mm: do batched scans for mem_cgroup
[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_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          * See swapfile.c:page_queue_congested().
370          */
371         if (!is_page_cache_freeable(page))
372                 return PAGE_KEEP;
373         if (!mapping) {
374                 /*
375                  * Some data journaling orphaned pages can have
376                  * page->mapping == NULL while being dirty with clean buffers.
377                  */
378                 if (page_has_private(page)) {
379                         if (try_to_free_buffers(page)) {
380                                 ClearPageDirty(page);
381                                 printk("%s: orphaned page\n", __func__);
382                                 return PAGE_CLEAN;
383                         }
384                 }
385                 return PAGE_KEEP;
386         }
387         if (mapping->a_ops->writepage == NULL)
388                 return PAGE_ACTIVATE;
389         if (!may_write_to_queue(mapping->backing_dev_info))
390                 return PAGE_KEEP;
391
392         if (clear_page_dirty_for_io(page)) {
393                 int res;
394                 struct writeback_control wbc = {
395                         .sync_mode = WB_SYNC_NONE,
396                         .nr_to_write = SWAP_CLUSTER_MAX,
397                         .range_start = 0,
398                         .range_end = LLONG_MAX,
399                         .nonblocking = 1,
400                         .for_reclaim = 1,
401                 };
402
403                 SetPageReclaim(page);
404                 res = mapping->a_ops->writepage(page, &wbc);
405                 if (res < 0)
406                         handle_write_error(mapping, page, res);
407                 if (res == AOP_WRITEPAGE_ACTIVATE) {
408                         ClearPageReclaim(page);
409                         return PAGE_ACTIVATE;
410                 }
411
412                 /*
413                  * Wait on writeback if requested to. This happens when
414                  * direct reclaiming a large contiguous area and the
415                  * first attempt to free a range of pages fails.
416                  */
417                 if (PageWriteback(page) && sync_writeback == PAGEOUT_IO_SYNC)
418                         wait_on_page_writeback(page);
419
420                 if (!PageWriteback(page)) {
421                         /* synchronous write or broken a_ops? */
422                         ClearPageReclaim(page);
423                 }
424                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
425                 return PAGE_SUCCESS;
426         }
427
428         return PAGE_CLEAN;
429 }
430
431 /*
432  * Same as remove_mapping, but if the page is removed from the mapping, it
433  * gets returned with a refcount of 0.
434  */
435 static int __remove_mapping(struct address_space *mapping, struct page *page)
436 {
437         BUG_ON(!PageLocked(page));
438         BUG_ON(mapping != page_mapping(page));
439
440         spin_lock_irq(&mapping->tree_lock);
441         /*
442          * The non racy check for a busy page.
443          *
444          * Must be careful with the order of the tests. When someone has
445          * a ref to the page, it may be possible that they dirty it then
446          * drop the reference. So if PageDirty is tested before page_count
447          * here, then the following race may occur:
448          *
449          * get_user_pages(&page);
450          * [user mapping goes away]
451          * write_to(page);
452          *                              !PageDirty(page)    [good]
453          * SetPageDirty(page);
454          * put_page(page);
455          *                              !page_count(page)   [good, discard it]
456          *
457          * [oops, our write_to data is lost]
458          *
459          * Reversing the order of the tests ensures such a situation cannot
460          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
461          * load is not satisfied before that of page->_count.
462          *
463          * Note that if SetPageDirty is always performed via set_page_dirty,
464          * and thus under tree_lock, then this ordering is not required.
465          */
466         if (!page_freeze_refs(page, 2))
467                 goto cannot_free;
468         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
469         if (unlikely(PageDirty(page))) {
470                 page_unfreeze_refs(page, 2);
471                 goto cannot_free;
472         }
473
474         if (PageSwapCache(page)) {
475                 swp_entry_t swap = { .val = page_private(page) };
476                 __delete_from_swap_cache(page);
477                 spin_unlock_irq(&mapping->tree_lock);
478                 swapcache_free(swap, page);
479         } else {
480                 __remove_from_page_cache(page);
481                 spin_unlock_irq(&mapping->tree_lock);
482                 mem_cgroup_uncharge_cache_page(page);
483         }
484
485         return 1;
486
487 cannot_free:
488         spin_unlock_irq(&mapping->tree_lock);
489         return 0;
490 }
491
492 /*
493  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
494  * someone else has a ref on the page, abort and return 0.  If it was
495  * successfully detached, return 1.  Assumes the caller has a single ref on
496  * this page.
497  */
498 int remove_mapping(struct address_space *mapping, struct page *page)
499 {
500         if (__remove_mapping(mapping, page)) {
501                 /*
502                  * Unfreezing the refcount with 1 rather than 2 effectively
503                  * drops the pagecache ref for us without requiring another
504                  * atomic operation.
505                  */
506                 page_unfreeze_refs(page, 1);
507                 return 1;
508         }
509         return 0;
510 }
511
512 /**
513  * putback_lru_page - put previously isolated page onto appropriate LRU list
514  * @page: page to be put back to appropriate lru list
515  *
516  * Add previously isolated @page to appropriate LRU list.
517  * Page may still be unevictable for other reasons.
518  *
519  * lru_lock must not be held, interrupts must be enabled.
520  */
521 void putback_lru_page(struct page *page)
522 {
523         int lru;
524         int active = !!TestClearPageActive(page);
525         int was_unevictable = PageUnevictable(page);
526
527         VM_BUG_ON(PageLRU(page));
528
529 redo:
530         ClearPageUnevictable(page);
531
532         if (page_evictable(page, NULL)) {
533                 /*
534                  * For evictable pages, we can use the cache.
535                  * In event of a race, worst case is we end up with an
536                  * unevictable page on [in]active list.
537                  * We know how to handle that.
538                  */
539                 lru = active + page_lru_base_type(page);
540                 lru_cache_add_lru(page, lru);
541         } else {
542                 /*
543                  * Put unevictable pages directly on zone's unevictable
544                  * list.
545                  */
546                 lru = LRU_UNEVICTABLE;
547                 add_page_to_unevictable_list(page);
548         }
549
550         /*
551          * page's status can change while we move it among lru. If an evictable
552          * page is on unevictable list, it never be freed. To avoid that,
553          * check after we added it to the list, again.
554          */
555         if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
556                 if (!isolate_lru_page(page)) {
557                         put_page(page);
558                         goto redo;
559                 }
560                 /* This means someone else dropped this page from LRU
561                  * So, it will be freed or putback to LRU again. There is
562                  * nothing to do here.
563                  */
564         }
565
566         if (was_unevictable && lru != LRU_UNEVICTABLE)
567                 count_vm_event(UNEVICTABLE_PGRESCUED);
568         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
569                 count_vm_event(UNEVICTABLE_PGCULLED);
570
571         put_page(page);         /* drop ref from isolate */
572 }
573
574 /*
575  * shrink_page_list() returns the number of reclaimed pages
576  */
577 static unsigned long shrink_page_list(struct list_head *page_list,
578                                         struct scan_control *sc,
579                                         enum pageout_io sync_writeback)
580 {
581         LIST_HEAD(ret_pages);
582         struct pagevec freed_pvec;
583         int pgactivate = 0;
584         unsigned long nr_reclaimed = 0;
585         unsigned long vm_flags;
586
587         cond_resched();
588
589         pagevec_init(&freed_pvec, 1);
590         while (!list_empty(page_list)) {
591                 struct address_space *mapping;
592                 struct page *page;
593                 int may_enter_fs;
594                 int referenced;
595
596                 cond_resched();
597
598                 page = lru_to_page(page_list);
599                 list_del(&page->lru);
600
601                 if (!trylock_page(page))
602                         goto keep;
603
604                 VM_BUG_ON(PageActive(page));
605
606                 sc->nr_scanned++;
607
608                 if (unlikely(!page_evictable(page, NULL)))
609                         goto cull_mlocked;
610
611                 if (!sc->may_unmap && page_mapped(page))
612                         goto keep_locked;
613
614                 /* Double the slab pressure for mapped and swapcache pages */
615                 if (page_mapped(page) || PageSwapCache(page))
616                         sc->nr_scanned++;
617
618                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
619                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
620
621                 if (PageWriteback(page)) {
622                         /*
623                          * Synchronous reclaim is performed in two passes,
624                          * first an asynchronous pass over the list to
625                          * start parallel writeback, and a second synchronous
626                          * pass to wait for the IO to complete.  Wait here
627                          * for any page for which writeback has already
628                          * started.
629                          */
630                         if (sync_writeback == PAGEOUT_IO_SYNC && may_enter_fs)
631                                 wait_on_page_writeback(page);
632                         else
633                                 goto keep_locked;
634                 }
635
636                 referenced = page_referenced(page, 1,
637                                                 sc->mem_cgroup, &vm_flags);
638                 /*
639                  * In active use or really unfreeable?  Activate it.
640                  * If page which have PG_mlocked lost isoltation race,
641                  * try_to_unmap moves it to unevictable list
642                  */
643                 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER &&
644                                         referenced && page_mapping_inuse(page)
645                                         && !(vm_flags & VM_LOCKED))
646                         goto activate_locked;
647
648                 /*
649                  * Anonymous process memory has backing store?
650                  * Try to allocate it some swap space here.
651                  */
652                 if (PageAnon(page) && !PageSwapCache(page)) {
653                         if (!(sc->gfp_mask & __GFP_IO))
654                                 goto keep_locked;
655                         if (!add_to_swap(page))
656                                 goto activate_locked;
657                         may_enter_fs = 1;
658                 }
659
660                 mapping = page_mapping(page);
661
662                 /*
663                  * The page is mapped into the page tables of one or more
664                  * processes. Try to unmap it here.
665                  */
666                 if (page_mapped(page) && mapping) {
667                         switch (try_to_unmap(page, 0)) {
668                         case SWAP_FAIL:
669                                 goto activate_locked;
670                         case SWAP_AGAIN:
671                                 goto keep_locked;
672                         case SWAP_MLOCK:
673                                 goto cull_mlocked;
674                         case SWAP_SUCCESS:
675                                 ; /* try to free the page below */
676                         }
677                 }
678
679                 if (PageDirty(page)) {
680                         if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && referenced)
681                                 goto keep_locked;
682                         if (!may_enter_fs)
683                                 goto keep_locked;
684                         if (!sc->may_writepage)
685                                 goto keep_locked;
686
687                         /* Page is dirty, try to write it out here */
688                         switch (pageout(page, mapping, sync_writeback)) {
689                         case PAGE_KEEP:
690                                 goto keep_locked;
691                         case PAGE_ACTIVATE:
692                                 goto activate_locked;
693                         case PAGE_SUCCESS:
694                                 if (PageWriteback(page) || PageDirty(page))
695                                         goto keep;
696                                 /*
697                                  * A synchronous write - probably a ramdisk.  Go
698                                  * ahead and try to reclaim the page.
699                                  */
700                                 if (!trylock_page(page))
701                                         goto keep;
702                                 if (PageDirty(page) || PageWriteback(page))
703                                         goto keep_locked;
704                                 mapping = page_mapping(page);
705                         case PAGE_CLEAN:
706                                 ; /* try to free the page below */
707                         }
708                 }
709
710                 /*
711                  * If the page has buffers, try to free the buffer mappings
712                  * associated with this page. If we succeed we try to free
713                  * the page as well.
714                  *
715                  * We do this even if the page is PageDirty().
716                  * try_to_release_page() does not perform I/O, but it is
717                  * possible for a page to have PageDirty set, but it is actually
718                  * clean (all its buffers are clean).  This happens if the
719                  * buffers were written out directly, with submit_bh(). ext3
720                  * will do this, as well as the blockdev mapping.
721                  * try_to_release_page() will discover that cleanness and will
722                  * drop the buffers and mark the page clean - it can be freed.
723                  *
724                  * Rarely, pages can have buffers and no ->mapping.  These are
725                  * the pages which were not successfully invalidated in
726                  * truncate_complete_page().  We try to drop those buffers here
727                  * and if that worked, and the page is no longer mapped into
728                  * process address space (page_count == 1) it can be freed.
729                  * Otherwise, leave the page on the LRU so it is swappable.
730                  */
731                 if (page_has_private(page)) {
732                         if (!try_to_release_page(page, sc->gfp_mask))
733                                 goto activate_locked;
734                         if (!mapping && page_count(page) == 1) {
735                                 unlock_page(page);
736                                 if (put_page_testzero(page))
737                                         goto free_it;
738                                 else {
739                                         /*
740                                          * rare race with speculative reference.
741                                          * the speculative reference will free
742                                          * this page shortly, so we may
743                                          * increment nr_reclaimed here (and
744                                          * leave it off the LRU).
745                                          */
746                                         nr_reclaimed++;
747                                         continue;
748                                 }
749                         }
750                 }
751
752                 if (!mapping || !__remove_mapping(mapping, page))
753                         goto keep_locked;
754
755                 /*
756                  * At this point, we have no other references and there is
757                  * no way to pick any more up (removed from LRU, removed
758                  * from pagecache). Can use non-atomic bitops now (and
759                  * we obviously don't have to worry about waking up a process
760                  * waiting on the page lock, because there are no references.
761                  */
762                 __clear_page_locked(page);
763 free_it:
764                 nr_reclaimed++;
765                 if (!pagevec_add(&freed_pvec, page)) {
766                         __pagevec_free(&freed_pvec);
767                         pagevec_reinit(&freed_pvec);
768                 }
769                 continue;
770
771 cull_mlocked:
772                 if (PageSwapCache(page))
773                         try_to_free_swap(page);
774                 unlock_page(page);
775                 putback_lru_page(page);
776                 continue;
777
778 activate_locked:
779                 /* Not a candidate for swapping, so reclaim swap space. */
780                 if (PageSwapCache(page) && vm_swap_full())
781                         try_to_free_swap(page);
782                 VM_BUG_ON(PageActive(page));
783                 SetPageActive(page);
784                 pgactivate++;
785 keep_locked:
786                 unlock_page(page);
787 keep:
788                 list_add(&page->lru, &ret_pages);
789                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
790         }
791         list_splice(&ret_pages, page_list);
792         if (pagevec_count(&freed_pvec))
793                 __pagevec_free(&freed_pvec);
794         count_vm_events(PGACTIVATE, pgactivate);
795         return nr_reclaimed;
796 }
797
798 /* LRU Isolation modes. */
799 #define ISOLATE_INACTIVE 0      /* Isolate inactive pages. */
800 #define ISOLATE_ACTIVE 1        /* Isolate active pages. */
801 #define ISOLATE_BOTH 2          /* Isolate both active and inactive pages. */
802
803 /*
804  * Attempt to remove the specified page from its LRU.  Only take this page
805  * if it is of the appropriate PageActive status.  Pages which are being
806  * freed elsewhere are also ignored.
807  *
808  * page:        page to consider
809  * mode:        one of the LRU isolation modes defined above
810  *
811  * returns 0 on success, -ve errno on failure.
812  */
813 int __isolate_lru_page(struct page *page, int mode, int file)
814 {
815         int ret = -EINVAL;
816
817         /* Only take pages on the LRU. */
818         if (!PageLRU(page))
819                 return ret;
820
821         /*
822          * When checking the active state, we need to be sure we are
823          * dealing with comparible boolean values.  Take the logical not
824          * of each.
825          */
826         if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
827                 return ret;
828
829         if (mode != ISOLATE_BOTH && page_is_file_cache(page) != file)
830                 return ret;
831
832         /*
833          * When this function is being called for lumpy reclaim, we
834          * initially look into all LRU pages, active, inactive and
835          * unevictable; only give shrink_page_list evictable pages.
836          */
837         if (PageUnevictable(page))
838                 return ret;
839
840         ret = -EBUSY;
841
842         if (likely(get_page_unless_zero(page))) {
843                 /*
844                  * Be careful not to clear PageLRU until after we're
845                  * sure the page is not being freed elsewhere -- the
846                  * page release code relies on it.
847                  */
848                 ClearPageLRU(page);
849                 ret = 0;
850         }
851
852         return ret;
853 }
854
855 /*
856  * zone->lru_lock is heavily contended.  Some of the functions that
857  * shrink the lists perform better by taking out a batch of pages
858  * and working on them outside the LRU lock.
859  *
860  * For pagecache intensive workloads, this function is the hottest
861  * spot in the kernel (apart from copy_*_user functions).
862  *
863  * Appropriate locks must be held before calling this function.
864  *
865  * @nr_to_scan: The number of pages to look through on the list.
866  * @src:        The LRU list to pull pages off.
867  * @dst:        The temp list to put pages on to.
868  * @scanned:    The number of pages that were scanned.
869  * @order:      The caller's attempted allocation order
870  * @mode:       One of the LRU isolation modes
871  * @file:       True [1] if isolating file [!anon] pages
872  *
873  * returns how many pages were moved onto *@dst.
874  */
875 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
876                 struct list_head *src, struct list_head *dst,
877                 unsigned long *scanned, int order, int mode, int file)
878 {
879         unsigned long nr_taken = 0;
880         unsigned long scan;
881
882         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
883                 struct page *page;
884                 unsigned long pfn;
885                 unsigned long end_pfn;
886                 unsigned long page_pfn;
887                 int zone_id;
888
889                 page = lru_to_page(src);
890                 prefetchw_prev_lru_page(page, src, flags);
891
892                 VM_BUG_ON(!PageLRU(page));
893
894                 switch (__isolate_lru_page(page, mode, file)) {
895                 case 0:
896                         list_move(&page->lru, dst);
897                         mem_cgroup_del_lru(page);
898                         nr_taken++;
899                         break;
900
901                 case -EBUSY:
902                         /* else it is being freed elsewhere */
903                         list_move(&page->lru, src);
904                         mem_cgroup_rotate_lru_list(page, page_lru(page));
905                         continue;
906
907                 default:
908                         BUG();
909                 }
910
911                 if (!order)
912                         continue;
913
914                 /*
915                  * Attempt to take all pages in the order aligned region
916                  * surrounding the tag page.  Only take those pages of
917                  * the same active state as that tag page.  We may safely
918                  * round the target page pfn down to the requested order
919                  * as the mem_map is guarenteed valid out to MAX_ORDER,
920                  * where that page is in a different zone we will detect
921                  * it from its zone id and abort this block scan.
922                  */
923                 zone_id = page_zone_id(page);
924                 page_pfn = page_to_pfn(page);
925                 pfn = page_pfn & ~((1 << order) - 1);
926                 end_pfn = pfn + (1 << order);
927                 for (; pfn < end_pfn; pfn++) {
928                         struct page *cursor_page;
929
930                         /* The target page is in the block, ignore it. */
931                         if (unlikely(pfn == page_pfn))
932                                 continue;
933
934                         /* Avoid holes within the zone. */
935                         if (unlikely(!pfn_valid_within(pfn)))
936                                 break;
937
938                         cursor_page = pfn_to_page(pfn);
939
940                         /* Check that we have not crossed a zone boundary. */
941                         if (unlikely(page_zone_id(cursor_page) != zone_id))
942                                 continue;
943
944                         /*
945                          * If we don't have enough swap space, reclaiming of
946                          * anon page which don't already have a swap slot is
947                          * pointless.
948                          */
949                         if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
950                                         !PageSwapCache(cursor_page))
951                                 continue;
952
953                         if (__isolate_lru_page(cursor_page, mode, file) == 0) {
954                                 list_move(&cursor_page->lru, dst);
955                                 mem_cgroup_del_lru(cursor_page);
956                                 nr_taken++;
957                                 scan++;
958                         }
959                 }
960         }
961
962         *scanned = scan;
963         return nr_taken;
964 }
965
966 static unsigned long isolate_pages_global(unsigned long nr,
967                                         struct list_head *dst,
968                                         unsigned long *scanned, int order,
969                                         int mode, struct zone *z,
970                                         struct mem_cgroup *mem_cont,
971                                         int active, int file)
972 {
973         int lru = LRU_BASE;
974         if (active)
975                 lru += LRU_ACTIVE;
976         if (file)
977                 lru += LRU_FILE;
978         return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
979                                                                 mode, file);
980 }
981
982 /*
983  * clear_active_flags() is a helper for shrink_active_list(), clearing
984  * any active bits from the pages in the list.
985  */
986 static unsigned long clear_active_flags(struct list_head *page_list,
987                                         unsigned int *count)
988 {
989         int nr_active = 0;
990         int lru;
991         struct page *page;
992
993         list_for_each_entry(page, page_list, lru) {
994                 lru = page_lru_base_type(page);
995                 if (PageActive(page)) {
996                         lru += LRU_ACTIVE;
997                         ClearPageActive(page);
998                         nr_active++;
999                 }
1000                 count[lru]++;
1001         }
1002
1003         return nr_active;
1004 }
1005
1006 /**
1007  * isolate_lru_page - tries to isolate a page from its LRU list
1008  * @page: page to isolate from its LRU list
1009  *
1010  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1011  * vmstat statistic corresponding to whatever LRU list the page was on.
1012  *
1013  * Returns 0 if the page was removed from an LRU list.
1014  * Returns -EBUSY if the page was not on an LRU list.
1015  *
1016  * The returned page will have PageLRU() cleared.  If it was found on
1017  * the active list, it will have PageActive set.  If it was found on
1018  * the unevictable list, it will have the PageUnevictable bit set. That flag
1019  * may need to be cleared by the caller before letting the page go.
1020  *
1021  * The vmstat statistic corresponding to the list on which the page was
1022  * found will be decremented.
1023  *
1024  * Restrictions:
1025  * (1) Must be called with an elevated refcount on the page. This is a
1026  *     fundamentnal difference from isolate_lru_pages (which is called
1027  *     without a stable reference).
1028  * (2) the lru_lock must not be held.
1029  * (3) interrupts must be enabled.
1030  */
1031 int isolate_lru_page(struct page *page)
1032 {
1033         int ret = -EBUSY;
1034
1035         if (PageLRU(page)) {
1036                 struct zone *zone = page_zone(page);
1037
1038                 spin_lock_irq(&zone->lru_lock);
1039                 if (PageLRU(page) && get_page_unless_zero(page)) {
1040                         int lru = page_lru(page);
1041                         ret = 0;
1042                         ClearPageLRU(page);
1043
1044                         del_page_from_lru_list(zone, page, lru);
1045                 }
1046                 spin_unlock_irq(&zone->lru_lock);
1047         }
1048         return ret;
1049 }
1050
1051 /*
1052  * Are there way too many processes in the direct reclaim path already?
1053  */
1054 static int too_many_isolated(struct zone *zone, int file,
1055                 struct scan_control *sc)
1056 {
1057         unsigned long inactive, isolated;
1058
1059         if (current_is_kswapd())
1060                 return 0;
1061
1062         if (!scanning_global_lru(sc))
1063                 return 0;
1064
1065         if (file) {
1066                 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1067                 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1068         } else {
1069                 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1070                 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1071         }
1072
1073         return isolated > inactive;
1074 }
1075
1076 /*
1077  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1078  * of reclaimed pages
1079  */
1080 static unsigned long shrink_inactive_list(unsigned long max_scan,
1081                         struct zone *zone, struct scan_control *sc,
1082                         int priority, int file)
1083 {
1084         LIST_HEAD(page_list);
1085         struct pagevec pvec;
1086         unsigned long nr_scanned = 0;
1087         unsigned long nr_reclaimed = 0;
1088         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1089         int lumpy_reclaim = 0;
1090
1091         while (unlikely(too_many_isolated(zone, file, sc))) {
1092                 congestion_wait(WRITE, HZ/10);
1093
1094                 /* We are about to die and free our memory. Return now. */
1095                 if (fatal_signal_pending(current))
1096                         return SWAP_CLUSTER_MAX;
1097         }
1098
1099         /*
1100          * If we need a large contiguous chunk of memory, or have
1101          * trouble getting a small set of contiguous pages, we
1102          * will reclaim both active and inactive pages.
1103          *
1104          * We use the same threshold as pageout congestion_wait below.
1105          */
1106         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1107                 lumpy_reclaim = 1;
1108         else if (sc->order && priority < DEF_PRIORITY - 2)
1109                 lumpy_reclaim = 1;
1110
1111         pagevec_init(&pvec, 1);
1112
1113         lru_add_drain();
1114         spin_lock_irq(&zone->lru_lock);
1115         do {
1116                 struct page *page;
1117                 unsigned long nr_taken;
1118                 unsigned long nr_scan;
1119                 unsigned long nr_freed;
1120                 unsigned long nr_active;
1121                 unsigned int count[NR_LRU_LISTS] = { 0, };
1122                 int mode = lumpy_reclaim ? ISOLATE_BOTH : ISOLATE_INACTIVE;
1123                 unsigned long nr_anon;
1124                 unsigned long nr_file;
1125
1126                 nr_taken = sc->isolate_pages(sc->swap_cluster_max,
1127                              &page_list, &nr_scan, sc->order, mode,
1128                                 zone, sc->mem_cgroup, 0, file);
1129
1130                 if (scanning_global_lru(sc)) {
1131                         zone->pages_scanned += nr_scan;
1132                         if (current_is_kswapd())
1133                                 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1134                                                        nr_scan);
1135                         else
1136                                 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1137                                                        nr_scan);
1138                 }
1139
1140                 if (nr_taken == 0)
1141                         goto done;
1142
1143                 nr_active = clear_active_flags(&page_list, count);
1144                 __count_vm_events(PGDEACTIVATE, nr_active);
1145
1146                 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1147                                                 -count[LRU_ACTIVE_FILE]);
1148                 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1149                                                 -count[LRU_INACTIVE_FILE]);
1150                 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1151                                                 -count[LRU_ACTIVE_ANON]);
1152                 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1153                                                 -count[LRU_INACTIVE_ANON]);
1154
1155                 nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1156                 nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1157                 __mod_zone_page_state(zone, NR_ISOLATED_ANON, nr_anon);
1158                 __mod_zone_page_state(zone, NR_ISOLATED_FILE, nr_file);
1159
1160                 reclaim_stat->recent_scanned[0] += count[LRU_INACTIVE_ANON];
1161                 reclaim_stat->recent_scanned[0] += count[LRU_ACTIVE_ANON];
1162                 reclaim_stat->recent_scanned[1] += count[LRU_INACTIVE_FILE];
1163                 reclaim_stat->recent_scanned[1] += count[LRU_ACTIVE_FILE];
1164
1165                 spin_unlock_irq(&zone->lru_lock);
1166
1167                 nr_scanned += nr_scan;
1168                 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
1169
1170                 /*
1171                  * If we are direct reclaiming for contiguous pages and we do
1172                  * not reclaim everything in the list, try again and wait
1173                  * for IO to complete. This will stall high-order allocations
1174                  * but that should be acceptable to the caller
1175                  */
1176                 if (nr_freed < nr_taken && !current_is_kswapd() &&
1177                     lumpy_reclaim) {
1178                         congestion_wait(BLK_RW_ASYNC, HZ/10);
1179
1180                         /*
1181                          * The attempt at page out may have made some
1182                          * of the pages active, mark them inactive again.
1183                          */
1184                         nr_active = clear_active_flags(&page_list, count);
1185                         count_vm_events(PGDEACTIVATE, nr_active);
1186
1187                         nr_freed += shrink_page_list(&page_list, sc,
1188                                                         PAGEOUT_IO_SYNC);
1189                 }
1190
1191                 nr_reclaimed += nr_freed;
1192
1193                 local_irq_disable();
1194                 if (current_is_kswapd())
1195                         __count_vm_events(KSWAPD_STEAL, nr_freed);
1196                 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
1197
1198                 spin_lock(&zone->lru_lock);
1199                 /*
1200                  * Put back any unfreeable pages.
1201                  */
1202                 while (!list_empty(&page_list)) {
1203                         int lru;
1204                         page = lru_to_page(&page_list);
1205                         VM_BUG_ON(PageLRU(page));
1206                         list_del(&page->lru);
1207                         if (unlikely(!page_evictable(page, NULL))) {
1208                                 spin_unlock_irq(&zone->lru_lock);
1209                                 putback_lru_page(page);
1210                                 spin_lock_irq(&zone->lru_lock);
1211                                 continue;
1212                         }
1213                         SetPageLRU(page);
1214                         lru = page_lru(page);
1215                         add_page_to_lru_list(zone, page, lru);
1216                         if (is_active_lru(lru)) {
1217                                 int file = is_file_lru(lru);
1218                                 reclaim_stat->recent_rotated[file]++;
1219                         }
1220                         if (!pagevec_add(&pvec, page)) {
1221                                 spin_unlock_irq(&zone->lru_lock);
1222                                 __pagevec_release(&pvec);
1223                                 spin_lock_irq(&zone->lru_lock);
1224                         }
1225                 }
1226                 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1227                 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1228
1229         } while (nr_scanned < max_scan);
1230
1231 done:
1232         spin_unlock_irq(&zone->lru_lock);
1233         pagevec_release(&pvec);
1234         return nr_reclaimed;
1235 }
1236
1237 /*
1238  * We are about to scan this zone at a certain priority level.  If that priority
1239  * level is smaller (ie: more urgent) than the previous priority, then note
1240  * that priority level within the zone.  This is done so that when the next
1241  * process comes in to scan this zone, it will immediately start out at this
1242  * priority level rather than having to build up its own scanning priority.
1243  * Here, this priority affects only the reclaim-mapped threshold.
1244  */
1245 static inline void note_zone_scanning_priority(struct zone *zone, int priority)
1246 {
1247         if (priority < zone->prev_priority)
1248                 zone->prev_priority = priority;
1249 }
1250
1251 /*
1252  * This moves pages from the active list to the inactive list.
1253  *
1254  * We move them the other way if the page is referenced by one or more
1255  * processes, from rmap.
1256  *
1257  * If the pages are mostly unmapped, the processing is fast and it is
1258  * appropriate to hold zone->lru_lock across the whole operation.  But if
1259  * the pages are mapped, the processing is slow (page_referenced()) so we
1260  * should drop zone->lru_lock around each page.  It's impossible to balance
1261  * this, so instead we remove the pages from the LRU while processing them.
1262  * It is safe to rely on PG_active against the non-LRU pages in here because
1263  * nobody will play with that bit on a non-LRU page.
1264  *
1265  * The downside is that we have to touch page->_count against each page.
1266  * But we had to alter page->flags anyway.
1267  */
1268
1269 static void move_active_pages_to_lru(struct zone *zone,
1270                                      struct list_head *list,
1271                                      enum lru_list lru)
1272 {
1273         unsigned long pgmoved = 0;
1274         struct pagevec pvec;
1275         struct page *page;
1276
1277         pagevec_init(&pvec, 1);
1278
1279         while (!list_empty(list)) {
1280                 page = lru_to_page(list);
1281
1282                 VM_BUG_ON(PageLRU(page));
1283                 SetPageLRU(page);
1284
1285                 list_move(&page->lru, &zone->lru[lru].list);
1286                 mem_cgroup_add_lru_list(page, lru);
1287                 pgmoved++;
1288
1289                 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1290                         spin_unlock_irq(&zone->lru_lock);
1291                         if (buffer_heads_over_limit)
1292                                 pagevec_strip(&pvec);
1293                         __pagevec_release(&pvec);
1294                         spin_lock_irq(&zone->lru_lock);
1295                 }
1296         }
1297         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1298         if (!is_active_lru(lru))
1299                 __count_vm_events(PGDEACTIVATE, pgmoved);
1300 }
1301
1302 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1303                         struct scan_control *sc, int priority, int file)
1304 {
1305         unsigned long nr_taken;
1306         unsigned long pgscanned;
1307         unsigned long vm_flags;
1308         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1309         LIST_HEAD(l_active);
1310         LIST_HEAD(l_inactive);
1311         struct page *page;
1312         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1313         unsigned long nr_rotated = 0;
1314
1315         lru_add_drain();
1316         spin_lock_irq(&zone->lru_lock);
1317         nr_taken = sc->isolate_pages(nr_pages, &l_hold, &pgscanned, sc->order,
1318                                         ISOLATE_ACTIVE, zone,
1319                                         sc->mem_cgroup, 1, file);
1320         /*
1321          * zone->pages_scanned is used for detect zone's oom
1322          * mem_cgroup remembers nr_scan by itself.
1323          */
1324         if (scanning_global_lru(sc)) {
1325                 zone->pages_scanned += pgscanned;
1326         }
1327         reclaim_stat->recent_scanned[file] += nr_taken;
1328
1329         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1330         if (file)
1331                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1332         else
1333                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1334         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1335         spin_unlock_irq(&zone->lru_lock);
1336
1337         while (!list_empty(&l_hold)) {
1338                 cond_resched();
1339                 page = lru_to_page(&l_hold);
1340                 list_del(&page->lru);
1341
1342                 if (unlikely(!page_evictable(page, NULL))) {
1343                         putback_lru_page(page);
1344                         continue;
1345                 }
1346
1347                 /* page_referenced clears PageReferenced */
1348                 if (page_mapping_inuse(page) &&
1349                     page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1350                         nr_rotated++;
1351                         /*
1352                          * Identify referenced, file-backed active pages and
1353                          * give them one more trip around the active list. So
1354                          * that executable code get better chances to stay in
1355                          * memory under moderate memory pressure.  Anon pages
1356                          * are not likely to be evicted by use-once streaming
1357                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1358                          * so we ignore them here.
1359                          */
1360                         if ((vm_flags & VM_EXEC) && !PageAnon(page)) {
1361                                 list_add(&page->lru, &l_active);
1362                                 continue;
1363                         }
1364                 }
1365
1366                 ClearPageActive(page);  /* we are de-activating */
1367                 list_add(&page->lru, &l_inactive);
1368         }
1369
1370         /*
1371          * Move pages back to the lru list.
1372          */
1373         spin_lock_irq(&zone->lru_lock);
1374         /*
1375          * Count referenced pages from currently used mappings as rotated,
1376          * even though only some of them are actually re-activated.  This
1377          * helps balance scan pressure between file and anonymous pages in
1378          * get_scan_ratio.
1379          */
1380         reclaim_stat->recent_rotated[file] += nr_rotated;
1381
1382         move_active_pages_to_lru(zone, &l_active,
1383                                                 LRU_ACTIVE + file * LRU_FILE);
1384         move_active_pages_to_lru(zone, &l_inactive,
1385                                                 LRU_BASE   + file * LRU_FILE);
1386         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1387         spin_unlock_irq(&zone->lru_lock);
1388 }
1389
1390 static int inactive_anon_is_low_global(struct zone *zone)
1391 {
1392         unsigned long active, inactive;
1393
1394         active = zone_page_state(zone, NR_ACTIVE_ANON);
1395         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1396
1397         if (inactive * zone->inactive_ratio < active)
1398                 return 1;
1399
1400         return 0;
1401 }
1402
1403 /**
1404  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1405  * @zone: zone to check
1406  * @sc:   scan control of this context
1407  *
1408  * Returns true if the zone does not have enough inactive anon pages,
1409  * meaning some active anon pages need to be deactivated.
1410  */
1411 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1412 {
1413         int low;
1414
1415         if (scanning_global_lru(sc))
1416                 low = inactive_anon_is_low_global(zone);
1417         else
1418                 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1419         return low;
1420 }
1421
1422 static int inactive_file_is_low_global(struct zone *zone)
1423 {
1424         unsigned long active, inactive;
1425
1426         active = zone_page_state(zone, NR_ACTIVE_FILE);
1427         inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1428
1429         return (active > inactive);
1430 }
1431
1432 /**
1433  * inactive_file_is_low - check if file pages need to be deactivated
1434  * @zone: zone to check
1435  * @sc:   scan control of this context
1436  *
1437  * When the system is doing streaming IO, memory pressure here
1438  * ensures that active file pages get deactivated, until more
1439  * than half of the file pages are on the inactive list.
1440  *
1441  * Once we get to that situation, protect the system's working
1442  * set from being evicted by disabling active file page aging.
1443  *
1444  * This uses a different ratio than the anonymous pages, because
1445  * the page cache uses a use-once replacement algorithm.
1446  */
1447 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1448 {
1449         int low;
1450
1451         if (scanning_global_lru(sc))
1452                 low = inactive_file_is_low_global(zone);
1453         else
1454                 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1455         return low;
1456 }
1457
1458 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1459         struct zone *zone, struct scan_control *sc, int priority)
1460 {
1461         int file = is_file_lru(lru);
1462
1463         if (lru == LRU_ACTIVE_FILE && inactive_file_is_low(zone, sc)) {
1464                 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1465                 return 0;
1466         }
1467
1468         if (lru == LRU_ACTIVE_ANON && inactive_anon_is_low(zone, sc)) {
1469                 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1470                 return 0;
1471         }
1472         return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1473 }
1474
1475 /*
1476  * Determine how aggressively the anon and file LRU lists should be
1477  * scanned.  The relative value of each set of LRU lists is determined
1478  * by looking at the fraction of the pages scanned we did rotate back
1479  * onto the active list instead of evict.
1480  *
1481  * percent[0] specifies how much pressure to put on ram/swap backed
1482  * memory, while percent[1] determines pressure on the file LRUs.
1483  */
1484 static void get_scan_ratio(struct zone *zone, struct scan_control *sc,
1485                                         unsigned long *percent)
1486 {
1487         unsigned long anon, file, free;
1488         unsigned long anon_prio, file_prio;
1489         unsigned long ap, fp;
1490         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1491
1492         anon  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1493                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1494         file  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1495                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1496
1497         if (scanning_global_lru(sc)) {
1498                 free  = zone_page_state(zone, NR_FREE_PAGES);
1499                 /* If we have very few page cache pages,
1500                    force-scan anon pages. */
1501                 if (unlikely(file + free <= high_wmark_pages(zone))) {
1502                         percent[0] = 100;
1503                         percent[1] = 0;
1504                         return;
1505                 }
1506         }
1507
1508         /*
1509          * OK, so we have swap space and a fair amount of page cache
1510          * pages.  We use the recently rotated / recently scanned
1511          * ratios to determine how valuable each cache is.
1512          *
1513          * Because workloads change over time (and to avoid overflow)
1514          * we keep these statistics as a floating average, which ends
1515          * up weighing recent references more than old ones.
1516          *
1517          * anon in [0], file in [1]
1518          */
1519         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1520                 spin_lock_irq(&zone->lru_lock);
1521                 reclaim_stat->recent_scanned[0] /= 2;
1522                 reclaim_stat->recent_rotated[0] /= 2;
1523                 spin_unlock_irq(&zone->lru_lock);
1524         }
1525
1526         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1527                 spin_lock_irq(&zone->lru_lock);
1528                 reclaim_stat->recent_scanned[1] /= 2;
1529                 reclaim_stat->recent_rotated[1] /= 2;
1530                 spin_unlock_irq(&zone->lru_lock);
1531         }
1532
1533         /*
1534          * With swappiness at 100, anonymous and file have the same priority.
1535          * This scanning priority is essentially the inverse of IO cost.
1536          */
1537         anon_prio = sc->swappiness;
1538         file_prio = 200 - sc->swappiness;
1539
1540         /*
1541          * The amount of pressure on anon vs file pages is inversely
1542          * proportional to the fraction of recently scanned pages on
1543          * each list that were recently referenced and in active use.
1544          */
1545         ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1546         ap /= reclaim_stat->recent_rotated[0] + 1;
1547
1548         fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1549         fp /= reclaim_stat->recent_rotated[1] + 1;
1550
1551         /* Normalize to percentages */
1552         percent[0] = 100 * ap / (ap + fp + 1);
1553         percent[1] = 100 - percent[0];
1554 }
1555
1556 /*
1557  * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1558  * until we collected @swap_cluster_max pages to scan.
1559  */
1560 static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
1561                                        unsigned long *nr_saved_scan,
1562                                        unsigned long swap_cluster_max)
1563 {
1564         unsigned long nr;
1565
1566         *nr_saved_scan += nr_to_scan;
1567         nr = *nr_saved_scan;
1568
1569         if (nr >= swap_cluster_max)
1570                 *nr_saved_scan = 0;
1571         else
1572                 nr = 0;
1573
1574         return nr;
1575 }
1576
1577 /*
1578  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1579  */
1580 static void shrink_zone(int priority, struct zone *zone,
1581                                 struct scan_control *sc)
1582 {
1583         unsigned long nr[NR_LRU_LISTS];
1584         unsigned long nr_to_scan;
1585         unsigned long percent[2];       /* anon @ 0; file @ 1 */
1586         enum lru_list l;
1587         unsigned long nr_reclaimed = sc->nr_reclaimed;
1588         unsigned long swap_cluster_max = sc->swap_cluster_max;
1589         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1590         int noswap = 0;
1591
1592         /* If we have no swap space, do not bother scanning anon pages. */
1593         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1594                 noswap = 1;
1595                 percent[0] = 0;
1596                 percent[1] = 100;
1597         } else
1598                 get_scan_ratio(zone, sc, percent);
1599
1600         for_each_evictable_lru(l) {
1601                 int file = is_file_lru(l);
1602                 unsigned long scan;
1603
1604                 scan = zone_nr_lru_pages(zone, sc, l);
1605                 if (priority || noswap) {
1606                         scan >>= priority;
1607                         scan = (scan * percent[file]) / 100;
1608                 }
1609                 nr[l] = nr_scan_try_batch(scan,
1610                                           &reclaim_stat->nr_saved_scan[l],
1611                                           swap_cluster_max);
1612         }
1613
1614         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1615                                         nr[LRU_INACTIVE_FILE]) {
1616                 for_each_evictable_lru(l) {
1617                         if (nr[l]) {
1618                                 nr_to_scan = min(nr[l], swap_cluster_max);
1619                                 nr[l] -= nr_to_scan;
1620
1621                                 nr_reclaimed += shrink_list(l, nr_to_scan,
1622                                                             zone, sc, priority);
1623                         }
1624                 }
1625                 /*
1626                  * On large memory systems, scan >> priority can become
1627                  * really large. This is fine for the starting priority;
1628                  * we want to put equal scanning pressure on each zone.
1629                  * However, if the VM has a harder time of freeing pages,
1630                  * with multiple processes reclaiming pages, the total
1631                  * freeing target can get unreasonably large.
1632                  */
1633                 if (nr_reclaimed > swap_cluster_max &&
1634                         priority < DEF_PRIORITY && !current_is_kswapd())
1635                         break;
1636         }
1637
1638         sc->nr_reclaimed = nr_reclaimed;
1639
1640         /*
1641          * Even if we did not try to evict anon pages at all, we want to
1642          * rebalance the anon lru active/inactive ratio.
1643          */
1644         if (inactive_anon_is_low(zone, sc) && nr_swap_pages > 0)
1645                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1646
1647         throttle_vm_writeout(sc->gfp_mask);
1648 }
1649
1650 /*
1651  * This is the direct reclaim path, for page-allocating processes.  We only
1652  * try to reclaim pages from zones which will satisfy the caller's allocation
1653  * request.
1654  *
1655  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1656  * Because:
1657  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1658  *    allocation or
1659  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1660  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1661  *    zone defense algorithm.
1662  *
1663  * If a zone is deemed to be full of pinned pages then just give it a light
1664  * scan then give up on it.
1665  */
1666 static void shrink_zones(int priority, struct zonelist *zonelist,
1667                                         struct scan_control *sc)
1668 {
1669         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1670         struct zoneref *z;
1671         struct zone *zone;
1672
1673         sc->all_unreclaimable = 1;
1674         for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1675                                         sc->nodemask) {
1676                 if (!populated_zone(zone))
1677                         continue;
1678                 /*
1679                  * Take care memory controller reclaiming has small influence
1680                  * to global LRU.
1681                  */
1682                 if (scanning_global_lru(sc)) {
1683                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1684                                 continue;
1685                         note_zone_scanning_priority(zone, priority);
1686
1687                         if (zone_is_all_unreclaimable(zone) &&
1688                                                 priority != DEF_PRIORITY)
1689                                 continue;       /* Let kswapd poll it */
1690                         sc->all_unreclaimable = 0;
1691                 } else {
1692                         /*
1693                          * Ignore cpuset limitation here. We just want to reduce
1694                          * # of used pages by us regardless of memory shortage.
1695                          */
1696                         sc->all_unreclaimable = 0;
1697                         mem_cgroup_note_reclaim_priority(sc->mem_cgroup,
1698                                                         priority);
1699                 }
1700
1701                 shrink_zone(priority, zone, sc);
1702         }
1703 }
1704
1705 /*
1706  * This is the main entry point to direct page reclaim.
1707  *
1708  * If a full scan of the inactive list fails to free enough memory then we
1709  * are "out of memory" and something needs to be killed.
1710  *
1711  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1712  * high - the zone may be full of dirty or under-writeback pages, which this
1713  * caller can't do much about.  We kick pdflush and take explicit naps in the
1714  * hope that some of these pages can be written.  But if the allocating task
1715  * holds filesystem locks which prevent writeout this might not work, and the
1716  * allocation attempt will fail.
1717  *
1718  * returns:     0, if no pages reclaimed
1719  *              else, the number of pages reclaimed
1720  */
1721 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
1722                                         struct scan_control *sc)
1723 {
1724         int priority;
1725         unsigned long ret = 0;
1726         unsigned long total_scanned = 0;
1727         struct reclaim_state *reclaim_state = current->reclaim_state;
1728         unsigned long lru_pages = 0;
1729         struct zoneref *z;
1730         struct zone *zone;
1731         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1732
1733         delayacct_freepages_start();
1734
1735         if (scanning_global_lru(sc))
1736                 count_vm_event(ALLOCSTALL);
1737         /*
1738          * mem_cgroup will not do shrink_slab.
1739          */
1740         if (scanning_global_lru(sc)) {
1741                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1742
1743                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1744                                 continue;
1745
1746                         lru_pages += zone_reclaimable_pages(zone);
1747                 }
1748         }
1749
1750         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1751                 sc->nr_scanned = 0;
1752                 if (!priority)
1753                         disable_swap_token();
1754                 shrink_zones(priority, zonelist, sc);
1755                 /*
1756                  * Don't shrink slabs when reclaiming memory from
1757                  * over limit cgroups
1758                  */
1759                 if (scanning_global_lru(sc)) {
1760                         shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
1761                         if (reclaim_state) {
1762                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
1763                                 reclaim_state->reclaimed_slab = 0;
1764                         }
1765                 }
1766                 total_scanned += sc->nr_scanned;
1767                 if (sc->nr_reclaimed >= sc->swap_cluster_max) {
1768                         ret = sc->nr_reclaimed;
1769                         goto out;
1770                 }
1771
1772                 /*
1773                  * Try to write back as many pages as we just scanned.  This
1774                  * tends to cause slow streaming writers to write data to the
1775                  * disk smoothly, at the dirtying rate, which is nice.   But
1776                  * that's undesirable in laptop mode, where we *want* lumpy
1777                  * writeout.  So in laptop mode, write out the whole world.
1778                  */
1779                 if (total_scanned > sc->swap_cluster_max +
1780                                         sc->swap_cluster_max / 2) {
1781                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
1782                         sc->may_writepage = 1;
1783                 }
1784
1785                 /* Take a nap, wait for some writeback to complete */
1786                 if (sc->nr_scanned && priority < DEF_PRIORITY - 2)
1787                         congestion_wait(BLK_RW_ASYNC, HZ/10);
1788         }
1789         /* top priority shrink_zones still had more to do? don't OOM, then */
1790         if (!sc->all_unreclaimable && scanning_global_lru(sc))
1791                 ret = sc->nr_reclaimed;
1792 out:
1793         /*
1794          * Now that we've scanned all the zones at this priority level, note
1795          * that level within the zone so that the next thread which performs
1796          * scanning of this zone will immediately start out at this priority
1797          * level.  This affects only the decision whether or not to bring
1798          * mapped pages onto the inactive list.
1799          */
1800         if (priority < 0)
1801                 priority = 0;
1802
1803         if (scanning_global_lru(sc)) {
1804                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1805
1806                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1807                                 continue;
1808
1809                         zone->prev_priority = priority;
1810                 }
1811         } else
1812                 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority);
1813
1814         delayacct_freepages_end();
1815
1816         return ret;
1817 }
1818
1819 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
1820                                 gfp_t gfp_mask, nodemask_t *nodemask)
1821 {
1822         struct scan_control sc = {
1823                 .gfp_mask = gfp_mask,
1824                 .may_writepage = !laptop_mode,
1825                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1826                 .may_unmap = 1,
1827                 .may_swap = 1,
1828                 .swappiness = vm_swappiness,
1829                 .order = order,
1830                 .mem_cgroup = NULL,
1831                 .isolate_pages = isolate_pages_global,
1832                 .nodemask = nodemask,
1833         };
1834
1835         return do_try_to_free_pages(zonelist, &sc);
1836 }
1837
1838 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
1839
1840 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
1841                                            gfp_t gfp_mask,
1842                                            bool noswap,
1843                                            unsigned int swappiness)
1844 {
1845         struct scan_control sc = {
1846                 .may_writepage = !laptop_mode,
1847                 .may_unmap = 1,
1848                 .may_swap = !noswap,
1849                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1850                 .swappiness = swappiness,
1851                 .order = 0,
1852                 .mem_cgroup = mem_cont,
1853                 .isolate_pages = mem_cgroup_isolate_pages,
1854                 .nodemask = NULL, /* we don't care the placement */
1855         };
1856         struct zonelist *zonelist;
1857
1858         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1859                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1860         zonelist = NODE_DATA(numa_node_id())->node_zonelists;
1861         return do_try_to_free_pages(zonelist, &sc);
1862 }
1863 #endif
1864
1865 /*
1866  * For kswapd, balance_pgdat() will work across all this node's zones until
1867  * they are all at high_wmark_pages(zone).
1868  *
1869  * Returns the number of pages which were actually freed.
1870  *
1871  * There is special handling here for zones which are full of pinned pages.
1872  * This can happen if the pages are all mlocked, or if they are all used by
1873  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1874  * What we do is to detect the case where all pages in the zone have been
1875  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1876  * dead and from now on, only perform a short scan.  Basically we're polling
1877  * the zone for when the problem goes away.
1878  *
1879  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
1880  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
1881  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
1882  * lower zones regardless of the number of free pages in the lower zones. This
1883  * interoperates with the page allocator fallback scheme to ensure that aging
1884  * of pages is balanced across the zones.
1885  */
1886 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
1887 {
1888         int all_zones_ok;
1889         int priority;
1890         int i;
1891         unsigned long total_scanned;
1892         struct reclaim_state *reclaim_state = current->reclaim_state;
1893         struct scan_control sc = {
1894                 .gfp_mask = GFP_KERNEL,
1895                 .may_unmap = 1,
1896                 .may_swap = 1,
1897                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1898                 .swappiness = vm_swappiness,
1899                 .order = order,
1900                 .mem_cgroup = NULL,
1901                 .isolate_pages = isolate_pages_global,
1902         };
1903         /*
1904          * temp_priority is used to remember the scanning priority at which
1905          * this zone was successfully refilled to
1906          * free_pages == high_wmark_pages(zone).
1907          */
1908         int temp_priority[MAX_NR_ZONES];
1909
1910 loop_again:
1911         total_scanned = 0;
1912         sc.nr_reclaimed = 0;
1913         sc.may_writepage = !laptop_mode;
1914         count_vm_event(PAGEOUTRUN);
1915
1916         for (i = 0; i < pgdat->nr_zones; i++)
1917                 temp_priority[i] = DEF_PRIORITY;
1918
1919         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1920                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1921                 unsigned long lru_pages = 0;
1922
1923                 /* The swap token gets in the way of swapout... */
1924                 if (!priority)
1925                         disable_swap_token();
1926
1927                 all_zones_ok = 1;
1928
1929                 /*
1930                  * Scan in the highmem->dma direction for the highest
1931                  * zone which needs scanning
1932                  */
1933                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1934                         struct zone *zone = pgdat->node_zones + i;
1935
1936                         if (!populated_zone(zone))
1937                                 continue;
1938
1939                         if (zone_is_all_unreclaimable(zone) &&
1940                             priority != DEF_PRIORITY)
1941                                 continue;
1942
1943                         /*
1944                          * Do some background aging of the anon list, to give
1945                          * pages a chance to be referenced before reclaiming.
1946                          */
1947                         if (inactive_anon_is_low(zone, &sc))
1948                                 shrink_active_list(SWAP_CLUSTER_MAX, zone,
1949                                                         &sc, priority, 0);
1950
1951                         if (!zone_watermark_ok(zone, order,
1952                                         high_wmark_pages(zone), 0, 0)) {
1953                                 end_zone = i;
1954                                 break;
1955                         }
1956                 }
1957                 if (i < 0)
1958                         goto out;
1959
1960                 for (i = 0; i <= end_zone; i++) {
1961                         struct zone *zone = pgdat->node_zones + i;
1962
1963                         lru_pages += zone_reclaimable_pages(zone);
1964                 }
1965
1966                 /*
1967                  * Now scan the zone in the dma->highmem direction, stopping
1968                  * at the last zone which needs scanning.
1969                  *
1970                  * We do this because the page allocator works in the opposite
1971                  * direction.  This prevents the page allocator from allocating
1972                  * pages behind kswapd's direction of progress, which would
1973                  * cause too much scanning of the lower zones.
1974                  */
1975                 for (i = 0; i <= end_zone; i++) {
1976                         struct zone *zone = pgdat->node_zones + i;
1977                         int nr_slab;
1978
1979                         if (!populated_zone(zone))
1980                                 continue;
1981
1982                         if (zone_is_all_unreclaimable(zone) &&
1983                                         priority != DEF_PRIORITY)
1984                                 continue;
1985
1986                         if (!zone_watermark_ok(zone, order,
1987                                         high_wmark_pages(zone), end_zone, 0))
1988                                 all_zones_ok = 0;
1989                         temp_priority[i] = priority;
1990                         sc.nr_scanned = 0;
1991                         note_zone_scanning_priority(zone, priority);
1992                         /*
1993                          * We put equal pressure on every zone, unless one
1994                          * zone has way too many pages free already.
1995                          */
1996                         if (!zone_watermark_ok(zone, order,
1997                                         8*high_wmark_pages(zone), end_zone, 0))
1998                                 shrink_zone(priority, zone, &sc);
1999                         reclaim_state->reclaimed_slab = 0;
2000                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
2001                                                 lru_pages);
2002                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2003                         total_scanned += sc.nr_scanned;
2004                         if (zone_is_all_unreclaimable(zone))
2005                                 continue;
2006                         if (nr_slab == 0 && zone->pages_scanned >=
2007                                         (zone_reclaimable_pages(zone) * 6))
2008                                         zone_set_flag(zone,
2009                                                       ZONE_ALL_UNRECLAIMABLE);
2010                         /*
2011                          * If we've done a decent amount of scanning and
2012                          * the reclaim ratio is low, start doing writepage
2013                          * even in laptop mode
2014                          */
2015                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2016                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2017                                 sc.may_writepage = 1;
2018                 }
2019                 if (all_zones_ok)
2020                         break;          /* kswapd: all done */
2021                 /*
2022                  * OK, kswapd is getting into trouble.  Take a nap, then take
2023                  * another pass across the zones.
2024                  */
2025                 if (total_scanned && priority < DEF_PRIORITY - 2)
2026                         congestion_wait(BLK_RW_ASYNC, HZ/10);
2027
2028                 /*
2029                  * We do this so kswapd doesn't build up large priorities for
2030                  * example when it is freeing in parallel with allocators. It
2031                  * matches the direct reclaim path behaviour in terms of impact
2032                  * on zone->*_priority.
2033                  */
2034                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2035                         break;
2036         }
2037 out:
2038         /*
2039          * Note within each zone the priority level at which this zone was
2040          * brought into a happy state.  So that the next thread which scans this
2041          * zone will start out at that priority level.
2042          */
2043         for (i = 0; i < pgdat->nr_zones; i++) {
2044                 struct zone *zone = pgdat->node_zones + i;
2045
2046                 zone->prev_priority = temp_priority[i];
2047         }
2048         if (!all_zones_ok) {
2049                 cond_resched();
2050
2051                 try_to_freeze();
2052
2053                 /*
2054                  * Fragmentation may mean that the system cannot be
2055                  * rebalanced for high-order allocations in all zones.
2056                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2057                  * it means the zones have been fully scanned and are still
2058                  * not balanced. For high-order allocations, there is
2059                  * little point trying all over again as kswapd may
2060                  * infinite loop.
2061                  *
2062                  * Instead, recheck all watermarks at order-0 as they
2063                  * are the most important. If watermarks are ok, kswapd will go
2064                  * back to sleep. High-order users can still perform direct
2065                  * reclaim if they wish.
2066                  */
2067                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2068                         order = sc.order = 0;
2069
2070                 goto loop_again;
2071         }
2072
2073         return sc.nr_reclaimed;
2074 }
2075
2076 /*
2077  * The background pageout daemon, started as a kernel thread
2078  * from the init process.
2079  *
2080  * This basically trickles out pages so that we have _some_
2081  * free memory available even if there is no other activity
2082  * that frees anything up. This is needed for things like routing
2083  * etc, where we otherwise might have all activity going on in
2084  * asynchronous contexts that cannot page things out.
2085  *
2086  * If there are applications that are active memory-allocators
2087  * (most normal use), this basically shouldn't matter.
2088  */
2089 static int kswapd(void *p)
2090 {
2091         unsigned long order;
2092         pg_data_t *pgdat = (pg_data_t*)p;
2093         struct task_struct *tsk = current;
2094         DEFINE_WAIT(wait);
2095         struct reclaim_state reclaim_state = {
2096                 .reclaimed_slab = 0,
2097         };
2098         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2099
2100         lockdep_set_current_reclaim_state(GFP_KERNEL);
2101
2102         if (!cpumask_empty(cpumask))
2103                 set_cpus_allowed_ptr(tsk, cpumask);
2104         current->reclaim_state = &reclaim_state;
2105
2106         /*
2107          * Tell the memory management that we're a "memory allocator",
2108          * and that if we need more memory we should get access to it
2109          * regardless (see "__alloc_pages()"). "kswapd" should
2110          * never get caught in the normal page freeing logic.
2111          *
2112          * (Kswapd normally doesn't need memory anyway, but sometimes
2113          * you need a small amount of memory in order to be able to
2114          * page out something else, and this flag essentially protects
2115          * us from recursively trying to free more memory as we're
2116          * trying to free the first piece of memory in the first place).
2117          */
2118         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2119         set_freezable();
2120
2121         order = 0;
2122         for ( ; ; ) {
2123                 unsigned long new_order;
2124
2125                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2126                 new_order = pgdat->kswapd_max_order;
2127                 pgdat->kswapd_max_order = 0;
2128                 if (order < new_order) {
2129                         /*
2130                          * Don't sleep if someone wants a larger 'order'
2131                          * allocation
2132                          */
2133                         order = new_order;
2134                 } else {
2135                         if (!freezing(current))
2136                                 schedule();
2137
2138                         order = pgdat->kswapd_max_order;
2139                 }
2140                 finish_wait(&pgdat->kswapd_wait, &wait);
2141
2142                 if (!try_to_freeze()) {
2143                         /* We can speed up thawing tasks if we don't call
2144                          * balance_pgdat after returning from the refrigerator
2145                          */
2146                         balance_pgdat(pgdat, order);
2147                 }
2148         }
2149         return 0;
2150 }
2151
2152 /*
2153  * A zone is low on free memory, so wake its kswapd task to service it.
2154  */
2155 void wakeup_kswapd(struct zone *zone, int order)
2156 {
2157         pg_data_t *pgdat;
2158
2159         if (!populated_zone(zone))
2160                 return;
2161
2162         pgdat = zone->zone_pgdat;
2163         if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
2164                 return;
2165         if (pgdat->kswapd_max_order < order)
2166                 pgdat->kswapd_max_order = order;
2167         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2168                 return;
2169         if (!waitqueue_active(&pgdat->kswapd_wait))
2170                 return;
2171         wake_up_interruptible(&pgdat->kswapd_wait);
2172 }
2173
2174 /*
2175  * The reclaimable count would be mostly accurate.
2176  * The less reclaimable pages may be
2177  * - mlocked pages, which will be moved to unevictable list when encountered
2178  * - mapped pages, which may require several travels to be reclaimed
2179  * - dirty pages, which is not "instantly" reclaimable
2180  */
2181 unsigned long global_reclaimable_pages(void)
2182 {
2183         int nr;
2184
2185         nr = global_page_state(NR_ACTIVE_FILE) +
2186              global_page_state(NR_INACTIVE_FILE);
2187
2188         if (nr_swap_pages > 0)
2189                 nr += global_page_state(NR_ACTIVE_ANON) +
2190                       global_page_state(NR_INACTIVE_ANON);
2191
2192         return nr;
2193 }
2194
2195 unsigned long zone_reclaimable_pages(struct zone *zone)
2196 {
2197         int nr;
2198
2199         nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2200              zone_page_state(zone, NR_INACTIVE_FILE);
2201
2202         if (nr_swap_pages > 0)
2203                 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2204                       zone_page_state(zone, NR_INACTIVE_ANON);
2205
2206         return nr;
2207 }
2208
2209 #ifdef CONFIG_HIBERNATION
2210 /*
2211  * Helper function for shrink_all_memory().  Tries to reclaim 'nr_pages' pages
2212  * from LRU lists system-wide, for given pass and priority.
2213  *
2214  * For pass > 3 we also try to shrink the LRU lists that contain a few pages
2215  */
2216 static void shrink_all_zones(unsigned long nr_pages, int prio,
2217                                       int pass, struct scan_control *sc)
2218 {
2219         struct zone *zone;
2220         unsigned long nr_reclaimed = 0;
2221         struct zone_reclaim_stat *reclaim_stat;
2222
2223         for_each_populated_zone(zone) {
2224                 enum lru_list l;
2225
2226                 if (zone_is_all_unreclaimable(zone) && prio != DEF_PRIORITY)
2227                         continue;
2228
2229                 for_each_evictable_lru(l) {
2230                         enum zone_stat_item ls = NR_LRU_BASE + l;
2231                         unsigned long lru_pages = zone_page_state(zone, ls);
2232
2233                         /* For pass = 0, we don't shrink the active list */
2234                         if (pass == 0 && (l == LRU_ACTIVE_ANON ||
2235                                                 l == LRU_ACTIVE_FILE))
2236                                 continue;
2237
2238                         reclaim_stat = get_reclaim_stat(zone, sc);
2239                         reclaim_stat->nr_saved_scan[l] +=
2240                                                 (lru_pages >> prio) + 1;
2241                         if (reclaim_stat->nr_saved_scan[l]
2242                                                 >= nr_pages || pass > 3) {
2243                                 unsigned long nr_to_scan;
2244
2245                                 reclaim_stat->nr_saved_scan[l] = 0;
2246                                 nr_to_scan = min(nr_pages, lru_pages);
2247                                 nr_reclaimed += shrink_list(l, nr_to_scan, zone,
2248                                                                 sc, prio);
2249                                 if (nr_reclaimed >= nr_pages) {
2250                                         sc->nr_reclaimed += nr_reclaimed;
2251                                         return;
2252                                 }
2253                         }
2254                 }
2255         }
2256         sc->nr_reclaimed += nr_reclaimed;
2257 }
2258
2259 /*
2260  * Try to free `nr_pages' of memory, system-wide, and return the number of
2261  * freed pages.
2262  *
2263  * Rather than trying to age LRUs the aim is to preserve the overall
2264  * LRU order by reclaiming preferentially
2265  * inactive > active > active referenced > active mapped
2266  */
2267 unsigned long shrink_all_memory(unsigned long nr_pages)
2268 {
2269         unsigned long lru_pages, nr_slab;
2270         int pass;
2271         struct reclaim_state reclaim_state;
2272         struct scan_control sc = {
2273                 .gfp_mask = GFP_KERNEL,
2274                 .may_unmap = 0,
2275                 .may_writepage = 1,
2276                 .isolate_pages = isolate_pages_global,
2277                 .nr_reclaimed = 0,
2278         };
2279
2280         current->reclaim_state = &reclaim_state;
2281
2282         lru_pages = global_reclaimable_pages();
2283         nr_slab = global_page_state(NR_SLAB_RECLAIMABLE);
2284         /* If slab caches are huge, it's better to hit them first */
2285         while (nr_slab >= lru_pages) {
2286                 reclaim_state.reclaimed_slab = 0;
2287                 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
2288                 if (!reclaim_state.reclaimed_slab)
2289                         break;
2290
2291                 sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2292                 if (sc.nr_reclaimed >= nr_pages)
2293                         goto out;
2294
2295                 nr_slab -= reclaim_state.reclaimed_slab;
2296         }
2297
2298         /*
2299          * We try to shrink LRUs in 5 passes:
2300          * 0 = Reclaim from inactive_list only
2301          * 1 = Reclaim from active list but don't reclaim mapped
2302          * 2 = 2nd pass of type 1
2303          * 3 = Reclaim mapped (normal reclaim)
2304          * 4 = 2nd pass of type 3
2305          */
2306         for (pass = 0; pass < 5; pass++) {
2307                 int prio;
2308
2309                 /* Force reclaiming mapped pages in the passes #3 and #4 */
2310                 if (pass > 2)
2311                         sc.may_unmap = 1;
2312
2313                 for (prio = DEF_PRIORITY; prio >= 0; prio--) {
2314                         unsigned long nr_to_scan = nr_pages - sc.nr_reclaimed;
2315
2316                         sc.nr_scanned = 0;
2317                         sc.swap_cluster_max = nr_to_scan;
2318                         shrink_all_zones(nr_to_scan, prio, pass, &sc);
2319                         if (sc.nr_reclaimed >= nr_pages)
2320                                 goto out;
2321
2322                         reclaim_state.reclaimed_slab = 0;
2323                         shrink_slab(sc.nr_scanned, sc.gfp_mask,
2324                                     global_reclaimable_pages());
2325                         sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2326                         if (sc.nr_reclaimed >= nr_pages)
2327                                 goto out;
2328
2329                         if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
2330                                 congestion_wait(BLK_RW_ASYNC, HZ / 10);
2331                 }
2332         }
2333
2334         /*
2335          * If sc.nr_reclaimed = 0, we could not shrink LRUs, but there may be
2336          * something in slab caches
2337          */
2338         if (!sc.nr_reclaimed) {
2339                 do {
2340                         reclaim_state.reclaimed_slab = 0;
2341                         shrink_slab(nr_pages, sc.gfp_mask,
2342                                     global_reclaimable_pages());
2343                         sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2344                 } while (sc.nr_reclaimed < nr_pages &&
2345                                 reclaim_state.reclaimed_slab > 0);
2346         }
2347
2348
2349 out:
2350         current->reclaim_state = NULL;
2351
2352         return sc.nr_reclaimed;
2353 }
2354 #endif /* CONFIG_HIBERNATION */
2355
2356 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2357    not required for correctness.  So if the last cpu in a node goes
2358    away, we get changed to run anywhere: as the first one comes back,
2359    restore their cpu bindings. */
2360 static int __devinit cpu_callback(struct notifier_block *nfb,
2361                                   unsigned long action, void *hcpu)
2362 {
2363         int nid;
2364
2365         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2366                 for_each_node_state(nid, N_HIGH_MEMORY) {
2367                         pg_data_t *pgdat = NODE_DATA(nid);
2368                         const struct cpumask *mask;
2369
2370                         mask = cpumask_of_node(pgdat->node_id);
2371
2372                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2373                                 /* One of our CPUs online: restore mask */
2374                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2375                 }
2376         }
2377         return NOTIFY_OK;
2378 }
2379
2380 /*
2381  * This kswapd start function will be called by init and node-hot-add.
2382  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2383  */
2384 int kswapd_run(int nid)
2385 {
2386         pg_data_t *pgdat = NODE_DATA(nid);
2387         int ret = 0;
2388
2389         if (pgdat->kswapd)
2390                 return 0;
2391
2392         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2393         if (IS_ERR(pgdat->kswapd)) {
2394                 /* failure at boot is fatal */
2395                 BUG_ON(system_state == SYSTEM_BOOTING);
2396                 printk("Failed to start kswapd on node %d\n",nid);
2397                 ret = -1;
2398         }
2399         return ret;
2400 }
2401
2402 static int __init kswapd_init(void)
2403 {
2404         int nid;
2405
2406         swap_setup();
2407         for_each_node_state(nid, N_HIGH_MEMORY)
2408                 kswapd_run(nid);
2409         hotcpu_notifier(cpu_callback, 0);
2410         return 0;
2411 }
2412
2413 module_init(kswapd_init)
2414
2415 #ifdef CONFIG_NUMA
2416 /*
2417  * Zone reclaim mode
2418  *
2419  * If non-zero call zone_reclaim when the number of free pages falls below
2420  * the watermarks.
2421  */
2422 int zone_reclaim_mode __read_mostly;
2423
2424 #define RECLAIM_OFF 0
2425 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
2426 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
2427 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
2428
2429 /*
2430  * Priority for ZONE_RECLAIM. This determines the fraction of pages
2431  * of a node considered for each zone_reclaim. 4 scans 1/16th of
2432  * a zone.
2433  */
2434 #define ZONE_RECLAIM_PRIORITY 4
2435
2436 /*
2437  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2438  * occur.
2439  */
2440 int sysctl_min_unmapped_ratio = 1;
2441
2442 /*
2443  * If the number of slab pages in a zone grows beyond this percentage then
2444  * slab reclaim needs to occur.
2445  */
2446 int sysctl_min_slab_ratio = 5;
2447
2448 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
2449 {
2450         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
2451         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
2452                 zone_page_state(zone, NR_ACTIVE_FILE);
2453
2454         /*
2455          * It's possible for there to be more file mapped pages than
2456          * accounted for by the pages on the file LRU lists because
2457          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2458          */
2459         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
2460 }
2461
2462 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
2463 static long zone_pagecache_reclaimable(struct zone *zone)
2464 {
2465         long nr_pagecache_reclaimable;
2466         long delta = 0;
2467
2468         /*
2469          * If RECLAIM_SWAP is set, then all file pages are considered
2470          * potentially reclaimable. Otherwise, we have to worry about
2471          * pages like swapcache and zone_unmapped_file_pages() provides
2472          * a better estimate
2473          */
2474         if (zone_reclaim_mode & RECLAIM_SWAP)
2475                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
2476         else
2477                 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
2478
2479         /* If we can't clean pages, remove dirty pages from consideration */
2480         if (!(zone_reclaim_mode & RECLAIM_WRITE))
2481                 delta += zone_page_state(zone, NR_FILE_DIRTY);
2482
2483         /* Watch for any possible underflows due to delta */
2484         if (unlikely(delta > nr_pagecache_reclaimable))
2485                 delta = nr_pagecache_reclaimable;
2486
2487         return nr_pagecache_reclaimable - delta;
2488 }
2489
2490 /*
2491  * Try to free up some pages from this zone through reclaim.
2492  */
2493 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2494 {
2495         /* Minimum pages needed in order to stay on node */
2496         const unsigned long nr_pages = 1 << order;
2497         struct task_struct *p = current;
2498         struct reclaim_state reclaim_state;
2499         int priority;
2500         struct scan_control sc = {
2501                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2502                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2503                 .may_swap = 1,
2504                 .swap_cluster_max = max_t(unsigned long, nr_pages,
2505                                         SWAP_CLUSTER_MAX),
2506                 .gfp_mask = gfp_mask,
2507                 .swappiness = vm_swappiness,
2508                 .order = order,
2509                 .isolate_pages = isolate_pages_global,
2510         };
2511         unsigned long slab_reclaimable;
2512
2513         disable_swap_token();
2514         cond_resched();
2515         /*
2516          * We need to be able to allocate from the reserves for RECLAIM_SWAP
2517          * and we also need to be able to write out pages for RECLAIM_WRITE
2518          * and RECLAIM_SWAP.
2519          */
2520         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2521         reclaim_state.reclaimed_slab = 0;
2522         p->reclaim_state = &reclaim_state;
2523
2524         if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
2525                 /*
2526                  * Free memory by calling shrink zone with increasing
2527                  * priorities until we have enough memory freed.
2528                  */
2529                 priority = ZONE_RECLAIM_PRIORITY;
2530                 do {
2531                         note_zone_scanning_priority(zone, priority);
2532                         shrink_zone(priority, zone, &sc);
2533                         priority--;
2534                 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
2535         }
2536
2537         slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2538         if (slab_reclaimable > zone->min_slab_pages) {
2539                 /*
2540                  * shrink_slab() does not currently allow us to determine how
2541                  * many pages were freed in this zone. So we take the current
2542                  * number of slab pages and shake the slab until it is reduced
2543                  * by the same nr_pages that we used for reclaiming unmapped
2544                  * pages.
2545                  *
2546                  * Note that shrink_slab will free memory on all zones and may
2547                  * take a long time.
2548                  */
2549                 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
2550                         zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
2551                                 slab_reclaimable - nr_pages)
2552                         ;
2553
2554                 /*
2555                  * Update nr_reclaimed by the number of slab pages we
2556                  * reclaimed from this zone.
2557                  */
2558                 sc.nr_reclaimed += slab_reclaimable -
2559                         zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2560         }
2561
2562         p->reclaim_state = NULL;
2563         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
2564         return sc.nr_reclaimed >= nr_pages;
2565 }
2566
2567 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2568 {
2569         int node_id;
2570         int ret;
2571
2572         /*
2573          * Zone reclaim reclaims unmapped file backed pages and
2574          * slab pages if we are over the defined limits.
2575          *
2576          * A small portion of unmapped file backed pages is needed for
2577          * file I/O otherwise pages read by file I/O will be immediately
2578          * thrown out if the zone is overallocated. So we do not reclaim
2579          * if less than a specified percentage of the zone is used by
2580          * unmapped file backed pages.
2581          */
2582         if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
2583             zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
2584                 return ZONE_RECLAIM_FULL;
2585
2586         if (zone_is_all_unreclaimable(zone))
2587                 return ZONE_RECLAIM_FULL;
2588
2589         /*
2590          * Do not scan if the allocation should not be delayed.
2591          */
2592         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
2593                 return ZONE_RECLAIM_NOSCAN;
2594
2595         /*
2596          * Only run zone reclaim on the local zone or on zones that do not
2597          * have associated processors. This will favor the local processor
2598          * over remote processors and spread off node memory allocations
2599          * as wide as possible.
2600          */
2601         node_id = zone_to_nid(zone);
2602         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
2603                 return ZONE_RECLAIM_NOSCAN;
2604
2605         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2606                 return ZONE_RECLAIM_NOSCAN;
2607
2608         ret = __zone_reclaim(zone, gfp_mask, order);
2609         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2610
2611         if (!ret)
2612                 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
2613
2614         return ret;
2615 }
2616 #endif
2617
2618 /*
2619  * page_evictable - test whether a page is evictable
2620  * @page: the page to test
2621  * @vma: the VMA in which the page is or will be mapped, may be NULL
2622  *
2623  * Test whether page is evictable--i.e., should be placed on active/inactive
2624  * lists vs unevictable list.  The vma argument is !NULL when called from the
2625  * fault path to determine how to instantate a new page.
2626  *
2627  * Reasons page might not be evictable:
2628  * (1) page's mapping marked unevictable
2629  * (2) page is part of an mlocked VMA
2630  *
2631  */
2632 int page_evictable(struct page *page, struct vm_area_struct *vma)
2633 {
2634
2635         if (mapping_unevictable(page_mapping(page)))
2636                 return 0;
2637
2638         if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2639                 return 0;
2640
2641         return 1;
2642 }
2643
2644 /**
2645  * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2646  * @page: page to check evictability and move to appropriate lru list
2647  * @zone: zone page is in
2648  *
2649  * Checks a page for evictability and moves the page to the appropriate
2650  * zone lru list.
2651  *
2652  * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2653  * have PageUnevictable set.
2654  */
2655 static void check_move_unevictable_page(struct page *page, struct zone *zone)
2656 {
2657         VM_BUG_ON(PageActive(page));
2658
2659 retry:
2660         ClearPageUnevictable(page);
2661         if (page_evictable(page, NULL)) {
2662                 enum lru_list l = page_lru_base_type(page);
2663
2664                 __dec_zone_state(zone, NR_UNEVICTABLE);
2665                 list_move(&page->lru, &zone->lru[l].list);
2666                 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
2667                 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
2668                 __count_vm_event(UNEVICTABLE_PGRESCUED);
2669         } else {
2670                 /*
2671                  * rotate unevictable list
2672                  */
2673                 SetPageUnevictable(page);
2674                 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
2675                 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
2676                 if (page_evictable(page, NULL))
2677                         goto retry;
2678         }
2679 }
2680
2681 /**
2682  * scan_mapping_unevictable_pages - scan an address space for evictable pages
2683  * @mapping: struct address_space to scan for evictable pages
2684  *
2685  * Scan all pages in mapping.  Check unevictable pages for
2686  * evictability and move them to the appropriate zone lru list.
2687  */
2688 void scan_mapping_unevictable_pages(struct address_space *mapping)
2689 {
2690         pgoff_t next = 0;
2691         pgoff_t end   = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
2692                          PAGE_CACHE_SHIFT;
2693         struct zone *zone;
2694         struct pagevec pvec;
2695
2696         if (mapping->nrpages == 0)
2697                 return;
2698
2699         pagevec_init(&pvec, 0);
2700         while (next < end &&
2701                 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
2702                 int i;
2703                 int pg_scanned = 0;
2704
2705                 zone = NULL;
2706
2707                 for (i = 0; i < pagevec_count(&pvec); i++) {
2708                         struct page *page = pvec.pages[i];
2709                         pgoff_t page_index = page->index;
2710                         struct zone *pagezone = page_zone(page);
2711
2712                         pg_scanned++;
2713                         if (page_index > next)
2714                                 next = page_index;
2715                         next++;
2716
2717                         if (pagezone != zone) {
2718                                 if (zone)
2719                                         spin_unlock_irq(&zone->lru_lock);
2720                                 zone = pagezone;
2721                                 spin_lock_irq(&zone->lru_lock);
2722                         }
2723
2724                         if (PageLRU(page) && PageUnevictable(page))
2725                                 check_move_unevictable_page(page, zone);
2726                 }
2727                 if (zone)
2728                         spin_unlock_irq(&zone->lru_lock);
2729                 pagevec_release(&pvec);
2730
2731                 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
2732         }
2733
2734 }
2735
2736 /**
2737  * scan_zone_unevictable_pages - check unevictable list for evictable pages
2738  * @zone - zone of which to scan the unevictable list
2739  *
2740  * Scan @zone's unevictable LRU lists to check for pages that have become
2741  * evictable.  Move those that have to @zone's inactive list where they
2742  * become candidates for reclaim, unless shrink_inactive_zone() decides
2743  * to reactivate them.  Pages that are still unevictable are rotated
2744  * back onto @zone's unevictable list.
2745  */
2746 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
2747 static void scan_zone_unevictable_pages(struct zone *zone)
2748 {
2749         struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
2750         unsigned long scan;
2751         unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
2752
2753         while (nr_to_scan > 0) {
2754                 unsigned long batch_size = min(nr_to_scan,
2755                                                 SCAN_UNEVICTABLE_BATCH_SIZE);
2756
2757                 spin_lock_irq(&zone->lru_lock);
2758                 for (scan = 0;  scan < batch_size; scan++) {
2759                         struct page *page = lru_to_page(l_unevictable);
2760
2761                         if (!trylock_page(page))
2762                                 continue;
2763
2764                         prefetchw_prev_lru_page(page, l_unevictable, flags);
2765
2766                         if (likely(PageLRU(page) && PageUnevictable(page)))
2767                                 check_move_unevictable_page(page, zone);
2768
2769                         unlock_page(page);
2770                 }
2771                 spin_unlock_irq(&zone->lru_lock);
2772
2773                 nr_to_scan -= batch_size;
2774         }
2775 }
2776
2777
2778 /**
2779  * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
2780  *
2781  * A really big hammer:  scan all zones' unevictable LRU lists to check for
2782  * pages that have become evictable.  Move those back to the zones'
2783  * inactive list where they become candidates for reclaim.
2784  * This occurs when, e.g., we have unswappable pages on the unevictable lists,
2785  * and we add swap to the system.  As such, it runs in the context of a task
2786  * that has possibly/probably made some previously unevictable pages
2787  * evictable.
2788  */
2789 static void scan_all_zones_unevictable_pages(void)
2790 {
2791         struct zone *zone;
2792
2793         for_each_zone(zone) {
2794                 scan_zone_unevictable_pages(zone);
2795         }
2796 }
2797
2798 /*
2799  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
2800  * all nodes' unevictable lists for evictable pages
2801  */
2802 unsigned long scan_unevictable_pages;
2803
2804 int scan_unevictable_handler(struct ctl_table *table, int write,
2805                            struct file *file, void __user *buffer,
2806                            size_t *length, loff_t *ppos)
2807 {
2808         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
2809
2810         if (write && *(unsigned long *)table->data)
2811                 scan_all_zones_unevictable_pages();
2812
2813         scan_unevictable_pages = 0;
2814         return 0;
2815 }
2816
2817 /*
2818  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
2819  * a specified node's per zone unevictable lists for evictable pages.
2820  */
2821
2822 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
2823                                           struct sysdev_attribute *attr,
2824                                           char *buf)
2825 {
2826         return sprintf(buf, "0\n");     /* always zero; should fit... */
2827 }
2828
2829 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
2830                                            struct sysdev_attribute *attr,
2831                                         const char *buf, size_t count)
2832 {
2833         struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
2834         struct zone *zone;
2835         unsigned long res;
2836         unsigned long req = strict_strtoul(buf, 10, &res);
2837
2838         if (!req)
2839                 return 1;       /* zero is no-op */
2840
2841         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2842                 if (!populated_zone(zone))
2843                         continue;
2844                 scan_zone_unevictable_pages(zone);
2845         }
2846         return 1;
2847 }
2848
2849
2850 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
2851                         read_scan_unevictable_node,
2852                         write_scan_unevictable_node);
2853
2854 int scan_unevictable_register_node(struct node *node)
2855 {
2856         return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
2857 }
2858
2859 void scan_unevictable_unregister_node(struct node *node)
2860 {
2861         sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
2862 }
2863