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