mm: unlockless reclaim
[linux-2.6.git] / mm / vmscan.c
1 /*
2  *  linux/mm/vmscan.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  *
6  *  Swap reorganised 29.12.95, Stephen Tweedie.
7  *  kswapd added: 7.1.96  sct
8  *  Removed kswapd_ctl limits, and swap out as many pages as needed
9  *  to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10  *  Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11  *  Multiqueue VM started 5.8.00, Rik van Riel.
12  */
13
14 #include <linux/mm.h>
15 #include <linux/module.h>
16 #include <linux/slab.h>
17 #include <linux/kernel_stat.h>
18 #include <linux/swap.h>
19 #include <linux/pagemap.h>
20 #include <linux/init.h>
21 #include <linux/highmem.h>
22 #include <linux/vmstat.h>
23 #include <linux/file.h>
24 #include <linux/writeback.h>
25 #include <linux/blkdev.h>
26 #include <linux/buffer_head.h>  /* for try_to_release_page(),
27                                         buffer_heads_over_limit */
28 #include <linux/mm_inline.h>
29 #include <linux/pagevec.h>
30 #include <linux/backing-dev.h>
31 #include <linux/rmap.h>
32 #include <linux/topology.h>
33 #include <linux/cpu.h>
34 #include <linux/cpuset.h>
35 #include <linux/notifier.h>
36 #include <linux/rwsem.h>
37 #include <linux/delay.h>
38 #include <linux/kthread.h>
39 #include <linux/freezer.h>
40 #include <linux/memcontrol.h>
41 #include <linux/delayacct.h>
42 #include <linux/sysctl.h>
43
44 #include <asm/tlbflush.h>
45 #include <asm/div64.h>
46
47 #include <linux/swapops.h>
48
49 #include "internal.h"
50
51 struct scan_control {
52         /* Incremented by the number of inactive pages that were scanned */
53         unsigned long nr_scanned;
54
55         /* 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                         switch (try_to_munlock(page)) {
627                         case SWAP_FAIL:         /* shouldn't happen */
628                         case SWAP_AGAIN:
629                                 goto keep_locked;
630                         case SWAP_MLOCK:
631                                 goto cull_mlocked;
632                         case SWAP_SUCCESS:
633                                 ; /* fall thru'; add to swap cache */
634                         }
635                         if (!add_to_swap(page, GFP_ATOMIC))
636                                 goto activate_locked;
637                 }
638 #endif /* CONFIG_SWAP */
639
640                 mapping = page_mapping(page);
641
642                 /*
643                  * The page is mapped into the page tables of one or more
644                  * processes. Try to unmap it here.
645                  */
646                 if (page_mapped(page) && mapping) {
647                         switch (try_to_unmap(page, 0)) {
648                         case SWAP_FAIL:
649                                 goto activate_locked;
650                         case SWAP_AGAIN:
651                                 goto keep_locked;
652                         case SWAP_MLOCK:
653                                 goto cull_mlocked;
654                         case SWAP_SUCCESS:
655                                 ; /* try to free the page below */
656                         }
657                 }
658
659                 if (PageDirty(page)) {
660                         if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && referenced)
661                                 goto keep_locked;
662                         if (!may_enter_fs)
663                                 goto keep_locked;
664                         if (!sc->may_writepage)
665                                 goto keep_locked;
666
667                         /* Page is dirty, try to write it out here */
668                         switch (pageout(page, mapping, sync_writeback)) {
669                         case PAGE_KEEP:
670                                 goto keep_locked;
671                         case PAGE_ACTIVATE:
672                                 goto activate_locked;
673                         case PAGE_SUCCESS:
674                                 if (PageWriteback(page) || PageDirty(page))
675                                         goto keep;
676                                 /*
677                                  * A synchronous write - probably a ramdisk.  Go
678                                  * ahead and try to reclaim the page.
679                                  */
680                                 if (!trylock_page(page))
681                                         goto keep;
682                                 if (PageDirty(page) || PageWriteback(page))
683                                         goto keep_locked;
684                                 mapping = page_mapping(page);
685                         case PAGE_CLEAN:
686                                 ; /* try to free the page below */
687                         }
688                 }
689
690                 /*
691                  * If the page has buffers, try to free the buffer mappings
692                  * associated with this page. If we succeed we try to free
693                  * the page as well.
694                  *
695                  * We do this even if the page is PageDirty().
696                  * try_to_release_page() does not perform I/O, but it is
697                  * possible for a page to have PageDirty set, but it is actually
698                  * clean (all its buffers are clean).  This happens if the
699                  * buffers were written out directly, with submit_bh(). ext3
700                  * will do this, as well as the blockdev mapping.
701                  * try_to_release_page() will discover that cleanness and will
702                  * drop the buffers and mark the page clean - it can be freed.
703                  *
704                  * Rarely, pages can have buffers and no ->mapping.  These are
705                  * the pages which were not successfully invalidated in
706                  * truncate_complete_page().  We try to drop those buffers here
707                  * and if that worked, and the page is no longer mapped into
708                  * process address space (page_count == 1) it can be freed.
709                  * Otherwise, leave the page on the LRU so it is swappable.
710                  */
711                 if (PagePrivate(page)) {
712                         if (!try_to_release_page(page, sc->gfp_mask))
713                                 goto activate_locked;
714                         if (!mapping && page_count(page) == 1) {
715                                 unlock_page(page);
716                                 if (put_page_testzero(page))
717                                         goto free_it;
718                                 else {
719                                         /*
720                                          * rare race with speculative reference.
721                                          * the speculative reference will free
722                                          * this page shortly, so we may
723                                          * increment nr_reclaimed here (and
724                                          * leave it off the LRU).
725                                          */
726                                         nr_reclaimed++;
727                                         continue;
728                                 }
729                         }
730                 }
731
732                 if (!mapping || !__remove_mapping(mapping, page))
733                         goto keep_locked;
734
735                 /*
736                  * At this point, we have no other references and there is
737                  * no way to pick any more up (removed from LRU, removed
738                  * from pagecache). Can use non-atomic bitops now (and
739                  * we obviously don't have to worry about waking up a process
740                  * waiting on the page lock, because there are no references.
741                  */
742                 __clear_page_locked(page);
743 free_it:
744                 nr_reclaimed++;
745                 if (!pagevec_add(&freed_pvec, page)) {
746                         __pagevec_free(&freed_pvec);
747                         pagevec_reinit(&freed_pvec);
748                 }
749                 continue;
750
751 cull_mlocked:
752                 unlock_page(page);
753                 putback_lru_page(page);
754                 continue;
755
756 activate_locked:
757                 /* Not a candidate for swapping, so reclaim swap space. */
758                 if (PageSwapCache(page) && vm_swap_full())
759                         remove_exclusive_swap_page_ref(page);
760                 VM_BUG_ON(PageActive(page));
761                 SetPageActive(page);
762                 pgactivate++;
763 keep_locked:
764                 unlock_page(page);
765 keep:
766                 list_add(&page->lru, &ret_pages);
767                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
768         }
769         list_splice(&ret_pages, page_list);
770         if (pagevec_count(&freed_pvec))
771                 __pagevec_free(&freed_pvec);
772         count_vm_events(PGACTIVATE, pgactivate);
773         return nr_reclaimed;
774 }
775
776 /* LRU Isolation modes. */
777 #define ISOLATE_INACTIVE 0      /* Isolate inactive pages. */
778 #define ISOLATE_ACTIVE 1        /* Isolate active pages. */
779 #define ISOLATE_BOTH 2          /* Isolate both active and inactive pages. */
780
781 /*
782  * Attempt to remove the specified page from its LRU.  Only take this page
783  * if it is of the appropriate PageActive status.  Pages which are being
784  * freed elsewhere are also ignored.
785  *
786  * page:        page to consider
787  * mode:        one of the LRU isolation modes defined above
788  *
789  * returns 0 on success, -ve errno on failure.
790  */
791 int __isolate_lru_page(struct page *page, int mode, int file)
792 {
793         int ret = -EINVAL;
794
795         /* Only take pages on the LRU. */
796         if (!PageLRU(page))
797                 return ret;
798
799         /*
800          * When checking the active state, we need to be sure we are
801          * dealing with comparible boolean values.  Take the logical not
802          * of each.
803          */
804         if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
805                 return ret;
806
807         if (mode != ISOLATE_BOTH && (!page_is_file_cache(page) != !file))
808                 return ret;
809
810         /*
811          * When this function is being called for lumpy reclaim, we
812          * initially look into all LRU pages, active, inactive and
813          * unevictable; only give shrink_page_list evictable pages.
814          */
815         if (PageUnevictable(page))
816                 return ret;
817
818         ret = -EBUSY;
819         if (likely(get_page_unless_zero(page))) {
820                 /*
821                  * Be careful not to clear PageLRU until after we're
822                  * sure the page is not being freed elsewhere -- the
823                  * page release code relies on it.
824                  */
825                 ClearPageLRU(page);
826                 ret = 0;
827         }
828
829         return ret;
830 }
831
832 /*
833  * zone->lru_lock is heavily contended.  Some of the functions that
834  * shrink the lists perform better by taking out a batch of pages
835  * and working on them outside the LRU lock.
836  *
837  * For pagecache intensive workloads, this function is the hottest
838  * spot in the kernel (apart from copy_*_user functions).
839  *
840  * Appropriate locks must be held before calling this function.
841  *
842  * @nr_to_scan: The number of pages to look through on the list.
843  * @src:        The LRU list to pull pages off.
844  * @dst:        The temp list to put pages on to.
845  * @scanned:    The number of pages that were scanned.
846  * @order:      The caller's attempted allocation order
847  * @mode:       One of the LRU isolation modes
848  * @file:       True [1] if isolating file [!anon] pages
849  *
850  * returns how many pages were moved onto *@dst.
851  */
852 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
853                 struct list_head *src, struct list_head *dst,
854                 unsigned long *scanned, int order, int mode, int file)
855 {
856         unsigned long nr_taken = 0;
857         unsigned long scan;
858
859         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
860                 struct page *page;
861                 unsigned long pfn;
862                 unsigned long end_pfn;
863                 unsigned long page_pfn;
864                 int zone_id;
865
866                 page = lru_to_page(src);
867                 prefetchw_prev_lru_page(page, src, flags);
868
869                 VM_BUG_ON(!PageLRU(page));
870
871                 switch (__isolate_lru_page(page, mode, file)) {
872                 case 0:
873                         list_move(&page->lru, dst);
874                         nr_taken++;
875                         break;
876
877                 case -EBUSY:
878                         /* else it is being freed elsewhere */
879                         list_move(&page->lru, src);
880                         continue;
881
882                 default:
883                         BUG();
884                 }
885
886                 if (!order)
887                         continue;
888
889                 /*
890                  * Attempt to take all pages in the order aligned region
891                  * surrounding the tag page.  Only take those pages of
892                  * the same active state as that tag page.  We may safely
893                  * round the target page pfn down to the requested order
894                  * as the mem_map is guarenteed valid out to MAX_ORDER,
895                  * where that page is in a different zone we will detect
896                  * it from its zone id and abort this block scan.
897                  */
898                 zone_id = page_zone_id(page);
899                 page_pfn = page_to_pfn(page);
900                 pfn = page_pfn & ~((1 << order) - 1);
901                 end_pfn = pfn + (1 << order);
902                 for (; pfn < end_pfn; pfn++) {
903                         struct page *cursor_page;
904
905                         /* The target page is in the block, ignore it. */
906                         if (unlikely(pfn == page_pfn))
907                                 continue;
908
909                         /* Avoid holes within the zone. */
910                         if (unlikely(!pfn_valid_within(pfn)))
911                                 break;
912
913                         cursor_page = pfn_to_page(pfn);
914
915                         /* Check that we have not crossed a zone boundary. */
916                         if (unlikely(page_zone_id(cursor_page) != zone_id))
917                                 continue;
918                         switch (__isolate_lru_page(cursor_page, mode, file)) {
919                         case 0:
920                                 list_move(&cursor_page->lru, dst);
921                                 nr_taken++;
922                                 scan++;
923                                 break;
924
925                         case -EBUSY:
926                                 /* else it is being freed elsewhere */
927                                 list_move(&cursor_page->lru, src);
928                         default:
929                                 break;  /* ! on LRU or wrong list */
930                         }
931                 }
932         }
933
934         *scanned = scan;
935         return nr_taken;
936 }
937
938 static unsigned long isolate_pages_global(unsigned long nr,
939                                         struct list_head *dst,
940                                         unsigned long *scanned, int order,
941                                         int mode, struct zone *z,
942                                         struct mem_cgroup *mem_cont,
943                                         int active, int file)
944 {
945         int lru = LRU_BASE;
946         if (active)
947                 lru += LRU_ACTIVE;
948         if (file)
949                 lru += LRU_FILE;
950         return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
951                                                                 mode, !!file);
952 }
953
954 /*
955  * clear_active_flags() is a helper for shrink_active_list(), clearing
956  * any active bits from the pages in the list.
957  */
958 static unsigned long clear_active_flags(struct list_head *page_list,
959                                         unsigned int *count)
960 {
961         int nr_active = 0;
962         int lru;
963         struct page *page;
964
965         list_for_each_entry(page, page_list, lru) {
966                 lru = page_is_file_cache(page);
967                 if (PageActive(page)) {
968                         lru += LRU_ACTIVE;
969                         ClearPageActive(page);
970                         nr_active++;
971                 }
972                 count[lru]++;
973         }
974
975         return nr_active;
976 }
977
978 /**
979  * isolate_lru_page - tries to isolate a page from its LRU list
980  * @page: page to isolate from its LRU list
981  *
982  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
983  * vmstat statistic corresponding to whatever LRU list the page was on.
984  *
985  * Returns 0 if the page was removed from an LRU list.
986  * Returns -EBUSY if the page was not on an LRU list.
987  *
988  * The returned page will have PageLRU() cleared.  If it was found on
989  * the active list, it will have PageActive set.  If it was found on
990  * the unevictable list, it will have the PageUnevictable bit set. That flag
991  * may need to be cleared by the caller before letting the page go.
992  *
993  * The vmstat statistic corresponding to the list on which the page was
994  * found will be decremented.
995  *
996  * Restrictions:
997  * (1) Must be called with an elevated refcount on the page. This is a
998  *     fundamentnal difference from isolate_lru_pages (which is called
999  *     without a stable reference).
1000  * (2) the lru_lock must not be held.
1001  * (3) interrupts must be enabled.
1002  */
1003 int isolate_lru_page(struct page *page)
1004 {
1005         int ret = -EBUSY;
1006
1007         if (PageLRU(page)) {
1008                 struct zone *zone = page_zone(page);
1009
1010                 spin_lock_irq(&zone->lru_lock);
1011                 if (PageLRU(page) && get_page_unless_zero(page)) {
1012                         int lru = page_lru(page);
1013                         ret = 0;
1014                         ClearPageLRU(page);
1015
1016                         del_page_from_lru_list(zone, page, lru);
1017                 }
1018                 spin_unlock_irq(&zone->lru_lock);
1019         }
1020         return ret;
1021 }
1022
1023 /*
1024  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1025  * of reclaimed pages
1026  */
1027 static unsigned long shrink_inactive_list(unsigned long max_scan,
1028                         struct zone *zone, struct scan_control *sc,
1029                         int priority, int file)
1030 {
1031         LIST_HEAD(page_list);
1032         struct pagevec pvec;
1033         unsigned long nr_scanned = 0;
1034         unsigned long nr_reclaimed = 0;
1035
1036         pagevec_init(&pvec, 1);
1037
1038         lru_add_drain();
1039         spin_lock_irq(&zone->lru_lock);
1040         do {
1041                 struct page *page;
1042                 unsigned long nr_taken;
1043                 unsigned long nr_scan;
1044                 unsigned long nr_freed;
1045                 unsigned long nr_active;
1046                 unsigned int count[NR_LRU_LISTS] = { 0, };
1047                 int mode = ISOLATE_INACTIVE;
1048
1049                 /*
1050                  * If we need a large contiguous chunk of memory, or have
1051                  * trouble getting a small set of contiguous pages, we
1052                  * will reclaim both active and inactive pages.
1053                  *
1054                  * We use the same threshold as pageout congestion_wait below.
1055                  */
1056                 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1057                         mode = ISOLATE_BOTH;
1058                 else if (sc->order && priority < DEF_PRIORITY - 2)
1059                         mode = ISOLATE_BOTH;
1060
1061                 nr_taken = sc->isolate_pages(sc->swap_cluster_max,
1062                              &page_list, &nr_scan, sc->order, mode,
1063                                 zone, sc->mem_cgroup, 0, file);
1064                 nr_active = clear_active_flags(&page_list, count);
1065                 __count_vm_events(PGDEACTIVATE, nr_active);
1066
1067                 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1068                                                 -count[LRU_ACTIVE_FILE]);
1069                 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1070                                                 -count[LRU_INACTIVE_FILE]);
1071                 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1072                                                 -count[LRU_ACTIVE_ANON]);
1073                 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1074                                                 -count[LRU_INACTIVE_ANON]);
1075
1076                 if (scan_global_lru(sc)) {
1077                         zone->pages_scanned += nr_scan;
1078                         zone->recent_scanned[0] += count[LRU_INACTIVE_ANON];
1079                         zone->recent_scanned[0] += count[LRU_ACTIVE_ANON];
1080                         zone->recent_scanned[1] += count[LRU_INACTIVE_FILE];
1081                         zone->recent_scanned[1] += count[LRU_ACTIVE_FILE];
1082                 }
1083                 spin_unlock_irq(&zone->lru_lock);
1084
1085                 nr_scanned += nr_scan;
1086                 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
1087
1088                 /*
1089                  * If we are direct reclaiming for contiguous pages and we do
1090                  * not reclaim everything in the list, try again and wait
1091                  * for IO to complete. This will stall high-order allocations
1092                  * but that should be acceptable to the caller
1093                  */
1094                 if (nr_freed < nr_taken && !current_is_kswapd() &&
1095                                         sc->order > PAGE_ALLOC_COSTLY_ORDER) {
1096                         congestion_wait(WRITE, HZ/10);
1097
1098                         /*
1099                          * The attempt at page out may have made some
1100                          * of the pages active, mark them inactive again.
1101                          */
1102                         nr_active = clear_active_flags(&page_list, count);
1103                         count_vm_events(PGDEACTIVATE, nr_active);
1104
1105                         nr_freed += shrink_page_list(&page_list, sc,
1106                                                         PAGEOUT_IO_SYNC);
1107                 }
1108
1109                 nr_reclaimed += nr_freed;
1110                 local_irq_disable();
1111                 if (current_is_kswapd()) {
1112                         __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scan);
1113                         __count_vm_events(KSWAPD_STEAL, nr_freed);
1114                 } else if (scan_global_lru(sc))
1115                         __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scan);
1116
1117                 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
1118
1119                 if (nr_taken == 0)
1120                         goto done;
1121
1122                 spin_lock(&zone->lru_lock);
1123                 /*
1124                  * Put back any unfreeable pages.
1125                  */
1126                 while (!list_empty(&page_list)) {
1127                         int lru;
1128                         page = lru_to_page(&page_list);
1129                         VM_BUG_ON(PageLRU(page));
1130                         list_del(&page->lru);
1131                         if (unlikely(!page_evictable(page, NULL))) {
1132                                 spin_unlock_irq(&zone->lru_lock);
1133                                 putback_lru_page(page);
1134                                 spin_lock_irq(&zone->lru_lock);
1135                                 continue;
1136                         }
1137                         SetPageLRU(page);
1138                         lru = page_lru(page);
1139                         add_page_to_lru_list(zone, page, lru);
1140                         mem_cgroup_move_lists(page, lru);
1141                         if (PageActive(page) && scan_global_lru(sc)) {
1142                                 int file = !!page_is_file_cache(page);
1143                                 zone->recent_rotated[file]++;
1144                         }
1145                         if (!pagevec_add(&pvec, page)) {
1146                                 spin_unlock_irq(&zone->lru_lock);
1147                                 __pagevec_release(&pvec);
1148                                 spin_lock_irq(&zone->lru_lock);
1149                         }
1150                 }
1151         } while (nr_scanned < max_scan);
1152         spin_unlock(&zone->lru_lock);
1153 done:
1154         local_irq_enable();
1155         pagevec_release(&pvec);
1156         return nr_reclaimed;
1157 }
1158
1159 /*
1160  * We are about to scan this zone at a certain priority level.  If that priority
1161  * level is smaller (ie: more urgent) than the previous priority, then note
1162  * that priority level within the zone.  This is done so that when the next
1163  * process comes in to scan this zone, it will immediately start out at this
1164  * priority level rather than having to build up its own scanning priority.
1165  * Here, this priority affects only the reclaim-mapped threshold.
1166  */
1167 static inline void note_zone_scanning_priority(struct zone *zone, int priority)
1168 {
1169         if (priority < zone->prev_priority)
1170                 zone->prev_priority = priority;
1171 }
1172
1173 static inline int zone_is_near_oom(struct zone *zone)
1174 {
1175         return zone->pages_scanned >= (zone_lru_pages(zone) * 3);
1176 }
1177
1178 /*
1179  * This moves pages from the active list to the inactive list.
1180  *
1181  * We move them the other way if the page is referenced by one or more
1182  * processes, from rmap.
1183  *
1184  * If the pages are mostly unmapped, the processing is fast and it is
1185  * appropriate to hold zone->lru_lock across the whole operation.  But if
1186  * the pages are mapped, the processing is slow (page_referenced()) so we
1187  * should drop zone->lru_lock around each page.  It's impossible to balance
1188  * this, so instead we remove the pages from the LRU while processing them.
1189  * It is safe to rely on PG_active against the non-LRU pages in here because
1190  * nobody will play with that bit on a non-LRU page.
1191  *
1192  * The downside is that we have to touch page->_count against each page.
1193  * But we had to alter page->flags anyway.
1194  */
1195
1196
1197 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1198                         struct scan_control *sc, int priority, int file)
1199 {
1200         unsigned long pgmoved;
1201         int pgdeactivate = 0;
1202         unsigned long pgscanned;
1203         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1204         LIST_HEAD(l_inactive);
1205         struct page *page;
1206         struct pagevec pvec;
1207         enum lru_list lru;
1208
1209         lru_add_drain();
1210         spin_lock_irq(&zone->lru_lock);
1211         pgmoved = sc->isolate_pages(nr_pages, &l_hold, &pgscanned, sc->order,
1212                                         ISOLATE_ACTIVE, zone,
1213                                         sc->mem_cgroup, 1, file);
1214         /*
1215          * zone->pages_scanned is used for detect zone's oom
1216          * mem_cgroup remembers nr_scan by itself.
1217          */
1218         if (scan_global_lru(sc)) {
1219                 zone->pages_scanned += pgscanned;
1220                 zone->recent_scanned[!!file] += pgmoved;
1221         }
1222
1223         if (file)
1224                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -pgmoved);
1225         else
1226                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -pgmoved);
1227         spin_unlock_irq(&zone->lru_lock);
1228
1229         pgmoved = 0;
1230         while (!list_empty(&l_hold)) {
1231                 cond_resched();
1232                 page = lru_to_page(&l_hold);
1233                 list_del(&page->lru);
1234
1235                 if (unlikely(!page_evictable(page, NULL))) {
1236                         putback_lru_page(page);
1237                         continue;
1238                 }
1239
1240                 /* page_referenced clears PageReferenced */
1241                 if (page_mapping_inuse(page) &&
1242                     page_referenced(page, 0, sc->mem_cgroup))
1243                         pgmoved++;
1244
1245                 list_add(&page->lru, &l_inactive);
1246         }
1247
1248         /*
1249          * Count referenced pages from currently used mappings as
1250          * rotated, even though they are moved to the inactive list.
1251          * This helps balance scan pressure between file and anonymous
1252          * pages in get_scan_ratio.
1253          */
1254         zone->recent_rotated[!!file] += pgmoved;
1255
1256         /*
1257          * Move the pages to the [file or anon] inactive list.
1258          */
1259         pagevec_init(&pvec, 1);
1260
1261         pgmoved = 0;
1262         lru = LRU_BASE + file * LRU_FILE;
1263         spin_lock_irq(&zone->lru_lock);
1264         while (!list_empty(&l_inactive)) {
1265                 page = lru_to_page(&l_inactive);
1266                 prefetchw_prev_lru_page(page, &l_inactive, flags);
1267                 VM_BUG_ON(PageLRU(page));
1268                 SetPageLRU(page);
1269                 VM_BUG_ON(!PageActive(page));
1270                 ClearPageActive(page);
1271
1272                 list_move(&page->lru, &zone->lru[lru].list);
1273                 mem_cgroup_move_lists(page, lru);
1274                 pgmoved++;
1275                 if (!pagevec_add(&pvec, page)) {
1276                         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1277                         spin_unlock_irq(&zone->lru_lock);
1278                         pgdeactivate += pgmoved;
1279                         pgmoved = 0;
1280                         if (buffer_heads_over_limit)
1281                                 pagevec_strip(&pvec);
1282                         __pagevec_release(&pvec);
1283                         spin_lock_irq(&zone->lru_lock);
1284                 }
1285         }
1286         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1287         pgdeactivate += pgmoved;
1288         if (buffer_heads_over_limit) {
1289                 spin_unlock_irq(&zone->lru_lock);
1290                 pagevec_strip(&pvec);
1291                 spin_lock_irq(&zone->lru_lock);
1292         }
1293         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1294         __count_vm_events(PGDEACTIVATE, pgdeactivate);
1295         spin_unlock_irq(&zone->lru_lock);
1296         if (vm_swap_full())
1297                 pagevec_swap_free(&pvec);
1298
1299         pagevec_release(&pvec);
1300 }
1301
1302 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1303         struct zone *zone, struct scan_control *sc, int priority)
1304 {
1305         int file = is_file_lru(lru);
1306
1307         if (lru == LRU_ACTIVE_FILE) {
1308                 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1309                 return 0;
1310         }
1311
1312         if (lru == LRU_ACTIVE_ANON &&
1313             (!scan_global_lru(sc) || inactive_anon_is_low(zone))) {
1314                 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1315                 return 0;
1316         }
1317         return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1318 }
1319
1320 /*
1321  * Determine how aggressively the anon and file LRU lists should be
1322  * scanned.  The relative value of each set of LRU lists is determined
1323  * by looking at the fraction of the pages scanned we did rotate back
1324  * onto the active list instead of evict.
1325  *
1326  * percent[0] specifies how much pressure to put on ram/swap backed
1327  * memory, while percent[1] determines pressure on the file LRUs.
1328  */
1329 static void get_scan_ratio(struct zone *zone, struct scan_control *sc,
1330                                         unsigned long *percent)
1331 {
1332         unsigned long anon, file, free;
1333         unsigned long anon_prio, file_prio;
1334         unsigned long ap, fp;
1335
1336         anon  = zone_page_state(zone, NR_ACTIVE_ANON) +
1337                 zone_page_state(zone, NR_INACTIVE_ANON);
1338         file  = zone_page_state(zone, NR_ACTIVE_FILE) +
1339                 zone_page_state(zone, NR_INACTIVE_FILE);
1340         free  = zone_page_state(zone, NR_FREE_PAGES);
1341
1342         /* If we have no swap space, do not bother scanning anon pages. */
1343         if (nr_swap_pages <= 0) {
1344                 percent[0] = 0;
1345                 percent[1] = 100;
1346                 return;
1347         }
1348
1349         /* If we have very few page cache pages, force-scan anon pages. */
1350         if (unlikely(file + free <= zone->pages_high)) {
1351                 percent[0] = 100;
1352                 percent[1] = 0;
1353                 return;
1354         }
1355
1356         /*
1357          * OK, so we have swap space and a fair amount of page cache
1358          * pages.  We use the recently rotated / recently scanned
1359          * ratios to determine how valuable each cache is.
1360          *
1361          * Because workloads change over time (and to avoid overflow)
1362          * we keep these statistics as a floating average, which ends
1363          * up weighing recent references more than old ones.
1364          *
1365          * anon in [0], file in [1]
1366          */
1367         if (unlikely(zone->recent_scanned[0] > anon / 4)) {
1368                 spin_lock_irq(&zone->lru_lock);
1369                 zone->recent_scanned[0] /= 2;
1370                 zone->recent_rotated[0] /= 2;
1371                 spin_unlock_irq(&zone->lru_lock);
1372         }
1373
1374         if (unlikely(zone->recent_scanned[1] > file / 4)) {
1375                 spin_lock_irq(&zone->lru_lock);
1376                 zone->recent_scanned[1] /= 2;
1377                 zone->recent_rotated[1] /= 2;
1378                 spin_unlock_irq(&zone->lru_lock);
1379         }
1380
1381         /*
1382          * With swappiness at 100, anonymous and file have the same priority.
1383          * This scanning priority is essentially the inverse of IO cost.
1384          */
1385         anon_prio = sc->swappiness;
1386         file_prio = 200 - sc->swappiness;
1387
1388         /*
1389          *                  anon       recent_rotated[0]
1390          * %anon = 100 * ----------- / ----------------- * IO cost
1391          *               anon + file      rotate_sum
1392          */
1393         ap = (anon_prio + 1) * (zone->recent_scanned[0] + 1);
1394         ap /= zone->recent_rotated[0] + 1;
1395
1396         fp = (file_prio + 1) * (zone->recent_scanned[1] + 1);
1397         fp /= zone->recent_rotated[1] + 1;
1398
1399         /* Normalize to percentages */
1400         percent[0] = 100 * ap / (ap + fp + 1);
1401         percent[1] = 100 - percent[0];
1402 }
1403
1404
1405 /*
1406  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1407  */
1408 static unsigned long shrink_zone(int priority, struct zone *zone,
1409                                 struct scan_control *sc)
1410 {
1411         unsigned long nr[NR_LRU_LISTS];
1412         unsigned long nr_to_scan;
1413         unsigned long nr_reclaimed = 0;
1414         unsigned long percent[2];       /* anon @ 0; file @ 1 */
1415         enum lru_list l;
1416
1417         get_scan_ratio(zone, sc, percent);
1418
1419         for_each_evictable_lru(l) {
1420                 if (scan_global_lru(sc)) {
1421                         int file = is_file_lru(l);
1422                         int scan;
1423
1424                         scan = zone_page_state(zone, NR_LRU_BASE + l);
1425                         if (priority) {
1426                                 scan >>= priority;
1427                                 scan = (scan * percent[file]) / 100;
1428                         }
1429                         zone->lru[l].nr_scan += scan;
1430                         nr[l] = zone->lru[l].nr_scan;
1431                         if (nr[l] >= sc->swap_cluster_max)
1432                                 zone->lru[l].nr_scan = 0;
1433                         else
1434                                 nr[l] = 0;
1435                 } else {
1436                         /*
1437                          * This reclaim occurs not because zone memory shortage
1438                          * but because memory controller hits its limit.
1439                          * Don't modify zone reclaim related data.
1440                          */
1441                         nr[l] = mem_cgroup_calc_reclaim(sc->mem_cgroup, zone,
1442                                                                 priority, l);
1443                 }
1444         }
1445
1446         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1447                                         nr[LRU_INACTIVE_FILE]) {
1448                 for_each_evictable_lru(l) {
1449                         if (nr[l]) {
1450                                 nr_to_scan = min(nr[l],
1451                                         (unsigned long)sc->swap_cluster_max);
1452                                 nr[l] -= nr_to_scan;
1453
1454                                 nr_reclaimed += shrink_list(l, nr_to_scan,
1455                                                         zone, sc, priority);
1456                         }
1457                 }
1458         }
1459
1460         /*
1461          * Even if we did not try to evict anon pages at all, we want to
1462          * rebalance the anon lru active/inactive ratio.
1463          */
1464         if (!scan_global_lru(sc) || inactive_anon_is_low(zone))
1465                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1466         else if (!scan_global_lru(sc))
1467                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1468
1469         throttle_vm_writeout(sc->gfp_mask);
1470         return nr_reclaimed;
1471 }
1472
1473 /*
1474  * This is the direct reclaim path, for page-allocating processes.  We only
1475  * try to reclaim pages from zones which will satisfy the caller's allocation
1476  * request.
1477  *
1478  * We reclaim from a zone even if that zone is over pages_high.  Because:
1479  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1480  *    allocation or
1481  * b) The zones may be over pages_high but they must go *over* pages_high to
1482  *    satisfy the `incremental min' zone defense algorithm.
1483  *
1484  * Returns the number of reclaimed pages.
1485  *
1486  * If a zone is deemed to be full of pinned pages then just give it a light
1487  * scan then give up on it.
1488  */
1489 static unsigned long shrink_zones(int priority, struct zonelist *zonelist,
1490                                         struct scan_control *sc)
1491 {
1492         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1493         unsigned long nr_reclaimed = 0;
1494         struct zoneref *z;
1495         struct zone *zone;
1496
1497         sc->all_unreclaimable = 1;
1498         for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1499                 if (!populated_zone(zone))
1500                         continue;
1501                 /*
1502                  * Take care memory controller reclaiming has small influence
1503                  * to global LRU.
1504                  */
1505                 if (scan_global_lru(sc)) {
1506                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1507                                 continue;
1508                         note_zone_scanning_priority(zone, priority);
1509
1510                         if (zone_is_all_unreclaimable(zone) &&
1511                                                 priority != DEF_PRIORITY)
1512                                 continue;       /* Let kswapd poll it */
1513                         sc->all_unreclaimable = 0;
1514                 } else {
1515                         /*
1516                          * Ignore cpuset limitation here. We just want to reduce
1517                          * # of used pages by us regardless of memory shortage.
1518                          */
1519                         sc->all_unreclaimable = 0;
1520                         mem_cgroup_note_reclaim_priority(sc->mem_cgroup,
1521                                                         priority);
1522                 }
1523
1524                 nr_reclaimed += shrink_zone(priority, zone, sc);
1525         }
1526
1527         return nr_reclaimed;
1528 }
1529
1530 /*
1531  * This is the main entry point to direct page reclaim.
1532  *
1533  * If a full scan of the inactive list fails to free enough memory then we
1534  * are "out of memory" and something needs to be killed.
1535  *
1536  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1537  * high - the zone may be full of dirty or under-writeback pages, which this
1538  * caller can't do much about.  We kick pdflush and take explicit naps in the
1539  * hope that some of these pages can be written.  But if the allocating task
1540  * holds filesystem locks which prevent writeout this might not work, and the
1541  * allocation attempt will fail.
1542  *
1543  * returns:     0, if no pages reclaimed
1544  *              else, the number of pages reclaimed
1545  */
1546 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
1547                                         struct scan_control *sc)
1548 {
1549         int priority;
1550         unsigned long ret = 0;
1551         unsigned long total_scanned = 0;
1552         unsigned long nr_reclaimed = 0;
1553         struct reclaim_state *reclaim_state = current->reclaim_state;
1554         unsigned long lru_pages = 0;
1555         struct zoneref *z;
1556         struct zone *zone;
1557         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1558
1559         delayacct_freepages_start();
1560
1561         if (scan_global_lru(sc))
1562                 count_vm_event(ALLOCSTALL);
1563         /*
1564          * mem_cgroup will not do shrink_slab.
1565          */
1566         if (scan_global_lru(sc)) {
1567                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1568
1569                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1570                                 continue;
1571
1572                         lru_pages += zone_lru_pages(zone);
1573                 }
1574         }
1575
1576         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1577                 sc->nr_scanned = 0;
1578                 if (!priority)
1579                         disable_swap_token();
1580                 nr_reclaimed += shrink_zones(priority, zonelist, sc);
1581                 /*
1582                  * Don't shrink slabs when reclaiming memory from
1583                  * over limit cgroups
1584                  */
1585                 if (scan_global_lru(sc)) {
1586                         shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
1587                         if (reclaim_state) {
1588                                 nr_reclaimed += reclaim_state->reclaimed_slab;
1589                                 reclaim_state->reclaimed_slab = 0;
1590                         }
1591                 }
1592                 total_scanned += sc->nr_scanned;
1593                 if (nr_reclaimed >= sc->swap_cluster_max) {
1594                         ret = nr_reclaimed;
1595                         goto out;
1596                 }
1597
1598                 /*
1599                  * Try to write back as many pages as we just scanned.  This
1600                  * tends to cause slow streaming writers to write data to the
1601                  * disk smoothly, at the dirtying rate, which is nice.   But
1602                  * that's undesirable in laptop mode, where we *want* lumpy
1603                  * writeout.  So in laptop mode, write out the whole world.
1604                  */
1605                 if (total_scanned > sc->swap_cluster_max +
1606                                         sc->swap_cluster_max / 2) {
1607                         wakeup_pdflush(laptop_mode ? 0 : total_scanned);
1608                         sc->may_writepage = 1;
1609                 }
1610
1611                 /* Take a nap, wait for some writeback to complete */
1612                 if (sc->nr_scanned && priority < DEF_PRIORITY - 2)
1613                         congestion_wait(WRITE, HZ/10);
1614         }
1615         /* top priority shrink_zones still had more to do? don't OOM, then */
1616         if (!sc->all_unreclaimable && scan_global_lru(sc))
1617                 ret = nr_reclaimed;
1618 out:
1619         /*
1620          * Now that we've scanned all the zones at this priority level, note
1621          * that level within the zone so that the next thread which performs
1622          * scanning of this zone will immediately start out at this priority
1623          * level.  This affects only the decision whether or not to bring
1624          * mapped pages onto the inactive list.
1625          */
1626         if (priority < 0)
1627                 priority = 0;
1628
1629         if (scan_global_lru(sc)) {
1630                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1631
1632                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1633                                 continue;
1634
1635                         zone->prev_priority = priority;
1636                 }
1637         } else
1638                 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority);
1639
1640         delayacct_freepages_end();
1641
1642         return ret;
1643 }
1644
1645 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
1646                                                                 gfp_t gfp_mask)
1647 {
1648         struct scan_control sc = {
1649                 .gfp_mask = gfp_mask,
1650                 .may_writepage = !laptop_mode,
1651                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1652                 .may_swap = 1,
1653                 .swappiness = vm_swappiness,
1654                 .order = order,
1655                 .mem_cgroup = NULL,
1656                 .isolate_pages = isolate_pages_global,
1657         };
1658
1659         return do_try_to_free_pages(zonelist, &sc);
1660 }
1661
1662 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
1663
1664 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
1665                                                 gfp_t gfp_mask)
1666 {
1667         struct scan_control sc = {
1668                 .may_writepage = !laptop_mode,
1669                 .may_swap = 1,
1670                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1671                 .swappiness = vm_swappiness,
1672                 .order = 0,
1673                 .mem_cgroup = mem_cont,
1674                 .isolate_pages = mem_cgroup_isolate_pages,
1675         };
1676         struct zonelist *zonelist;
1677
1678         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1679                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1680         zonelist = NODE_DATA(numa_node_id())->node_zonelists;
1681         return do_try_to_free_pages(zonelist, &sc);
1682 }
1683 #endif
1684
1685 /*
1686  * For kswapd, balance_pgdat() will work across all this node's zones until
1687  * they are all at pages_high.
1688  *
1689  * Returns the number of pages which were actually freed.
1690  *
1691  * There is special handling here for zones which are full of pinned pages.
1692  * This can happen if the pages are all mlocked, or if they are all used by
1693  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1694  * What we do is to detect the case where all pages in the zone have been
1695  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1696  * dead and from now on, only perform a short scan.  Basically we're polling
1697  * the zone for when the problem goes away.
1698  *
1699  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
1700  * zones which have free_pages > pages_high, but once a zone is found to have
1701  * free_pages <= pages_high, we scan that zone and the lower zones regardless
1702  * of the number of free pages in the lower zones.  This interoperates with
1703  * the page allocator fallback scheme to ensure that aging of pages is balanced
1704  * across the zones.
1705  */
1706 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
1707 {
1708         int all_zones_ok;
1709         int priority;
1710         int i;
1711         unsigned long total_scanned;
1712         unsigned long nr_reclaimed;
1713         struct reclaim_state *reclaim_state = current->reclaim_state;
1714         struct scan_control sc = {
1715                 .gfp_mask = GFP_KERNEL,
1716                 .may_swap = 1,
1717                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1718                 .swappiness = vm_swappiness,
1719                 .order = order,
1720                 .mem_cgroup = NULL,
1721                 .isolate_pages = isolate_pages_global,
1722         };
1723         /*
1724          * temp_priority is used to remember the scanning priority at which
1725          * this zone was successfully refilled to free_pages == pages_high.
1726          */
1727         int temp_priority[MAX_NR_ZONES];
1728
1729 loop_again:
1730         total_scanned = 0;
1731         nr_reclaimed = 0;
1732         sc.may_writepage = !laptop_mode;
1733         count_vm_event(PAGEOUTRUN);
1734
1735         for (i = 0; i < pgdat->nr_zones; i++)
1736                 temp_priority[i] = DEF_PRIORITY;
1737
1738         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1739                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1740                 unsigned long lru_pages = 0;
1741
1742                 /* The swap token gets in the way of swapout... */
1743                 if (!priority)
1744                         disable_swap_token();
1745
1746                 all_zones_ok = 1;
1747
1748                 /*
1749                  * Scan in the highmem->dma direction for the highest
1750                  * zone which needs scanning
1751                  */
1752                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1753                         struct zone *zone = pgdat->node_zones + i;
1754
1755                         if (!populated_zone(zone))
1756                                 continue;
1757
1758                         if (zone_is_all_unreclaimable(zone) &&
1759                             priority != DEF_PRIORITY)
1760                                 continue;
1761
1762                         /*
1763                          * Do some background aging of the anon list, to give
1764                          * pages a chance to be referenced before reclaiming.
1765                          */
1766                         if (inactive_anon_is_low(zone))
1767                                 shrink_active_list(SWAP_CLUSTER_MAX, zone,
1768                                                         &sc, priority, 0);
1769
1770                         if (!zone_watermark_ok(zone, order, zone->pages_high,
1771                                                0, 0)) {
1772                                 end_zone = i;
1773                                 break;
1774                         }
1775                 }
1776                 if (i < 0)
1777                         goto out;
1778
1779                 for (i = 0; i <= end_zone; i++) {
1780                         struct zone *zone = pgdat->node_zones + i;
1781
1782                         lru_pages += zone_lru_pages(zone);
1783                 }
1784
1785                 /*
1786                  * Now scan the zone in the dma->highmem direction, stopping
1787                  * at the last zone which needs scanning.
1788                  *
1789                  * We do this because the page allocator works in the opposite
1790                  * direction.  This prevents the page allocator from allocating
1791                  * pages behind kswapd's direction of progress, which would
1792                  * cause too much scanning of the lower zones.
1793                  */
1794                 for (i = 0; i <= end_zone; i++) {
1795                         struct zone *zone = pgdat->node_zones + i;
1796                         int nr_slab;
1797
1798                         if (!populated_zone(zone))
1799                                 continue;
1800
1801                         if (zone_is_all_unreclaimable(zone) &&
1802                                         priority != DEF_PRIORITY)
1803                                 continue;
1804
1805                         if (!zone_watermark_ok(zone, order, zone->pages_high,
1806                                                end_zone, 0))
1807                                 all_zones_ok = 0;
1808                         temp_priority[i] = priority;
1809                         sc.nr_scanned = 0;
1810                         note_zone_scanning_priority(zone, priority);
1811                         /*
1812                          * We put equal pressure on every zone, unless one
1813                          * zone has way too many pages free already.
1814                          */
1815                         if (!zone_watermark_ok(zone, order, 8*zone->pages_high,
1816                                                 end_zone, 0))
1817                                 nr_reclaimed += shrink_zone(priority, zone, &sc);
1818                         reclaim_state->reclaimed_slab = 0;
1819                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1820                                                 lru_pages);
1821                         nr_reclaimed += reclaim_state->reclaimed_slab;
1822                         total_scanned += sc.nr_scanned;
1823                         if (zone_is_all_unreclaimable(zone))
1824                                 continue;
1825                         if (nr_slab == 0 && zone->pages_scanned >=
1826                                                 (zone_lru_pages(zone) * 6))
1827                                         zone_set_flag(zone,
1828                                                       ZONE_ALL_UNRECLAIMABLE);
1829                         /*
1830                          * If we've done a decent amount of scanning and
1831                          * the reclaim ratio is low, start doing writepage
1832                          * even in laptop mode
1833                          */
1834                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1835                             total_scanned > nr_reclaimed + nr_reclaimed / 2)
1836                                 sc.may_writepage = 1;
1837                 }
1838                 if (all_zones_ok)
1839                         break;          /* kswapd: all done */
1840                 /*
1841                  * OK, kswapd is getting into trouble.  Take a nap, then take
1842                  * another pass across the zones.
1843                  */
1844                 if (total_scanned && priority < DEF_PRIORITY - 2)
1845                         congestion_wait(WRITE, HZ/10);
1846
1847                 /*
1848                  * We do this so kswapd doesn't build up large priorities for
1849                  * example when it is freeing in parallel with allocators. It
1850                  * matches the direct reclaim path behaviour in terms of impact
1851                  * on zone->*_priority.
1852                  */
1853                 if (nr_reclaimed >= SWAP_CLUSTER_MAX)
1854                         break;
1855         }
1856 out:
1857         /*
1858          * Note within each zone the priority level at which this zone was
1859          * brought into a happy state.  So that the next thread which scans this
1860          * zone will start out at that priority level.
1861          */
1862         for (i = 0; i < pgdat->nr_zones; i++) {
1863                 struct zone *zone = pgdat->node_zones + i;
1864
1865                 zone->prev_priority = temp_priority[i];
1866         }
1867         if (!all_zones_ok) {
1868                 cond_resched();
1869
1870                 try_to_freeze();
1871
1872                 goto loop_again;
1873         }
1874
1875         return nr_reclaimed;
1876 }
1877
1878 /*
1879  * The background pageout daemon, started as a kernel thread
1880  * from the init process.
1881  *
1882  * This basically trickles out pages so that we have _some_
1883  * free memory available even if there is no other activity
1884  * that frees anything up. This is needed for things like routing
1885  * etc, where we otherwise might have all activity going on in
1886  * asynchronous contexts that cannot page things out.
1887  *
1888  * If there are applications that are active memory-allocators
1889  * (most normal use), this basically shouldn't matter.
1890  */
1891 static int kswapd(void *p)
1892 {
1893         unsigned long order;
1894         pg_data_t *pgdat = (pg_data_t*)p;
1895         struct task_struct *tsk = current;
1896         DEFINE_WAIT(wait);
1897         struct reclaim_state reclaim_state = {
1898                 .reclaimed_slab = 0,
1899         };
1900         node_to_cpumask_ptr(cpumask, pgdat->node_id);
1901
1902         if (!cpus_empty(*cpumask))
1903                 set_cpus_allowed_ptr(tsk, cpumask);
1904         current->reclaim_state = &reclaim_state;
1905
1906         /*
1907          * Tell the memory management that we're a "memory allocator",
1908          * and that if we need more memory we should get access to it
1909          * regardless (see "__alloc_pages()"). "kswapd" should
1910          * never get caught in the normal page freeing logic.
1911          *
1912          * (Kswapd normally doesn't need memory anyway, but sometimes
1913          * you need a small amount of memory in order to be able to
1914          * page out something else, and this flag essentially protects
1915          * us from recursively trying to free more memory as we're
1916          * trying to free the first piece of memory in the first place).
1917          */
1918         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
1919         set_freezable();
1920
1921         order = 0;
1922         for ( ; ; ) {
1923                 unsigned long new_order;
1924
1925                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1926                 new_order = pgdat->kswapd_max_order;
1927                 pgdat->kswapd_max_order = 0;
1928                 if (order < new_order) {
1929                         /*
1930                          * Don't sleep if someone wants a larger 'order'
1931                          * allocation
1932                          */
1933                         order = new_order;
1934                 } else {
1935                         if (!freezing(current))
1936                                 schedule();
1937
1938                         order = pgdat->kswapd_max_order;
1939                 }
1940                 finish_wait(&pgdat->kswapd_wait, &wait);
1941
1942                 if (!try_to_freeze()) {
1943                         /* We can speed up thawing tasks if we don't call
1944                          * balance_pgdat after returning from the refrigerator
1945                          */
1946                         balance_pgdat(pgdat, order);
1947                 }
1948         }
1949         return 0;
1950 }
1951
1952 /*
1953  * A zone is low on free memory, so wake its kswapd task to service it.
1954  */
1955 void wakeup_kswapd(struct zone *zone, int order)
1956 {
1957         pg_data_t *pgdat;
1958
1959         if (!populated_zone(zone))
1960                 return;
1961
1962         pgdat = zone->zone_pgdat;
1963         if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0))
1964                 return;
1965         if (pgdat->kswapd_max_order < order)
1966                 pgdat->kswapd_max_order = order;
1967         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1968                 return;
1969         if (!waitqueue_active(&pgdat->kswapd_wait))
1970                 return;
1971         wake_up_interruptible(&pgdat->kswapd_wait);
1972 }
1973
1974 unsigned long global_lru_pages(void)
1975 {
1976         return global_page_state(NR_ACTIVE_ANON)
1977                 + global_page_state(NR_ACTIVE_FILE)
1978                 + global_page_state(NR_INACTIVE_ANON)
1979                 + global_page_state(NR_INACTIVE_FILE);
1980 }
1981
1982 #ifdef CONFIG_PM
1983 /*
1984  * Helper function for shrink_all_memory().  Tries to reclaim 'nr_pages' pages
1985  * from LRU lists system-wide, for given pass and priority, and returns the
1986  * number of reclaimed pages
1987  *
1988  * For pass > 3 we also try to shrink the LRU lists that contain a few pages
1989  */
1990 static unsigned long shrink_all_zones(unsigned long nr_pages, int prio,
1991                                       int pass, struct scan_control *sc)
1992 {
1993         struct zone *zone;
1994         unsigned long nr_to_scan, ret = 0;
1995         enum lru_list l;
1996
1997         for_each_zone(zone) {
1998
1999                 if (!populated_zone(zone))
2000                         continue;
2001
2002                 if (zone_is_all_unreclaimable(zone) && prio != DEF_PRIORITY)
2003                         continue;
2004
2005                 for_each_evictable_lru(l) {
2006                         /* For pass = 0, we don't shrink the active list */
2007                         if (pass == 0 &&
2008                                 (l == LRU_ACTIVE || l == LRU_ACTIVE_FILE))
2009                                 continue;
2010
2011                         zone->lru[l].nr_scan +=
2012                                 (zone_page_state(zone, NR_LRU_BASE + l)
2013                                                                 >> prio) + 1;
2014                         if (zone->lru[l].nr_scan >= nr_pages || pass > 3) {
2015                                 zone->lru[l].nr_scan = 0;
2016                                 nr_to_scan = min(nr_pages,
2017                                         zone_page_state(zone,
2018                                                         NR_LRU_BASE + l));
2019                                 ret += shrink_list(l, nr_to_scan, zone,
2020                                                                 sc, prio);
2021                                 if (ret >= nr_pages)
2022                                         return ret;
2023                         }
2024                 }
2025         }
2026
2027         return ret;
2028 }
2029
2030 /*
2031  * Try to free `nr_pages' of memory, system-wide, and return the number of
2032  * freed pages.
2033  *
2034  * Rather than trying to age LRUs the aim is to preserve the overall
2035  * LRU order by reclaiming preferentially
2036  * inactive > active > active referenced > active mapped
2037  */
2038 unsigned long shrink_all_memory(unsigned long nr_pages)
2039 {
2040         unsigned long lru_pages, nr_slab;
2041         unsigned long ret = 0;
2042         int pass;
2043         struct reclaim_state reclaim_state;
2044         struct scan_control sc = {
2045                 .gfp_mask = GFP_KERNEL,
2046                 .may_swap = 0,
2047                 .swap_cluster_max = nr_pages,
2048                 .may_writepage = 1,
2049                 .swappiness = vm_swappiness,
2050                 .isolate_pages = isolate_pages_global,
2051         };
2052
2053         current->reclaim_state = &reclaim_state;
2054
2055         lru_pages = global_lru_pages();
2056         nr_slab = global_page_state(NR_SLAB_RECLAIMABLE);
2057         /* If slab caches are huge, it's better to hit them first */
2058         while (nr_slab >= lru_pages) {
2059                 reclaim_state.reclaimed_slab = 0;
2060                 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
2061                 if (!reclaim_state.reclaimed_slab)
2062                         break;
2063
2064                 ret += reclaim_state.reclaimed_slab;
2065                 if (ret >= nr_pages)
2066                         goto out;
2067
2068                 nr_slab -= reclaim_state.reclaimed_slab;
2069         }
2070
2071         /*
2072          * We try to shrink LRUs in 5 passes:
2073          * 0 = Reclaim from inactive_list only
2074          * 1 = Reclaim from active list but don't reclaim mapped
2075          * 2 = 2nd pass of type 1
2076          * 3 = Reclaim mapped (normal reclaim)
2077          * 4 = 2nd pass of type 3
2078          */
2079         for (pass = 0; pass < 5; pass++) {
2080                 int prio;
2081
2082                 /* Force reclaiming mapped pages in the passes #3 and #4 */
2083                 if (pass > 2) {
2084                         sc.may_swap = 1;
2085                         sc.swappiness = 100;
2086                 }
2087
2088                 for (prio = DEF_PRIORITY; prio >= 0; prio--) {
2089                         unsigned long nr_to_scan = nr_pages - ret;
2090
2091                         sc.nr_scanned = 0;
2092                         ret += shrink_all_zones(nr_to_scan, prio, pass, &sc);
2093                         if (ret >= nr_pages)
2094                                 goto out;
2095
2096                         reclaim_state.reclaimed_slab = 0;
2097                         shrink_slab(sc.nr_scanned, sc.gfp_mask,
2098                                         global_lru_pages());
2099                         ret += reclaim_state.reclaimed_slab;
2100                         if (ret >= nr_pages)
2101                                 goto out;
2102
2103                         if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
2104                                 congestion_wait(WRITE, HZ / 10);
2105                 }
2106         }
2107
2108         /*
2109          * If ret = 0, we could not shrink LRUs, but there may be something
2110          * in slab caches
2111          */
2112         if (!ret) {
2113                 do {
2114                         reclaim_state.reclaimed_slab = 0;
2115                         shrink_slab(nr_pages, sc.gfp_mask, global_lru_pages());
2116                         ret += reclaim_state.reclaimed_slab;
2117                 } while (ret < nr_pages && reclaim_state.reclaimed_slab > 0);
2118         }
2119
2120 out:
2121         current->reclaim_state = NULL;
2122
2123         return ret;
2124 }
2125 #endif
2126
2127 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2128    not required for correctness.  So if the last cpu in a node goes
2129    away, we get changed to run anywhere: as the first one comes back,
2130    restore their cpu bindings. */
2131 static int __devinit cpu_callback(struct notifier_block *nfb,
2132                                   unsigned long action, void *hcpu)
2133 {
2134         int nid;
2135
2136         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2137                 for_each_node_state(nid, N_HIGH_MEMORY) {
2138                         pg_data_t *pgdat = NODE_DATA(nid);
2139                         node_to_cpumask_ptr(mask, pgdat->node_id);
2140
2141                         if (any_online_cpu(*mask) < nr_cpu_ids)
2142                                 /* One of our CPUs online: restore mask */
2143                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2144                 }
2145         }
2146         return NOTIFY_OK;
2147 }
2148
2149 /*
2150  * This kswapd start function will be called by init and node-hot-add.
2151  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2152  */
2153 int kswapd_run(int nid)
2154 {
2155         pg_data_t *pgdat = NODE_DATA(nid);
2156         int ret = 0;
2157
2158         if (pgdat->kswapd)
2159                 return 0;
2160
2161         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2162         if (IS_ERR(pgdat->kswapd)) {
2163                 /* failure at boot is fatal */
2164                 BUG_ON(system_state == SYSTEM_BOOTING);
2165                 printk("Failed to start kswapd on node %d\n",nid);
2166                 ret = -1;
2167         }
2168         return ret;
2169 }
2170
2171 static int __init kswapd_init(void)
2172 {
2173         int nid;
2174
2175         swap_setup();
2176         for_each_node_state(nid, N_HIGH_MEMORY)
2177                 kswapd_run(nid);
2178         hotcpu_notifier(cpu_callback, 0);
2179         return 0;
2180 }
2181
2182 module_init(kswapd_init)
2183
2184 #ifdef CONFIG_NUMA
2185 /*
2186  * Zone reclaim mode
2187  *
2188  * If non-zero call zone_reclaim when the number of free pages falls below
2189  * the watermarks.
2190  */
2191 int zone_reclaim_mode __read_mostly;
2192
2193 #define RECLAIM_OFF 0
2194 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
2195 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
2196 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
2197
2198 /*
2199  * Priority for ZONE_RECLAIM. This determines the fraction of pages
2200  * of a node considered for each zone_reclaim. 4 scans 1/16th of
2201  * a zone.
2202  */
2203 #define ZONE_RECLAIM_PRIORITY 4
2204
2205 /*
2206  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2207  * occur.
2208  */
2209 int sysctl_min_unmapped_ratio = 1;
2210
2211 /*
2212  * If the number of slab pages in a zone grows beyond this percentage then
2213  * slab reclaim needs to occur.
2214  */
2215 int sysctl_min_slab_ratio = 5;
2216
2217 /*
2218  * Try to free up some pages from this zone through reclaim.
2219  */
2220 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2221 {
2222         /* Minimum pages needed in order to stay on node */
2223         const unsigned long nr_pages = 1 << order;
2224         struct task_struct *p = current;
2225         struct reclaim_state reclaim_state;
2226         int priority;
2227         unsigned long nr_reclaimed = 0;
2228         struct scan_control sc = {
2229                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2230                 .may_swap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2231                 .swap_cluster_max = max_t(unsigned long, nr_pages,
2232                                         SWAP_CLUSTER_MAX),
2233                 .gfp_mask = gfp_mask,
2234                 .swappiness = vm_swappiness,
2235                 .isolate_pages = isolate_pages_global,
2236         };
2237         unsigned long slab_reclaimable;
2238
2239         disable_swap_token();
2240         cond_resched();
2241         /*
2242          * We need to be able to allocate from the reserves for RECLAIM_SWAP
2243          * and we also need to be able to write out pages for RECLAIM_WRITE
2244          * and RECLAIM_SWAP.
2245          */
2246         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2247         reclaim_state.reclaimed_slab = 0;
2248         p->reclaim_state = &reclaim_state;
2249
2250         if (zone_page_state(zone, NR_FILE_PAGES) -
2251                 zone_page_state(zone, NR_FILE_MAPPED) >
2252                 zone->min_unmapped_pages) {
2253                 /*
2254                  * Free memory by calling shrink zone with increasing
2255                  * priorities until we have enough memory freed.
2256                  */
2257                 priority = ZONE_RECLAIM_PRIORITY;
2258                 do {
2259                         note_zone_scanning_priority(zone, priority);
2260                         nr_reclaimed += shrink_zone(priority, zone, &sc);
2261                         priority--;
2262                 } while (priority >= 0 && nr_reclaimed < nr_pages);
2263         }
2264
2265         slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2266         if (slab_reclaimable > zone->min_slab_pages) {
2267                 /*
2268                  * shrink_slab() does not currently allow us to determine how
2269                  * many pages were freed in this zone. So we take the current
2270                  * number of slab pages and shake the slab until it is reduced
2271                  * by the same nr_pages that we used for reclaiming unmapped
2272                  * pages.
2273                  *
2274                  * Note that shrink_slab will free memory on all zones and may
2275                  * take a long time.
2276                  */
2277                 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
2278                         zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
2279                                 slab_reclaimable - nr_pages)
2280                         ;
2281
2282                 /*
2283                  * Update nr_reclaimed by the number of slab pages we
2284                  * reclaimed from this zone.
2285                  */
2286                 nr_reclaimed += slab_reclaimable -
2287                         zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2288         }
2289
2290         p->reclaim_state = NULL;
2291         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
2292         return nr_reclaimed >= nr_pages;
2293 }
2294
2295 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2296 {
2297         int node_id;
2298         int ret;
2299
2300         /*
2301          * Zone reclaim reclaims unmapped file backed pages and
2302          * slab pages if we are over the defined limits.
2303          *
2304          * A small portion of unmapped file backed pages is needed for
2305          * file I/O otherwise pages read by file I/O will be immediately
2306          * thrown out if the zone is overallocated. So we do not reclaim
2307          * if less than a specified percentage of the zone is used by
2308          * unmapped file backed pages.
2309          */
2310         if (zone_page_state(zone, NR_FILE_PAGES) -
2311             zone_page_state(zone, NR_FILE_MAPPED) <= zone->min_unmapped_pages
2312             && zone_page_state(zone, NR_SLAB_RECLAIMABLE)
2313                         <= zone->min_slab_pages)
2314                 return 0;
2315
2316         if (zone_is_all_unreclaimable(zone))
2317                 return 0;
2318
2319         /*
2320          * Do not scan if the allocation should not be delayed.
2321          */
2322         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
2323                         return 0;
2324
2325         /*
2326          * Only run zone reclaim on the local zone or on zones that do not
2327          * have associated processors. This will favor the local processor
2328          * over remote processors and spread off node memory allocations
2329          * as wide as possible.
2330          */
2331         node_id = zone_to_nid(zone);
2332         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
2333                 return 0;
2334
2335         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2336                 return 0;
2337         ret = __zone_reclaim(zone, gfp_mask, order);
2338         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2339
2340         return ret;
2341 }
2342 #endif
2343
2344 #ifdef CONFIG_UNEVICTABLE_LRU
2345 /*
2346  * page_evictable - test whether a page is evictable
2347  * @page: the page to test
2348  * @vma: the VMA in which the page is or will be mapped, may be NULL
2349  *
2350  * Test whether page is evictable--i.e., should be placed on active/inactive
2351  * lists vs unevictable list.  The vma argument is !NULL when called from the
2352  * fault path to determine how to instantate a new page.
2353  *
2354  * Reasons page might not be evictable:
2355  * (1) page's mapping marked unevictable
2356  * (2) page is part of an mlocked VMA
2357  *
2358  */
2359 int page_evictable(struct page *page, struct vm_area_struct *vma)
2360 {
2361
2362         if (mapping_unevictable(page_mapping(page)))
2363                 return 0;
2364
2365         if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2366                 return 0;
2367
2368         return 1;
2369 }
2370
2371 static void show_page_path(struct page *page)
2372 {
2373         char buf[256];
2374         if (page_is_file_cache(page)) {
2375                 struct address_space *mapping = page->mapping;
2376                 struct dentry *dentry;
2377                 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
2378
2379                 spin_lock(&mapping->i_mmap_lock);
2380                 dentry = d_find_alias(mapping->host);
2381                 printk(KERN_INFO "rescued: %s %lu\n",
2382                        dentry_path(dentry, buf, 256), pgoff);
2383                 spin_unlock(&mapping->i_mmap_lock);
2384         } else {
2385 #if defined(CONFIG_MM_OWNER) && defined(CONFIG_MMU)
2386                 struct anon_vma *anon_vma;
2387                 struct vm_area_struct *vma;
2388
2389                 anon_vma = page_lock_anon_vma(page);
2390                 if (!anon_vma)
2391                         return;
2392
2393                 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
2394                         printk(KERN_INFO "rescued: anon %s\n",
2395                                vma->vm_mm->owner->comm);
2396                         break;
2397                 }
2398                 page_unlock_anon_vma(anon_vma);
2399 #endif
2400         }
2401 }
2402
2403
2404 /**
2405  * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2406  * @page: page to check evictability and move to appropriate lru list
2407  * @zone: zone page is in
2408  *
2409  * Checks a page for evictability and moves the page to the appropriate
2410  * zone lru list.
2411  *
2412  * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2413  * have PageUnevictable set.
2414  */
2415 static void check_move_unevictable_page(struct page *page, struct zone *zone)
2416 {
2417         VM_BUG_ON(PageActive(page));
2418
2419 retry:
2420         ClearPageUnevictable(page);
2421         if (page_evictable(page, NULL)) {
2422                 enum lru_list l = LRU_INACTIVE_ANON + page_is_file_cache(page);
2423
2424                 show_page_path(page);
2425
2426                 __dec_zone_state(zone, NR_UNEVICTABLE);
2427                 list_move(&page->lru, &zone->lru[l].list);
2428                 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
2429                 __count_vm_event(UNEVICTABLE_PGRESCUED);
2430         } else {
2431                 /*
2432                  * rotate unevictable list
2433                  */
2434                 SetPageUnevictable(page);
2435                 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
2436                 if (page_evictable(page, NULL))
2437                         goto retry;
2438         }
2439 }
2440
2441 /**
2442  * scan_mapping_unevictable_pages - scan an address space for evictable pages
2443  * @mapping: struct address_space to scan for evictable pages
2444  *
2445  * Scan all pages in mapping.  Check unevictable pages for
2446  * evictability and move them to the appropriate zone lru list.
2447  */
2448 void scan_mapping_unevictable_pages(struct address_space *mapping)
2449 {
2450         pgoff_t next = 0;
2451         pgoff_t end   = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
2452                          PAGE_CACHE_SHIFT;
2453         struct zone *zone;
2454         struct pagevec pvec;
2455
2456         if (mapping->nrpages == 0)
2457                 return;
2458
2459         pagevec_init(&pvec, 0);
2460         while (next < end &&
2461                 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
2462                 int i;
2463                 int pg_scanned = 0;
2464
2465                 zone = NULL;
2466
2467                 for (i = 0; i < pagevec_count(&pvec); i++) {
2468                         struct page *page = pvec.pages[i];
2469                         pgoff_t page_index = page->index;
2470                         struct zone *pagezone = page_zone(page);
2471
2472                         pg_scanned++;
2473                         if (page_index > next)
2474                                 next = page_index;
2475                         next++;
2476
2477                         if (pagezone != zone) {
2478                                 if (zone)
2479                                         spin_unlock_irq(&zone->lru_lock);
2480                                 zone = pagezone;
2481                                 spin_lock_irq(&zone->lru_lock);
2482                         }
2483
2484                         if (PageLRU(page) && PageUnevictable(page))
2485                                 check_move_unevictable_page(page, zone);
2486                 }
2487                 if (zone)
2488                         spin_unlock_irq(&zone->lru_lock);
2489                 pagevec_release(&pvec);
2490
2491                 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
2492         }
2493
2494 }
2495
2496 /**
2497  * scan_zone_unevictable_pages - check unevictable list for evictable pages
2498  * @zone - zone of which to scan the unevictable list
2499  *
2500  * Scan @zone's unevictable LRU lists to check for pages that have become
2501  * evictable.  Move those that have to @zone's inactive list where they
2502  * become candidates for reclaim, unless shrink_inactive_zone() decides
2503  * to reactivate them.  Pages that are still unevictable are rotated
2504  * back onto @zone's unevictable list.
2505  */
2506 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
2507 void scan_zone_unevictable_pages(struct zone *zone)
2508 {
2509         struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
2510         unsigned long scan;
2511         unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
2512
2513         while (nr_to_scan > 0) {
2514                 unsigned long batch_size = min(nr_to_scan,
2515                                                 SCAN_UNEVICTABLE_BATCH_SIZE);
2516
2517                 spin_lock_irq(&zone->lru_lock);
2518                 for (scan = 0;  scan < batch_size; scan++) {
2519                         struct page *page = lru_to_page(l_unevictable);
2520
2521                         if (!trylock_page(page))
2522                                 continue;
2523
2524                         prefetchw_prev_lru_page(page, l_unevictable, flags);
2525
2526                         if (likely(PageLRU(page) && PageUnevictable(page)))
2527                                 check_move_unevictable_page(page, zone);
2528
2529                         unlock_page(page);
2530                 }
2531                 spin_unlock_irq(&zone->lru_lock);
2532
2533                 nr_to_scan -= batch_size;
2534         }
2535 }
2536
2537
2538 /**
2539  * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
2540  *
2541  * A really big hammer:  scan all zones' unevictable LRU lists to check for
2542  * pages that have become evictable.  Move those back to the zones'
2543  * inactive list where they become candidates for reclaim.
2544  * This occurs when, e.g., we have unswappable pages on the unevictable lists,
2545  * and we add swap to the system.  As such, it runs in the context of a task
2546  * that has possibly/probably made some previously unevictable pages
2547  * evictable.
2548  */
2549 void scan_all_zones_unevictable_pages(void)
2550 {
2551         struct zone *zone;
2552
2553         for_each_zone(zone) {
2554                 scan_zone_unevictable_pages(zone);
2555         }
2556 }
2557
2558 /*
2559  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
2560  * all nodes' unevictable lists for evictable pages
2561  */
2562 unsigned long scan_unevictable_pages;
2563
2564 int scan_unevictable_handler(struct ctl_table *table, int write,
2565                            struct file *file, void __user *buffer,
2566                            size_t *length, loff_t *ppos)
2567 {
2568         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
2569
2570         if (write && *(unsigned long *)table->data)
2571                 scan_all_zones_unevictable_pages();
2572
2573         scan_unevictable_pages = 0;
2574         return 0;
2575 }
2576
2577 /*
2578  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
2579  * a specified node's per zone unevictable lists for evictable pages.
2580  */
2581
2582 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
2583                                           struct sysdev_attribute *attr,
2584                                           char *buf)
2585 {
2586         return sprintf(buf, "0\n");     /* always zero; should fit... */
2587 }
2588
2589 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
2590                                            struct sysdev_attribute *attr,
2591                                         const char *buf, size_t count)
2592 {
2593         struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
2594         struct zone *zone;
2595         unsigned long res;
2596         unsigned long req = strict_strtoul(buf, 10, &res);
2597
2598         if (!req)
2599                 return 1;       /* zero is no-op */
2600
2601         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2602                 if (!populated_zone(zone))
2603                         continue;
2604                 scan_zone_unevictable_pages(zone);
2605         }
2606         return 1;
2607 }
2608
2609
2610 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
2611                         read_scan_unevictable_node,
2612                         write_scan_unevictable_node);
2613
2614 int scan_unevictable_register_node(struct node *node)
2615 {
2616         return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
2617 }
2618
2619 void scan_unevictable_unregister_node(struct node *node)
2620 {
2621         sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
2622 }
2623
2624 #endif