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