vmscan: check high watermark after shrink zone
[linux-2.6.git] / mm / vmscan.c
1 /*
2  *  linux/mm/vmscan.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  *
6  *  Swap reorganised 29.12.95, Stephen Tweedie.
7  *  kswapd added: 7.1.96  sct
8  *  Removed kswapd_ctl limits, and swap out as many pages as needed
9  *  to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10  *  Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11  *  Multiqueue VM started 5.8.00, Rik van Riel.
12  */
13
14 #include <linux/mm.h>
15 #include <linux/module.h>
16 #include <linux/slab.h>
17 #include <linux/kernel_stat.h>
18 #include <linux/swap.h>
19 #include <linux/pagemap.h>
20 #include <linux/init.h>
21 #include <linux/highmem.h>
22 #include <linux/vmstat.h>
23 #include <linux/file.h>
24 #include <linux/writeback.h>
25 #include <linux/blkdev.h>
26 #include <linux/buffer_head.h>  /* for try_to_release_page(),
27                                         buffer_heads_over_limit */
28 #include <linux/mm_inline.h>
29 #include <linux/pagevec.h>
30 #include <linux/backing-dev.h>
31 #include <linux/rmap.h>
32 #include <linux/topology.h>
33 #include <linux/cpu.h>
34 #include <linux/cpuset.h>
35 #include <linux/notifier.h>
36 #include <linux/rwsem.h>
37 #include <linux/delay.h>
38 #include <linux/kthread.h>
39 #include <linux/freezer.h>
40 #include <linux/memcontrol.h>
41 #include <linux/delayacct.h>
42 #include <linux/sysctl.h>
43
44 #include <asm/tlbflush.h>
45 #include <asm/div64.h>
46
47 #include <linux/swapops.h>
48
49 #include "internal.h"
50
51 struct scan_control {
52         /* Incremented by the number of inactive pages that were scanned */
53         unsigned long nr_scanned;
54
55         /* Number of pages freed so far during a call to shrink_zones() */
56         unsigned long nr_reclaimed;
57
58         /* How many pages shrink_list() should reclaim */
59         unsigned long nr_to_reclaim;
60
61         unsigned long hibernation_mode;
62
63         /* This context's GFP mask */
64         gfp_t gfp_mask;
65
66         int may_writepage;
67
68         /* Can mapped pages be reclaimed? */
69         int may_unmap;
70
71         /* Can pages be swapped as part of reclaim? */
72         int may_swap;
73
74         int swappiness;
75
76         int all_unreclaimable;
77
78         int order;
79
80         /* Which cgroup do we reclaim from */
81         struct mem_cgroup *mem_cgroup;
82
83         /*
84          * Nodemask of nodes allowed by the caller. If NULL, all nodes
85          * are scanned.
86          */
87         nodemask_t      *nodemask;
88
89         /* Pluggable isolate pages callback */
90         unsigned long (*isolate_pages)(unsigned long nr, struct list_head *dst,
91                         unsigned long *scanned, int order, int mode,
92                         struct zone *z, struct mem_cgroup *mem_cont,
93                         int active, int file);
94 };
95
96 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
97
98 #ifdef ARCH_HAS_PREFETCH
99 #define prefetch_prev_lru_page(_page, _base, _field)                    \
100         do {                                                            \
101                 if ((_page)->lru.prev != _base) {                       \
102                         struct page *prev;                              \
103                                                                         \
104                         prev = lru_to_page(&(_page->lru));              \
105                         prefetch(&prev->_field);                        \
106                 }                                                       \
107         } while (0)
108 #else
109 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
110 #endif
111
112 #ifdef ARCH_HAS_PREFETCHW
113 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
114         do {                                                            \
115                 if ((_page)->lru.prev != _base) {                       \
116                         struct page *prev;                              \
117                                                                         \
118                         prev = lru_to_page(&(_page->lru));              \
119                         prefetchw(&prev->_field);                       \
120                 }                                                       \
121         } while (0)
122 #else
123 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
124 #endif
125
126 /*
127  * From 0 .. 100.  Higher means more swappy.
128  */
129 int vm_swappiness = 60;
130 long vm_total_pages;    /* The total number of pages which the VM controls */
131
132 static LIST_HEAD(shrinker_list);
133 static DECLARE_RWSEM(shrinker_rwsem);
134
135 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
136 #define scanning_global_lru(sc) (!(sc)->mem_cgroup)
137 #else
138 #define scanning_global_lru(sc) (1)
139 #endif
140
141 static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
142                                                   struct scan_control *sc)
143 {
144         if (!scanning_global_lru(sc))
145                 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
146
147         return &zone->reclaim_stat;
148 }
149
150 static unsigned long zone_nr_lru_pages(struct zone *zone,
151                                 struct scan_control *sc, enum lru_list lru)
152 {
153         if (!scanning_global_lru(sc))
154                 return mem_cgroup_zone_nr_pages(sc->mem_cgroup, zone, lru);
155
156         return zone_page_state(zone, NR_LRU_BASE + lru);
157 }
158
159
160 /*
161  * Add a shrinker callback to be called from the vm
162  */
163 void register_shrinker(struct shrinker *shrinker)
164 {
165         shrinker->nr = 0;
166         down_write(&shrinker_rwsem);
167         list_add_tail(&shrinker->list, &shrinker_list);
168         up_write(&shrinker_rwsem);
169 }
170 EXPORT_SYMBOL(register_shrinker);
171
172 /*
173  * Remove one
174  */
175 void unregister_shrinker(struct shrinker *shrinker)
176 {
177         down_write(&shrinker_rwsem);
178         list_del(&shrinker->list);
179         up_write(&shrinker_rwsem);
180 }
181 EXPORT_SYMBOL(unregister_shrinker);
182
183 #define SHRINK_BATCH 128
184 /*
185  * Call the shrink functions to age shrinkable caches
186  *
187  * Here we assume it costs one seek to replace a lru page and that it also
188  * takes a seek to recreate a cache object.  With this in mind we age equal
189  * percentages of the lru and ageable caches.  This should balance the seeks
190  * generated by these structures.
191  *
192  * If the vm encountered mapped pages on the LRU it increase the pressure on
193  * slab to avoid swapping.
194  *
195  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
196  *
197  * `lru_pages' represents the number of on-LRU pages in all the zones which
198  * are eligible for the caller's allocation attempt.  It is used for balancing
199  * slab reclaim versus page reclaim.
200  *
201  * Returns the number of slab objects which we shrunk.
202  */
203 unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
204                         unsigned long lru_pages)
205 {
206         struct shrinker *shrinker;
207         unsigned long ret = 0;
208
209         if (scanned == 0)
210                 scanned = SWAP_CLUSTER_MAX;
211
212         if (!down_read_trylock(&shrinker_rwsem))
213                 return 1;       /* Assume we'll be able to shrink next time */
214
215         list_for_each_entry(shrinker, &shrinker_list, list) {
216                 unsigned long long delta;
217                 unsigned long total_scan;
218                 unsigned long max_pass = (*shrinker->shrink)(0, gfp_mask);
219
220                 delta = (4 * scanned) / shrinker->seeks;
221                 delta *= max_pass;
222                 do_div(delta, lru_pages + 1);
223                 shrinker->nr += delta;
224                 if (shrinker->nr < 0) {
225                         printk(KERN_ERR "shrink_slab: %pF negative objects to "
226                                "delete nr=%ld\n",
227                                shrinker->shrink, shrinker->nr);
228                         shrinker->nr = max_pass;
229                 }
230
231                 /*
232                  * Avoid risking looping forever due to too large nr value:
233                  * never try to free more than twice the estimate number of
234                  * freeable entries.
235                  */
236                 if (shrinker->nr > max_pass * 2)
237                         shrinker->nr = max_pass * 2;
238
239                 total_scan = shrinker->nr;
240                 shrinker->nr = 0;
241
242                 while (total_scan >= SHRINK_BATCH) {
243                         long this_scan = SHRINK_BATCH;
244                         int shrink_ret;
245                         int nr_before;
246
247                         nr_before = (*shrinker->shrink)(0, gfp_mask);
248                         shrink_ret = (*shrinker->shrink)(this_scan, gfp_mask);
249                         if (shrink_ret == -1)
250                                 break;
251                         if (shrink_ret < nr_before)
252                                 ret += nr_before - shrink_ret;
253                         count_vm_events(SLABS_SCANNED, this_scan);
254                         total_scan -= this_scan;
255
256                         cond_resched();
257                 }
258
259                 shrinker->nr += total_scan;
260         }
261         up_read(&shrinker_rwsem);
262         return ret;
263 }
264
265 /* Called without lock on whether page is mapped, so answer is unstable */
266 static inline int page_mapping_inuse(struct page *page)
267 {
268         struct address_space *mapping;
269
270         /* Page is in somebody's page tables. */
271         if (page_mapped(page))
272                 return 1;
273
274         /* Be more reluctant to reclaim swapcache than pagecache */
275         if (PageSwapCache(page))
276                 return 1;
277
278         mapping = page_mapping(page);
279         if (!mapping)
280                 return 0;
281
282         /* File is mmap'd by somebody? */
283         return mapping_mapped(mapping);
284 }
285
286 static inline int is_page_cache_freeable(struct page *page)
287 {
288         /*
289          * A freeable page cache page is referenced only by the caller
290          * that isolated the page, the page cache radix tree and
291          * optional buffer heads at page->private.
292          */
293         return page_count(page) - page_has_private(page) == 2;
294 }
295
296 static int may_write_to_queue(struct backing_dev_info *bdi)
297 {
298         if (current->flags & PF_SWAPWRITE)
299                 return 1;
300         if (!bdi_write_congested(bdi))
301                 return 1;
302         if (bdi == current->backing_dev_info)
303                 return 1;
304         return 0;
305 }
306
307 /*
308  * We detected a synchronous write error writing a page out.  Probably
309  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
310  * fsync(), msync() or close().
311  *
312  * The tricky part is that after writepage we cannot touch the mapping: nothing
313  * prevents it from being freed up.  But we have a ref on the page and once
314  * that page is locked, the mapping is pinned.
315  *
316  * We're allowed to run sleeping lock_page() here because we know the caller has
317  * __GFP_FS.
318  */
319 static void handle_write_error(struct address_space *mapping,
320                                 struct page *page, int error)
321 {
322         lock_page(page);
323         if (page_mapping(page) == mapping)
324                 mapping_set_error(mapping, error);
325         unlock_page(page);
326 }
327
328 /* Request for sync pageout. */
329 enum pageout_io {
330         PAGEOUT_IO_ASYNC,
331         PAGEOUT_IO_SYNC,
332 };
333
334 /* possible outcome of pageout() */
335 typedef enum {
336         /* failed to write page out, page is locked */
337         PAGE_KEEP,
338         /* move page to the active list, page is locked */
339         PAGE_ACTIVATE,
340         /* page has been sent to the disk successfully, page is unlocked */
341         PAGE_SUCCESS,
342         /* page is clean and locked */
343         PAGE_CLEAN,
344 } pageout_t;
345
346 /*
347  * pageout is called by shrink_page_list() for each dirty page.
348  * Calls ->writepage().
349  */
350 static pageout_t pageout(struct page *page, struct address_space *mapping,
351                                                 enum pageout_io sync_writeback)
352 {
353         /*
354          * If the page is dirty, only perform writeback if that write
355          * will be non-blocking.  To prevent this allocation from being
356          * stalled by pagecache activity.  But note that there may be
357          * stalls if we need to run get_block().  We could test
358          * PagePrivate for that.
359          *
360          * If this process is currently in __generic_file_aio_write() against
361          * this page's queue, we can perform writeback even if that
362          * will block.
363          *
364          * If the page is swapcache, write it back even if that would
365          * block, for some throttling. This happens by accident, because
366          * swap_backing_dev_info is bust: it doesn't reflect the
367          * congestion state of the swapdevs.  Easy to fix, if needed.
368          */
369         if (!is_page_cache_freeable(page))
370                 return PAGE_KEEP;
371         if (!mapping) {
372                 /*
373                  * Some data journaling orphaned pages can have
374                  * page->mapping == NULL while being dirty with clean buffers.
375                  */
376                 if (page_has_private(page)) {
377                         if (try_to_free_buffers(page)) {
378                                 ClearPageDirty(page);
379                                 printk("%s: orphaned page\n", __func__);
380                                 return PAGE_CLEAN;
381                         }
382                 }
383                 return PAGE_KEEP;
384         }
385         if (mapping->a_ops->writepage == NULL)
386                 return PAGE_ACTIVATE;
387         if (!may_write_to_queue(mapping->backing_dev_info))
388                 return PAGE_KEEP;
389
390         if (clear_page_dirty_for_io(page)) {
391                 int res;
392                 struct writeback_control wbc = {
393                         .sync_mode = WB_SYNC_NONE,
394                         .nr_to_write = SWAP_CLUSTER_MAX,
395                         .range_start = 0,
396                         .range_end = LLONG_MAX,
397                         .nonblocking = 1,
398                         .for_reclaim = 1,
399                 };
400
401                 SetPageReclaim(page);
402                 res = mapping->a_ops->writepage(page, &wbc);
403                 if (res < 0)
404                         handle_write_error(mapping, page, res);
405                 if (res == AOP_WRITEPAGE_ACTIVATE) {
406                         ClearPageReclaim(page);
407                         return PAGE_ACTIVATE;
408                 }
409
410                 /*
411                  * Wait on writeback if requested to. This happens when
412                  * direct reclaiming a large contiguous area and the
413                  * first attempt to free a range of pages fails.
414                  */
415                 if (PageWriteback(page) && sync_writeback == PAGEOUT_IO_SYNC)
416                         wait_on_page_writeback(page);
417
418                 if (!PageWriteback(page)) {
419                         /* synchronous write or broken a_ops? */
420                         ClearPageReclaim(page);
421                 }
422                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
423                 return PAGE_SUCCESS;
424         }
425
426         return PAGE_CLEAN;
427 }
428
429 /*
430  * Same as remove_mapping, but if the page is removed from the mapping, it
431  * gets returned with a refcount of 0.
432  */
433 static int __remove_mapping(struct address_space *mapping, struct page *page)
434 {
435         BUG_ON(!PageLocked(page));
436         BUG_ON(mapping != page_mapping(page));
437
438         spin_lock_irq(&mapping->tree_lock);
439         /*
440          * The non racy check for a busy page.
441          *
442          * Must be careful with the order of the tests. When someone has
443          * a ref to the page, it may be possible that they dirty it then
444          * drop the reference. So if PageDirty is tested before page_count
445          * here, then the following race may occur:
446          *
447          * get_user_pages(&page);
448          * [user mapping goes away]
449          * write_to(page);
450          *                              !PageDirty(page)    [good]
451          * SetPageDirty(page);
452          * put_page(page);
453          *                              !page_count(page)   [good, discard it]
454          *
455          * [oops, our write_to data is lost]
456          *
457          * Reversing the order of the tests ensures such a situation cannot
458          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
459          * load is not satisfied before that of page->_count.
460          *
461          * Note that if SetPageDirty is always performed via set_page_dirty,
462          * and thus under tree_lock, then this ordering is not required.
463          */
464         if (!page_freeze_refs(page, 2))
465                 goto cannot_free;
466         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
467         if (unlikely(PageDirty(page))) {
468                 page_unfreeze_refs(page, 2);
469                 goto cannot_free;
470         }
471
472         if (PageSwapCache(page)) {
473                 swp_entry_t swap = { .val = page_private(page) };
474                 __delete_from_swap_cache(page);
475                 spin_unlock_irq(&mapping->tree_lock);
476                 swapcache_free(swap, page);
477         } else {
478                 __remove_from_page_cache(page);
479                 spin_unlock_irq(&mapping->tree_lock);
480                 mem_cgroup_uncharge_cache_page(page);
481         }
482
483         return 1;
484
485 cannot_free:
486         spin_unlock_irq(&mapping->tree_lock);
487         return 0;
488 }
489
490 /*
491  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
492  * someone else has a ref on the page, abort and return 0.  If it was
493  * successfully detached, return 1.  Assumes the caller has a single ref on
494  * this page.
495  */
496 int remove_mapping(struct address_space *mapping, struct page *page)
497 {
498         if (__remove_mapping(mapping, page)) {
499                 /*
500                  * Unfreezing the refcount with 1 rather than 2 effectively
501                  * drops the pagecache ref for us without requiring another
502                  * atomic operation.
503                  */
504                 page_unfreeze_refs(page, 1);
505                 return 1;
506         }
507         return 0;
508 }
509
510 /**
511  * putback_lru_page - put previously isolated page onto appropriate LRU list
512  * @page: page to be put back to appropriate lru list
513  *
514  * Add previously isolated @page to appropriate LRU list.
515  * Page may still be unevictable for other reasons.
516  *
517  * lru_lock must not be held, interrupts must be enabled.
518  */
519 void putback_lru_page(struct page *page)
520 {
521         int lru;
522         int active = !!TestClearPageActive(page);
523         int was_unevictable = PageUnevictable(page);
524
525         VM_BUG_ON(PageLRU(page));
526
527 redo:
528         ClearPageUnevictable(page);
529
530         if (page_evictable(page, NULL)) {
531                 /*
532                  * For evictable pages, we can use the cache.
533                  * In event of a race, worst case is we end up with an
534                  * unevictable page on [in]active list.
535                  * We know how to handle that.
536                  */
537                 lru = active + page_lru_base_type(page);
538                 lru_cache_add_lru(page, lru);
539         } else {
540                 /*
541                  * Put unevictable pages directly on zone's unevictable
542                  * list.
543                  */
544                 lru = LRU_UNEVICTABLE;
545                 add_page_to_unevictable_list(page);
546                 /*
547                  * When racing with an mlock clearing (page is
548                  * unlocked), make sure that if the other thread does
549                  * not observe our setting of PG_lru and fails
550                  * isolation, we see PG_mlocked cleared below and move
551                  * the page back to the evictable list.
552                  *
553                  * The other side is TestClearPageMlocked().
554                  */
555                 smp_mb();
556         }
557
558         /*
559          * page's status can change while we move it among lru. If an evictable
560          * page is on unevictable list, it never be freed. To avoid that,
561          * check after we added it to the list, again.
562          */
563         if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
564                 if (!isolate_lru_page(page)) {
565                         put_page(page);
566                         goto redo;
567                 }
568                 /* This means someone else dropped this page from LRU
569                  * So, it will be freed or putback to LRU again. There is
570                  * nothing to do here.
571                  */
572         }
573
574         if (was_unevictable && lru != LRU_UNEVICTABLE)
575                 count_vm_event(UNEVICTABLE_PGRESCUED);
576         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
577                 count_vm_event(UNEVICTABLE_PGCULLED);
578
579         put_page(page);         /* drop ref from isolate */
580 }
581
582 /*
583  * shrink_page_list() returns the number of reclaimed pages
584  */
585 static unsigned long shrink_page_list(struct list_head *page_list,
586                                         struct scan_control *sc,
587                                         enum pageout_io sync_writeback)
588 {
589         LIST_HEAD(ret_pages);
590         struct pagevec freed_pvec;
591         int pgactivate = 0;
592         unsigned long nr_reclaimed = 0;
593         unsigned long vm_flags;
594
595         cond_resched();
596
597         pagevec_init(&freed_pvec, 1);
598         while (!list_empty(page_list)) {
599                 struct address_space *mapping;
600                 struct page *page;
601                 int may_enter_fs;
602                 int referenced;
603
604                 cond_resched();
605
606                 page = lru_to_page(page_list);
607                 list_del(&page->lru);
608
609                 if (!trylock_page(page))
610                         goto keep;
611
612                 VM_BUG_ON(PageActive(page));
613
614                 sc->nr_scanned++;
615
616                 if (unlikely(!page_evictable(page, NULL)))
617                         goto cull_mlocked;
618
619                 if (!sc->may_unmap && page_mapped(page))
620                         goto keep_locked;
621
622                 /* Double the slab pressure for mapped and swapcache pages */
623                 if (page_mapped(page) || PageSwapCache(page))
624                         sc->nr_scanned++;
625
626                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
627                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
628
629                 if (PageWriteback(page)) {
630                         /*
631                          * Synchronous reclaim is performed in two passes,
632                          * first an asynchronous pass over the list to
633                          * start parallel writeback, and a second synchronous
634                          * pass to wait for the IO to complete.  Wait here
635                          * for any page for which writeback has already
636                          * started.
637                          */
638                         if (sync_writeback == PAGEOUT_IO_SYNC && may_enter_fs)
639                                 wait_on_page_writeback(page);
640                         else
641                                 goto keep_locked;
642                 }
643
644                 referenced = page_referenced(page, 1,
645                                                 sc->mem_cgroup, &vm_flags);
646                 /*
647                  * In active use or really unfreeable?  Activate it.
648                  * If page which have PG_mlocked lost isoltation race,
649                  * try_to_unmap moves it to unevictable list
650                  */
651                 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER &&
652                                         referenced && page_mapping_inuse(page)
653                                         && !(vm_flags & VM_LOCKED))
654                         goto activate_locked;
655
656                 /*
657                  * Anonymous process memory has backing store?
658                  * Try to allocate it some swap space here.
659                  */
660                 if (PageAnon(page) && !PageSwapCache(page)) {
661                         if (!(sc->gfp_mask & __GFP_IO))
662                                 goto keep_locked;
663                         if (!add_to_swap(page))
664                                 goto activate_locked;
665                         may_enter_fs = 1;
666                 }
667
668                 mapping = page_mapping(page);
669
670                 /*
671                  * The page is mapped into the page tables of one or more
672                  * processes. Try to unmap it here.
673                  */
674                 if (page_mapped(page) && mapping) {
675                         switch (try_to_unmap(page, TTU_UNMAP)) {
676                         case SWAP_FAIL:
677                                 goto activate_locked;
678                         case SWAP_AGAIN:
679                                 goto keep_locked;
680                         case SWAP_MLOCK:
681                                 goto cull_mlocked;
682                         case SWAP_SUCCESS:
683                                 ; /* try to free the page below */
684                         }
685                 }
686
687                 if (PageDirty(page)) {
688                         if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && referenced)
689                                 goto keep_locked;
690                         if (!may_enter_fs)
691                                 goto keep_locked;
692                         if (!sc->may_writepage)
693                                 goto keep_locked;
694
695                         /* Page is dirty, try to write it out here */
696                         switch (pageout(page, mapping, sync_writeback)) {
697                         case PAGE_KEEP:
698                                 goto keep_locked;
699                         case PAGE_ACTIVATE:
700                                 goto activate_locked;
701                         case PAGE_SUCCESS:
702                                 if (PageWriteback(page) || PageDirty(page))
703                                         goto keep;
704                                 /*
705                                  * A synchronous write - probably a ramdisk.  Go
706                                  * ahead and try to reclaim the page.
707                                  */
708                                 if (!trylock_page(page))
709                                         goto keep;
710                                 if (PageDirty(page) || PageWriteback(page))
711                                         goto keep_locked;
712                                 mapping = page_mapping(page);
713                         case PAGE_CLEAN:
714                                 ; /* try to free the page below */
715                         }
716                 }
717
718                 /*
719                  * If the page has buffers, try to free the buffer mappings
720                  * associated with this page. If we succeed we try to free
721                  * the page as well.
722                  *
723                  * We do this even if the page is PageDirty().
724                  * try_to_release_page() does not perform I/O, but it is
725                  * possible for a page to have PageDirty set, but it is actually
726                  * clean (all its buffers are clean).  This happens if the
727                  * buffers were written out directly, with submit_bh(). ext3
728                  * will do this, as well as the blockdev mapping.
729                  * try_to_release_page() will discover that cleanness and will
730                  * drop the buffers and mark the page clean - it can be freed.
731                  *
732                  * Rarely, pages can have buffers and no ->mapping.  These are
733                  * the pages which were not successfully invalidated in
734                  * truncate_complete_page().  We try to drop those buffers here
735                  * and if that worked, and the page is no longer mapped into
736                  * process address space (page_count == 1) it can be freed.
737                  * Otherwise, leave the page on the LRU so it is swappable.
738                  */
739                 if (page_has_private(page)) {
740                         if (!try_to_release_page(page, sc->gfp_mask))
741                                 goto activate_locked;
742                         if (!mapping && page_count(page) == 1) {
743                                 unlock_page(page);
744                                 if (put_page_testzero(page))
745                                         goto free_it;
746                                 else {
747                                         /*
748                                          * rare race with speculative reference.
749                                          * the speculative reference will free
750                                          * this page shortly, so we may
751                                          * increment nr_reclaimed here (and
752                                          * leave it off the LRU).
753                                          */
754                                         nr_reclaimed++;
755                                         continue;
756                                 }
757                         }
758                 }
759
760                 if (!mapping || !__remove_mapping(mapping, page))
761                         goto keep_locked;
762
763                 /*
764                  * At this point, we have no other references and there is
765                  * no way to pick any more up (removed from LRU, removed
766                  * from pagecache). Can use non-atomic bitops now (and
767                  * we obviously don't have to worry about waking up a process
768                  * waiting on the page lock, because there are no references.
769                  */
770                 __clear_page_locked(page);
771 free_it:
772                 nr_reclaimed++;
773                 if (!pagevec_add(&freed_pvec, page)) {
774                         __pagevec_free(&freed_pvec);
775                         pagevec_reinit(&freed_pvec);
776                 }
777                 continue;
778
779 cull_mlocked:
780                 if (PageSwapCache(page))
781                         try_to_free_swap(page);
782                 unlock_page(page);
783                 putback_lru_page(page);
784                 continue;
785
786 activate_locked:
787                 /* Not a candidate for swapping, so reclaim swap space. */
788                 if (PageSwapCache(page) && vm_swap_full())
789                         try_to_free_swap(page);
790                 VM_BUG_ON(PageActive(page));
791                 SetPageActive(page);
792                 pgactivate++;
793 keep_locked:
794                 unlock_page(page);
795 keep:
796                 list_add(&page->lru, &ret_pages);
797                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
798         }
799         list_splice(&ret_pages, page_list);
800         if (pagevec_count(&freed_pvec))
801                 __pagevec_free(&freed_pvec);
802         count_vm_events(PGACTIVATE, pgactivate);
803         return nr_reclaimed;
804 }
805
806 /* LRU Isolation modes. */
807 #define ISOLATE_INACTIVE 0      /* Isolate inactive pages. */
808 #define ISOLATE_ACTIVE 1        /* Isolate active pages. */
809 #define ISOLATE_BOTH 2          /* Isolate both active and inactive pages. */
810
811 /*
812  * Attempt to remove the specified page from its LRU.  Only take this page
813  * if it is of the appropriate PageActive status.  Pages which are being
814  * freed elsewhere are also ignored.
815  *
816  * page:        page to consider
817  * mode:        one of the LRU isolation modes defined above
818  *
819  * returns 0 on success, -ve errno on failure.
820  */
821 int __isolate_lru_page(struct page *page, int mode, int file)
822 {
823         int ret = -EINVAL;
824
825         /* Only take pages on the LRU. */
826         if (!PageLRU(page))
827                 return ret;
828
829         /*
830          * When checking the active state, we need to be sure we are
831          * dealing with comparible boolean values.  Take the logical not
832          * of each.
833          */
834         if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
835                 return ret;
836
837         if (mode != ISOLATE_BOTH && page_is_file_cache(page) != file)
838                 return ret;
839
840         /*
841          * When this function is being called for lumpy reclaim, we
842          * initially look into all LRU pages, active, inactive and
843          * unevictable; only give shrink_page_list evictable pages.
844          */
845         if (PageUnevictable(page))
846                 return ret;
847
848         ret = -EBUSY;
849
850         if (likely(get_page_unless_zero(page))) {
851                 /*
852                  * Be careful not to clear PageLRU until after we're
853                  * sure the page is not being freed elsewhere -- the
854                  * page release code relies on it.
855                  */
856                 ClearPageLRU(page);
857                 ret = 0;
858         }
859
860         return ret;
861 }
862
863 /*
864  * zone->lru_lock is heavily contended.  Some of the functions that
865  * shrink the lists perform better by taking out a batch of pages
866  * and working on them outside the LRU lock.
867  *
868  * For pagecache intensive workloads, this function is the hottest
869  * spot in the kernel (apart from copy_*_user functions).
870  *
871  * Appropriate locks must be held before calling this function.
872  *
873  * @nr_to_scan: The number of pages to look through on the list.
874  * @src:        The LRU list to pull pages off.
875  * @dst:        The temp list to put pages on to.
876  * @scanned:    The number of pages that were scanned.
877  * @order:      The caller's attempted allocation order
878  * @mode:       One of the LRU isolation modes
879  * @file:       True [1] if isolating file [!anon] pages
880  *
881  * returns how many pages were moved onto *@dst.
882  */
883 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
884                 struct list_head *src, struct list_head *dst,
885                 unsigned long *scanned, int order, int mode, int file)
886 {
887         unsigned long nr_taken = 0;
888         unsigned long scan;
889
890         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
891                 struct page *page;
892                 unsigned long pfn;
893                 unsigned long end_pfn;
894                 unsigned long page_pfn;
895                 int zone_id;
896
897                 page = lru_to_page(src);
898                 prefetchw_prev_lru_page(page, src, flags);
899
900                 VM_BUG_ON(!PageLRU(page));
901
902                 switch (__isolate_lru_page(page, mode, file)) {
903                 case 0:
904                         list_move(&page->lru, dst);
905                         mem_cgroup_del_lru(page);
906                         nr_taken++;
907                         break;
908
909                 case -EBUSY:
910                         /* else it is being freed elsewhere */
911                         list_move(&page->lru, src);
912                         mem_cgroup_rotate_lru_list(page, page_lru(page));
913                         continue;
914
915                 default:
916                         BUG();
917                 }
918
919                 if (!order)
920                         continue;
921
922                 /*
923                  * Attempt to take all pages in the order aligned region
924                  * surrounding the tag page.  Only take those pages of
925                  * the same active state as that tag page.  We may safely
926                  * round the target page pfn down to the requested order
927                  * as the mem_map is guarenteed valid out to MAX_ORDER,
928                  * where that page is in a different zone we will detect
929                  * it from its zone id and abort this block scan.
930                  */
931                 zone_id = page_zone_id(page);
932                 page_pfn = page_to_pfn(page);
933                 pfn = page_pfn & ~((1 << order) - 1);
934                 end_pfn = pfn + (1 << order);
935                 for (; pfn < end_pfn; pfn++) {
936                         struct page *cursor_page;
937
938                         /* The target page is in the block, ignore it. */
939                         if (unlikely(pfn == page_pfn))
940                                 continue;
941
942                         /* Avoid holes within the zone. */
943                         if (unlikely(!pfn_valid_within(pfn)))
944                                 break;
945
946                         cursor_page = pfn_to_page(pfn);
947
948                         /* Check that we have not crossed a zone boundary. */
949                         if (unlikely(page_zone_id(cursor_page) != zone_id))
950                                 continue;
951
952                         /*
953                          * If we don't have enough swap space, reclaiming of
954                          * anon page which don't already have a swap slot is
955                          * pointless.
956                          */
957                         if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
958                                         !PageSwapCache(cursor_page))
959                                 continue;
960
961                         if (__isolate_lru_page(cursor_page, mode, file) == 0) {
962                                 list_move(&cursor_page->lru, dst);
963                                 mem_cgroup_del_lru(cursor_page);
964                                 nr_taken++;
965                                 scan++;
966                         }
967                 }
968         }
969
970         *scanned = scan;
971         return nr_taken;
972 }
973
974 static unsigned long isolate_pages_global(unsigned long nr,
975                                         struct list_head *dst,
976                                         unsigned long *scanned, int order,
977                                         int mode, struct zone *z,
978                                         struct mem_cgroup *mem_cont,
979                                         int active, int file)
980 {
981         int lru = LRU_BASE;
982         if (active)
983                 lru += LRU_ACTIVE;
984         if (file)
985                 lru += LRU_FILE;
986         return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
987                                                                 mode, file);
988 }
989
990 /*
991  * clear_active_flags() is a helper for shrink_active_list(), clearing
992  * any active bits from the pages in the list.
993  */
994 static unsigned long clear_active_flags(struct list_head *page_list,
995                                         unsigned int *count)
996 {
997         int nr_active = 0;
998         int lru;
999         struct page *page;
1000
1001         list_for_each_entry(page, page_list, lru) {
1002                 lru = page_lru_base_type(page);
1003                 if (PageActive(page)) {
1004                         lru += LRU_ACTIVE;
1005                         ClearPageActive(page);
1006                         nr_active++;
1007                 }
1008                 count[lru]++;
1009         }
1010
1011         return nr_active;
1012 }
1013
1014 /**
1015  * isolate_lru_page - tries to isolate a page from its LRU list
1016  * @page: page to isolate from its LRU list
1017  *
1018  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1019  * vmstat statistic corresponding to whatever LRU list the page was on.
1020  *
1021  * Returns 0 if the page was removed from an LRU list.
1022  * Returns -EBUSY if the page was not on an LRU list.
1023  *
1024  * The returned page will have PageLRU() cleared.  If it was found on
1025  * the active list, it will have PageActive set.  If it was found on
1026  * the unevictable list, it will have the PageUnevictable bit set. That flag
1027  * may need to be cleared by the caller before letting the page go.
1028  *
1029  * The vmstat statistic corresponding to the list on which the page was
1030  * found will be decremented.
1031  *
1032  * Restrictions:
1033  * (1) Must be called with an elevated refcount on the page. This is a
1034  *     fundamentnal difference from isolate_lru_pages (which is called
1035  *     without a stable reference).
1036  * (2) the lru_lock must not be held.
1037  * (3) interrupts must be enabled.
1038  */
1039 int isolate_lru_page(struct page *page)
1040 {
1041         int ret = -EBUSY;
1042
1043         if (PageLRU(page)) {
1044                 struct zone *zone = page_zone(page);
1045
1046                 spin_lock_irq(&zone->lru_lock);
1047                 if (PageLRU(page) && get_page_unless_zero(page)) {
1048                         int lru = page_lru(page);
1049                         ret = 0;
1050                         ClearPageLRU(page);
1051
1052                         del_page_from_lru_list(zone, page, lru);
1053                 }
1054                 spin_unlock_irq(&zone->lru_lock);
1055         }
1056         return ret;
1057 }
1058
1059 /*
1060  * Are there way too many processes in the direct reclaim path already?
1061  */
1062 static int too_many_isolated(struct zone *zone, int file,
1063                 struct scan_control *sc)
1064 {
1065         unsigned long inactive, isolated;
1066
1067         if (current_is_kswapd())
1068                 return 0;
1069
1070         if (!scanning_global_lru(sc))
1071                 return 0;
1072
1073         if (file) {
1074                 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1075                 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1076         } else {
1077                 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1078                 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1079         }
1080
1081         return isolated > inactive;
1082 }
1083
1084 /*
1085  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1086  * of reclaimed pages
1087  */
1088 static unsigned long shrink_inactive_list(unsigned long max_scan,
1089                         struct zone *zone, struct scan_control *sc,
1090                         int priority, int file)
1091 {
1092         LIST_HEAD(page_list);
1093         struct pagevec pvec;
1094         unsigned long nr_scanned = 0;
1095         unsigned long nr_reclaimed = 0;
1096         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1097         int lumpy_reclaim = 0;
1098
1099         while (unlikely(too_many_isolated(zone, file, sc))) {
1100                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1101
1102                 /* We are about to die and free our memory. Return now. */
1103                 if (fatal_signal_pending(current))
1104                         return SWAP_CLUSTER_MAX;
1105         }
1106
1107         /*
1108          * If we need a large contiguous chunk of memory, or have
1109          * trouble getting a small set of contiguous pages, we
1110          * will reclaim both active and inactive pages.
1111          *
1112          * We use the same threshold as pageout congestion_wait below.
1113          */
1114         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1115                 lumpy_reclaim = 1;
1116         else if (sc->order && priority < DEF_PRIORITY - 2)
1117                 lumpy_reclaim = 1;
1118
1119         pagevec_init(&pvec, 1);
1120
1121         lru_add_drain();
1122         spin_lock_irq(&zone->lru_lock);
1123         do {
1124                 struct page *page;
1125                 unsigned long nr_taken;
1126                 unsigned long nr_scan;
1127                 unsigned long nr_freed;
1128                 unsigned long nr_active;
1129                 unsigned int count[NR_LRU_LISTS] = { 0, };
1130                 int mode = lumpy_reclaim ? ISOLATE_BOTH : ISOLATE_INACTIVE;
1131                 unsigned long nr_anon;
1132                 unsigned long nr_file;
1133
1134                 nr_taken = sc->isolate_pages(SWAP_CLUSTER_MAX,
1135                              &page_list, &nr_scan, sc->order, mode,
1136                                 zone, sc->mem_cgroup, 0, file);
1137
1138                 if (scanning_global_lru(sc)) {
1139                         zone->pages_scanned += nr_scan;
1140                         if (current_is_kswapd())
1141                                 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1142                                                        nr_scan);
1143                         else
1144                                 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1145                                                        nr_scan);
1146                 }
1147
1148                 if (nr_taken == 0)
1149                         goto done;
1150
1151                 nr_active = clear_active_flags(&page_list, count);
1152                 __count_vm_events(PGDEACTIVATE, nr_active);
1153
1154                 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1155                                                 -count[LRU_ACTIVE_FILE]);
1156                 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1157                                                 -count[LRU_INACTIVE_FILE]);
1158                 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1159                                                 -count[LRU_ACTIVE_ANON]);
1160                 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1161                                                 -count[LRU_INACTIVE_ANON]);
1162
1163                 nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1164                 nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1165                 __mod_zone_page_state(zone, NR_ISOLATED_ANON, nr_anon);
1166                 __mod_zone_page_state(zone, NR_ISOLATED_FILE, nr_file);
1167
1168                 reclaim_stat->recent_scanned[0] += nr_anon;
1169                 reclaim_stat->recent_scanned[1] += nr_file;
1170
1171                 spin_unlock_irq(&zone->lru_lock);
1172
1173                 nr_scanned += nr_scan;
1174                 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
1175
1176                 /*
1177                  * If we are direct reclaiming for contiguous pages and we do
1178                  * not reclaim everything in the list, try again and wait
1179                  * for IO to complete. This will stall high-order allocations
1180                  * but that should be acceptable to the caller
1181                  */
1182                 if (nr_freed < nr_taken && !current_is_kswapd() &&
1183                     lumpy_reclaim) {
1184                         congestion_wait(BLK_RW_ASYNC, HZ/10);
1185
1186                         /*
1187                          * The attempt at page out may have made some
1188                          * of the pages active, mark them inactive again.
1189                          */
1190                         nr_active = clear_active_flags(&page_list, count);
1191                         count_vm_events(PGDEACTIVATE, nr_active);
1192
1193                         nr_freed += shrink_page_list(&page_list, sc,
1194                                                         PAGEOUT_IO_SYNC);
1195                 }
1196
1197                 nr_reclaimed += nr_freed;
1198
1199                 local_irq_disable();
1200                 if (current_is_kswapd())
1201                         __count_vm_events(KSWAPD_STEAL, nr_freed);
1202                 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
1203
1204                 spin_lock(&zone->lru_lock);
1205                 /*
1206                  * Put back any unfreeable pages.
1207                  */
1208                 while (!list_empty(&page_list)) {
1209                         int lru;
1210                         page = lru_to_page(&page_list);
1211                         VM_BUG_ON(PageLRU(page));
1212                         list_del(&page->lru);
1213                         if (unlikely(!page_evictable(page, NULL))) {
1214                                 spin_unlock_irq(&zone->lru_lock);
1215                                 putback_lru_page(page);
1216                                 spin_lock_irq(&zone->lru_lock);
1217                                 continue;
1218                         }
1219                         SetPageLRU(page);
1220                         lru = page_lru(page);
1221                         add_page_to_lru_list(zone, page, lru);
1222                         if (is_active_lru(lru)) {
1223                                 int file = is_file_lru(lru);
1224                                 reclaim_stat->recent_rotated[file]++;
1225                         }
1226                         if (!pagevec_add(&pvec, page)) {
1227                                 spin_unlock_irq(&zone->lru_lock);
1228                                 __pagevec_release(&pvec);
1229                                 spin_lock_irq(&zone->lru_lock);
1230                         }
1231                 }
1232                 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1233                 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1234
1235         } while (nr_scanned < max_scan);
1236
1237 done:
1238         spin_unlock_irq(&zone->lru_lock);
1239         pagevec_release(&pvec);
1240         return nr_reclaimed;
1241 }
1242
1243 /*
1244  * We are about to scan this zone at a certain priority level.  If that priority
1245  * level is smaller (ie: more urgent) than the previous priority, then note
1246  * that priority level within the zone.  This is done so that when the next
1247  * process comes in to scan this zone, it will immediately start out at this
1248  * priority level rather than having to build up its own scanning priority.
1249  * Here, this priority affects only the reclaim-mapped threshold.
1250  */
1251 static inline void note_zone_scanning_priority(struct zone *zone, int priority)
1252 {
1253         if (priority < zone->prev_priority)
1254                 zone->prev_priority = priority;
1255 }
1256
1257 /*
1258  * This moves pages from the active list to the inactive list.
1259  *
1260  * We move them the other way if the page is referenced by one or more
1261  * processes, from rmap.
1262  *
1263  * If the pages are mostly unmapped, the processing is fast and it is
1264  * appropriate to hold zone->lru_lock across the whole operation.  But if
1265  * the pages are mapped, the processing is slow (page_referenced()) so we
1266  * should drop zone->lru_lock around each page.  It's impossible to balance
1267  * this, so instead we remove the pages from the LRU while processing them.
1268  * It is safe to rely on PG_active against the non-LRU pages in here because
1269  * nobody will play with that bit on a non-LRU page.
1270  *
1271  * The downside is that we have to touch page->_count against each page.
1272  * But we had to alter page->flags anyway.
1273  */
1274
1275 static void move_active_pages_to_lru(struct zone *zone,
1276                                      struct list_head *list,
1277                                      enum lru_list lru)
1278 {
1279         unsigned long pgmoved = 0;
1280         struct pagevec pvec;
1281         struct page *page;
1282
1283         pagevec_init(&pvec, 1);
1284
1285         while (!list_empty(list)) {
1286                 page = lru_to_page(list);
1287
1288                 VM_BUG_ON(PageLRU(page));
1289                 SetPageLRU(page);
1290
1291                 list_move(&page->lru, &zone->lru[lru].list);
1292                 mem_cgroup_add_lru_list(page, lru);
1293                 pgmoved++;
1294
1295                 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1296                         spin_unlock_irq(&zone->lru_lock);
1297                         if (buffer_heads_over_limit)
1298                                 pagevec_strip(&pvec);
1299                         __pagevec_release(&pvec);
1300                         spin_lock_irq(&zone->lru_lock);
1301                 }
1302         }
1303         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1304         if (!is_active_lru(lru))
1305                 __count_vm_events(PGDEACTIVATE, pgmoved);
1306 }
1307
1308 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1309                         struct scan_control *sc, int priority, int file)
1310 {
1311         unsigned long nr_taken;
1312         unsigned long pgscanned;
1313         unsigned long vm_flags;
1314         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1315         LIST_HEAD(l_active);
1316         LIST_HEAD(l_inactive);
1317         struct page *page;
1318         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1319         unsigned long nr_rotated = 0;
1320
1321         lru_add_drain();
1322         spin_lock_irq(&zone->lru_lock);
1323         nr_taken = sc->isolate_pages(nr_pages, &l_hold, &pgscanned, sc->order,
1324                                         ISOLATE_ACTIVE, zone,
1325                                         sc->mem_cgroup, 1, file);
1326         /*
1327          * zone->pages_scanned is used for detect zone's oom
1328          * mem_cgroup remembers nr_scan by itself.
1329          */
1330         if (scanning_global_lru(sc)) {
1331                 zone->pages_scanned += pgscanned;
1332         }
1333         reclaim_stat->recent_scanned[file] += nr_taken;
1334
1335         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1336         if (file)
1337                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1338         else
1339                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1340         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1341         spin_unlock_irq(&zone->lru_lock);
1342
1343         while (!list_empty(&l_hold)) {
1344                 cond_resched();
1345                 page = lru_to_page(&l_hold);
1346                 list_del(&page->lru);
1347
1348                 if (unlikely(!page_evictable(page, NULL))) {
1349                         putback_lru_page(page);
1350                         continue;
1351                 }
1352
1353                 /* page_referenced clears PageReferenced */
1354                 if (page_mapping_inuse(page) &&
1355                     page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1356                         nr_rotated++;
1357                         /*
1358                          * Identify referenced, file-backed active pages and
1359                          * give them one more trip around the active list. So
1360                          * that executable code get better chances to stay in
1361                          * memory under moderate memory pressure.  Anon pages
1362                          * are not likely to be evicted by use-once streaming
1363                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1364                          * so we ignore them here.
1365                          */
1366                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1367                                 list_add(&page->lru, &l_active);
1368                                 continue;
1369                         }
1370                 }
1371
1372                 ClearPageActive(page);  /* we are de-activating */
1373                 list_add(&page->lru, &l_inactive);
1374         }
1375
1376         /*
1377          * Move pages back to the lru list.
1378          */
1379         spin_lock_irq(&zone->lru_lock);
1380         /*
1381          * Count referenced pages from currently used mappings as rotated,
1382          * even though only some of them are actually re-activated.  This
1383          * helps balance scan pressure between file and anonymous pages in
1384          * get_scan_ratio.
1385          */
1386         reclaim_stat->recent_rotated[file] += nr_rotated;
1387
1388         move_active_pages_to_lru(zone, &l_active,
1389                                                 LRU_ACTIVE + file * LRU_FILE);
1390         move_active_pages_to_lru(zone, &l_inactive,
1391                                                 LRU_BASE   + file * LRU_FILE);
1392         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1393         spin_unlock_irq(&zone->lru_lock);
1394 }
1395
1396 static int inactive_anon_is_low_global(struct zone *zone)
1397 {
1398         unsigned long active, inactive;
1399
1400         active = zone_page_state(zone, NR_ACTIVE_ANON);
1401         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1402
1403         if (inactive * zone->inactive_ratio < active)
1404                 return 1;
1405
1406         return 0;
1407 }
1408
1409 /**
1410  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1411  * @zone: zone to check
1412  * @sc:   scan control of this context
1413  *
1414  * Returns true if the zone does not have enough inactive anon pages,
1415  * meaning some active anon pages need to be deactivated.
1416  */
1417 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1418 {
1419         int low;
1420
1421         if (scanning_global_lru(sc))
1422                 low = inactive_anon_is_low_global(zone);
1423         else
1424                 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1425         return low;
1426 }
1427
1428 static int inactive_file_is_low_global(struct zone *zone)
1429 {
1430         unsigned long active, inactive;
1431
1432         active = zone_page_state(zone, NR_ACTIVE_FILE);
1433         inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1434
1435         return (active > inactive);
1436 }
1437
1438 /**
1439  * inactive_file_is_low - check if file pages need to be deactivated
1440  * @zone: zone to check
1441  * @sc:   scan control of this context
1442  *
1443  * When the system is doing streaming IO, memory pressure here
1444  * ensures that active file pages get deactivated, until more
1445  * than half of the file pages are on the inactive list.
1446  *
1447  * Once we get to that situation, protect the system's working
1448  * set from being evicted by disabling active file page aging.
1449  *
1450  * This uses a different ratio than the anonymous pages, because
1451  * the page cache uses a use-once replacement algorithm.
1452  */
1453 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1454 {
1455         int low;
1456
1457         if (scanning_global_lru(sc))
1458                 low = inactive_file_is_low_global(zone);
1459         else
1460                 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1461         return low;
1462 }
1463
1464 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1465                                 int file)
1466 {
1467         if (file)
1468                 return inactive_file_is_low(zone, sc);
1469         else
1470                 return inactive_anon_is_low(zone, sc);
1471 }
1472
1473 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1474         struct zone *zone, struct scan_control *sc, int priority)
1475 {
1476         int file = is_file_lru(lru);
1477
1478         if (is_active_lru(lru)) {
1479                 if (inactive_list_is_low(zone, sc, file))
1480                     shrink_active_list(nr_to_scan, zone, sc, priority, file);
1481                 return 0;
1482         }
1483
1484         return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1485 }
1486
1487 /*
1488  * Determine how aggressively the anon and file LRU lists should be
1489  * scanned.  The relative value of each set of LRU lists is determined
1490  * by looking at the fraction of the pages scanned we did rotate back
1491  * onto the active list instead of evict.
1492  *
1493  * percent[0] specifies how much pressure to put on ram/swap backed
1494  * memory, while percent[1] determines pressure on the file LRUs.
1495  */
1496 static void get_scan_ratio(struct zone *zone, struct scan_control *sc,
1497                                         unsigned long *percent)
1498 {
1499         unsigned long anon, file, free;
1500         unsigned long anon_prio, file_prio;
1501         unsigned long ap, fp;
1502         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1503
1504         anon  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1505                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1506         file  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1507                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1508
1509         if (scanning_global_lru(sc)) {
1510                 free  = zone_page_state(zone, NR_FREE_PAGES);
1511                 /* If we have very few page cache pages,
1512                    force-scan anon pages. */
1513                 if (unlikely(file + free <= high_wmark_pages(zone))) {
1514                         percent[0] = 100;
1515                         percent[1] = 0;
1516                         return;
1517                 }
1518         }
1519
1520         /*
1521          * OK, so we have swap space and a fair amount of page cache
1522          * pages.  We use the recently rotated / recently scanned
1523          * ratios to determine how valuable each cache is.
1524          *
1525          * Because workloads change over time (and to avoid overflow)
1526          * we keep these statistics as a floating average, which ends
1527          * up weighing recent references more than old ones.
1528          *
1529          * anon in [0], file in [1]
1530          */
1531         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1532                 spin_lock_irq(&zone->lru_lock);
1533                 reclaim_stat->recent_scanned[0] /= 2;
1534                 reclaim_stat->recent_rotated[0] /= 2;
1535                 spin_unlock_irq(&zone->lru_lock);
1536         }
1537
1538         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1539                 spin_lock_irq(&zone->lru_lock);
1540                 reclaim_stat->recent_scanned[1] /= 2;
1541                 reclaim_stat->recent_rotated[1] /= 2;
1542                 spin_unlock_irq(&zone->lru_lock);
1543         }
1544
1545         /*
1546          * With swappiness at 100, anonymous and file have the same priority.
1547          * This scanning priority is essentially the inverse of IO cost.
1548          */
1549         anon_prio = sc->swappiness;
1550         file_prio = 200 - sc->swappiness;
1551
1552         /*
1553          * The amount of pressure on anon vs file pages is inversely
1554          * proportional to the fraction of recently scanned pages on
1555          * each list that were recently referenced and in active use.
1556          */
1557         ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1558         ap /= reclaim_stat->recent_rotated[0] + 1;
1559
1560         fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1561         fp /= reclaim_stat->recent_rotated[1] + 1;
1562
1563         /* Normalize to percentages */
1564         percent[0] = 100 * ap / (ap + fp + 1);
1565         percent[1] = 100 - percent[0];
1566 }
1567
1568 /*
1569  * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1570  * until we collected @swap_cluster_max pages to scan.
1571  */
1572 static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
1573                                        unsigned long *nr_saved_scan)
1574 {
1575         unsigned long nr;
1576
1577         *nr_saved_scan += nr_to_scan;
1578         nr = *nr_saved_scan;
1579
1580         if (nr >= SWAP_CLUSTER_MAX)
1581                 *nr_saved_scan = 0;
1582         else
1583                 nr = 0;
1584
1585         return nr;
1586 }
1587
1588 /*
1589  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1590  */
1591 static void shrink_zone(int priority, struct zone *zone,
1592                                 struct scan_control *sc)
1593 {
1594         unsigned long nr[NR_LRU_LISTS];
1595         unsigned long nr_to_scan;
1596         unsigned long percent[2];       /* anon @ 0; file @ 1 */
1597         enum lru_list l;
1598         unsigned long nr_reclaimed = sc->nr_reclaimed;
1599         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1600         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1601         int noswap = 0;
1602
1603         /* If we have no swap space, do not bother scanning anon pages. */
1604         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1605                 noswap = 1;
1606                 percent[0] = 0;
1607                 percent[1] = 100;
1608         } else
1609                 get_scan_ratio(zone, sc, percent);
1610
1611         for_each_evictable_lru(l) {
1612                 int file = is_file_lru(l);
1613                 unsigned long scan;
1614
1615                 scan = zone_nr_lru_pages(zone, sc, l);
1616                 if (priority || noswap) {
1617                         scan >>= priority;
1618                         scan = (scan * percent[file]) / 100;
1619                 }
1620                 nr[l] = nr_scan_try_batch(scan,
1621                                           &reclaim_stat->nr_saved_scan[l]);
1622         }
1623
1624         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1625                                         nr[LRU_INACTIVE_FILE]) {
1626                 for_each_evictable_lru(l) {
1627                         if (nr[l]) {
1628                                 nr_to_scan = min_t(unsigned long,
1629                                                    nr[l], SWAP_CLUSTER_MAX);
1630                                 nr[l] -= nr_to_scan;
1631
1632                                 nr_reclaimed += shrink_list(l, nr_to_scan,
1633                                                             zone, sc, priority);
1634                         }
1635                 }
1636                 /*
1637                  * On large memory systems, scan >> priority can become
1638                  * really large. This is fine for the starting priority;
1639                  * we want to put equal scanning pressure on each zone.
1640                  * However, if the VM has a harder time of freeing pages,
1641                  * with multiple processes reclaiming pages, the total
1642                  * freeing target can get unreasonably large.
1643                  */
1644                 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
1645                         break;
1646         }
1647
1648         sc->nr_reclaimed = nr_reclaimed;
1649
1650         /*
1651          * Even if we did not try to evict anon pages at all, we want to
1652          * rebalance the anon lru active/inactive ratio.
1653          */
1654         if (inactive_anon_is_low(zone, sc) && nr_swap_pages > 0)
1655                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1656
1657         throttle_vm_writeout(sc->gfp_mask);
1658 }
1659
1660 /*
1661  * This is the direct reclaim path, for page-allocating processes.  We only
1662  * try to reclaim pages from zones which will satisfy the caller's allocation
1663  * request.
1664  *
1665  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1666  * Because:
1667  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1668  *    allocation or
1669  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1670  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1671  *    zone defense algorithm.
1672  *
1673  * If a zone is deemed to be full of pinned pages then just give it a light
1674  * scan then give up on it.
1675  */
1676 static void shrink_zones(int priority, struct zonelist *zonelist,
1677                                         struct scan_control *sc)
1678 {
1679         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1680         struct zoneref *z;
1681         struct zone *zone;
1682
1683         sc->all_unreclaimable = 1;
1684         for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1685                                         sc->nodemask) {
1686                 if (!populated_zone(zone))
1687                         continue;
1688                 /*
1689                  * Take care memory controller reclaiming has small influence
1690                  * to global LRU.
1691                  */
1692                 if (scanning_global_lru(sc)) {
1693                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1694                                 continue;
1695                         note_zone_scanning_priority(zone, priority);
1696
1697                         if (zone_is_all_unreclaimable(zone) &&
1698                                                 priority != DEF_PRIORITY)
1699                                 continue;       /* Let kswapd poll it */
1700                         sc->all_unreclaimable = 0;
1701                 } else {
1702                         /*
1703                          * Ignore cpuset limitation here. We just want to reduce
1704                          * # of used pages by us regardless of memory shortage.
1705                          */
1706                         sc->all_unreclaimable = 0;
1707                         mem_cgroup_note_reclaim_priority(sc->mem_cgroup,
1708                                                         priority);
1709                 }
1710
1711                 shrink_zone(priority, zone, sc);
1712         }
1713 }
1714
1715 /*
1716  * This is the main entry point to direct page reclaim.
1717  *
1718  * If a full scan of the inactive list fails to free enough memory then we
1719  * are "out of memory" and something needs to be killed.
1720  *
1721  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1722  * high - the zone may be full of dirty or under-writeback pages, which this
1723  * caller can't do much about.  We kick the writeback threads and take explicit
1724  * naps in the hope that some of these pages can be written.  But if the
1725  * allocating task holds filesystem locks which prevent writeout this might not
1726  * work, and the allocation attempt will fail.
1727  *
1728  * returns:     0, if no pages reclaimed
1729  *              else, the number of pages reclaimed
1730  */
1731 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
1732                                         struct scan_control *sc)
1733 {
1734         int priority;
1735         unsigned long ret = 0;
1736         unsigned long total_scanned = 0;
1737         struct reclaim_state *reclaim_state = current->reclaim_state;
1738         unsigned long lru_pages = 0;
1739         struct zoneref *z;
1740         struct zone *zone;
1741         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1742         unsigned long writeback_threshold;
1743
1744         delayacct_freepages_start();
1745
1746         if (scanning_global_lru(sc))
1747                 count_vm_event(ALLOCSTALL);
1748         /*
1749          * mem_cgroup will not do shrink_slab.
1750          */
1751         if (scanning_global_lru(sc)) {
1752                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1753
1754                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1755                                 continue;
1756
1757                         lru_pages += zone_reclaimable_pages(zone);
1758                 }
1759         }
1760
1761         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1762                 sc->nr_scanned = 0;
1763                 if (!priority)
1764                         disable_swap_token();
1765                 shrink_zones(priority, zonelist, sc);
1766                 /*
1767                  * Don't shrink slabs when reclaiming memory from
1768                  * over limit cgroups
1769                  */
1770                 if (scanning_global_lru(sc)) {
1771                         shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
1772                         if (reclaim_state) {
1773                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
1774                                 reclaim_state->reclaimed_slab = 0;
1775                         }
1776                 }
1777                 total_scanned += sc->nr_scanned;
1778                 if (sc->nr_reclaimed >= sc->nr_to_reclaim) {
1779                         ret = sc->nr_reclaimed;
1780                         goto out;
1781                 }
1782
1783                 /*
1784                  * Try to write back as many pages as we just scanned.  This
1785                  * tends to cause slow streaming writers to write data to the
1786                  * disk smoothly, at the dirtying rate, which is nice.   But
1787                  * that's undesirable in laptop mode, where we *want* lumpy
1788                  * writeout.  So in laptop mode, write out the whole world.
1789                  */
1790                 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
1791                 if (total_scanned > writeback_threshold) {
1792                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
1793                         sc->may_writepage = 1;
1794                 }
1795
1796                 /* Take a nap, wait for some writeback to complete */
1797                 if (!sc->hibernation_mode && sc->nr_scanned &&
1798                     priority < DEF_PRIORITY - 2)
1799                         congestion_wait(BLK_RW_ASYNC, HZ/10);
1800         }
1801         /* top priority shrink_zones still had more to do? don't OOM, then */
1802         if (!sc->all_unreclaimable && scanning_global_lru(sc))
1803                 ret = sc->nr_reclaimed;
1804 out:
1805         /*
1806          * Now that we've scanned all the zones at this priority level, note
1807          * that level within the zone so that the next thread which performs
1808          * scanning of this zone will immediately start out at this priority
1809          * level.  This affects only the decision whether or not to bring
1810          * mapped pages onto the inactive list.
1811          */
1812         if (priority < 0)
1813                 priority = 0;
1814
1815         if (scanning_global_lru(sc)) {
1816                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1817
1818                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1819                                 continue;
1820
1821                         zone->prev_priority = priority;
1822                 }
1823         } else
1824                 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority);
1825
1826         delayacct_freepages_end();
1827
1828         return ret;
1829 }
1830
1831 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
1832                                 gfp_t gfp_mask, nodemask_t *nodemask)
1833 {
1834         struct scan_control sc = {
1835                 .gfp_mask = gfp_mask,
1836                 .may_writepage = !laptop_mode,
1837                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
1838                 .may_unmap = 1,
1839                 .may_swap = 1,
1840                 .swappiness = vm_swappiness,
1841                 .order = order,
1842                 .mem_cgroup = NULL,
1843                 .isolate_pages = isolate_pages_global,
1844                 .nodemask = nodemask,
1845         };
1846
1847         return do_try_to_free_pages(zonelist, &sc);
1848 }
1849
1850 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
1851
1852 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
1853                                                 gfp_t gfp_mask, bool noswap,
1854                                                 unsigned int swappiness,
1855                                                 struct zone *zone, int nid)
1856 {
1857         struct scan_control sc = {
1858                 .may_writepage = !laptop_mode,
1859                 .may_unmap = 1,
1860                 .may_swap = !noswap,
1861                 .swappiness = swappiness,
1862                 .order = 0,
1863                 .mem_cgroup = mem,
1864                 .isolate_pages = mem_cgroup_isolate_pages,
1865         };
1866         nodemask_t nm  = nodemask_of_node(nid);
1867
1868         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1869                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1870         sc.nodemask = &nm;
1871         sc.nr_reclaimed = 0;
1872         sc.nr_scanned = 0;
1873         /*
1874          * NOTE: Although we can get the priority field, using it
1875          * here is not a good idea, since it limits the pages we can scan.
1876          * if we don't reclaim here, the shrink_zone from balance_pgdat
1877          * will pick up pages from other mem cgroup's as well. We hack
1878          * the priority and make it zero.
1879          */
1880         shrink_zone(0, zone, &sc);
1881         return sc.nr_reclaimed;
1882 }
1883
1884 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
1885                                            gfp_t gfp_mask,
1886                                            bool noswap,
1887                                            unsigned int swappiness)
1888 {
1889         struct zonelist *zonelist;
1890         struct scan_control sc = {
1891                 .may_writepage = !laptop_mode,
1892                 .may_unmap = 1,
1893                 .may_swap = !noswap,
1894                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
1895                 .swappiness = swappiness,
1896                 .order = 0,
1897                 .mem_cgroup = mem_cont,
1898                 .isolate_pages = mem_cgroup_isolate_pages,
1899                 .nodemask = NULL, /* we don't care the placement */
1900         };
1901
1902         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1903                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1904         zonelist = NODE_DATA(numa_node_id())->node_zonelists;
1905         return do_try_to_free_pages(zonelist, &sc);
1906 }
1907 #endif
1908
1909 /* is kswapd sleeping prematurely? */
1910 static int sleeping_prematurely(pg_data_t *pgdat, int order, long remaining)
1911 {
1912         int i;
1913
1914         /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
1915         if (remaining)
1916                 return 1;
1917
1918         /* If after HZ/10, a zone is below the high mark, it's premature */
1919         for (i = 0; i < pgdat->nr_zones; i++) {
1920                 struct zone *zone = pgdat->node_zones + i;
1921
1922                 if (!populated_zone(zone))
1923                         continue;
1924
1925                 if (zone_is_all_unreclaimable(zone))
1926                         continue;
1927
1928                 if (!zone_watermark_ok(zone, order, high_wmark_pages(zone),
1929                                                                 0, 0))
1930                         return 1;
1931         }
1932
1933         return 0;
1934 }
1935
1936 /*
1937  * For kswapd, balance_pgdat() will work across all this node's zones until
1938  * they are all at high_wmark_pages(zone).
1939  *
1940  * Returns the number of pages which were actually freed.
1941  *
1942  * There is special handling here for zones which are full of pinned pages.
1943  * This can happen if the pages are all mlocked, or if they are all used by
1944  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1945  * What we do is to detect the case where all pages in the zone have been
1946  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1947  * dead and from now on, only perform a short scan.  Basically we're polling
1948  * the zone for when the problem goes away.
1949  *
1950  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
1951  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
1952  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
1953  * lower zones regardless of the number of free pages in the lower zones. This
1954  * interoperates with the page allocator fallback scheme to ensure that aging
1955  * of pages is balanced across the zones.
1956  */
1957 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
1958 {
1959         int all_zones_ok;
1960         int priority;
1961         int i;
1962         unsigned long total_scanned;
1963         struct reclaim_state *reclaim_state = current->reclaim_state;
1964         struct scan_control sc = {
1965                 .gfp_mask = GFP_KERNEL,
1966                 .may_unmap = 1,
1967                 .may_swap = 1,
1968                 /*
1969                  * kswapd doesn't want to be bailed out while reclaim. because
1970                  * we want to put equal scanning pressure on each zone.
1971                  */
1972                 .nr_to_reclaim = ULONG_MAX,
1973                 .swappiness = vm_swappiness,
1974                 .order = order,
1975                 .mem_cgroup = NULL,
1976                 .isolate_pages = isolate_pages_global,
1977         };
1978         /*
1979          * temp_priority is used to remember the scanning priority at which
1980          * this zone was successfully refilled to
1981          * free_pages == high_wmark_pages(zone).
1982          */
1983         int temp_priority[MAX_NR_ZONES];
1984
1985 loop_again:
1986         total_scanned = 0;
1987         sc.nr_reclaimed = 0;
1988         sc.may_writepage = !laptop_mode;
1989         count_vm_event(PAGEOUTRUN);
1990
1991         for (i = 0; i < pgdat->nr_zones; i++)
1992                 temp_priority[i] = DEF_PRIORITY;
1993
1994         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1995                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1996                 unsigned long lru_pages = 0;
1997                 int has_under_min_watermark_zone = 0;
1998
1999                 /* The swap token gets in the way of swapout... */
2000                 if (!priority)
2001                         disable_swap_token();
2002
2003                 all_zones_ok = 1;
2004
2005                 /*
2006                  * Scan in the highmem->dma direction for the highest
2007                  * zone which needs scanning
2008                  */
2009                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2010                         struct zone *zone = pgdat->node_zones + i;
2011
2012                         if (!populated_zone(zone))
2013                                 continue;
2014
2015                         if (zone_is_all_unreclaimable(zone) &&
2016                             priority != DEF_PRIORITY)
2017                                 continue;
2018
2019                         /*
2020                          * Do some background aging of the anon list, to give
2021                          * pages a chance to be referenced before reclaiming.
2022                          */
2023                         if (inactive_anon_is_low(zone, &sc))
2024                                 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2025                                                         &sc, priority, 0);
2026
2027                         if (!zone_watermark_ok(zone, order,
2028                                         high_wmark_pages(zone), 0, 0)) {
2029                                 end_zone = i;
2030                                 break;
2031                         }
2032                 }
2033                 if (i < 0)
2034                         goto out;
2035
2036                 for (i = 0; i <= end_zone; i++) {
2037                         struct zone *zone = pgdat->node_zones + i;
2038
2039                         lru_pages += zone_reclaimable_pages(zone);
2040                 }
2041
2042                 /*
2043                  * Now scan the zone in the dma->highmem direction, stopping
2044                  * at the last zone which needs scanning.
2045                  *
2046                  * We do this because the page allocator works in the opposite
2047                  * direction.  This prevents the page allocator from allocating
2048                  * pages behind kswapd's direction of progress, which would
2049                  * cause too much scanning of the lower zones.
2050                  */
2051                 for (i = 0; i <= end_zone; i++) {
2052                         struct zone *zone = pgdat->node_zones + i;
2053                         int nr_slab;
2054                         int nid, zid;
2055
2056                         if (!populated_zone(zone))
2057                                 continue;
2058
2059                         if (zone_is_all_unreclaimable(zone) &&
2060                                         priority != DEF_PRIORITY)
2061                                 continue;
2062
2063                         temp_priority[i] = priority;
2064                         sc.nr_scanned = 0;
2065                         note_zone_scanning_priority(zone, priority);
2066
2067                         nid = pgdat->node_id;
2068                         zid = zone_idx(zone);
2069                         /*
2070                          * Call soft limit reclaim before calling shrink_zone.
2071                          * For now we ignore the return value
2072                          */
2073                         mem_cgroup_soft_limit_reclaim(zone, order, sc.gfp_mask,
2074                                                         nid, zid);
2075                         /*
2076                          * We put equal pressure on every zone, unless one
2077                          * zone has way too many pages free already.
2078                          */
2079                         if (!zone_watermark_ok(zone, order,
2080                                         8*high_wmark_pages(zone), end_zone, 0))
2081                                 shrink_zone(priority, zone, &sc);
2082                         reclaim_state->reclaimed_slab = 0;
2083                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
2084                                                 lru_pages);
2085                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2086                         total_scanned += sc.nr_scanned;
2087                         if (zone_is_all_unreclaimable(zone))
2088                                 continue;
2089                         if (nr_slab == 0 && zone->pages_scanned >=
2090                                         (zone_reclaimable_pages(zone) * 6))
2091                                         zone_set_flag(zone,
2092                                                       ZONE_ALL_UNRECLAIMABLE);
2093                         /*
2094                          * If we've done a decent amount of scanning and
2095                          * the reclaim ratio is low, start doing writepage
2096                          * even in laptop mode
2097                          */
2098                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2099                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2100                                 sc.may_writepage = 1;
2101
2102                         if (!zone_watermark_ok(zone, order,
2103                                         high_wmark_pages(zone), end_zone, 0)) {
2104                                 all_zones_ok = 0;
2105                                 /*
2106                                  * We are still under min water mark.  This
2107                                  * means that we have a GFP_ATOMIC allocation
2108                                  * failure risk. Hurry up!
2109                                  */
2110                                 if (!zone_watermark_ok(zone, order,
2111                                             min_wmark_pages(zone), end_zone, 0))
2112                                         has_under_min_watermark_zone = 1;
2113                         }
2114
2115                 }
2116                 if (all_zones_ok)
2117                         break;          /* kswapd: all done */
2118                 /*
2119                  * OK, kswapd is getting into trouble.  Take a nap, then take
2120                  * another pass across the zones.
2121                  */
2122                 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2123                         if (has_under_min_watermark_zone)
2124                                 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2125                         else
2126                                 congestion_wait(BLK_RW_ASYNC, HZ/10);
2127                 }
2128
2129                 /*
2130                  * We do this so kswapd doesn't build up large priorities for
2131                  * example when it is freeing in parallel with allocators. It
2132                  * matches the direct reclaim path behaviour in terms of impact
2133                  * on zone->*_priority.
2134                  */
2135                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2136                         break;
2137         }
2138 out:
2139         /*
2140          * Note within each zone the priority level at which this zone was
2141          * brought into a happy state.  So that the next thread which scans this
2142          * zone will start out at that priority level.
2143          */
2144         for (i = 0; i < pgdat->nr_zones; i++) {
2145                 struct zone *zone = pgdat->node_zones + i;
2146
2147                 zone->prev_priority = temp_priority[i];
2148         }
2149         if (!all_zones_ok) {
2150                 cond_resched();
2151
2152                 try_to_freeze();
2153
2154                 /*
2155                  * Fragmentation may mean that the system cannot be
2156                  * rebalanced for high-order allocations in all zones.
2157                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2158                  * it means the zones have been fully scanned and are still
2159                  * not balanced. For high-order allocations, there is
2160                  * little point trying all over again as kswapd may
2161                  * infinite loop.
2162                  *
2163                  * Instead, recheck all watermarks at order-0 as they
2164                  * are the most important. If watermarks are ok, kswapd will go
2165                  * back to sleep. High-order users can still perform direct
2166                  * reclaim if they wish.
2167                  */
2168                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2169                         order = sc.order = 0;
2170
2171                 goto loop_again;
2172         }
2173
2174         return sc.nr_reclaimed;
2175 }
2176
2177 /*
2178  * The background pageout daemon, started as a kernel thread
2179  * from the init process.
2180  *
2181  * This basically trickles out pages so that we have _some_
2182  * free memory available even if there is no other activity
2183  * that frees anything up. This is needed for things like routing
2184  * etc, where we otherwise might have all activity going on in
2185  * asynchronous contexts that cannot page things out.
2186  *
2187  * If there are applications that are active memory-allocators
2188  * (most normal use), this basically shouldn't matter.
2189  */
2190 static int kswapd(void *p)
2191 {
2192         unsigned long order;
2193         pg_data_t *pgdat = (pg_data_t*)p;
2194         struct task_struct *tsk = current;
2195         DEFINE_WAIT(wait);
2196         struct reclaim_state reclaim_state = {
2197                 .reclaimed_slab = 0,
2198         };
2199         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2200
2201         lockdep_set_current_reclaim_state(GFP_KERNEL);
2202
2203         if (!cpumask_empty(cpumask))
2204                 set_cpus_allowed_ptr(tsk, cpumask);
2205         current->reclaim_state = &reclaim_state;
2206
2207         /*
2208          * Tell the memory management that we're a "memory allocator",
2209          * and that if we need more memory we should get access to it
2210          * regardless (see "__alloc_pages()"). "kswapd" should
2211          * never get caught in the normal page freeing logic.
2212          *
2213          * (Kswapd normally doesn't need memory anyway, but sometimes
2214          * you need a small amount of memory in order to be able to
2215          * page out something else, and this flag essentially protects
2216          * us from recursively trying to free more memory as we're
2217          * trying to free the first piece of memory in the first place).
2218          */
2219         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2220         set_freezable();
2221
2222         order = 0;
2223         for ( ; ; ) {
2224                 unsigned long new_order;
2225                 int ret;
2226
2227                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2228                 new_order = pgdat->kswapd_max_order;
2229                 pgdat->kswapd_max_order = 0;
2230                 if (order < new_order) {
2231                         /*
2232                          * Don't sleep if someone wants a larger 'order'
2233                          * allocation
2234                          */
2235                         order = new_order;
2236                 } else {
2237                         if (!freezing(current) && !kthread_should_stop()) {
2238                                 long remaining = 0;
2239
2240                                 /* Try to sleep for a short interval */
2241                                 if (!sleeping_prematurely(pgdat, order, remaining)) {
2242                                         remaining = schedule_timeout(HZ/10);
2243                                         finish_wait(&pgdat->kswapd_wait, &wait);
2244                                         prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2245                                 }
2246
2247                                 /*
2248                                  * After a short sleep, check if it was a
2249                                  * premature sleep. If not, then go fully
2250                                  * to sleep until explicitly woken up
2251                                  */
2252                                 if (!sleeping_prematurely(pgdat, order, remaining))
2253                                         schedule();
2254                                 else {
2255                                         if (remaining)
2256                                                 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2257                                         else
2258                                                 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2259                                 }
2260                         }
2261
2262                         order = pgdat->kswapd_max_order;
2263                 }
2264                 finish_wait(&pgdat->kswapd_wait, &wait);
2265
2266                 ret = try_to_freeze();
2267                 if (kthread_should_stop())
2268                         break;
2269
2270                 /*
2271                  * We can speed up thawing tasks if we don't call balance_pgdat
2272                  * after returning from the refrigerator
2273                  */
2274                 if (!ret)
2275                         balance_pgdat(pgdat, order);
2276         }
2277         return 0;
2278 }
2279
2280 /*
2281  * A zone is low on free memory, so wake its kswapd task to service it.
2282  */
2283 void wakeup_kswapd(struct zone *zone, int order)
2284 {
2285         pg_data_t *pgdat;
2286
2287         if (!populated_zone(zone))
2288                 return;
2289
2290         pgdat = zone->zone_pgdat;
2291         if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
2292                 return;
2293         if (pgdat->kswapd_max_order < order)
2294                 pgdat->kswapd_max_order = order;
2295         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2296                 return;
2297         if (!waitqueue_active(&pgdat->kswapd_wait))
2298                 return;
2299         wake_up_interruptible(&pgdat->kswapd_wait);
2300 }
2301
2302 /*
2303  * The reclaimable count would be mostly accurate.
2304  * The less reclaimable pages may be
2305  * - mlocked pages, which will be moved to unevictable list when encountered
2306  * - mapped pages, which may require several travels to be reclaimed
2307  * - dirty pages, which is not "instantly" reclaimable
2308  */
2309 unsigned long global_reclaimable_pages(void)
2310 {
2311         int nr;
2312
2313         nr = global_page_state(NR_ACTIVE_FILE) +
2314              global_page_state(NR_INACTIVE_FILE);
2315
2316         if (nr_swap_pages > 0)
2317                 nr += global_page_state(NR_ACTIVE_ANON) +
2318                       global_page_state(NR_INACTIVE_ANON);
2319
2320         return nr;
2321 }
2322
2323 unsigned long zone_reclaimable_pages(struct zone *zone)
2324 {
2325         int nr;
2326
2327         nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2328              zone_page_state(zone, NR_INACTIVE_FILE);
2329
2330         if (nr_swap_pages > 0)
2331                 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2332                       zone_page_state(zone, NR_INACTIVE_ANON);
2333
2334         return nr;
2335 }
2336
2337 #ifdef CONFIG_HIBERNATION
2338 /*
2339  * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2340  * freed pages.
2341  *
2342  * Rather than trying to age LRUs the aim is to preserve the overall
2343  * LRU order by reclaiming preferentially
2344  * inactive > active > active referenced > active mapped
2345  */
2346 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
2347 {
2348         struct reclaim_state reclaim_state;
2349         struct scan_control sc = {
2350                 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2351                 .may_swap = 1,
2352                 .may_unmap = 1,
2353                 .may_writepage = 1,
2354                 .nr_to_reclaim = nr_to_reclaim,
2355                 .hibernation_mode = 1,
2356                 .swappiness = vm_swappiness,
2357                 .order = 0,
2358                 .isolate_pages = isolate_pages_global,
2359         };
2360         struct zonelist * zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2361         struct task_struct *p = current;
2362         unsigned long nr_reclaimed;
2363
2364         p->flags |= PF_MEMALLOC;
2365         lockdep_set_current_reclaim_state(sc.gfp_mask);
2366         reclaim_state.reclaimed_slab = 0;
2367         p->reclaim_state = &reclaim_state;
2368
2369         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2370
2371         p->reclaim_state = NULL;
2372         lockdep_clear_current_reclaim_state();
2373         p->flags &= ~PF_MEMALLOC;
2374
2375         return nr_reclaimed;
2376 }
2377 #endif /* CONFIG_HIBERNATION */
2378
2379 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2380    not required for correctness.  So if the last cpu in a node goes
2381    away, we get changed to run anywhere: as the first one comes back,
2382    restore their cpu bindings. */
2383 static int __devinit cpu_callback(struct notifier_block *nfb,
2384                                   unsigned long action, void *hcpu)
2385 {
2386         int nid;
2387
2388         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2389                 for_each_node_state(nid, N_HIGH_MEMORY) {
2390                         pg_data_t *pgdat = NODE_DATA(nid);
2391                         const struct cpumask *mask;
2392
2393                         mask = cpumask_of_node(pgdat->node_id);
2394
2395                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2396                                 /* One of our CPUs online: restore mask */
2397                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2398                 }
2399         }
2400         return NOTIFY_OK;
2401 }
2402
2403 /*
2404  * This kswapd start function will be called by init and node-hot-add.
2405  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2406  */
2407 int kswapd_run(int nid)
2408 {
2409         pg_data_t *pgdat = NODE_DATA(nid);
2410         int ret = 0;
2411
2412         if (pgdat->kswapd)
2413                 return 0;
2414
2415         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2416         if (IS_ERR(pgdat->kswapd)) {
2417                 /* failure at boot is fatal */
2418                 BUG_ON(system_state == SYSTEM_BOOTING);
2419                 printk("Failed to start kswapd on node %d\n",nid);
2420                 ret = -1;
2421         }
2422         return ret;
2423 }
2424
2425 /*
2426  * Called by memory hotplug when all memory in a node is offlined.
2427  */
2428 void kswapd_stop(int nid)
2429 {
2430         struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
2431
2432         if (kswapd)
2433                 kthread_stop(kswapd);
2434 }
2435
2436 static int __init kswapd_init(void)
2437 {
2438         int nid;
2439
2440         swap_setup();
2441         for_each_node_state(nid, N_HIGH_MEMORY)
2442                 kswapd_run(nid);
2443         hotcpu_notifier(cpu_callback, 0);
2444         return 0;
2445 }
2446
2447 module_init(kswapd_init)
2448
2449 #ifdef CONFIG_NUMA
2450 /*
2451  * Zone reclaim mode
2452  *
2453  * If non-zero call zone_reclaim when the number of free pages falls below
2454  * the watermarks.
2455  */
2456 int zone_reclaim_mode __read_mostly;
2457
2458 #define RECLAIM_OFF 0
2459 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
2460 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
2461 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
2462
2463 /*
2464  * Priority for ZONE_RECLAIM. This determines the fraction of pages
2465  * of a node considered for each zone_reclaim. 4 scans 1/16th of
2466  * a zone.
2467  */
2468 #define ZONE_RECLAIM_PRIORITY 4
2469
2470 /*
2471  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2472  * occur.
2473  */
2474 int sysctl_min_unmapped_ratio = 1;
2475
2476 /*
2477  * If the number of slab pages in a zone grows beyond this percentage then
2478  * slab reclaim needs to occur.
2479  */
2480 int sysctl_min_slab_ratio = 5;
2481
2482 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
2483 {
2484         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
2485         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
2486                 zone_page_state(zone, NR_ACTIVE_FILE);
2487
2488         /*
2489          * It's possible for there to be more file mapped pages than
2490          * accounted for by the pages on the file LRU lists because
2491          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2492          */
2493         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
2494 }
2495
2496 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
2497 static long zone_pagecache_reclaimable(struct zone *zone)
2498 {
2499         long nr_pagecache_reclaimable;
2500         long delta = 0;
2501
2502         /*
2503          * If RECLAIM_SWAP is set, then all file pages are considered
2504          * potentially reclaimable. Otherwise, we have to worry about
2505          * pages like swapcache and zone_unmapped_file_pages() provides
2506          * a better estimate
2507          */
2508         if (zone_reclaim_mode & RECLAIM_SWAP)
2509                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
2510         else
2511                 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
2512
2513         /* If we can't clean pages, remove dirty pages from consideration */
2514         if (!(zone_reclaim_mode & RECLAIM_WRITE))
2515                 delta += zone_page_state(zone, NR_FILE_DIRTY);
2516
2517         /* Watch for any possible underflows due to delta */
2518         if (unlikely(delta > nr_pagecache_reclaimable))
2519                 delta = nr_pagecache_reclaimable;
2520
2521         return nr_pagecache_reclaimable - delta;
2522 }
2523
2524 /*
2525  * Try to free up some pages from this zone through reclaim.
2526  */
2527 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2528 {
2529         /* Minimum pages needed in order to stay on node */
2530         const unsigned long nr_pages = 1 << order;
2531         struct task_struct *p = current;
2532         struct reclaim_state reclaim_state;
2533         int priority;
2534         struct scan_control sc = {
2535                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2536                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2537                 .may_swap = 1,
2538                 .nr_to_reclaim = max_t(unsigned long, nr_pages,
2539                                        SWAP_CLUSTER_MAX),
2540                 .gfp_mask = gfp_mask,
2541                 .swappiness = vm_swappiness,
2542                 .order = order,
2543                 .isolate_pages = isolate_pages_global,
2544         };
2545         unsigned long slab_reclaimable;
2546
2547         disable_swap_token();
2548         cond_resched();
2549         /*
2550          * We need to be able to allocate from the reserves for RECLAIM_SWAP
2551          * and we also need to be able to write out pages for RECLAIM_WRITE
2552          * and RECLAIM_SWAP.
2553          */
2554         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2555         reclaim_state.reclaimed_slab = 0;
2556         p->reclaim_state = &reclaim_state;
2557
2558         if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
2559                 /*
2560                  * Free memory by calling shrink zone with increasing
2561                  * priorities until we have enough memory freed.
2562                  */
2563                 priority = ZONE_RECLAIM_PRIORITY;
2564                 do {
2565                         note_zone_scanning_priority(zone, priority);
2566                         shrink_zone(priority, zone, &sc);
2567                         priority--;
2568                 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
2569         }
2570
2571         slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2572         if (slab_reclaimable > zone->min_slab_pages) {
2573                 /*
2574                  * shrink_slab() does not currently allow us to determine how
2575                  * many pages were freed in this zone. So we take the current
2576                  * number of slab pages and shake the slab until it is reduced
2577                  * by the same nr_pages that we used for reclaiming unmapped
2578                  * pages.
2579                  *
2580                  * Note that shrink_slab will free memory on all zones and may
2581                  * take a long time.
2582                  */
2583                 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
2584                         zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
2585                                 slab_reclaimable - nr_pages)
2586                         ;
2587
2588                 /*
2589                  * Update nr_reclaimed by the number of slab pages we
2590                  * reclaimed from this zone.
2591                  */
2592                 sc.nr_reclaimed += slab_reclaimable -
2593                         zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2594         }
2595
2596         p->reclaim_state = NULL;
2597         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
2598         return sc.nr_reclaimed >= nr_pages;
2599 }
2600
2601 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2602 {
2603         int node_id;
2604         int ret;
2605
2606         /*
2607          * Zone reclaim reclaims unmapped file backed pages and
2608          * slab pages if we are over the defined limits.
2609          *
2610          * A small portion of unmapped file backed pages is needed for
2611          * file I/O otherwise pages read by file I/O will be immediately
2612          * thrown out if the zone is overallocated. So we do not reclaim
2613          * if less than a specified percentage of the zone is used by
2614          * unmapped file backed pages.
2615          */
2616         if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
2617             zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
2618                 return ZONE_RECLAIM_FULL;
2619
2620         if (zone_is_all_unreclaimable(zone))
2621                 return ZONE_RECLAIM_FULL;
2622
2623         /*
2624          * Do not scan if the allocation should not be delayed.
2625          */
2626         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
2627                 return ZONE_RECLAIM_NOSCAN;
2628
2629         /*
2630          * Only run zone reclaim on the local zone or on zones that do not
2631          * have associated processors. This will favor the local processor
2632          * over remote processors and spread off node memory allocations
2633          * as wide as possible.
2634          */
2635         node_id = zone_to_nid(zone);
2636         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
2637                 return ZONE_RECLAIM_NOSCAN;
2638
2639         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2640                 return ZONE_RECLAIM_NOSCAN;
2641
2642         ret = __zone_reclaim(zone, gfp_mask, order);
2643         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2644
2645         if (!ret)
2646                 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
2647
2648         return ret;
2649 }
2650 #endif
2651
2652 /*
2653  * page_evictable - test whether a page is evictable
2654  * @page: the page to test
2655  * @vma: the VMA in which the page is or will be mapped, may be NULL
2656  *
2657  * Test whether page is evictable--i.e., should be placed on active/inactive
2658  * lists vs unevictable list.  The vma argument is !NULL when called from the
2659  * fault path to determine how to instantate a new page.
2660  *
2661  * Reasons page might not be evictable:
2662  * (1) page's mapping marked unevictable
2663  * (2) page is part of an mlocked VMA
2664  *
2665  */
2666 int page_evictable(struct page *page, struct vm_area_struct *vma)
2667 {
2668
2669         if (mapping_unevictable(page_mapping(page)))
2670                 return 0;
2671
2672         if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2673                 return 0;
2674
2675         return 1;
2676 }
2677
2678 /**
2679  * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2680  * @page: page to check evictability and move to appropriate lru list
2681  * @zone: zone page is in
2682  *
2683  * Checks a page for evictability and moves the page to the appropriate
2684  * zone lru list.
2685  *
2686  * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2687  * have PageUnevictable set.
2688  */
2689 static void check_move_unevictable_page(struct page *page, struct zone *zone)
2690 {
2691         VM_BUG_ON(PageActive(page));
2692
2693 retry:
2694         ClearPageUnevictable(page);
2695         if (page_evictable(page, NULL)) {
2696                 enum lru_list l = page_lru_base_type(page);
2697
2698                 __dec_zone_state(zone, NR_UNEVICTABLE);
2699                 list_move(&page->lru, &zone->lru[l].list);
2700                 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
2701                 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
2702                 __count_vm_event(UNEVICTABLE_PGRESCUED);
2703         } else {
2704                 /*
2705                  * rotate unevictable list
2706                  */
2707                 SetPageUnevictable(page);
2708                 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
2709                 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
2710                 if (page_evictable(page, NULL))
2711                         goto retry;
2712         }
2713 }
2714
2715 /**
2716  * scan_mapping_unevictable_pages - scan an address space for evictable pages
2717  * @mapping: struct address_space to scan for evictable pages
2718  *
2719  * Scan all pages in mapping.  Check unevictable pages for
2720  * evictability and move them to the appropriate zone lru list.
2721  */
2722 void scan_mapping_unevictable_pages(struct address_space *mapping)
2723 {
2724         pgoff_t next = 0;
2725         pgoff_t end   = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
2726                          PAGE_CACHE_SHIFT;
2727         struct zone *zone;
2728         struct pagevec pvec;
2729
2730         if (mapping->nrpages == 0)
2731                 return;
2732
2733         pagevec_init(&pvec, 0);
2734         while (next < end &&
2735                 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
2736                 int i;
2737                 int pg_scanned = 0;
2738
2739                 zone = NULL;
2740
2741                 for (i = 0; i < pagevec_count(&pvec); i++) {
2742                         struct page *page = pvec.pages[i];
2743                         pgoff_t page_index = page->index;
2744                         struct zone *pagezone = page_zone(page);
2745
2746                         pg_scanned++;
2747                         if (page_index > next)
2748                                 next = page_index;
2749                         next++;
2750
2751                         if (pagezone != zone) {
2752                                 if (zone)
2753                                         spin_unlock_irq(&zone->lru_lock);
2754                                 zone = pagezone;
2755                                 spin_lock_irq(&zone->lru_lock);
2756                         }
2757
2758                         if (PageLRU(page) && PageUnevictable(page))
2759                                 check_move_unevictable_page(page, zone);
2760                 }
2761                 if (zone)
2762                         spin_unlock_irq(&zone->lru_lock);
2763                 pagevec_release(&pvec);
2764
2765                 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
2766         }
2767
2768 }
2769
2770 /**
2771  * scan_zone_unevictable_pages - check unevictable list for evictable pages
2772  * @zone - zone of which to scan the unevictable list
2773  *
2774  * Scan @zone's unevictable LRU lists to check for pages that have become
2775  * evictable.  Move those that have to @zone's inactive list where they
2776  * become candidates for reclaim, unless shrink_inactive_zone() decides
2777  * to reactivate them.  Pages that are still unevictable are rotated
2778  * back onto @zone's unevictable list.
2779  */
2780 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
2781 static void scan_zone_unevictable_pages(struct zone *zone)
2782 {
2783         struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
2784         unsigned long scan;
2785         unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
2786
2787         while (nr_to_scan > 0) {
2788                 unsigned long batch_size = min(nr_to_scan,
2789                                                 SCAN_UNEVICTABLE_BATCH_SIZE);
2790
2791                 spin_lock_irq(&zone->lru_lock);
2792                 for (scan = 0;  scan < batch_size; scan++) {
2793                         struct page *page = lru_to_page(l_unevictable);
2794
2795                         if (!trylock_page(page))
2796                                 continue;
2797
2798                         prefetchw_prev_lru_page(page, l_unevictable, flags);
2799
2800                         if (likely(PageLRU(page) && PageUnevictable(page)))
2801                                 check_move_unevictable_page(page, zone);
2802
2803                         unlock_page(page);
2804                 }
2805                 spin_unlock_irq(&zone->lru_lock);
2806
2807                 nr_to_scan -= batch_size;
2808         }
2809 }
2810
2811
2812 /**
2813  * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
2814  *
2815  * A really big hammer:  scan all zones' unevictable LRU lists to check for
2816  * pages that have become evictable.  Move those back to the zones'
2817  * inactive list where they become candidates for reclaim.
2818  * This occurs when, e.g., we have unswappable pages on the unevictable lists,
2819  * and we add swap to the system.  As such, it runs in the context of a task
2820  * that has possibly/probably made some previously unevictable pages
2821  * evictable.
2822  */
2823 static void scan_all_zones_unevictable_pages(void)
2824 {
2825         struct zone *zone;
2826
2827         for_each_zone(zone) {
2828                 scan_zone_unevictable_pages(zone);
2829         }
2830 }
2831
2832 /*
2833  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
2834  * all nodes' unevictable lists for evictable pages
2835  */
2836 unsigned long scan_unevictable_pages;
2837
2838 int scan_unevictable_handler(struct ctl_table *table, int write,
2839                            void __user *buffer,
2840                            size_t *length, loff_t *ppos)
2841 {
2842         proc_doulongvec_minmax(table, write, buffer, length, ppos);
2843
2844         if (write && *(unsigned long *)table->data)
2845                 scan_all_zones_unevictable_pages();
2846
2847         scan_unevictable_pages = 0;
2848         return 0;
2849 }
2850
2851 /*
2852  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
2853  * a specified node's per zone unevictable lists for evictable pages.
2854  */
2855
2856 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
2857                                           struct sysdev_attribute *attr,
2858                                           char *buf)
2859 {
2860         return sprintf(buf, "0\n");     /* always zero; should fit... */
2861 }
2862
2863 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
2864                                            struct sysdev_attribute *attr,
2865                                         const char *buf, size_t count)
2866 {
2867         struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
2868         struct zone *zone;
2869         unsigned long res;
2870         unsigned long req = strict_strtoul(buf, 10, &res);
2871
2872         if (!req)
2873                 return 1;       /* zero is no-op */
2874
2875         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2876                 if (!populated_zone(zone))
2877                         continue;
2878                 scan_zone_unevictable_pages(zone);
2879         }
2880         return 1;
2881 }
2882
2883
2884 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
2885                         read_scan_unevictable_node,
2886                         write_scan_unevictable_node);
2887
2888 int scan_unevictable_register_node(struct node *node)
2889 {
2890         return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
2891 }
2892
2893 void scan_unevictable_unregister_node(struct node *node)
2894 {
2895         sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
2896 }
2897