memcg: synchronized LRU
[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
516         /*
517          * page's status can change while we move it among lru. If an evictable
518          * page is on unevictable list, it never be freed. To avoid that,
519          * check after we added it to the list, again.
520          */
521         if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
522                 if (!isolate_lru_page(page)) {
523                         put_page(page);
524                         goto redo;
525                 }
526                 /* This means someone else dropped this page from LRU
527                  * So, it will be freed or putback to LRU again. There is
528                  * nothing to do here.
529                  */
530         }
531
532         if (was_unevictable && lru != LRU_UNEVICTABLE)
533                 count_vm_event(UNEVICTABLE_PGRESCUED);
534         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
535                 count_vm_event(UNEVICTABLE_PGCULLED);
536
537         put_page(page);         /* drop ref from isolate */
538 }
539
540 #else /* CONFIG_UNEVICTABLE_LRU */
541
542 void putback_lru_page(struct page *page)
543 {
544         int lru;
545         VM_BUG_ON(PageLRU(page));
546
547         lru = !!TestClearPageActive(page) + page_is_file_cache(page);
548         lru_cache_add_lru(page, lru);
549         put_page(page);
550 }
551 #endif /* CONFIG_UNEVICTABLE_LRU */
552
553
554 /*
555  * shrink_page_list() returns the number of reclaimed pages
556  */
557 static unsigned long shrink_page_list(struct list_head *page_list,
558                                         struct scan_control *sc,
559                                         enum pageout_io sync_writeback)
560 {
561         LIST_HEAD(ret_pages);
562         struct pagevec freed_pvec;
563         int pgactivate = 0;
564         unsigned long nr_reclaimed = 0;
565
566         cond_resched();
567
568         pagevec_init(&freed_pvec, 1);
569         while (!list_empty(page_list)) {
570                 struct address_space *mapping;
571                 struct page *page;
572                 int may_enter_fs;
573                 int referenced;
574
575                 cond_resched();
576
577                 page = lru_to_page(page_list);
578                 list_del(&page->lru);
579
580                 if (!trylock_page(page))
581                         goto keep;
582
583                 VM_BUG_ON(PageActive(page));
584
585                 sc->nr_scanned++;
586
587                 if (unlikely(!page_evictable(page, NULL)))
588                         goto cull_mlocked;
589
590                 if (!sc->may_swap && page_mapped(page))
591                         goto keep_locked;
592
593                 /* Double the slab pressure for mapped and swapcache pages */
594                 if (page_mapped(page) || PageSwapCache(page))
595                         sc->nr_scanned++;
596
597                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
598                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
599
600                 if (PageWriteback(page)) {
601                         /*
602                          * Synchronous reclaim is performed in two passes,
603                          * first an asynchronous pass over the list to
604                          * start parallel writeback, and a second synchronous
605                          * pass to wait for the IO to complete.  Wait here
606                          * for any page for which writeback has already
607                          * started.
608                          */
609                         if (sync_writeback == PAGEOUT_IO_SYNC && may_enter_fs)
610                                 wait_on_page_writeback(page);
611                         else
612                                 goto keep_locked;
613                 }
614
615                 referenced = page_referenced(page, 1, sc->mem_cgroup);
616                 /* In active use or really unfreeable?  Activate it. */
617                 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER &&
618                                         referenced && page_mapping_inuse(page))
619                         goto activate_locked;
620
621                 /*
622                  * Anonymous process memory has backing store?
623                  * Try to allocate it some swap space here.
624                  */
625                 if (PageAnon(page) && !PageSwapCache(page)) {
626                         if (!(sc->gfp_mask & __GFP_IO))
627                                 goto keep_locked;
628                         if (!add_to_swap(page))
629                                 goto activate_locked;
630                         may_enter_fs = 1;
631                 }
632
633                 mapping = page_mapping(page);
634
635                 /*
636                  * The page is mapped into the page tables of one or more
637                  * processes. Try to unmap it here.
638                  */
639                 if (page_mapped(page) && mapping) {
640                         switch (try_to_unmap(page, 0)) {
641                         case SWAP_FAIL:
642                                 goto activate_locked;
643                         case SWAP_AGAIN:
644                                 goto keep_locked;
645                         case SWAP_MLOCK:
646                                 goto cull_mlocked;
647                         case SWAP_SUCCESS:
648                                 ; /* try to free the page below */
649                         }
650                 }
651
652                 if (PageDirty(page)) {
653                         if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && referenced)
654                                 goto keep_locked;
655                         if (!may_enter_fs)
656                                 goto keep_locked;
657                         if (!sc->may_writepage)
658                                 goto keep_locked;
659
660                         /* Page is dirty, try to write it out here */
661                         switch (pageout(page, mapping, sync_writeback)) {
662                         case PAGE_KEEP:
663                                 goto keep_locked;
664                         case PAGE_ACTIVATE:
665                                 goto activate_locked;
666                         case PAGE_SUCCESS:
667                                 if (PageWriteback(page) || PageDirty(page))
668                                         goto keep;
669                                 /*
670                                  * A synchronous write - probably a ramdisk.  Go
671                                  * ahead and try to reclaim the page.
672                                  */
673                                 if (!trylock_page(page))
674                                         goto keep;
675                                 if (PageDirty(page) || PageWriteback(page))
676                                         goto keep_locked;
677                                 mapping = page_mapping(page);
678                         case PAGE_CLEAN:
679                                 ; /* try to free the page below */
680                         }
681                 }
682
683                 /*
684                  * If the page has buffers, try to free the buffer mappings
685                  * associated with this page. If we succeed we try to free
686                  * the page as well.
687                  *
688                  * We do this even if the page is PageDirty().
689                  * try_to_release_page() does not perform I/O, but it is
690                  * possible for a page to have PageDirty set, but it is actually
691                  * clean (all its buffers are clean).  This happens if the
692                  * buffers were written out directly, with submit_bh(). ext3
693                  * will do this, as well as the blockdev mapping.
694                  * try_to_release_page() will discover that cleanness and will
695                  * drop the buffers and mark the page clean - it can be freed.
696                  *
697                  * Rarely, pages can have buffers and no ->mapping.  These are
698                  * the pages which were not successfully invalidated in
699                  * truncate_complete_page().  We try to drop those buffers here
700                  * and if that worked, and the page is no longer mapped into
701                  * process address space (page_count == 1) it can be freed.
702                  * Otherwise, leave the page on the LRU so it is swappable.
703                  */
704                 if (PagePrivate(page)) {
705                         if (!try_to_release_page(page, sc->gfp_mask))
706                                 goto activate_locked;
707                         if (!mapping && page_count(page) == 1) {
708                                 unlock_page(page);
709                                 if (put_page_testzero(page))
710                                         goto free_it;
711                                 else {
712                                         /*
713                                          * rare race with speculative reference.
714                                          * the speculative reference will free
715                                          * this page shortly, so we may
716                                          * increment nr_reclaimed here (and
717                                          * leave it off the LRU).
718                                          */
719                                         nr_reclaimed++;
720                                         continue;
721                                 }
722                         }
723                 }
724
725                 if (!mapping || !__remove_mapping(mapping, page))
726                         goto keep_locked;
727
728                 /*
729                  * At this point, we have no other references and there is
730                  * no way to pick any more up (removed from LRU, removed
731                  * from pagecache). Can use non-atomic bitops now (and
732                  * we obviously don't have to worry about waking up a process
733                  * waiting on the page lock, because there are no references.
734                  */
735                 __clear_page_locked(page);
736 free_it:
737                 nr_reclaimed++;
738                 if (!pagevec_add(&freed_pvec, page)) {
739                         __pagevec_free(&freed_pvec);
740                         pagevec_reinit(&freed_pvec);
741                 }
742                 continue;
743
744 cull_mlocked:
745                 if (PageSwapCache(page))
746                         try_to_free_swap(page);
747                 unlock_page(page);
748                 putback_lru_page(page);
749                 continue;
750
751 activate_locked:
752                 /* Not a candidate for swapping, so reclaim swap space. */
753                 if (PageSwapCache(page) && vm_swap_full())
754                         try_to_free_swap(page);
755                 VM_BUG_ON(PageActive(page));
756                 SetPageActive(page);
757                 pgactivate++;
758 keep_locked:
759                 unlock_page(page);
760 keep:
761                 list_add(&page->lru, &ret_pages);
762                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
763         }
764         list_splice(&ret_pages, page_list);
765         if (pagevec_count(&freed_pvec))
766                 __pagevec_free(&freed_pvec);
767         count_vm_events(PGACTIVATE, pgactivate);
768         return nr_reclaimed;
769 }
770
771 /* LRU Isolation modes. */
772 #define ISOLATE_INACTIVE 0      /* Isolate inactive pages. */
773 #define ISOLATE_ACTIVE 1        /* Isolate active pages. */
774 #define ISOLATE_BOTH 2          /* Isolate both active and inactive pages. */
775
776 /*
777  * Attempt to remove the specified page from its LRU.  Only take this page
778  * if it is of the appropriate PageActive status.  Pages which are being
779  * freed elsewhere are also ignored.
780  *
781  * page:        page to consider
782  * mode:        one of the LRU isolation modes defined above
783  *
784  * returns 0 on success, -ve errno on failure.
785  */
786 int __isolate_lru_page(struct page *page, int mode, int file)
787 {
788         int ret = -EINVAL;
789
790         /* Only take pages on the LRU. */
791         if (!PageLRU(page))
792                 return ret;
793
794         /*
795          * When checking the active state, we need to be sure we are
796          * dealing with comparible boolean values.  Take the logical not
797          * of each.
798          */
799         if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
800                 return ret;
801
802         if (mode != ISOLATE_BOTH && (!page_is_file_cache(page) != !file))
803                 return ret;
804
805         /*
806          * When this function is being called for lumpy reclaim, we
807          * initially look into all LRU pages, active, inactive and
808          * unevictable; only give shrink_page_list evictable pages.
809          */
810         if (PageUnevictable(page))
811                 return ret;
812
813         ret = -EBUSY;
814
815         if (likely(get_page_unless_zero(page))) {
816                 /*
817                  * Be careful not to clear PageLRU until after we're
818                  * sure the page is not being freed elsewhere -- the
819                  * page release code relies on it.
820                  */
821                 ClearPageLRU(page);
822                 ret = 0;
823                 mem_cgroup_del_lru(page);
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                         if (PageActive(page) && scan_global_lru(sc)) {
1138                                 int file = !!page_is_file_cache(page);
1139                                 zone->recent_rotated[file]++;
1140                         }
1141                         if (!pagevec_add(&pvec, page)) {
1142                                 spin_unlock_irq(&zone->lru_lock);
1143                                 __pagevec_release(&pvec);
1144                                 spin_lock_irq(&zone->lru_lock);
1145                         }
1146                 }
1147         } while (nr_scanned < max_scan);
1148         spin_unlock(&zone->lru_lock);
1149 done:
1150         local_irq_enable();
1151         pagevec_release(&pvec);
1152         return nr_reclaimed;
1153 }
1154
1155 /*
1156  * We are about to scan this zone at a certain priority level.  If that priority
1157  * level is smaller (ie: more urgent) than the previous priority, then note
1158  * that priority level within the zone.  This is done so that when the next
1159  * process comes in to scan this zone, it will immediately start out at this
1160  * priority level rather than having to build up its own scanning priority.
1161  * Here, this priority affects only the reclaim-mapped threshold.
1162  */
1163 static inline void note_zone_scanning_priority(struct zone *zone, int priority)
1164 {
1165         if (priority < zone->prev_priority)
1166                 zone->prev_priority = priority;
1167 }
1168
1169 /*
1170  * This moves pages from the active list to the inactive list.
1171  *
1172  * We move them the other way if the page is referenced by one or more
1173  * processes, from rmap.
1174  *
1175  * If the pages are mostly unmapped, the processing is fast and it is
1176  * appropriate to hold zone->lru_lock across the whole operation.  But if
1177  * the pages are mapped, the processing is slow (page_referenced()) so we
1178  * should drop zone->lru_lock around each page.  It's impossible to balance
1179  * this, so instead we remove the pages from the LRU while processing them.
1180  * It is safe to rely on PG_active against the non-LRU pages in here because
1181  * nobody will play with that bit on a non-LRU page.
1182  *
1183  * The downside is that we have to touch page->_count against each page.
1184  * But we had to alter page->flags anyway.
1185  */
1186
1187
1188 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1189                         struct scan_control *sc, int priority, int file)
1190 {
1191         unsigned long pgmoved;
1192         int pgdeactivate = 0;
1193         unsigned long pgscanned;
1194         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1195         LIST_HEAD(l_inactive);
1196         struct page *page;
1197         struct pagevec pvec;
1198         enum lru_list lru;
1199
1200         lru_add_drain();
1201         spin_lock_irq(&zone->lru_lock);
1202         pgmoved = sc->isolate_pages(nr_pages, &l_hold, &pgscanned, sc->order,
1203                                         ISOLATE_ACTIVE, zone,
1204                                         sc->mem_cgroup, 1, file);
1205         /*
1206          * zone->pages_scanned is used for detect zone's oom
1207          * mem_cgroup remembers nr_scan by itself.
1208          */
1209         if (scan_global_lru(sc)) {
1210                 zone->pages_scanned += pgscanned;
1211                 zone->recent_scanned[!!file] += pgmoved;
1212         }
1213
1214         if (file)
1215                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -pgmoved);
1216         else
1217                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -pgmoved);
1218         spin_unlock_irq(&zone->lru_lock);
1219
1220         pgmoved = 0;
1221         while (!list_empty(&l_hold)) {
1222                 cond_resched();
1223                 page = lru_to_page(&l_hold);
1224                 list_del(&page->lru);
1225
1226                 if (unlikely(!page_evictable(page, NULL))) {
1227                         putback_lru_page(page);
1228                         continue;
1229                 }
1230
1231                 /* page_referenced clears PageReferenced */
1232                 if (page_mapping_inuse(page) &&
1233                     page_referenced(page, 0, sc->mem_cgroup))
1234                         pgmoved++;
1235
1236                 list_add(&page->lru, &l_inactive);
1237         }
1238
1239         /*
1240          * Move the pages to the [file or anon] inactive list.
1241          */
1242         pagevec_init(&pvec, 1);
1243         pgmoved = 0;
1244         lru = LRU_BASE + file * LRU_FILE;
1245
1246         spin_lock_irq(&zone->lru_lock);
1247         /*
1248          * Count referenced pages from currently used mappings as
1249          * rotated, even though they are moved to the inactive list.
1250          * This helps balance scan pressure between file and anonymous
1251          * pages in get_scan_ratio.
1252          */
1253         if (scan_global_lru(sc))
1254                 zone->recent_rotated[!!file] += pgmoved;
1255
1256         while (!list_empty(&l_inactive)) {
1257                 page = lru_to_page(&l_inactive);
1258                 prefetchw_prev_lru_page(page, &l_inactive, flags);
1259                 VM_BUG_ON(PageLRU(page));
1260                 SetPageLRU(page);
1261                 VM_BUG_ON(!PageActive(page));
1262                 ClearPageActive(page);
1263
1264                 list_move(&page->lru, &zone->lru[lru].list);
1265                 mem_cgroup_add_lru_list(page, lru);
1266                 pgmoved++;
1267                 if (!pagevec_add(&pvec, page)) {
1268                         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1269                         spin_unlock_irq(&zone->lru_lock);
1270                         pgdeactivate += pgmoved;
1271                         pgmoved = 0;
1272                         if (buffer_heads_over_limit)
1273                                 pagevec_strip(&pvec);
1274                         __pagevec_release(&pvec);
1275                         spin_lock_irq(&zone->lru_lock);
1276                 }
1277         }
1278         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1279         pgdeactivate += pgmoved;
1280         if (buffer_heads_over_limit) {
1281                 spin_unlock_irq(&zone->lru_lock);
1282                 pagevec_strip(&pvec);
1283                 spin_lock_irq(&zone->lru_lock);
1284         }
1285         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1286         __count_vm_events(PGDEACTIVATE, pgdeactivate);
1287         spin_unlock_irq(&zone->lru_lock);
1288         if (vm_swap_full())
1289                 pagevec_swap_free(&pvec);
1290
1291         pagevec_release(&pvec);
1292 }
1293
1294 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1295         struct zone *zone, struct scan_control *sc, int priority)
1296 {
1297         int file = is_file_lru(lru);
1298
1299         if (lru == LRU_ACTIVE_FILE) {
1300                 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1301                 return 0;
1302         }
1303
1304         if (lru == LRU_ACTIVE_ANON &&
1305             (!scan_global_lru(sc) || inactive_anon_is_low(zone))) {
1306                 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1307                 return 0;
1308         }
1309         return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1310 }
1311
1312 /*
1313  * Determine how aggressively the anon and file LRU lists should be
1314  * scanned.  The relative value of each set of LRU lists is determined
1315  * by looking at the fraction of the pages scanned we did rotate back
1316  * onto the active list instead of evict.
1317  *
1318  * percent[0] specifies how much pressure to put on ram/swap backed
1319  * memory, while percent[1] determines pressure on the file LRUs.
1320  */
1321 static void get_scan_ratio(struct zone *zone, struct scan_control *sc,
1322                                         unsigned long *percent)
1323 {
1324         unsigned long anon, file, free;
1325         unsigned long anon_prio, file_prio;
1326         unsigned long ap, fp;
1327
1328         /* If we have no swap space, do not bother scanning anon pages. */
1329         if (nr_swap_pages <= 0) {
1330                 percent[0] = 0;
1331                 percent[1] = 100;
1332                 return;
1333         }
1334
1335         anon  = zone_page_state(zone, NR_ACTIVE_ANON) +
1336                 zone_page_state(zone, NR_INACTIVE_ANON);
1337         file  = zone_page_state(zone, NR_ACTIVE_FILE) +
1338                 zone_page_state(zone, NR_INACTIVE_FILE);
1339         free  = zone_page_state(zone, NR_FREE_PAGES);
1340
1341         /* If we have very few page cache pages, force-scan anon pages. */
1342         if (unlikely(file + free <= zone->pages_high)) {
1343                 percent[0] = 100;
1344                 percent[1] = 0;
1345                 return;
1346         }
1347
1348         /*
1349          * OK, so we have swap space and a fair amount of page cache
1350          * pages.  We use the recently rotated / recently scanned
1351          * ratios to determine how valuable each cache is.
1352          *
1353          * Because workloads change over time (and to avoid overflow)
1354          * we keep these statistics as a floating average, which ends
1355          * up weighing recent references more than old ones.
1356          *
1357          * anon in [0], file in [1]
1358          */
1359         if (unlikely(zone->recent_scanned[0] > anon / 4)) {
1360                 spin_lock_irq(&zone->lru_lock);
1361                 zone->recent_scanned[0] /= 2;
1362                 zone->recent_rotated[0] /= 2;
1363                 spin_unlock_irq(&zone->lru_lock);
1364         }
1365
1366         if (unlikely(zone->recent_scanned[1] > file / 4)) {
1367                 spin_lock_irq(&zone->lru_lock);
1368                 zone->recent_scanned[1] /= 2;
1369                 zone->recent_rotated[1] /= 2;
1370                 spin_unlock_irq(&zone->lru_lock);
1371         }
1372
1373         /*
1374          * With swappiness at 100, anonymous and file have the same priority.
1375          * This scanning priority is essentially the inverse of IO cost.
1376          */
1377         anon_prio = sc->swappiness;
1378         file_prio = 200 - sc->swappiness;
1379
1380         /*
1381          * The amount of pressure on anon vs file pages is inversely
1382          * proportional to the fraction of recently scanned pages on
1383          * each list that were recently referenced and in active use.
1384          */
1385         ap = (anon_prio + 1) * (zone->recent_scanned[0] + 1);
1386         ap /= zone->recent_rotated[0] + 1;
1387
1388         fp = (file_prio + 1) * (zone->recent_scanned[1] + 1);
1389         fp /= zone->recent_rotated[1] + 1;
1390
1391         /* Normalize to percentages */
1392         percent[0] = 100 * ap / (ap + fp + 1);
1393         percent[1] = 100 - percent[0];
1394 }
1395
1396
1397 /*
1398  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1399  */
1400 static void shrink_zone(int priority, struct zone *zone,
1401                                 struct scan_control *sc)
1402 {
1403         unsigned long nr[NR_LRU_LISTS];
1404         unsigned long nr_to_scan;
1405         unsigned long percent[2];       /* anon @ 0; file @ 1 */
1406         enum lru_list l;
1407         unsigned long nr_reclaimed = sc->nr_reclaimed;
1408         unsigned long swap_cluster_max = sc->swap_cluster_max;
1409
1410         get_scan_ratio(zone, sc, percent);
1411
1412         for_each_evictable_lru(l) {
1413                 if (scan_global_lru(sc)) {
1414                         int file = is_file_lru(l);
1415                         int scan;
1416
1417                         scan = zone_page_state(zone, NR_LRU_BASE + l);
1418                         if (priority) {
1419                                 scan >>= priority;
1420                                 scan = (scan * percent[file]) / 100;
1421                         }
1422                         zone->lru[l].nr_scan += scan;
1423                         nr[l] = zone->lru[l].nr_scan;
1424                         if (nr[l] >= swap_cluster_max)
1425                                 zone->lru[l].nr_scan = 0;
1426                         else
1427                                 nr[l] = 0;
1428                 } else {
1429                         /*
1430                          * This reclaim occurs not because zone memory shortage
1431                          * but because memory controller hits its limit.
1432                          * Don't modify zone reclaim related data.
1433                          */
1434                         nr[l] = mem_cgroup_calc_reclaim(sc->mem_cgroup, zone,
1435                                                                 priority, l);
1436                 }
1437         }
1438
1439         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1440                                         nr[LRU_INACTIVE_FILE]) {
1441                 for_each_evictable_lru(l) {
1442                         if (nr[l]) {
1443                                 nr_to_scan = min(nr[l], swap_cluster_max);
1444                                 nr[l] -= nr_to_scan;
1445
1446                                 nr_reclaimed += shrink_list(l, nr_to_scan,
1447                                                             zone, sc, priority);
1448                         }
1449                 }
1450                 /*
1451                  * On large memory systems, scan >> priority can become
1452                  * really large. This is fine for the starting priority;
1453                  * we want to put equal scanning pressure on each zone.
1454                  * However, if the VM has a harder time of freeing pages,
1455                  * with multiple processes reclaiming pages, the total
1456                  * freeing target can get unreasonably large.
1457                  */
1458                 if (nr_reclaimed > swap_cluster_max &&
1459                         priority < DEF_PRIORITY && !current_is_kswapd())
1460                         break;
1461         }
1462
1463         sc->nr_reclaimed = nr_reclaimed;
1464
1465         /*
1466          * Even if we did not try to evict anon pages at all, we want to
1467          * rebalance the anon lru active/inactive ratio.
1468          */
1469         if (!scan_global_lru(sc) || inactive_anon_is_low(zone))
1470                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1471         else if (!scan_global_lru(sc))
1472                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1473
1474         throttle_vm_writeout(sc->gfp_mask);
1475 }
1476
1477 /*
1478  * This is the direct reclaim path, for page-allocating processes.  We only
1479  * try to reclaim pages from zones which will satisfy the caller's allocation
1480  * request.
1481  *
1482  * We reclaim from a zone even if that zone is over pages_high.  Because:
1483  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1484  *    allocation or
1485  * b) The zones may be over pages_high but they must go *over* pages_high to
1486  *    satisfy the `incremental min' zone defense algorithm.
1487  *
1488  * If a zone is deemed to be full of pinned pages then just give it a light
1489  * scan then give up on it.
1490  */
1491 static void shrink_zones(int priority, struct zonelist *zonelist,
1492                                         struct scan_control *sc)
1493 {
1494         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1495         struct zoneref *z;
1496         struct zone *zone;
1497
1498         sc->all_unreclaimable = 1;
1499         for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1500                 if (!populated_zone(zone))
1501                         continue;
1502                 /*
1503                  * Take care memory controller reclaiming has small influence
1504                  * to global LRU.
1505                  */
1506                 if (scan_global_lru(sc)) {
1507                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1508                                 continue;
1509                         note_zone_scanning_priority(zone, priority);
1510
1511                         if (zone_is_all_unreclaimable(zone) &&
1512                                                 priority != DEF_PRIORITY)
1513                                 continue;       /* Let kswapd poll it */
1514                         sc->all_unreclaimable = 0;
1515                 } else {
1516                         /*
1517                          * Ignore cpuset limitation here. We just want to reduce
1518                          * # of used pages by us regardless of memory shortage.
1519                          */
1520                         sc->all_unreclaimable = 0;
1521                         mem_cgroup_note_reclaim_priority(sc->mem_cgroup,
1522                                                         priority);
1523                 }
1524
1525                 shrink_zone(priority, zone, sc);
1526         }
1527 }
1528
1529 /*
1530  * This is the main entry point to direct page reclaim.
1531  *
1532  * If a full scan of the inactive list fails to free enough memory then we
1533  * are "out of memory" and something needs to be killed.
1534  *
1535  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1536  * high - the zone may be full of dirty or under-writeback pages, which this
1537  * caller can't do much about.  We kick pdflush and take explicit naps in the
1538  * hope that some of these pages can be written.  But if the allocating task
1539  * holds filesystem locks which prevent writeout this might not work, and the
1540  * allocation attempt will fail.
1541  *
1542  * returns:     0, if no pages reclaimed
1543  *              else, the number of pages reclaimed
1544  */
1545 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
1546                                         struct scan_control *sc)
1547 {
1548         int priority;
1549         unsigned long ret = 0;
1550         unsigned long total_scanned = 0;
1551         struct reclaim_state *reclaim_state = current->reclaim_state;
1552         unsigned long lru_pages = 0;
1553         struct zoneref *z;
1554         struct zone *zone;
1555         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1556
1557         delayacct_freepages_start();
1558
1559         if (scan_global_lru(sc))
1560                 count_vm_event(ALLOCSTALL);
1561         /*
1562          * mem_cgroup will not do shrink_slab.
1563          */
1564         if (scan_global_lru(sc)) {
1565                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1566
1567                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1568                                 continue;
1569
1570                         lru_pages += zone_lru_pages(zone);
1571                 }
1572         }
1573
1574         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1575                 sc->nr_scanned = 0;
1576                 if (!priority)
1577                         disable_swap_token();
1578                 shrink_zones(priority, zonelist, sc);
1579                 /*
1580                  * Don't shrink slabs when reclaiming memory from
1581                  * over limit cgroups
1582                  */
1583                 if (scan_global_lru(sc)) {
1584                         shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
1585                         if (reclaim_state) {
1586                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
1587                                 reclaim_state->reclaimed_slab = 0;
1588                         }
1589                 }
1590                 total_scanned += sc->nr_scanned;
1591                 if (sc->nr_reclaimed >= sc->swap_cluster_max) {
1592                         ret = sc->nr_reclaimed;
1593                         goto out;
1594                 }
1595
1596                 /*
1597                  * Try to write back as many pages as we just scanned.  This
1598                  * tends to cause slow streaming writers to write data to the
1599                  * disk smoothly, at the dirtying rate, which is nice.   But
1600                  * that's undesirable in laptop mode, where we *want* lumpy
1601                  * writeout.  So in laptop mode, write out the whole world.
1602                  */
1603                 if (total_scanned > sc->swap_cluster_max +
1604                                         sc->swap_cluster_max / 2) {
1605                         wakeup_pdflush(laptop_mode ? 0 : total_scanned);
1606                         sc->may_writepage = 1;
1607                 }
1608
1609                 /* Take a nap, wait for some writeback to complete */
1610                 if (sc->nr_scanned && priority < DEF_PRIORITY - 2)
1611                         congestion_wait(WRITE, HZ/10);
1612         }
1613         /* top priority shrink_zones still had more to do? don't OOM, then */
1614         if (!sc->all_unreclaimable && scan_global_lru(sc))
1615                 ret = sc->nr_reclaimed;
1616 out:
1617         /*
1618          * Now that we've scanned all the zones at this priority level, note
1619          * that level within the zone so that the next thread which performs
1620          * scanning of this zone will immediately start out at this priority
1621          * level.  This affects only the decision whether or not to bring
1622          * mapped pages onto the inactive list.
1623          */
1624         if (priority < 0)
1625                 priority = 0;
1626
1627         if (scan_global_lru(sc)) {
1628                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1629
1630                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1631                                 continue;
1632
1633                         zone->prev_priority = priority;
1634                 }
1635         } else
1636                 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority);
1637
1638         delayacct_freepages_end();
1639
1640         return ret;
1641 }
1642
1643 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
1644                                                                 gfp_t gfp_mask)
1645 {
1646         struct scan_control sc = {
1647                 .gfp_mask = gfp_mask,
1648                 .may_writepage = !laptop_mode,
1649                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1650                 .may_swap = 1,
1651                 .swappiness = vm_swappiness,
1652                 .order = order,
1653                 .mem_cgroup = NULL,
1654                 .isolate_pages = isolate_pages_global,
1655         };
1656
1657         return do_try_to_free_pages(zonelist, &sc);
1658 }
1659
1660 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
1661
1662 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
1663                                                 gfp_t gfp_mask,
1664                                            bool noswap)
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         if (noswap)
1678                 sc.may_swap = 0;
1679
1680         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1681                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1682         zonelist = NODE_DATA(numa_node_id())->node_zonelists;
1683         return do_try_to_free_pages(zonelist, &sc);
1684 }
1685 #endif
1686
1687 /*
1688  * For kswapd, balance_pgdat() will work across all this node's zones until
1689  * they are all at pages_high.
1690  *
1691  * Returns the number of pages which were actually freed.
1692  *
1693  * There is special handling here for zones which are full of pinned pages.
1694  * This can happen if the pages are all mlocked, or if they are all used by
1695  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1696  * What we do is to detect the case where all pages in the zone have been
1697  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1698  * dead and from now on, only perform a short scan.  Basically we're polling
1699  * the zone for when the problem goes away.
1700  *
1701  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
1702  * zones which have free_pages > pages_high, but once a zone is found to have
1703  * free_pages <= pages_high, we scan that zone and the lower zones regardless
1704  * of the number of free pages in the lower zones.  This interoperates with
1705  * the page allocator fallback scheme to ensure that aging of pages is balanced
1706  * across the zones.
1707  */
1708 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
1709 {
1710         int all_zones_ok;
1711         int priority;
1712         int i;
1713         unsigned long total_scanned;
1714         struct reclaim_state *reclaim_state = current->reclaim_state;
1715         struct scan_control sc = {
1716                 .gfp_mask = GFP_KERNEL,
1717                 .may_swap = 1,
1718                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1719                 .swappiness = vm_swappiness,
1720                 .order = order,
1721                 .mem_cgroup = NULL,
1722                 .isolate_pages = isolate_pages_global,
1723         };
1724         /*
1725          * temp_priority is used to remember the scanning priority at which
1726          * this zone was successfully refilled to free_pages == pages_high.
1727          */
1728         int temp_priority[MAX_NR_ZONES];
1729
1730 loop_again:
1731         total_scanned = 0;
1732         sc.nr_reclaimed = 0;
1733         sc.may_writepage = !laptop_mode;
1734         count_vm_event(PAGEOUTRUN);
1735
1736         for (i = 0; i < pgdat->nr_zones; i++)
1737                 temp_priority[i] = DEF_PRIORITY;
1738
1739         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1740                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1741                 unsigned long lru_pages = 0;
1742
1743                 /* The swap token gets in the way of swapout... */
1744                 if (!priority)
1745                         disable_swap_token();
1746
1747                 all_zones_ok = 1;
1748
1749                 /*
1750                  * Scan in the highmem->dma direction for the highest
1751                  * zone which needs scanning
1752                  */
1753                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1754                         struct zone *zone = pgdat->node_zones + i;
1755
1756                         if (!populated_zone(zone))
1757                                 continue;
1758
1759                         if (zone_is_all_unreclaimable(zone) &&
1760                             priority != DEF_PRIORITY)
1761                                 continue;
1762
1763                         /*
1764                          * Do some background aging of the anon list, to give
1765                          * pages a chance to be referenced before reclaiming.
1766                          */
1767                         if (inactive_anon_is_low(zone))
1768                                 shrink_active_list(SWAP_CLUSTER_MAX, zone,
1769                                                         &sc, priority, 0);
1770
1771                         if (!zone_watermark_ok(zone, order, zone->pages_high,
1772                                                0, 0)) {
1773                                 end_zone = i;
1774                                 break;
1775                         }
1776                 }
1777                 if (i < 0)
1778                         goto out;
1779
1780                 for (i = 0; i <= end_zone; i++) {
1781                         struct zone *zone = pgdat->node_zones + i;
1782
1783                         lru_pages += zone_lru_pages(zone);
1784                 }
1785
1786                 /*
1787                  * Now scan the zone in the dma->highmem direction, stopping
1788                  * at the last zone which needs scanning.
1789                  *
1790                  * We do this because the page allocator works in the opposite
1791                  * direction.  This prevents the page allocator from allocating
1792                  * pages behind kswapd's direction of progress, which would
1793                  * cause too much scanning of the lower zones.
1794                  */
1795                 for (i = 0; i <= end_zone; i++) {
1796                         struct zone *zone = pgdat->node_zones + i;
1797                         int nr_slab;
1798
1799                         if (!populated_zone(zone))
1800                                 continue;
1801
1802                         if (zone_is_all_unreclaimable(zone) &&
1803                                         priority != DEF_PRIORITY)
1804                                 continue;
1805
1806                         if (!zone_watermark_ok(zone, order, zone->pages_high,
1807                                                end_zone, 0))
1808                                 all_zones_ok = 0;
1809                         temp_priority[i] = priority;
1810                         sc.nr_scanned = 0;
1811                         note_zone_scanning_priority(zone, priority);
1812                         /*
1813                          * We put equal pressure on every zone, unless one
1814                          * zone has way too many pages free already.
1815                          */
1816                         if (!zone_watermark_ok(zone, order, 8*zone->pages_high,
1817                                                 end_zone, 0))
1818                                 shrink_zone(priority, zone, &sc);
1819                         reclaim_state->reclaimed_slab = 0;
1820                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1821                                                 lru_pages);
1822                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1823                         total_scanned += sc.nr_scanned;
1824                         if (zone_is_all_unreclaimable(zone))
1825                                 continue;
1826                         if (nr_slab == 0 && zone->pages_scanned >=
1827                                                 (zone_lru_pages(zone) * 6))
1828                                         zone_set_flag(zone,
1829                                                       ZONE_ALL_UNRECLAIMABLE);
1830                         /*
1831                          * If we've done a decent amount of scanning and
1832                          * the reclaim ratio is low, start doing writepage
1833                          * even in laptop mode
1834                          */
1835                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1836                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
1837                                 sc.may_writepage = 1;
1838                 }
1839                 if (all_zones_ok)
1840                         break;          /* kswapd: all done */
1841                 /*
1842                  * OK, kswapd is getting into trouble.  Take a nap, then take
1843                  * another pass across the zones.
1844                  */
1845                 if (total_scanned && priority < DEF_PRIORITY - 2)
1846                         congestion_wait(WRITE, HZ/10);
1847
1848                 /*
1849                  * We do this so kswapd doesn't build up large priorities for
1850                  * example when it is freeing in parallel with allocators. It
1851                  * matches the direct reclaim path behaviour in terms of impact
1852                  * on zone->*_priority.
1853                  */
1854                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
1855                         break;
1856         }
1857 out:
1858         /*
1859          * Note within each zone the priority level at which this zone was
1860          * brought into a happy state.  So that the next thread which scans this
1861          * zone will start out at that priority level.
1862          */
1863         for (i = 0; i < pgdat->nr_zones; i++) {
1864                 struct zone *zone = pgdat->node_zones + i;
1865
1866                 zone->prev_priority = temp_priority[i];
1867         }
1868         if (!all_zones_ok) {
1869                 cond_resched();
1870
1871                 try_to_freeze();
1872
1873                 /*
1874                  * Fragmentation may mean that the system cannot be
1875                  * rebalanced for high-order allocations in all zones.
1876                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
1877                  * it means the zones have been fully scanned and are still
1878                  * not balanced. For high-order allocations, there is
1879                  * little point trying all over again as kswapd may
1880                  * infinite loop.
1881                  *
1882                  * Instead, recheck all watermarks at order-0 as they
1883                  * are the most important. If watermarks are ok, kswapd will go
1884                  * back to sleep. High-order users can still perform direct
1885                  * reclaim if they wish.
1886                  */
1887                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
1888                         order = sc.order = 0;
1889
1890                 goto loop_again;
1891         }
1892
1893         return sc.nr_reclaimed;
1894 }
1895
1896 /*
1897  * The background pageout daemon, started as a kernel thread
1898  * from the init process.
1899  *
1900  * This basically trickles out pages so that we have _some_
1901  * free memory available even if there is no other activity
1902  * that frees anything up. This is needed for things like routing
1903  * etc, where we otherwise might have all activity going on in
1904  * asynchronous contexts that cannot page things out.
1905  *
1906  * If there are applications that are active memory-allocators
1907  * (most normal use), this basically shouldn't matter.
1908  */
1909 static int kswapd(void *p)
1910 {
1911         unsigned long order;
1912         pg_data_t *pgdat = (pg_data_t*)p;
1913         struct task_struct *tsk = current;
1914         DEFINE_WAIT(wait);
1915         struct reclaim_state reclaim_state = {
1916                 .reclaimed_slab = 0,
1917         };
1918         node_to_cpumask_ptr(cpumask, pgdat->node_id);
1919
1920         if (!cpumask_empty(cpumask))
1921                 set_cpus_allowed_ptr(tsk, cpumask);
1922         current->reclaim_state = &reclaim_state;
1923
1924         /*
1925          * Tell the memory management that we're a "memory allocator",
1926          * and that if we need more memory we should get access to it
1927          * regardless (see "__alloc_pages()"). "kswapd" should
1928          * never get caught in the normal page freeing logic.
1929          *
1930          * (Kswapd normally doesn't need memory anyway, but sometimes
1931          * you need a small amount of memory in order to be able to
1932          * page out something else, and this flag essentially protects
1933          * us from recursively trying to free more memory as we're
1934          * trying to free the first piece of memory in the first place).
1935          */
1936         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
1937         set_freezable();
1938
1939         order = 0;
1940         for ( ; ; ) {
1941                 unsigned long new_order;
1942
1943                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1944                 new_order = pgdat->kswapd_max_order;
1945                 pgdat->kswapd_max_order = 0;
1946                 if (order < new_order) {
1947                         /*
1948                          * Don't sleep if someone wants a larger 'order'
1949                          * allocation
1950                          */
1951                         order = new_order;
1952                 } else {
1953                         if (!freezing(current))
1954                                 schedule();
1955
1956                         order = pgdat->kswapd_max_order;
1957                 }
1958                 finish_wait(&pgdat->kswapd_wait, &wait);
1959
1960                 if (!try_to_freeze()) {
1961                         /* We can speed up thawing tasks if we don't call
1962                          * balance_pgdat after returning from the refrigerator
1963                          */
1964                         balance_pgdat(pgdat, order);
1965                 }
1966         }
1967         return 0;
1968 }
1969
1970 /*
1971  * A zone is low on free memory, so wake its kswapd task to service it.
1972  */
1973 void wakeup_kswapd(struct zone *zone, int order)
1974 {
1975         pg_data_t *pgdat;
1976
1977         if (!populated_zone(zone))
1978                 return;
1979
1980         pgdat = zone->zone_pgdat;
1981         if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0))
1982                 return;
1983         if (pgdat->kswapd_max_order < order)
1984                 pgdat->kswapd_max_order = order;
1985         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1986                 return;
1987         if (!waitqueue_active(&pgdat->kswapd_wait))
1988                 return;
1989         wake_up_interruptible(&pgdat->kswapd_wait);
1990 }
1991
1992 unsigned long global_lru_pages(void)
1993 {
1994         return global_page_state(NR_ACTIVE_ANON)
1995                 + global_page_state(NR_ACTIVE_FILE)
1996                 + global_page_state(NR_INACTIVE_ANON)
1997                 + global_page_state(NR_INACTIVE_FILE);
1998 }
1999
2000 #ifdef CONFIG_PM
2001 /*
2002  * Helper function for shrink_all_memory().  Tries to reclaim 'nr_pages' pages
2003  * from LRU lists system-wide, for given pass and priority, and returns the
2004  * number of reclaimed pages
2005  *
2006  * For pass > 3 we also try to shrink the LRU lists that contain a few pages
2007  */
2008 static unsigned long shrink_all_zones(unsigned long nr_pages, int prio,
2009                                       int pass, struct scan_control *sc)
2010 {
2011         struct zone *zone;
2012         unsigned long nr_to_scan, ret = 0;
2013         enum lru_list l;
2014
2015         for_each_zone(zone) {
2016
2017                 if (!populated_zone(zone))
2018                         continue;
2019
2020                 if (zone_is_all_unreclaimable(zone) && prio != DEF_PRIORITY)
2021                         continue;
2022
2023                 for_each_evictable_lru(l) {
2024                         /* For pass = 0, we don't shrink the active list */
2025                         if (pass == 0 &&
2026                                 (l == LRU_ACTIVE || l == LRU_ACTIVE_FILE))
2027                                 continue;
2028
2029                         zone->lru[l].nr_scan +=
2030                                 (zone_page_state(zone, NR_LRU_BASE + l)
2031                                                                 >> prio) + 1;
2032                         if (zone->lru[l].nr_scan >= nr_pages || pass > 3) {
2033                                 zone->lru[l].nr_scan = 0;
2034                                 nr_to_scan = min(nr_pages,
2035                                         zone_page_state(zone,
2036                                                         NR_LRU_BASE + l));
2037                                 ret += shrink_list(l, nr_to_scan, zone,
2038                                                                 sc, prio);
2039                                 if (ret >= nr_pages)
2040                                         return ret;
2041                         }
2042                 }
2043         }
2044
2045         return ret;
2046 }
2047
2048 /*
2049  * Try to free `nr_pages' of memory, system-wide, and return the number of
2050  * freed pages.
2051  *
2052  * Rather than trying to age LRUs the aim is to preserve the overall
2053  * LRU order by reclaiming preferentially
2054  * inactive > active > active referenced > active mapped
2055  */
2056 unsigned long shrink_all_memory(unsigned long nr_pages)
2057 {
2058         unsigned long lru_pages, nr_slab;
2059         unsigned long ret = 0;
2060         int pass;
2061         struct reclaim_state reclaim_state;
2062         struct scan_control sc = {
2063                 .gfp_mask = GFP_KERNEL,
2064                 .may_swap = 0,
2065                 .swap_cluster_max = nr_pages,
2066                 .may_writepage = 1,
2067                 .swappiness = vm_swappiness,
2068                 .isolate_pages = isolate_pages_global,
2069         };
2070
2071         current->reclaim_state = &reclaim_state;
2072
2073         lru_pages = global_lru_pages();
2074         nr_slab = global_page_state(NR_SLAB_RECLAIMABLE);
2075         /* If slab caches are huge, it's better to hit them first */
2076         while (nr_slab >= lru_pages) {
2077                 reclaim_state.reclaimed_slab = 0;
2078                 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
2079                 if (!reclaim_state.reclaimed_slab)
2080                         break;
2081
2082                 ret += reclaim_state.reclaimed_slab;
2083                 if (ret >= nr_pages)
2084                         goto out;
2085
2086                 nr_slab -= reclaim_state.reclaimed_slab;
2087         }
2088
2089         /*
2090          * We try to shrink LRUs in 5 passes:
2091          * 0 = Reclaim from inactive_list only
2092          * 1 = Reclaim from active list but don't reclaim mapped
2093          * 2 = 2nd pass of type 1
2094          * 3 = Reclaim mapped (normal reclaim)
2095          * 4 = 2nd pass of type 3
2096          */
2097         for (pass = 0; pass < 5; pass++) {
2098                 int prio;
2099
2100                 /* Force reclaiming mapped pages in the passes #3 and #4 */
2101                 if (pass > 2) {
2102                         sc.may_swap = 1;
2103                         sc.swappiness = 100;
2104                 }
2105
2106                 for (prio = DEF_PRIORITY; prio >= 0; prio--) {
2107                         unsigned long nr_to_scan = nr_pages - ret;
2108
2109                         sc.nr_scanned = 0;
2110                         ret += shrink_all_zones(nr_to_scan, prio, pass, &sc);
2111                         if (ret >= nr_pages)
2112                                 goto out;
2113
2114                         reclaim_state.reclaimed_slab = 0;
2115                         shrink_slab(sc.nr_scanned, sc.gfp_mask,
2116                                         global_lru_pages());
2117                         ret += reclaim_state.reclaimed_slab;
2118                         if (ret >= nr_pages)
2119                                 goto out;
2120
2121                         if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
2122                                 congestion_wait(WRITE, HZ / 10);
2123                 }
2124         }
2125
2126         /*
2127          * If ret = 0, we could not shrink LRUs, but there may be something
2128          * in slab caches
2129          */
2130         if (!ret) {
2131                 do {
2132                         reclaim_state.reclaimed_slab = 0;
2133                         shrink_slab(nr_pages, sc.gfp_mask, global_lru_pages());
2134                         ret += reclaim_state.reclaimed_slab;
2135                 } while (ret < nr_pages && reclaim_state.reclaimed_slab > 0);
2136         }
2137
2138 out:
2139         current->reclaim_state = NULL;
2140
2141         return ret;
2142 }
2143 #endif
2144
2145 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2146    not required for correctness.  So if the last cpu in a node goes
2147    away, we get changed to run anywhere: as the first one comes back,
2148    restore their cpu bindings. */
2149 static int __devinit cpu_callback(struct notifier_block *nfb,
2150                                   unsigned long action, void *hcpu)
2151 {
2152         int nid;
2153
2154         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2155                 for_each_node_state(nid, N_HIGH_MEMORY) {
2156                         pg_data_t *pgdat = NODE_DATA(nid);
2157                         node_to_cpumask_ptr(mask, pgdat->node_id);
2158
2159                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2160                                 /* One of our CPUs online: restore mask */
2161                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2162                 }
2163         }
2164         return NOTIFY_OK;
2165 }
2166
2167 /*
2168  * This kswapd start function will be called by init and node-hot-add.
2169  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2170  */
2171 int kswapd_run(int nid)
2172 {
2173         pg_data_t *pgdat = NODE_DATA(nid);
2174         int ret = 0;
2175
2176         if (pgdat->kswapd)
2177                 return 0;
2178
2179         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2180         if (IS_ERR(pgdat->kswapd)) {
2181                 /* failure at boot is fatal */
2182                 BUG_ON(system_state == SYSTEM_BOOTING);
2183                 printk("Failed to start kswapd on node %d\n",nid);
2184                 ret = -1;
2185         }
2186         return ret;
2187 }
2188
2189 static int __init kswapd_init(void)
2190 {
2191         int nid;
2192
2193         swap_setup();
2194         for_each_node_state(nid, N_HIGH_MEMORY)
2195                 kswapd_run(nid);
2196         hotcpu_notifier(cpu_callback, 0);
2197         return 0;
2198 }
2199
2200 module_init(kswapd_init)
2201
2202 #ifdef CONFIG_NUMA
2203 /*
2204  * Zone reclaim mode
2205  *
2206  * If non-zero call zone_reclaim when the number of free pages falls below
2207  * the watermarks.
2208  */
2209 int zone_reclaim_mode __read_mostly;
2210
2211 #define RECLAIM_OFF 0
2212 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
2213 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
2214 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
2215
2216 /*
2217  * Priority for ZONE_RECLAIM. This determines the fraction of pages
2218  * of a node considered for each zone_reclaim. 4 scans 1/16th of
2219  * a zone.
2220  */
2221 #define ZONE_RECLAIM_PRIORITY 4
2222
2223 /*
2224  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2225  * occur.
2226  */
2227 int sysctl_min_unmapped_ratio = 1;
2228
2229 /*
2230  * If the number of slab pages in a zone grows beyond this percentage then
2231  * slab reclaim needs to occur.
2232  */
2233 int sysctl_min_slab_ratio = 5;
2234
2235 /*
2236  * Try to free up some pages from this zone through reclaim.
2237  */
2238 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2239 {
2240         /* Minimum pages needed in order to stay on node */
2241         const unsigned long nr_pages = 1 << order;
2242         struct task_struct *p = current;
2243         struct reclaim_state reclaim_state;
2244         int priority;
2245         struct scan_control sc = {
2246                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2247                 .may_swap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2248                 .swap_cluster_max = max_t(unsigned long, nr_pages,
2249                                         SWAP_CLUSTER_MAX),
2250                 .gfp_mask = gfp_mask,
2251                 .swappiness = vm_swappiness,
2252                 .isolate_pages = isolate_pages_global,
2253         };
2254         unsigned long slab_reclaimable;
2255
2256         disable_swap_token();
2257         cond_resched();
2258         /*
2259          * We need to be able to allocate from the reserves for RECLAIM_SWAP
2260          * and we also need to be able to write out pages for RECLAIM_WRITE
2261          * and RECLAIM_SWAP.
2262          */
2263         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2264         reclaim_state.reclaimed_slab = 0;
2265         p->reclaim_state = &reclaim_state;
2266
2267         if (zone_page_state(zone, NR_FILE_PAGES) -
2268                 zone_page_state(zone, NR_FILE_MAPPED) >
2269                 zone->min_unmapped_pages) {
2270                 /*
2271                  * Free memory by calling shrink zone with increasing
2272                  * priorities until we have enough memory freed.
2273                  */
2274                 priority = ZONE_RECLAIM_PRIORITY;
2275                 do {
2276                         note_zone_scanning_priority(zone, priority);
2277                         shrink_zone(priority, zone, &sc);
2278                         priority--;
2279                 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
2280         }
2281
2282         slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2283         if (slab_reclaimable > zone->min_slab_pages) {
2284                 /*
2285                  * shrink_slab() does not currently allow us to determine how
2286                  * many pages were freed in this zone. So we take the current
2287                  * number of slab pages and shake the slab until it is reduced
2288                  * by the same nr_pages that we used for reclaiming unmapped
2289                  * pages.
2290                  *
2291                  * Note that shrink_slab will free memory on all zones and may
2292                  * take a long time.
2293                  */
2294                 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
2295                         zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
2296                                 slab_reclaimable - nr_pages)
2297                         ;
2298
2299                 /*
2300                  * Update nr_reclaimed by the number of slab pages we
2301                  * reclaimed from this zone.
2302                  */
2303                 sc.nr_reclaimed += slab_reclaimable -
2304                         zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2305         }
2306
2307         p->reclaim_state = NULL;
2308         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
2309         return sc.nr_reclaimed >= nr_pages;
2310 }
2311
2312 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2313 {
2314         int node_id;
2315         int ret;
2316
2317         /*
2318          * Zone reclaim reclaims unmapped file backed pages and
2319          * slab pages if we are over the defined limits.
2320          *
2321          * A small portion of unmapped file backed pages is needed for
2322          * file I/O otherwise pages read by file I/O will be immediately
2323          * thrown out if the zone is overallocated. So we do not reclaim
2324          * if less than a specified percentage of the zone is used by
2325          * unmapped file backed pages.
2326          */
2327         if (zone_page_state(zone, NR_FILE_PAGES) -
2328             zone_page_state(zone, NR_FILE_MAPPED) <= zone->min_unmapped_pages
2329             && zone_page_state(zone, NR_SLAB_RECLAIMABLE)
2330                         <= zone->min_slab_pages)
2331                 return 0;
2332
2333         if (zone_is_all_unreclaimable(zone))
2334                 return 0;
2335
2336         /*
2337          * Do not scan if the allocation should not be delayed.
2338          */
2339         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
2340                         return 0;
2341
2342         /*
2343          * Only run zone reclaim on the local zone or on zones that do not
2344          * have associated processors. This will favor the local processor
2345          * over remote processors and spread off node memory allocations
2346          * as wide as possible.
2347          */
2348         node_id = zone_to_nid(zone);
2349         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
2350                 return 0;
2351
2352         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2353                 return 0;
2354         ret = __zone_reclaim(zone, gfp_mask, order);
2355         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2356
2357         return ret;
2358 }
2359 #endif
2360
2361 #ifdef CONFIG_UNEVICTABLE_LRU
2362 /*
2363  * page_evictable - test whether a page is evictable
2364  * @page: the page to test
2365  * @vma: the VMA in which the page is or will be mapped, may be NULL
2366  *
2367  * Test whether page is evictable--i.e., should be placed on active/inactive
2368  * lists vs unevictable list.  The vma argument is !NULL when called from the
2369  * fault path to determine how to instantate a new page.
2370  *
2371  * Reasons page might not be evictable:
2372  * (1) page's mapping marked unevictable
2373  * (2) page is part of an mlocked VMA
2374  *
2375  */
2376 int page_evictable(struct page *page, struct vm_area_struct *vma)
2377 {
2378
2379         if (mapping_unevictable(page_mapping(page)))
2380                 return 0;
2381
2382         if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2383                 return 0;
2384
2385         return 1;
2386 }
2387
2388 /**
2389  * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2390  * @page: page to check evictability and move to appropriate lru list
2391  * @zone: zone page is in
2392  *
2393  * Checks a page for evictability and moves the page to the appropriate
2394  * zone lru list.
2395  *
2396  * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2397  * have PageUnevictable set.
2398  */
2399 static void check_move_unevictable_page(struct page *page, struct zone *zone)
2400 {
2401         VM_BUG_ON(PageActive(page));
2402
2403 retry:
2404         ClearPageUnevictable(page);
2405         if (page_evictable(page, NULL)) {
2406                 enum lru_list l = LRU_INACTIVE_ANON + page_is_file_cache(page);
2407
2408                 __dec_zone_state(zone, NR_UNEVICTABLE);
2409                 list_move(&page->lru, &zone->lru[l].list);
2410                 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
2411                 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
2412                 __count_vm_event(UNEVICTABLE_PGRESCUED);
2413         } else {
2414                 /*
2415                  * rotate unevictable list
2416                  */
2417                 SetPageUnevictable(page);
2418                 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
2419                 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
2420                 if (page_evictable(page, NULL))
2421                         goto retry;
2422         }
2423 }
2424
2425 /**
2426  * scan_mapping_unevictable_pages - scan an address space for evictable pages
2427  * @mapping: struct address_space to scan for evictable pages
2428  *
2429  * Scan all pages in mapping.  Check unevictable pages for
2430  * evictability and move them to the appropriate zone lru list.
2431  */
2432 void scan_mapping_unevictable_pages(struct address_space *mapping)
2433 {
2434         pgoff_t next = 0;
2435         pgoff_t end   = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
2436                          PAGE_CACHE_SHIFT;
2437         struct zone *zone;
2438         struct pagevec pvec;
2439
2440         if (mapping->nrpages == 0)
2441                 return;
2442
2443         pagevec_init(&pvec, 0);
2444         while (next < end &&
2445                 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
2446                 int i;
2447                 int pg_scanned = 0;
2448
2449                 zone = NULL;
2450
2451                 for (i = 0; i < pagevec_count(&pvec); i++) {
2452                         struct page *page = pvec.pages[i];
2453                         pgoff_t page_index = page->index;
2454                         struct zone *pagezone = page_zone(page);
2455
2456                         pg_scanned++;
2457                         if (page_index > next)
2458                                 next = page_index;
2459                         next++;
2460
2461                         if (pagezone != zone) {
2462                                 if (zone)
2463                                         spin_unlock_irq(&zone->lru_lock);
2464                                 zone = pagezone;
2465                                 spin_lock_irq(&zone->lru_lock);
2466                         }
2467
2468                         if (PageLRU(page) && PageUnevictable(page))
2469                                 check_move_unevictable_page(page, zone);
2470                 }
2471                 if (zone)
2472                         spin_unlock_irq(&zone->lru_lock);
2473                 pagevec_release(&pvec);
2474
2475                 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
2476         }
2477
2478 }
2479
2480 /**
2481  * scan_zone_unevictable_pages - check unevictable list for evictable pages
2482  * @zone - zone of which to scan the unevictable list
2483  *
2484  * Scan @zone's unevictable LRU lists to check for pages that have become
2485  * evictable.  Move those that have to @zone's inactive list where they
2486  * become candidates for reclaim, unless shrink_inactive_zone() decides
2487  * to reactivate them.  Pages that are still unevictable are rotated
2488  * back onto @zone's unevictable list.
2489  */
2490 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
2491 static void scan_zone_unevictable_pages(struct zone *zone)
2492 {
2493         struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
2494         unsigned long scan;
2495         unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
2496
2497         while (nr_to_scan > 0) {
2498                 unsigned long batch_size = min(nr_to_scan,
2499                                                 SCAN_UNEVICTABLE_BATCH_SIZE);
2500
2501                 spin_lock_irq(&zone->lru_lock);
2502                 for (scan = 0;  scan < batch_size; scan++) {
2503                         struct page *page = lru_to_page(l_unevictable);
2504
2505                         if (!trylock_page(page))
2506                                 continue;
2507
2508                         prefetchw_prev_lru_page(page, l_unevictable, flags);
2509
2510                         if (likely(PageLRU(page) && PageUnevictable(page)))
2511                                 check_move_unevictable_page(page, zone);
2512
2513                         unlock_page(page);
2514                 }
2515                 spin_unlock_irq(&zone->lru_lock);
2516
2517                 nr_to_scan -= batch_size;
2518         }
2519 }
2520
2521
2522 /**
2523  * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
2524  *
2525  * A really big hammer:  scan all zones' unevictable LRU lists to check for
2526  * pages that have become evictable.  Move those back to the zones'
2527  * inactive list where they become candidates for reclaim.
2528  * This occurs when, e.g., we have unswappable pages on the unevictable lists,
2529  * and we add swap to the system.  As such, it runs in the context of a task
2530  * that has possibly/probably made some previously unevictable pages
2531  * evictable.
2532  */
2533 static void scan_all_zones_unevictable_pages(void)
2534 {
2535         struct zone *zone;
2536
2537         for_each_zone(zone) {
2538                 scan_zone_unevictable_pages(zone);
2539         }
2540 }
2541
2542 /*
2543  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
2544  * all nodes' unevictable lists for evictable pages
2545  */
2546 unsigned long scan_unevictable_pages;
2547
2548 int scan_unevictable_handler(struct ctl_table *table, int write,
2549                            struct file *file, void __user *buffer,
2550                            size_t *length, loff_t *ppos)
2551 {
2552         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
2553
2554         if (write && *(unsigned long *)table->data)
2555                 scan_all_zones_unevictable_pages();
2556
2557         scan_unevictable_pages = 0;
2558         return 0;
2559 }
2560
2561 /*
2562  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
2563  * a specified node's per zone unevictable lists for evictable pages.
2564  */
2565
2566 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
2567                                           struct sysdev_attribute *attr,
2568                                           char *buf)
2569 {
2570         return sprintf(buf, "0\n");     /* always zero; should fit... */
2571 }
2572
2573 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
2574                                            struct sysdev_attribute *attr,
2575                                         const char *buf, size_t count)
2576 {
2577         struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
2578         struct zone *zone;
2579         unsigned long res;
2580         unsigned long req = strict_strtoul(buf, 10, &res);
2581
2582         if (!req)
2583                 return 1;       /* zero is no-op */
2584
2585         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2586                 if (!populated_zone(zone))
2587                         continue;
2588                 scan_zone_unevictable_pages(zone);
2589         }
2590         return 1;
2591 }
2592
2593
2594 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
2595                         read_scan_unevictable_node,
2596                         write_scan_unevictable_node);
2597
2598 int scan_unevictable_register_node(struct node *node)
2599 {
2600         return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
2601 }
2602
2603 void scan_unevictable_unregister_node(struct node *node)
2604 {
2605         sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
2606 }
2607
2608 #endif