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