d\n",nid);
2514                 ret = -1;
2515         }
2516         return ret;
2517 }
2518
2519 /*
2520  * Called by memory hotplug when all memory in a node is offlined.
2521  */
2522 void kswapd_stop(int nid)
2523 {
2524         struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
2525
2526         if (kswapd)
2527                 kthread_stop(kswapd);
2528 }
2529
2530 static int __init kswapd_init(void)
2531 {
2532         int nid;
2533
2534         swap_setup();
2535         for_each_node_state(nid, N_HIGH_MEMORY)
2536                 kswapd_run(nid);
2537         hotcpu_notifier(cpu_callback, 0);
2538         return 0;
2539 }
2540
2541 module_init(kswapd_init)
2542
2543 #ifdef CONFIG_NUMA
2544 /*
2545  * Zone reclaim mode
2546  *
2547  * If non-zero call zone_reclaim when the number of free pages falls below
2548  * the watermarks.
2549  */
2550 int zone_reclaim_mode __read_mostly;
2551
2552 #define RECLAIM_OFF 0
2553 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
2554 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
2555 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
2556
2557 /*
2558  * Priority for ZONE_RECLAIM. This determines the fraction of pages
2559  * of a node considered for each zone_reclaim. 4 scans 1/16th of
2560  * a zone.
2561  */
2562 #define ZONE_RECLAIM_PRIORITY 4
2563
2564 /*
2565  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2566  * occur.
2567  */
2568 int sysctl_min_unmapped_ratio = 1;
2569
2570 /*
2571  * If the number of slab pages in a zone grows beyond this percentage then
2572  * slab reclaim needs to occur.
2573  */
2574 int sysctl_min_slab_ratio = 5;
2575
2576 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
2577 {
2578         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
2579         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
2580                 zone_page_state(zone, NR_ACTIVE_FILE);
2581
2582         /*
2583          * It's possible for there to be more file mapped pages than
2584          * accounted for by the pages on the file LRU lists because
2585          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2586          */
2587         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
2588 }
2589
2590 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
2591 static long zone_pagecache_reclaimable(struct zone *zone)
2592 {
2593         long nr_pagecache_reclaimable;
2594         long delta = 0;
2595
2596         /*
2597          * If RECLAIM_SWAP is set, then all file pages are considered
2598          * potentially reclaimable. Otherwise, we have to worry about
2599          * pages like swapcache and zone_unmapped_file_pages() provides
2600          * a better estimate
2601          */
2602         if (zone_reclaim_mode & RECLAIM_SWAP)
2603                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
2604         else
2605                 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
2606
2607         /* If we can't clean pages, remove dirty pages from consideration */
2608         if (!(zone_reclaim_mode & RECLAIM_WRITE))
2609                 delta += zone_page_state(zone, NR_FILE_DIRTY);
2610
2611         /* Watch for any possible underflows due to delta */
2612         if (unlikely(delta > nr_pagecache_reclaimable))
2613                 delta = nr_pagecache_reclaimable;
2614
2615         return nr_pagecache_reclaimable - delta;
2616 }
2617
2618 /*
2619  * Try to free up some pages from this zone through reclaim.
2620  */
2621 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2622 {
2623         /* Minimum pages needed in order to stay on node */
2624         const unsigned long nr_pages = 1 << order;
2625         struct task_struct *p = current;
2626         struct reclaim_state reclaim_state;
2627         int priority;
2628         struct scan_control sc = {
2629                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2630                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2631                 .may_swap = 1,
2632                 .swap_cluster_max = max_t(unsigned long, nr_pages,
2633                                         SWAP_CLUSTER_MAX),
2634                 .gfp_mask = gfp_mask,
2635                 .swappiness = vm_swappiness,
2636                 .order = order,
2637                 .isolate_pages = isolate_pages_global,
2638         };
2639         unsigned long slab_reclaimable;
2640
2641         disable_swap_token();
2642         cond_resched();
2643         /*
2644          * We need to be able to allocate from the reserves for RECLAIM_SWAP
2645          * and we also need to be able to write out pages for RECLAIM_WRITE
2646          * and RECLAIM_SWAP.
2647          */
2648         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2649         reclaim_state.reclaimed_slab = 0;
2650         p->reclaim_state = &reclaim_state;
2651
2652         if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
2653                 /*
2654                  * Free memory by calling shrink zone with increasing
2655                  * priorities until we have enough memory freed.
2656                  */
2657                 priority = ZONE_RECLAIM_PRIORITY;
2658                 do {
2659                         note_zone_scanning_priority(zone, priority);
2660                         shrink_zone(priority, zone, &sc);
2661                         priority--;
2662                 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
2663         }
2664
2665         slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2666         if (slab_reclaimable > zone->min_slab_pages) {
2667                 /*
2668                  * shrink_slab() does not currently allow us to determine how
2669                  * many pages were freed in this zone. So we take the current
2670                  * number of slab pages and shake the slab until it is reduced
2671                  * by the same nr_pages that we used for reclaiming unmapped
2672                  * pages.
2673                  *
2674                  * Note that shrink_slab will free memory on all zones and may
2675                  * take a long time.
2676                  */
2677                 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
2678                         zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
2679                                 slab_reclaimable - nr_pages)
2680                         ;
2681
2682                 /*
2683                  * Update nr_reclaimed by the number of slab pages we
2684                  * reclaimed from this zone.
2685                  */
2686                 sc.nr_reclaimed += slab_reclaimable -
2687                         zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2688         }
2689
2690         p->reclaim_state = NULL;
2691         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
2692         return sc.nr_reclaimed >= nr_pages;
2693 }
2694
2695 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2696 {
2697         int node_id;
2698         int ret;
2699
2700         /*
2701          * Zone reclaim reclaims unmapped file backed pages and
2702          * slab pages if we are over the defined limits.
2703          *
2704          * A small portion of unmapped file backed pages is needed for
2705          * file I/O otherwise pages read by file I/O will be immediately
2706          * thrown out if the zone is overallocated. So we do not reclaim
2707          * if less than a specified percentage of the zone is used by
2708          * unmapped file backed pages.
2709          */
2710         if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
2711             zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
2712                 return ZONE_RECLAIM_FULL;
2713
2714         if (zone_is_all_unreclaimable(zone))
2715                 return ZONE_RECLAIM_FULL;
2716
2717         /*
2718          * Do not scan if the allocation should not be delayed.
2719          */
2720         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
2721                 return ZONE_RECLAIM_NOSCAN;
2722
2723         /*
2724          * Only run zone reclaim on the local zone or on zones that do not
2725          * have associated processors. This will favor the local processor
2726          * over remote processors and spread off node memory allocations
2727          * as wide as possible.
2728          */
2729         node_id = zone_to_nid(zone);
2730         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
2731                 return ZONE_RECLAIM_NOSCAN;
2732
2733         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2734                 return ZONE_RECLAIM_NOSCAN;
2735
2736         ret = __zone_reclaim(zone, gfp_mask, order);
2737         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2738
2739         if (!ret)
2740                 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
2741
2742         return ret;
2743 }
2744 #endif
2745
2746 /*
2747  * page_evictable - test whether a page is evictable
2748  * @page: the page to test
2749  * @vma: the VMA in which the page is or will be mapped, may be NULL
2750  *
2751  * Test whether page is evictable--i.e., should be placed on active/inactive
2752  * lists vs unevictable list.  The vma argument is !NULL when called from the
2753  * fault path to determine how to instantate a new page.
2754  *
2755  * Reasons page might not be evictable:
2756  * (1) page's mapping marked unevictable
2757  * (2) page is part of an mlocked VMA
2758  *
2759  */
2760 int page_evictable(struct page *page, struct vm_area_struct *vma)
2761 {
2762
2763         if (mapping_unevictable(page_mapping(page)))
2764                 return 0;
2765
2766         if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2767                 return 0;
2768
2769         return 1;
2770 }
2771
2772 /**
2773  * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2774  * @page: page to check evictability and move to appropriate lru list
2775  * @zone: zone page is in
2776  *
2777  * Checks a page for evictability and moves the page to the appropriate
2778  * zone lru list.
2779  *
2780  * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2781  * have PageUnevictable set.
2782  */
2783 static void check_move_unevictable_page(struct page *page, struct zone *zone)
2784 {
2785         VM_BUG_ON(PageActive(page));
2786
2787 retry:
2788         ClearPageUnevictable(page);
2789         if (page_evictable(page, NULL)) {
2790                 enum lru_list l = page_lru_base_type(page);
2791
2792                 __dec_zone_state(zone, NR_UNEVICTABLE);
2793                 list_move(&page->lru, &zone->lru[l].list);
2794                 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
2795                 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
2796                 __count_vm_event(UNEVICTABLE_PGRESCUED);
2797         } else {
2798                 /*
2799                  * rotate unevictable list
2800                  */
2801                 SetPageUnevictable(page);
2802                 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
2803                 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
2804                 if (page_evictable(page, NULL))
2805                         goto retry;
2806         }
2807 }
2808
2809 /**
2810  * scan_mapping_unevictable_pages - scan an address space for evictable pages
2811  * @mapping: struct address_space to scan for evictable pages
2812  *
2813  * Scan all pages in mapping.  Check unevictable pages for
2814  * evictability and move them to the appropriate zone lru list.
2815  */
2816 void scan_mapping_unevictable_pages(struct address_space *mapping)
2817 {
2818         pgoff_t next = 0;
2819         pgoff_t end   = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
2820                          PAGE_CACHE_SHIFT;
2821         struct zone *zone;
2822         struct pagevec pvec;
2823
2824         if (mapping->nrpages == 0)
2825                 return;
2826
2827         pagevec_init(&pvec, 0);
2828         while (next < end &&
2829                 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
2830                 int i;
2831                 int pg_scanned = 0;
2832
2833                 zone = NULL;
2834
2835                 for (i = 0; i < pagevec_count(&pvec); i++) {
2836                         struct page *page = pvec.pages[i];
2837                         pgoff_t page_index = page->index;
2838                         struct zone *pagezone = page_zone(page);
2839
2840                         pg_scanned++;
2841                         if (page_index > next)
2842                                 next = page_index;
2843                         next++;
2844
2845                         if (pagezone != zone) {
2846                                 if (zone)
2847                                         spin_unlock_irq(&zone->lru_lock);
2848                                 zone = pagezone;
2849                                 spin_lock_irq(&zone->lru_lock);
2850                         }
2851
2852                         if (PageLRU(page) && PageUnevictable(page))
2853                                 check_move_unevictable_page(page, zone);
2854                 }
2855                 if (zone)
2856                         spin_unlock_irq(&zone->lru_lock);
2857                 pagevec_release(&pvec);
2858
2859                 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
2860         }
2861
2862 }
2863
2864 /**
2865  * scan_zone_unevictable_pages - check unevictable list for evictable pages
2866  * @zone - zone of which to scan the unevictable list
2867  *
2868  * Scan @zone's unevictable LRU lists to check for pages that have become
2869  * evictable.  Move those that have to @zone's inactive list where they
2870  * become candidates for reclaim, unless shrink_inactive_zone() decides
2871  * to reactivate them.  Pages that are still unevictable are rotated
2872  * back onto @zone's unevictable list.
2873  */
2874 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
2875 static void scan_zone_unevictable_pages(struct zone *zone)
2876 {
2877         struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
2878         unsigned long scan;
2879         unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
2880
2881         while (nr_to_scan > 0) {
2882                 unsigned long batch_size = min(nr_to_scan,
2883                                                 SCAN_UNEVICTABLE_BATCH_SIZE);
2884
2885                 spin_lock_irq(&zone->lru_lock);
2886                 for (scan = 0;  scan < batch_size; scan++) {
2887                         struct page *page = lru_to_page(l_unevictable);
2888
2889                         if (!trylock_page(page))
2890                                 continue;
2891
2892                         prefetchw_prev_lru_page(page, l_unevictable, flags);
2893
2894                         if (likely(PageLRU(page) && PageUnevictable(page)))
2895                                 check_move_unevictable_page(page, zone);
2896
2897                         unlock_page(page);
2898                 }
2899                 spin_unlock_irq(&zone->lru_lock);
2900
2901                 nr_to_scan -= batch_size;
2902         }
2903 }
2904
2905
2906 /**
2907  * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
2908  *
2909  * A really big hammer:  scan all zones' unevictable LRU lists to check for
2910  * pages that have become evictable.  Move those back to the zones'
2911  * inactive list where they become candidates for reclaim.
2912  * This occurs when, e.g., we have unswappable pages on the unevictable lists,
2913  * and we add swap to the system.  As such, it runs in the context of a task
2914  * that has possibly/probably made some previously unevictable pages
2915  * evictable.
2916  */
2917 static void scan_all_zones_unevictable_pages(void)
2918 {
2919         struct zone *zone;
2920
2921         for_each_zone(zone) {
2922                 scan_zone_unevictable_pages(zone);
2923         }
2924 }
2925
2926 /*
2927  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
2928  * all nodes' unevictable lists for evictable pages
2929  */
2930 unsigned long scan_unevictable_pages;
2931
2932 int scan_unevictable_handler(struct ctl_table *table, int write,
2933                            void __user *buffer,
2934                            size_t *length, loff_t *ppos)
2935 {
2936         proc_doulongvec_minmax(table, write, buffer, length, ppos);
2937
2938         if (write && *(unsigned long *)table->data)
2939                 scan_all_zones_unevictable_pages();
2940
2941         scan_unevictable_pages = 0;
2942         return 0;
2943 }
2944
2945 /*
2946  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
2947  * a specified node's per zone unevictable lists for evictable pages.
2948  */
2949
2950 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
2951                                           struct sysdev_attribute *attr,
2952                                           char *buf)
2953 {
2954         return sprintf(buf, "0\n");     /* always zero; should fit... */
2955 }
2956
2957 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
2958                                            struct sysdev_attribute *attr,
2959                                         const char *buf, size_t count)
2960 {
2961         struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
2962         struct zone *zone;
2963         unsigned long res;
2964         unsigned long req = strict_strtoul(buf, 10, &res);
2965
2966         if (!req)
2967                 return 1;       /* zero is no-op */
2968
2969         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2970                 if (!populated_zone(zone))
2971                         continue;
2972                 scan_zone_unevictable_pages(zone);
2973         }
2974         return 1;
2975 }
2976
2977
2978 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
2979                         read_scan_unevictable_node,
2980                         write_scan_unevictable_node);
2981
2982 int scan_unevictable_register_node(struct node *node)
2983 {
2984         return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
2985 }
2986
2987 void scan_unevictable_unregister_node(struct node *node)
2988 {
2989         sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
2990 }
2991