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