412d7872fc75af2e33444470951029db21b95b6d
[linux-2.6.git] / mm / vmscan.c
1 /*
2  *  linux/mm/vmscan.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  *
6  *  Swap reorganised 29.12.95, Stephen Tweedie.
7  *  kswapd added: 7.1.96  sct
8  *  Removed kswapd_ctl limits, and swap out as many pages as needed
9  *  to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10  *  Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11  *  Multiqueue VM started 5.8.00, Rik van Riel.
12  */
13
14 #include <linux/mm.h>
15 #include <linux/module.h>
16 #include <linux/slab.h>
17 #include <linux/kernel_stat.h>
18 #include <linux/swap.h>
19 #include <linux/pagemap.h>
20 #include <linux/init.h>
21 #include <linux/highmem.h>
22 #include <linux/vmstat.h>
23 #include <linux/file.h>
24 #include <linux/writeback.h>
25 #include <linux/blkdev.h>
26 #include <linux/buffer_head.h>  /* for try_to_release_page(),
27                                         buffer_heads_over_limit */
28 #include <linux/mm_inline.h>
29 #include <linux/pagevec.h>
30 #include <linux/backing-dev.h>
31 #include <linux/rmap.h>
32 #include <linux/topology.h>
33 #include <linux/cpu.h>
34 #include <linux/cpuset.h>
35 #include <linux/notifier.h>
36 #include <linux/rwsem.h>
37 #include <linux/delay.h>
38 #include <linux/kthread.h>
39 #include <linux/freezer.h>
40 #include <linux/memcontrol.h>
41 #include <linux/delayacct.h>
42 #include <linux/sysctl.h>
43
44 #include <asm/tlbflush.h>
45 #include <asm/div64.h>
46
47 #include <linux/swapops.h>
48
49 #include "internal.h"
50
51 struct scan_control {
52         /* Incremented by the number of inactive pages that were scanned */
53         unsigned long nr_scanned;
54
55         /* This context's GFP mask */
56         gfp_t gfp_mask;
57
58         int may_writepage;
59
60         /* Can pages be swapped as part of reclaim? */
61         int may_swap;
62
63         /* This context's SWAP_CLUSTER_MAX. If freeing memory for
64          * suspend, we effectively ignore SWAP_CLUSTER_MAX.
65          * In this context, it doesn't matter that we scan the
66          * whole list at once. */
67         int swap_cluster_max;
68
69         int swappiness;
70
71         int all_unreclaimable;
72
73         int order;
74
75         /* Which cgroup do we reclaim from */
76         struct mem_cgroup *mem_cgroup;
77
78         /* Pluggable isolate pages callback */
79         unsigned long (*isolate_pages)(unsigned long nr, struct list_head *dst,
80                         unsigned long *scanned, int order, int mode,
81                         struct zone *z, struct mem_cgroup *mem_cont,
82                         int active, int file);
83 };
84
85 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
86
87 #ifdef ARCH_HAS_PREFETCH
88 #define prefetch_prev_lru_page(_page, _base, _field)                    \
89         do {                                                            \
90                 if ((_page)->lru.prev != _base) {                       \
91                         struct page *prev;                              \
92                                                                         \
93                         prev = lru_to_page(&(_page->lru));              \
94                         prefetch(&prev->_field);                        \
95                 }                                                       \
96         } while (0)
97 #else
98 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
99 #endif
100
101 #ifdef ARCH_HAS_PREFETCHW
102 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
103         do {                                                            \
104                 if ((_page)->lru.prev != _base) {                       \
105                         struct page *prev;                              \
106                                                                         \
107                         prev = lru_to_page(&(_page->lru));              \
108                         prefetchw(&prev->_field);                       \
109                 }                                                       \
110         } while (0)
111 #else
112 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
113 #endif
114
115 /*
116  * From 0 .. 100.  Higher means more swappy.
117  */
118 int vm_swappiness = 60;
119 long vm_total_pages;    /* The total number of pages which the VM controls */
120
121 static LIST_HEAD(shrinker_list);
122 static DECLARE_RWSEM(shrinker_rwsem);
123
124 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
125 #define scan_global_lru(sc)     (!(sc)->mem_cgroup)
126 #else
127 #define scan_global_lru(sc)     (1)
128 #endif
129
130 /*
131  * Add a shrinker callback to be called from the vm
132  */
133 void register_shrinker(struct shrinker *shrinker)
134 {
135         shrinker->nr = 0;
136         down_write(&shrinker_rwsem);
137         list_add_tail(&shrinker->list, &shrinker_list);
138         up_write(&shrinker_rwsem);
139 }
140 EXPORT_SYMBOL(register_shrinker);
141
142 /*
143  * Remove one
144  */
145 void unregister_shrinker(struct shrinker *shrinker)
146 {
147         down_write(&shrinker_rwsem);
148         list_del(&shrinker->list);
149         up_write(&shrinker_rwsem);
150 }
151 EXPORT_SYMBOL(unregister_shrinker);
152
153 #define SHRINK_BATCH 128
154 /*
155  * Call the shrink functions to age shrinkable caches
156  *
157  * Here we assume it costs one seek to replace a lru page and that it also
158  * takes a seek to recreate a cache object.  With this in mind we age equal
159  * percentages of the lru and ageable caches.  This should balance the seeks
160  * generated by these structures.
161  *
162  * If the vm encountered mapped pages on the LRU it increase the pressure on
163  * slab to avoid swapping.
164  *
165  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
166  *
167  * `lru_pages' represents the number of on-LRU pages in all the zones which
168  * are eligible for the caller's allocation attempt.  It is used for balancing
169  * slab reclaim versus page reclaim.
170  *
171  * Returns the number of slab objects which we shrunk.
172  */
173 unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
174                         unsigned long lru_pages)
175 {
176         struct shrinker *shrinker;
177         unsigned long ret = 0;
178
179         if (scanned == 0)
180                 scanned = SWAP_CLUSTER_MAX;
181
182         if (!down_read_trylock(&shrinker_rwsem))
183                 return 1;       /* Assume we'll be able to shrink next time */
184
185         list_for_each_entry(shrinker, &shrinker_list, list) {
186                 unsigned long long delta;
187                 unsigned long total_scan;
188                 unsigned long max_pass = (*shrinker->shrink)(0, gfp_mask);
189
190                 delta = (4 * scanned) / shrinker->seeks;
191                 delta *= max_pass;
192                 do_div(delta, lru_pages + 1);
193                 shrinker->nr += delta;
194                 if (shrinker->nr < 0) {
195                         printk(KERN_ERR "%s: nr=%ld\n",
196                                         __func__, shrinker->nr);
197                         shrinker->nr = max_pass;
198                 }
199
200                 /*
201                  * Avoid risking looping forever due to too large nr value:
202                  * never try to free more than twice the estimate number of
203                  * freeable entries.
204                  */
205                 if (shrinker->nr > max_pass * 2)
206                         shrinker->nr = max_pass * 2;
207
208                 total_scan = shrinker->nr;
209                 shrinker->nr = 0;
210
211                 while (total_scan >= SHRINK_BATCH) {
212                         long this_scan = SHRINK_BATCH;
213                         int shrink_ret;
214                         int nr_before;
215
216                         nr_before = (*shrinker->shrink)(0, gfp_mask);
217                         shrink_ret = (*shrinker->shrink)(this_scan, gfp_mask);
218                         if (shrink_ret == -1)
219                                 break;
220                         if (shrink_ret < nr_before)
221                                 ret += nr_before - shrink_ret;
222                         count_vm_events(SLABS_SCANNED, this_scan);
223                         total_scan -= this_scan;
224
225                         cond_resched();
226                 }
227
228                 shrinker->nr += total_scan;
229         }
230         up_read(&shrinker_rwsem);
231         return ret;
232 }
233
234 /* Called without lock on whether page is mapped, so answer is unstable */
235 static inline int page_mapping_inuse(struct page *page)
236 {
237         struct address_space *mapping;
238
239         /* Page is in somebody's page tables. */
240         if (page_mapped(page))
241                 return 1;
242
243         /* Be more reluctant to reclaim swapcache than pagecache */
244         if (PageSwapCache(page))
245                 return 1;
246
247         mapping = page_mapping(page);
248         if (!mapping)
249                 return 0;
250
251         /* File is mmap'd by somebody? */
252         return mapping_mapped(mapping);
253 }
254
255 static inline int is_page_cache_freeable(struct page *page)
256 {
257         return page_count(page) - !!PagePrivate(page) == 2;
258 }
259
260 static int may_write_to_queue(struct backing_dev_info *bdi)
261 {
262         if (current->flags & PF_SWAPWRITE)
263                 return 1;
264         if (!bdi_write_congested(bdi))
265                 return 1;
266         if (bdi == current->backing_dev_info)
267                 return 1;
268         return 0;
269 }
270
271 /*
272  * We detected a synchronous write error writing a page out.  Probably
273  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
274  * fsync(), msync() or close().
275  *
276  * The tricky part is that after writepage we cannot touch the mapping: nothing
277  * prevents it from being freed up.  But we have a ref on the page and once
278  * that page is locked, the mapping is pinned.
279  *
280  * We're allowed to run sleeping lock_page() here because we know the caller has
281  * __GFP_FS.
282  */
283 static void handle_write_error(struct address_space *mapping,
284                                 struct page *page, int error)
285 {
286         lock_page(page);
287         if (page_mapping(page) == mapping)
288                 mapping_set_error(mapping, error);
289         unlock_page(page);
290 }
291
292 /* Request for sync pageout. */
293 enum pageout_io {
294         PAGEOUT_IO_ASYNC,
295         PAGEOUT_IO_SYNC,
296 };
297
298 /* possible outcome of pageout() */
299 typedef enum {
300         /* failed to write page out, page is locked */
301         PAGE_KEEP,
302         /* move page to the active list, page is locked */
303         PAGE_ACTIVATE,
304         /* page has been sent to the disk successfully, page is unlocked */
305         PAGE_SUCCESS,
306         /* page is clean and locked */
307         PAGE_CLEAN,
308 } pageout_t;
309
310 /*
311  * pageout is called by shrink_page_list() for each dirty page.
312  * Calls ->writepage().
313  */
314 static pageout_t pageout(struct page *page, struct address_space *mapping,
315                                                 enum pageout_io sync_writeback)
316 {
317         /*
318          * If the page is dirty, only perform writeback if that write
319          * will be non-blocking.  To prevent this allocation from being
320          * stalled by pagecache activity.  But note that there may be
321          * stalls if we need to run get_block().  We could test
322          * PagePrivate for that.
323          *
324          * If this process is currently in generic_file_write() against
325          * this page's queue, we can perform writeback even if that
326          * will block.
327          *
328          * If the page is swapcache, write it back even if that would
329          * block, for some throttling. This happens by accident, because
330          * swap_backing_dev_info is bust: it doesn't reflect the
331          * congestion state of the swapdevs.  Easy to fix, if needed.
332          * See swapfile.c:page_queue_congested().
333          */
334         if (!is_page_cache_freeable(page))
335                 return PAGE_KEEP;
336         if (!mapping) {
337                 /*
338                  * Some data journaling orphaned pages can have
339                  * page->mapping == NULL while being dirty with clean buffers.
340                  */
341                 if (PagePrivate(page)) {
342                         if (try_to_free_buffers(page)) {
343                                 ClearPageDirty(page);
344                                 printk("%s: orphaned page\n", __func__);
345                                 return PAGE_CLEAN;
346                         }
347                 }
348                 return PAGE_KEEP;
349         }
350         if (mapping->a_ops->writepage == NULL)
351                 return PAGE_ACTIVATE;
352         if (!may_write_to_queue(mapping->backing_dev_info))
353                 return PAGE_KEEP;
354
355         if (clear_page_dirty_for_io(page)) {
356                 int res;
357                 struct writeback_control wbc = {
358                         .sync_mode = WB_SYNC_NONE,
359                         .nr_to_write = SWAP_CLUSTER_MAX,
360                         .range_start = 0,
361                         .range_end = LLONG_MAX,
362                         .nonblocking = 1,
363                         .for_reclaim = 1,
364                 };
365
366                 SetPageReclaim(page);
367                 res = mapping->a_ops->writepage(page, &wbc);
368                 if (res < 0)
369                         handle_write_error(mapping, page, res);
370                 if (res == AOP_WRITEPAGE_ACTIVATE) {
371                         ClearPageReclaim(page);
372                         return PAGE_ACTIVATE;
373                 }
374
375                 /*
376                  * Wait on writeback if requested to. This happens when
377                  * direct reclaiming a large contiguous area and the
378                  * first attempt to free a range of pages fails.
379                  */
380                 if (PageWriteback(page) && sync_writeback == PAGEOUT_IO_SYNC)
381                         wait_on_page_writeback(page);
382
383                 if (!PageWriteback(page)) {
384                         /* synchronous write or broken a_ops? */
385                         ClearPageReclaim(page);
386                 }
387                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
388                 return PAGE_SUCCESS;
389         }
390
391         return PAGE_CLEAN;
392 }
393
394 /*
395  * Same as remove_mapping, but if the page is removed from the mapping, it
396  * gets returned with a refcount of 0.
397  */
398 static int __remove_mapping(struct address_space *mapping, struct page *page)
399 {
400         BUG_ON(!PageLocked(page));
401         BUG_ON(mapping != page_mapping(page));
402
403         spin_lock_irq(&mapping->tree_lock);
404         /*
405          * The non racy check for a busy page.
406          *
407          * Must be careful with the order of the tests. When someone has
408          * a ref to the page, it may be possible that they dirty it then
409          * drop the reference. So if PageDirty is tested before page_count
410          * here, then the following race may occur:
411          *
412          * get_user_pages(&page);
413          * [user mapping goes away]
414          * write_to(page);
415          *                              !PageDirty(page)    [good]
416          * SetPageDirty(page);
417          * put_page(page);
418          *                              !page_count(page)   [good, discard it]
419          *
420          * [oops, our write_to data is lost]
421          *
422          * Reversing the order of the tests ensures such a situation cannot
423          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
424          * load is not satisfied before that of page->_count.
425          *
426          * Note that if SetPageDirty is always performed via set_page_dirty,
427          * and thus under tree_lock, then this ordering is not required.
428          */
429         if (!page_freeze_refs(page, 2))
430                 goto cannot_free;
431         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
432         if (unlikely(PageDirty(page))) {
433                 page_unfreeze_refs(page, 2);
434                 goto cannot_free;
435         }
436
437         if (PageSwapCache(page)) {
438                 swp_entry_t swap = { .val = page_private(page) };
439                 __delete_from_swap_cache(page);
440                 spin_unlock_irq(&mapping->tree_lock);
441                 swap_free(swap);
442         } else {
443                 __remove_from_page_cache(page);
444                 spin_unlock_irq(&mapping->tree_lock);
445         }
446
447         return 1;
448
449 cannot_free:
450         spin_unlock_irq(&mapping->tree_lock);
451         return 0;
452 }
453
454 /*
455  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
456  * someone else has a ref on the page, abort and return 0.  If it was
457  * successfully detached, return 1.  Assumes the caller has a single ref on
458  * this page.
459  */
460 int remove_mapping(struct address_space *mapping, struct page *page)
461 {
462         if (__remove_mapping(mapping, page)) {
463                 /*
464                  * Unfreezing the refcount with 1 rather than 2 effectively
465                  * drops the pagecache ref for us without requiring another
466                  * atomic operation.
467                  */
468                 page_unfreeze_refs(page, 1);
469                 return 1;
470         }
471         return 0;
472 }
473
474 /**
475  * putback_lru_page - put previously isolated page onto appropriate LRU list
476  * @page: page to be put back to appropriate lru list
477  *
478  * Add previously isolated @page to appropriate LRU list.
479  * Page may still be unevictable for other reasons.
480  *
481  * lru_lock must not be held, interrupts must be enabled.
482  */
483 #ifdef CONFIG_UNEVICTABLE_LRU
484 void putback_lru_page(struct page *page)
485 {
486         int lru;
487         int active = !!TestClearPageActive(page);
488         int was_unevictable = PageUnevictable(page);
489
490         VM_BUG_ON(PageLRU(page));
491
492 redo:
493         ClearPageUnevictable(page);
494
495         if (page_evictable(page, NULL)) {
496                 /*
497                  * For evictable pages, we can use the cache.
498                  * In event of a race, worst case is we end up with an
499                  * unevictable page on [in]active list.
500                  * We know how to handle that.
501                  */
502                 lru = active + page_is_file_cache(page);
503                 lru_cache_add_lru(page, lru);
504         } else {
505                 /*
506                  * Put unevictable pages directly on zone's unevictable
507                  * list.
508                  */
509                 lru = LRU_UNEVICTABLE;
510                 add_page_to_unevictable_list(page);
511         }
512         mem_cgroup_move_lists(page, lru);
513
514         /*
515          * page's status can change while we move it among lru. If an evictable
516          * page is on unevictable list, it never be freed. To avoid that,
517          * check after we added it to the list, again.
518          */
519         if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
520                 if (!isolate_lru_page(page)) {
521                         put_page(page);
522                         goto redo;
523                 }
524                 /* This means someone else dropped this page from LRU
525                  * So, it will be freed or putback to LRU again. There is
526                  * nothing to do here.
527                  */
528         }
529
530         if (was_unevictable && lru != LRU_UNEVICTABLE)
531                 count_vm_event(UNEVICTABLE_PGRESCUED);
532         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
533                 count_vm_event(UNEVICTABLE_PGCULLED);
534
535         put_page(page);         /* drop ref from isolate */
536 }
537
538 #else /* CONFIG_UNEVICTABLE_LRU */
539
540 void putback_lru_page(struct page *page)
541 {
542         int lru;
543         VM_BUG_ON(PageLRU(page));
544
545         lru = !!TestClearPageActive(page) + page_is_file_cache(page);
546         lru_cache_add_lru(page, lru);
547         mem_cgroup_move_lists(page, lru);
548         put_page(page);
549 }
550 #endif /* CONFIG_UNEVICTABLE_LRU */
551
552
553 /*
554  * shrink_page_list() returns the number of reclaimed pages
555  */
556 static unsigned long shrink_page_list(struct list_head *page_list,
557                                         struct scan_control *sc,
558                                         enum pageout_io sync_writeback)
559 {
560         LIST_HEAD(ret_pages);
561         struct pagevec freed_pvec;
562         int pgactivate = 0;
563         unsigned long nr_reclaimed = 0;
564
565         cond_resched();
566
567         pagevec_init(&freed_pvec, 1);
568         while (!list_empty(page_list)) {
569                 struct address_space *mapping;
570                 struct page *page;
571                 int may_enter_fs;
572                 int referenced;
573
574                 cond_resched();
575
576                 page = lru_to_page(page_list);
577                 list_del(&page->lru);
578
579                 if (!trylock_page(page))
580                         goto keep;
581
582                 VM_BUG_ON(PageActive(page));
583
584                 sc->nr_scanned++;
585
586                 if (unlikely(!page_evictable(page, NULL)))
587                         goto cull_mlocked;
588
589                 if (!sc->may_swap && page_mapped(page))
590                         goto keep_locked;
591
592                 /* Double the slab pressure for mapped and swapcache pages */
593                 if (page_mapped(page) || PageSwapCache(page))
594                         sc->nr_scanned++;
595
596                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
597                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
598
599                 if (PageWriteback(page)) {
600                         /*
601                          * Synchronous reclaim is performed in two passes,
602                          * first an asynchronous pass over the list to
603                          * start parallel writeback, and a second synchronous
604                          * pass to wait for the IO to complete.  Wait here
605                          * for any page for which writeback has already
606                          * started.
607                          */
608                         if (sync_writeback == PAGEOUT_IO_SYNC && may_enter_fs)
609                                 wait_on_page_writeback(page);
610                         else
611                                 goto keep_locked;
612                 }
613
614                 referenced = page_referenced(page, 1, sc->mem_cgroup);
615                 /* In active use or really unfreeable?  Activate it. */
616                 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER &&
617                                         referenced && page_mapping_inuse(page))
618                         goto activate_locked;
619
620 #ifdef CONFIG_SWAP
621                 /*
622                  * Anonymous process memory has backing store?
623                  * Try to allocate it some swap space here.
624                  */
625                 if (PageAnon(page) && !PageSwapCache(page)) {
626                         switch (try_to_munlock(page)) {
627                         case SWAP_FAIL:         /* shouldn't happen */
628                         case SWAP_AGAIN:
629                                 goto keep_locked;
630                         case SWAP_MLOCK:
631                                 goto cull_mlocked;
632                         case SWAP_SUCCESS:
633                                 ; /* fall thru'; add to swap cache */
634                         }
635                         if (!add_to_swap(page, GFP_ATOMIC))
636                                 goto activate_locked;
637                 }
638 #endif /* CONFIG_SWAP */
639
640                 mapping = page_mapping(page);
641
642                 /*
643                  * The page is mapped into the page tables of one or more
644                  * processes. Try to unmap it here.
645                  */
646                 if (page_mapped(page) && mapping) {
647                         switch (try_to_unmap(page, 0)) {
648                         case SWAP_FAIL:
649                                 goto activate_locked;
650                         case SWAP_AGAIN:
651                                 goto keep_locked;
652                         case SWAP_MLOCK:
653                                 goto cull_mlocked;
654                         case SWAP_SUCCESS:
655                                 ; /* try to free the page below */
656                         }
657                 }
658
659                 if (PageDirty(page)) {
660                         if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && referenced)
661                                 goto keep_locked;
662                         if (!may_enter_fs)
663                                 goto keep_locked;
664                         if (!sc->may_writepage)
665                                 goto keep_locked;
666
667                         /* Page is dirty, try to write it out here */
668                         switch (pageout(page, mapping, sync_writeback)) {
669                         case PAGE_KEEP:
670                                 goto keep_locked;
671                         case PAGE_ACTIVATE:
672                                 goto activate_locked;
673                         case PAGE_SUCCESS:
674                                 if (PageWriteback(page) || PageDirty(page))
675                                         goto keep;
676                                 /*
677                                  * A synchronous write - probably a ramdisk.  Go
678                                  * ahead and try to reclaim the page.
679                                  */
680                                 if (!trylock_page(page))
681                                         goto keep;
682                                 if (PageDirty(page) || PageWriteback(page))
683                                         goto keep_locked;
684                                 mapping = page_mapping(page);
685                         case PAGE_CLEAN:
686                                 ; /* try to free the page below */
687                         }
688                 }
689
690                 /*
691                  * If the page has buffers, try to free the buffer mappings
692                  * associated with this page. If we succeed we try to free
693                  * the page as well.
694                  *
695                  * We do this even if the page is PageDirty().
696                  * try_to_release_page() does not perform I/O, but it is
697                  * possible for a page to have PageDirty set, but it is actually
698                  * clean (all its buffers are clean).  This happens if the
699                  * buffers were written out directly, with submit_bh(). ext3
700                  * will do this, as well as the blockdev mapping.
701                  * try_to_release_page() will discover that cleanness and will
702                  * drop the buffers and mark the page clean - it can be freed.
703                  *
704                  * Rarely, pages can have buffers and no ->mapping.  These are
705                  * the pages which were not successfully invalidated in
706                  * truncate_complete_page().  We try to drop those buffers here
707                  * and if that worked, and the page is no longer mapped into
708                  * process address space (page_count == 1) it can be freed.
709                  * Otherwise, leave the page on the LRU so it is swappable.
710                  */
711                 if (PagePrivate(page)) {
712                         if (!try_to_release_page(page, sc->gfp_mask))
713                                 goto activate_locked;
714                         if (!mapping && page_count(page) == 1) {
715                                 unlock_page(page);
716                                 if (put_page_testzero(page))
717                                         goto free_it;
718                                 else {
719                                         /*
720                                          * rare race with speculative reference.
721                                          * the speculative reference will free
722                                          * this page shortly, so we may
723                                          * increment nr_reclaimed here (and
724                                          * leave it off the LRU).
725                                          */
726                                         nr_reclaimed++;
727                                         continue;
728                                 }
729                         }
730                 }
731
732                 if (!mapping || !__remove_mapping(mapping, page))
733                         goto keep_locked;
734
735                 unlock_page(page);
736 free_it:
737                 nr_reclaimed++;
738                 if (!pagevec_add(&freed_pvec, page)) {
739                         __pagevec_free(&freed_pvec);
740                         pagevec_reinit(&freed_pvec);
741                 }
742                 continue;
743
744 cull_mlocked:
745                 unlock_page(page);
746                 putback_lru_page(page);
747                 continue;
748
749 activate_locked:
750                 /* Not a candidate for swapping, so reclaim swap space. */
751                 if (PageSwapCache(page) && vm_swap_full())
752                         remove_exclusive_swap_page_ref(page);
753                 VM_BUG_ON(PageActive(page));
754                 SetPageActive(page);
755                 pgactivate++;
756 keep_locked:
757                 unlock_page(page);
758 keep:
759                 list_add(&page->lru, &ret_pages);
760                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
761         }
762         list_splice(&ret_pages, page_list);
763         if (pagevec_count(&freed_pvec))
764                 __pagevec_free(&freed_pvec);
765         count_vm_events(PGACTIVATE, pgactivate);
766         return nr_reclaimed;
767 }
768
769 /* LRU Isolation modes. */
770 #define ISOLATE_INACTIVE 0      /* Isolate inactive pages. */
771 #define ISOLATE_ACTIVE 1        /* Isolate active pages. */
772 #define ISOLATE_BOTH 2          /* Isolate both active and inactive pages. */
773
774 /*
775  * Attempt to remove the specified page from its LRU.  Only take this page
776  * if it is of the appropriate PageActive status.  Pages which are being
777  * freed elsewhere are also ignored.
778  *
779  * page:        page to consider
780  * mode:        one of the LRU isolation modes defined above
781  *
782  * returns 0 on success, -ve errno on failure.
783  */
784 int __isolate_lru_page(struct page *page, int mode, int file)
785 {
786         int ret = -EINVAL;
787
788         /* Only take pages on the LRU. */
789         if (!PageLRU(page))
790                 return ret;
791
792         /*
793          * When checking the active state, we need to be sure we are
794          * dealing with comparible boolean values.  Take the logical not
795          * of each.
796          */
797         if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
798                 return ret;
799
800         if (mode != ISOLATE_BOTH && (!page_is_file_cache(page) != !file))
801                 return ret;
802
803         /*
804          * When this function is being called for lumpy reclaim, we
805          * initially look into all LRU pages, active, inactive and
806          * unevictable; only give shrink_page_list evictable pages.
807          */
808         if (PageUnevictable(page))
809                 return ret;
810
811         ret = -EBUSY;
812         if (likely(get_page_unless_zero(page))) {
813                 /*
814                  * Be careful not to clear PageLRU until after we're
815                  * sure the page is not being freed elsewhere -- the
816                  * page release code relies on it.
817                  */
818                 ClearPageLRU(page);
819                 ret = 0;
820         }
821
822         return ret;
823 }
824
825 /*
826  * zone->lru_lock is heavily contended.  Some of the functions that
827  * shrink the lists perform better by taking out a batch of pages
828  * and working on them outside the LRU lock.
829  *
830  * For pagecache intensive workloads, this function is the hottest
831  * spot in the kernel (apart from copy_*_user functions).
832  *
833  * Appropriate locks must be held before calling this function.
834  *
835  * @nr_to_scan: The number of pages to look through on the list.
836  * @src:        The LRU list to pull pages off.
837  * @dst:        The temp list to put pages on to.
838  * @scanned:    The number of pages that were scanned.
839  * @order:      The caller's attempted allocation order
840  * @mode:       One of the LRU isolation modes
841  * @file:       True [1] if isolating file [!anon] pages
842  *
843  * returns how many pages were moved onto *@dst.
844  */
845 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
846                 struct list_head *src, struct list_head *dst,
847                 unsigned long *scanned, int order, int mode, int file)
848 {
849         unsigned long nr_taken = 0;
850         unsigned long scan;
851
852         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
853                 struct page *page;
854                 unsigned long pfn;
855                 unsigned long end_pfn;
856                 unsigned long page_pfn;
857                 int zone_id;
858
859                 page = lru_to_page(src);
860                 prefetchw_prev_lru_page(page, src, flags);
861
862                 VM_BUG_ON(!PageLRU(page));
863
864                 switch (__isolate_lru_page(page, mode, file)) {
865                 case 0:
866                         list_move(&page->lru, dst);
867                         nr_taken++;
868                         break;
869
870                 case -EBUSY:
871                         /* else it is being freed elsewhere */
872                         list_move(&page->lru, src);
873                         continue;
874
875                 default:
876                         BUG();
877                 }
878
879                 if (!order)
880                         continue;
881
882                 /*
883                  * Attempt to take all pages in the order aligned region
884                  * surrounding the tag page.  Only take those pages of
885                  * the same active state as that tag page.  We may safely
886                  * round the target page pfn down to the requested order
887                  * as the mem_map is guarenteed valid out to MAX_ORDER,
888                  * where that page is in a different zone we will detect
889                  * it from its zone id and abort this block scan.
890                  */
891                 zone_id = page_zone_id(page);
892                 page_pfn = page_to_pfn(page);
893                 pfn = page_pfn & ~((1 << order) - 1);
894                 end_pfn = pfn + (1 << order);
895                 for (; pfn < end_pfn; pfn++) {
896                         struct page *cursor_page;
897
898                         /* The target page is in the block, ignore it. */
899                         if (unlikely(pfn == page_pfn))
900                                 continue;
901
902                         /* Avoid holes within the zone. */
903                         if (unlikely(!pfn_valid_within(pfn)))
904                                 break;
905
906                         cursor_page = pfn_to_page(pfn);
907
908                         /* Check that we have not crossed a zone boundary. */
909                         if (unlikely(page_zone_id(cursor_page) != zone_id))
910                                 continue;
911                         switch (__isolate_lru_page(cursor_page, mode, file)) {
912                         case 0:
913                                 list_move(&cursor_page->lru, dst);
914                                 nr_taken++;
915                                 scan++;
916                                 break;
917
918                         case -EBUSY:
919                                 /* else it is being freed elsewhere */
920                                 list_move(&cursor_page->lru, src);
921                         default:
922                                 break;  /* ! on LRU or wrong list */
923                         }
924                 }
925         }
926
927         *scanned = scan;
928         return nr_taken;
929 }
930
931 static unsigned long isolate_pages_global(unsigned long nr,
932                                         struct list_head *dst,
933                                         unsigned long *scanned, int order,
934                                         int mode, struct zone *z,
935                                         struct mem_cgroup *mem_cont,
936                                         int active, int file)
937 {
938         int lru = LRU_BASE;
939         if (active)
940                 lru += LRU_ACTIVE;
941         if (file)
942                 lru += LRU_FILE;
943         return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
944                                                                 mode, !!file);
945 }
946
947 /*
948  * clear_active_flags() is a helper for shrink_active_list(), clearing
949  * any active bits from the pages in the list.
950  */
951 static unsigned long clear_active_flags(struct list_head *page_list,
952                                         unsigned int *count)
953 {
954         int nr_active = 0;
955         int lru;
956         struct page *page;
957
958         list_for_each_entry(page, page_list, lru) {
959                 lru = page_is_file_cache(page);
960                 if (PageActive(page)) {
961                         lru += LRU_ACTIVE;
962                         ClearPageActive(page);
963                         nr_active++;
964                 }
965                 count[lru]++;
966         }
967
968         return nr_active;
969 }
970
971 /**
972  * isolate_lru_page - tries to isolate a page from its LRU list
973  * @page: page to isolate from its LRU list
974  *
975  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
976  * vmstat statistic corresponding to whatever LRU list the page was on.
977  *
978  * Returns 0 if the page was removed from an LRU list.
979  * Returns -EBUSY if the page was not on an LRU list.
980  *
981  * The returned page will have PageLRU() cleared.  If it was found on
982  * the active list, it will have PageActive set.  If it was found on
983  * the unevictable list, it will have the PageUnevictable bit set. That flag
984  * may need to be cleared by the caller before letting the page go.
985  *
986  * The vmstat statistic corresponding to the list on which the page was
987  * found will be decremented.
988  *
989  * Restrictions:
990  * (1) Must be called with an elevated refcount on the page. This is a
991  *     fundamentnal difference from isolate_lru_pages (which is called
992  *     without a stable reference).
993  * (2) the lru_lock must not be held.
994  * (3) interrupts must be enabled.
995  */
996 int isolate_lru_page(struct page *page)
997 {
998         int ret = -EBUSY;
999
1000         if (PageLRU(page)) {
1001                 struct zone *zone = page_zone(page);
1002
1003                 spin_lock_irq(&zone->lru_lock);
1004                 if (PageLRU(page) && get_page_unless_zero(page)) {
1005                         int lru = page_lru(page);
1006                         ret = 0;
1007                         ClearPageLRU(page);
1008
1009                         del_page_from_lru_list(zone, page, lru);
1010                 }
1011                 spin_unlock_irq(&zone->lru_lock);
1012         }
1013         return ret;
1014 }
1015
1016 /*
1017  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1018  * of reclaimed pages
1019  */
1020 static unsigned long shrink_inactive_list(unsigned long max_scan,
1021                         struct zone *zone, struct scan_control *sc,
1022                         int priority, int file)
1023 {
1024         LIST_HEAD(page_list);
1025         struct pagevec pvec;
1026         unsigned long nr_scanned = 0;
1027         unsigned long nr_reclaimed = 0;
1028
1029         pagevec_init(&pvec, 1);
1030
1031         lru_add_drain();
1032         spin_lock_irq(&zone->lru_lock);
1033         do {
1034                 struct page *page;
1035                 unsigned long nr_taken;
1036                 unsigned long nr_scan;
1037                 unsigned long nr_freed;
1038                 unsigned long nr_active;
1039                 unsigned int count[NR_LRU_LISTS] = { 0, };
1040                 int mode = ISOLATE_INACTIVE;
1041
1042                 /*
1043                  * If we need a large contiguous chunk of memory, or have
1044                  * trouble getting a small set of contiguous pages, we
1045                  * will reclaim both active and inactive pages.
1046                  *
1047                  * We use the same threshold as pageout congestion_wait below.
1048                  */
1049                 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1050                         mode = ISOLATE_BOTH;
1051                 else if (sc->order && priority < DEF_PRIORITY - 2)
1052                         mode = ISOLATE_BOTH;
1053
1054                 nr_taken = sc->isolate_pages(sc->swap_cluster_max,
1055                              &page_list, &nr_scan, sc->order, mode,
1056                                 zone, sc->mem_cgroup, 0, file);
1057                 nr_active = clear_active_flags(&page_list, count);
1058                 __count_vm_events(PGDEACTIVATE, nr_active);
1059
1060                 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1061                                                 -count[LRU_ACTIVE_FILE]);
1062                 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1063                                                 -count[LRU_INACTIVE_FILE]);
1064                 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1065                                                 -count[LRU_ACTIVE_ANON]);
1066                 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1067                                                 -count[LRU_INACTIVE_ANON]);
1068
1069                 if (scan_global_lru(sc)) {
1070                         zone->pages_scanned += nr_scan;
1071                         zone->recent_scanned[0] += count[LRU_INACTIVE_ANON];
1072                         zone->recent_scanned[0] += count[LRU_ACTIVE_ANON];
1073                         zone->recent_scanned[1] += count[LRU_INACTIVE_FILE];
1074                         zone->recent_scanned[1] += count[LRU_ACTIVE_FILE];
1075                 }
1076                 spin_unlock_irq(&zone->lru_lock);
1077
1078                 nr_scanned += nr_scan;
1079                 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
1080
1081                 /*
1082                  * If we are direct reclaiming for contiguous pages and we do
1083                  * not reclaim everything in the list, try again and wait
1084                  * for IO to complete. This will stall high-order allocations
1085                  * but that should be acceptable to the caller
1086                  */
1087                 if (nr_freed < nr_taken && !current_is_kswapd() &&
1088                                         sc->order > PAGE_ALLOC_COSTLY_ORDER) {
1089                         congestion_wait(WRITE, HZ/10);
1090
1091                         /*
1092                          * The attempt at page out may have made some
1093                          * of the pages active, mark them inactive again.
1094                          */
1095                         nr_active = clear_active_flags(&page_list, count);
1096                         count_vm_events(PGDEACTIVATE, nr_active);
1097
1098                         nr_freed += shrink_page_list(&page_list, sc,
1099                                                         PAGEOUT_IO_SYNC);
1100                 }
1101
1102                 nr_reclaimed += nr_freed;
1103                 local_irq_disable();
1104                 if (current_is_kswapd()) {
1105                         __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scan);
1106                         __count_vm_events(KSWAPD_STEAL, nr_freed);
1107                 } else if (scan_global_lru(sc))
1108                         __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scan);
1109
1110                 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
1111
1112                 if (nr_taken == 0)
1113                         goto done;
1114
1115                 spin_lock(&zone->lru_lock);
1116                 /*
1117                  * Put back any unfreeable pages.
1118                  */
1119                 while (!list_empty(&page_list)) {
1120                         int lru;
1121                         page = lru_to_page(&page_list);
1122                         VM_BUG_ON(PageLRU(page));
1123                         list_del(&page->lru);
1124                         if (unlikely(!page_evictable(page, NULL))) {
1125                                 spin_unlock_irq(&zone->lru_lock);
1126                                 putback_lru_page(page);
1127                                 spin_lock_irq(&zone->lru_lock);
1128                                 continue;
1129                         }
1130                         SetPageLRU(page);
1131                         lru = page_lru(page);
1132                         add_page_to_lru_list(zone, page, lru);
1133                         mem_cgroup_move_lists(page, lru);
1134                         if (PageActive(page) && scan_global_lru(sc)) {
1135                                 int file = !!page_is_file_cache(page);
1136                                 zone->recent_rotated[file]++;
1137                         }
1138                         if (!pagevec_add(&pvec, page)) {
1139                                 spin_unlock_irq(&zone->lru_lock);
1140                                 __pagevec_release(&pvec);
1141                                 spin_lock_irq(&zone->lru_lock);
1142                         }
1143                 }
1144         } while (nr_scanned < max_scan);
1145         spin_unlock(&zone->lru_lock);
1146 done:
1147         local_irq_enable();
1148         pagevec_release(&pvec);
1149         return nr_reclaimed;
1150 }
1151
1152 /*
1153  * We are about to scan this zone at a certain priority level.  If that priority
1154  * level is smaller (ie: more urgent) than the previous priority, then note
1155  * that priority level within the zone.  This is done so that when the next
1156  * process comes in to scan this zone, it will immediately start out at this
1157  * priority level rather than having to build up its own scanning priority.
1158  * Here, this priority affects only the reclaim-mapped threshold.
1159  */
1160 static inline void note_zone_scanning_priority(struct zone *zone, int priority)
1161 {
1162         if (priority < zone->prev_priority)
1163                 zone->prev_priority = priority;
1164 }
1165
1166 static inline int zone_is_near_oom(struct zone *zone)
1167 {
1168         return zone->pages_scanned >= (zone_lru_pages(zone) * 3);
1169 }
1170
1171 /*
1172  * This moves pages from the active list to the inactive list.
1173  *
1174  * We move them the other way if the page is referenced by one or more
1175  * processes, from rmap.
1176  *
1177  * If the pages are mostly unmapped, the processing is fast and it is
1178  * appropriate to hold zone->lru_lock across the whole operation.  But if
1179  * the pages are mapped, the processing is slow (page_referenced()) so we
1180  * should drop zone->lru_lock around each page.  It's impossible to balance
1181  * this, so instead we remove the pages from the LRU while processing them.
1182  * It is safe to rely on PG_active against the non-LRU pages in here because
1183  * nobody will play with that bit on a non-LRU page.
1184  *
1185  * The downside is that we have to touch page->_count against each page.
1186  * But we had to alter page->flags anyway.
1187  */
1188
1189
1190 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1191                         struct scan_control *sc, int priority, int file)
1192 {
1193         unsigned long pgmoved;
1194         int pgdeactivate = 0;
1195         unsigned long pgscanned;
1196         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1197         LIST_HEAD(l_inactive);
1198         struct page *page;
1199         struct pagevec pvec;
1200         enum lru_list lru;
1201
1202         lru_add_drain();
1203         spin_lock_irq(&zone->lru_lock);
1204         pgmoved = sc->isolate_pages(nr_pages, &l_hold, &pgscanned, sc->order,
1205                                         ISOLATE_ACTIVE, zone,
1206                                         sc->mem_cgroup, 1, file);
1207         /*
1208          * zone->pages_scanned is used for detect zone's oom
1209          * mem_cgroup remembers nr_scan by itself.
1210          */
1211         if (scan_global_lru(sc)) {
1212                 zone->pages_scanned += pgscanned;
1213                 zone->recent_scanned[!!file] += pgmoved;
1214         }
1215
1216         if (file)
1217                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -pgmoved);
1218         else
1219                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -pgmoved);
1220         spin_unlock_irq(&zone->lru_lock);
1221
1222         pgmoved = 0;
1223         while (!list_empty(&l_hold)) {
1224                 cond_resched();
1225                 page = lru_to_page(&l_hold);
1226                 list_del(&page->lru);
1227
1228                 if (unlikely(!page_evictable(page, NULL))) {
1229                         putback_lru_page(page);
1230                         continue;
1231                 }
1232
1233                 /* page_referenced clears PageReferenced */
1234                 if (page_mapping_inuse(page) &&
1235                     page_referenced(page, 0, sc->mem_cgroup))
1236                         pgmoved++;
1237
1238                 list_add(&page->lru, &l_inactive);
1239         }
1240
1241         /*
1242          * Count referenced pages from currently used mappings as
1243          * rotated, even though they are moved to the inactive list.
1244          * This helps balance scan pressure between file and anonymous
1245          * pages in get_scan_ratio.
1246          */
1247         zone->recent_rotated[!!file] += pgmoved;
1248
1249         /*
1250          * Move the pages to the [file or anon] inactive list.
1251          */
1252         pagevec_init(&pvec, 1);
1253
1254         pgmoved = 0;
1255         lru = LRU_BASE + file * LRU_FILE;
1256         spin_lock_irq(&zone->lru_lock);
1257         while (!list_empty(&l_inactive)) {
1258                 page = lru_to_page(&l_inactive);
1259                 prefetchw_prev_lru_page(page, &l_inactive, flags);
1260                 VM_BUG_ON(PageLRU(page));
1261                 SetPageLRU(page);
1262                 VM_BUG_ON(!PageActive(page));
1263                 ClearPageActive(page);
1264
1265                 list_move(&page->lru, &zone->lru[lru].list);
1266                 mem_cgroup_move_lists(page, lru);
1267                 pgmoved++;
1268                 if (!pagevec_add(&pvec, page)) {
1269                         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1270                         spin_unlock_irq(&zone->lru_lock);
1271                         pgdeactivate += pgmoved;
1272                         pgmoved = 0;
1273                         if (buffer_heads_over_limit)
1274                                 pagevec_strip(&pvec);
1275                         __pagevec_release(&pvec);
1276                         spin_lock_irq(&zone->lru_lock);
1277                 }
1278         }
1279         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1280         pgdeactivate += pgmoved;
1281         if (buffer_heads_over_limit) {
1282                 spin_unlock_irq(&zone->lru_lock);
1283                 pagevec_strip(&pvec);
1284                 spin_lock_irq(&zone->lru_lock);
1285         }
1286         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1287         __count_vm_events(PGDEACTIVATE, pgdeactivate);
1288         spin_unlock_irq(&zone->lru_lock);
1289         if (vm_swap_full())
1290                 pagevec_swap_free(&pvec);
1291
1292         pagevec_release(&pvec);
1293 }
1294
1295 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1296         struct zone *zone, struct scan_control *sc, int priority)
1297 {
1298         int file = is_file_lru(lru);
1299
1300         if (lru == LRU_ACTIVE_FILE) {
1301                 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1302                 return 0;
1303         }
1304
1305         if (lru == LRU_ACTIVE_ANON &&
1306             (!scan_global_lru(sc) || inactive_anon_is_low(zone))) {
1307                 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1308                 return 0;
1309         }
1310         return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1311 }
1312
1313 /*
1314  * Determine how aggressively the anon and file LRU lists should be
1315  * scanned.  The relative value of each set of LRU lists is determined
1316  * by looking at the fraction of the pages scanned we did rotate back
1317  * onto the active list instead of evict.
1318  *
1319  * percent[0] specifies how much pressure to put on ram/swap backed
1320  * memory, while percent[1] determines pressure on the file LRUs.
1321  */
1322 static void get_scan_ratio(struct zone *zone, struct scan_control *sc,
1323                                         unsigned long *percent)
1324 {
1325         unsigned long anon, file, free;
1326         unsigned long anon_prio, file_prio;
1327         unsigned long ap, fp;
1328
1329         anon  = zone_page_state(zone, NR_ACTIVE_ANON) +
1330                 zone_page_state(zone, NR_INACTIVE_ANON);
1331         file  = zone_page_state(zone, NR_ACTIVE_FILE) +
1332                 zone_page_state(zone, NR_INACTIVE_FILE);
1333         free  = zone_page_state(zone, NR_FREE_PAGES);
1334
1335         /* If we have no swap space, do not bother scanning anon pages. */
1336         if (nr_swap_pages <= 0) {
1337                 percent[0] = 0;
1338                 percent[1] = 100;
1339                 return;
1340         }
1341
1342         /* If we have very few page cache pages, force-scan anon pages. */
1343         if (unlikely(file + free <= zone->pages_high)) {
1344                 percent[0] = 100;
1345                 percent[1] = 0;
1346                 return;
1347         }
1348
1349         /*
1350          * OK, so we have swap space and a fair amount of page cache
1351          * pages.  We use the recently rotated / recently scanned
1352          * ratios to determine how valuable each cache is.
1353          *
1354          * Because workloads change over time (and to avoid overflow)
1355          * we keep these statistics as a floating average, which ends
1356          * up weighing recent references more than old ones.
1357          *
1358          * anon in [0], file in [1]
1359          */
1360         if (unlikely(zone->recent_scanned[0] > anon / 4)) {
1361                 spin_lock_irq(&zone->lru_lock);
1362                 zone->recent_scanned[0] /= 2;
1363                 zone->recent_rotated[0] /= 2;
1364                 spin_unlock_irq(&zone->lru_lock);
1365         }
1366
1367         if (unlikely(zone->recent_scanned[1] > file / 4)) {
1368                 spin_lock_irq(&zone->lru_lock);
1369                 zone->recent_scanned[1] /= 2;
1370                 zone->recent_rotated[1] /= 2;
1371                 spin_unlock_irq(&zone->lru_lock);
1372         }
1373
1374         /*
1375          * With swappiness at 100, anonymous and file have the same priority.
1376          * This scanning priority is essentially the inverse of IO cost.
1377          */
1378         anon_prio = sc->swappiness;
1379         file_prio = 200 - sc->swappiness;
1380
1381         /*
1382          *                  anon       recent_rotated[0]
1383          * %anon = 100 * ----------- / ----------------- * IO cost
1384          *               anon + file      rotate_sum
1385          */
1386         ap = (anon_prio + 1) * (zone->recent_scanned[0] + 1);
1387         ap /= zone->recent_rotated[0] + 1;
1388
1389         fp = (file_prio + 1) * (zone->recent_scanned[1] + 1);
1390         fp /= zone->recent_rotated[1] + 1;
1391
1392         /* Normalize to percentages */
1393         percent[0] = 100 * ap / (ap + fp + 1);
1394         percent[1] = 100 - percent[0];
1395 }
1396
1397
1398 /*
1399  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1400  */
1401 static unsigned long shrink_zone(int priority, struct zone *zone,
1402                                 struct scan_control *sc)
1403 {
1404         unsigned long nr[NR_LRU_LISTS];
1405         unsigned long nr_to_scan;
1406         unsigned long nr_reclaimed = 0;
1407         unsigned long percent[2];       /* anon @ 0; file @ 1 */
1408         enum lru_list l;
1409
1410         get_scan_ratio(zone, sc, percent);
1411
1412         for_each_evictable_lru(l) {
1413                 if (scan_global_lru(sc)) {
1414                         int file = is_file_lru(l);
1415                         int scan;
1416
1417                         scan = zone_page_state(zone, NR_LRU_BASE + l);
1418                         if (priority) {
1419                                 scan >>= priority;
1420                                 scan = (scan * percent[file]) / 100;
1421                         }
1422                         zone->lru[l].nr_scan += scan;
1423                         nr[l] = zone->lru[l].nr_scan;
1424                         if (nr[l] >= sc->swap_cluster_max)
1425                                 zone->lru[l].nr_scan = 0;
1426                         else
1427                                 nr[l] = 0;
1428                 } else {
1429                         /*
1430                          * This reclaim occurs not because zone memory shortage
1431                          * but because memory controller hits its limit.
1432                          * Don't modify zone reclaim related data.
1433                          */
1434                         nr[l] = mem_cgroup_calc_reclaim(sc->mem_cgroup, zone,
1435                                                                 priority, l);
1436                 }
1437         }
1438
1439         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1440                                         nr[LRU_INACTIVE_FILE]) {
1441                 for_each_evictable_lru(l) {
1442                         if (nr[l]) {
1443                                 nr_to_scan = min(nr[l],
1444                                         (unsigned long)sc->swap_cluster_max);
1445                                 nr[l] -= nr_to_scan;
1446
1447                                 nr_reclaimed += shrink_list(l, nr_to_scan,
1448                                                         zone, sc, priority);
1449                         }
1450                 }
1451         }
1452
1453         /*
1454          * Even if we did not try to evict anon pages at all, we want to
1455          * rebalance the anon lru active/inactive ratio.
1456          */
1457         if (!scan_global_lru(sc) || inactive_anon_is_low(zone))
1458                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1459         else if (!scan_global_lru(sc))
1460                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1461
1462         throttle_vm_writeout(sc->gfp_mask);
1463         return nr_reclaimed;
1464 }
1465
1466 /*
1467  * This is the direct reclaim path, for page-allocating processes.  We only
1468  * try to reclaim pages from zones which will satisfy the caller's allocation
1469  * request.
1470  *
1471  * We reclaim from a zone even if that zone is over pages_high.  Because:
1472  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1473  *    allocation or
1474  * b) The zones may be over pages_high but they must go *over* pages_high to
1475  *    satisfy the `incremental min' zone defense algorithm.
1476  *
1477  * Returns the number of reclaimed pages.
1478  *
1479  * If a zone is deemed to be full of pinned pages then just give it a light
1480  * scan then give up on it.
1481  */
1482 static unsigned long shrink_zones(int priority, struct zonelist *zonelist,
1483                                         struct scan_control *sc)
1484 {
1485         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1486         unsigned long nr_reclaimed = 0;
1487         struct zoneref *z;
1488         struct zone *zone;
1489
1490         sc->all_unreclaimable = 1;
1491         for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1492                 if (!populated_zone(zone))
1493                         continue;
1494                 /*
1495                  * Take care memory controller reclaiming has small influence
1496                  * to global LRU.
1497                  */
1498                 if (scan_global_lru(sc)) {
1499                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1500                                 continue;
1501                         note_zone_scanning_priority(zone, priority);
1502
1503                         if (zone_is_all_unreclaimable(zone) &&
1504                                                 priority != DEF_PRIORITY)
1505                                 continue;       /* Let kswapd poll it */
1506                         sc->all_unreclaimable = 0;
1507                 } else {
1508                         /*
1509                          * Ignore cpuset limitation here. We just want to reduce
1510                          * # of used pages by us regardless of memory shortage.
1511                          */
1512                         sc->all_unreclaimable = 0;
1513                         mem_cgroup_note_reclaim_priority(sc->mem_cgroup,
1514                                                         priority);
1515                 }
1516
1517                 nr_reclaimed += shrink_zone(priority, zone, sc);
1518         }
1519
1520         return nr_reclaimed;
1521 }
1522
1523 /*
1524  * This is the main entry point to direct page reclaim.
1525  *
1526  * If a full scan of the inactive list fails to free enough memory then we
1527  * are "out of memory" and something needs to be killed.
1528  *
1529  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1530  * high - the zone may be full of dirty or under-writeback pages, which this
1531  * caller can't do much about.  We kick pdflush and take explicit naps in the
1532  * hope that some of these pages can be written.  But if the allocating task
1533  * holds filesystem locks which prevent writeout this might not work, and the
1534  * allocation attempt will fail.
1535  *
1536  * returns:     0, if no pages reclaimed
1537  *              else, the number of pages reclaimed
1538  */
1539 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
1540                                         struct scan_control *sc)
1541 {
1542         int priority;
1543         unsigned long ret = 0;
1544         unsigned long total_scanned = 0;
1545         unsigned long nr_reclaimed = 0;
1546         struct reclaim_state *reclaim_state = current->reclaim_state;
1547         unsigned long lru_pages = 0;
1548         struct zoneref *z;
1549         struct zone *zone;
1550         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1551
1552         delayacct_freepages_start();
1553
1554         if (scan_global_lru(sc))
1555                 count_vm_event(ALLOCSTALL);
1556         /*
1557          * mem_cgroup will not do shrink_slab.
1558          */
1559         if (scan_global_lru(sc)) {
1560                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1561
1562                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1563                                 continue;
1564
1565                         lru_pages += zone_lru_pages(zone);
1566                 }
1567         }
1568
1569         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1570                 sc->nr_scanned = 0;
1571                 if (!priority)
1572                         disable_swap_token();
1573                 nr_reclaimed += shrink_zones(priority, zonelist, sc);
1574                 /*
1575                  * Don't shrink slabs when reclaiming memory from
1576                  * over limit cgroups
1577                  */
1578                 if (scan_global_lru(sc)) {
1579                         shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
1580                         if (reclaim_state) {
1581                                 nr_reclaimed += reclaim_state->reclaimed_slab;
1582                                 reclaim_state->reclaimed_slab = 0;
1583                         }
1584                 }
1585                 total_scanned += sc->nr_scanned;
1586                 if (nr_reclaimed >= sc->swap_cluster_max) {
1587                         ret = nr_reclaimed;
1588                         goto out;
1589                 }
1590
1591                 /*
1592                  * Try to write back as many pages as we just scanned.  This
1593                  * tends to cause slow streaming writers to write data to the
1594                  * disk smoothly, at the dirtying rate, which is nice.   But
1595                  * that's undesirable in laptop mode, where we *want* lumpy
1596                  * writeout.  So in laptop mode, write out the whole world.
1597                  */
1598                 if (total_scanned > sc->swap_cluster_max +
1599                                         sc->swap_cluster_max / 2) {
1600                         wakeup_pdflush(laptop_mode ? 0 : total_scanned);
1601                         sc->may_writepage = 1;
1602                 }
1603
1604                 /* Take a nap, wait for some writeback to complete */
1605                 if (sc->nr_scanned && priority < DEF_PRIORITY - 2)
1606                         congestion_wait(WRITE, HZ/10);
1607         }
1608         /* top priority shrink_zones still had more to do? don't OOM, then */
1609         if (!sc->all_unreclaimable && scan_global_lru(sc))
1610                 ret = nr_reclaimed;
1611 out:
1612         /*
1613          * Now that we've scanned all the zones at this priority level, note
1614          * that level within the zone so that the next thread which performs
1615          * scanning of this zone will immediately start out at this priority
1616          * level.  This affects only the decision whether or not to bring
1617          * mapped pages onto the inactive list.
1618          */
1619         if (priority < 0)
1620                 priority = 0;
1621
1622         if (scan_global_lru(sc)) {
1623                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1624
1625                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1626                                 continue;
1627
1628                         zone->prev_priority = priority;
1629                 }
1630         } else
1631                 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority);
1632
1633         delayacct_freepages_end();
1634
1635         return ret;
1636 }
1637
1638 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
1639                                                                 gfp_t gfp_mask)
1640 {
1641         struct scan_control sc = {
1642                 .gfp_mask = gfp_mask,
1643                 .may_writepage = !laptop_mode,
1644                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1645                 .may_swap = 1,
1646                 .swappiness = vm_swappiness,
1647                 .order = order,
1648                 .mem_cgroup = NULL,
1649                 .isolate_pages = isolate_pages_global,
1650         };
1651
1652         return do_try_to_free_pages(zonelist, &sc);
1653 }
1654
1655 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
1656
1657 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
1658                                                 gfp_t gfp_mask)
1659 {
1660         struct scan_control sc = {
1661                 .may_writepage = !laptop_mode,
1662                 .may_swap = 1,
1663                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1664                 .swappiness = vm_swappiness,
1665                 .order = 0,
1666                 .mem_cgroup = mem_cont,
1667                 .isolate_pages = mem_cgroup_isolate_pages,
1668         };
1669         struct zonelist *zonelist;
1670
1671         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1672                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1673         zonelist = NODE_DATA(numa_node_id())->node_zonelists;
1674         return do_try_to_free_pages(zonelist, &sc);
1675 }
1676 #endif
1677
1678 /*
1679  * For kswapd, balance_pgdat() will work across all this node's zones until
1680  * they are all at pages_high.
1681  *
1682  * Returns the number of pages which were actually freed.
1683  *
1684  * There is special handling here for zones which are full of pinned pages.
1685  * This can happen if the pages are all mlocked, or if they are all used by
1686  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1687  * What we do is to detect the case where all pages in the zone have been
1688  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1689  * dead and from now on, only perform a short scan.  Basically we're polling
1690  * the zone for when the problem goes away.
1691  *
1692  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
1693  * zones which have free_pages > pages_high, but once a zone is found to have
1694  * free_pages <= pages_high, we scan that zone and the lower zones regardless
1695  * of the number of free pages in the lower zones.  This interoperates with
1696  * the page allocator fallback scheme to ensure that aging of pages is balanced
1697  * across the zones.
1698  */
1699 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
1700 {
1701         int all_zones_ok;
1702         int priority;
1703         int i;
1704         unsigned long total_scanned;
1705         unsigned long nr_reclaimed;
1706         struct reclaim_state *reclaim_state = current->reclaim_state;
1707         struct scan_control sc = {
1708                 .gfp_mask = GFP_KERNEL,
1709                 .may_swap = 1,
1710                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1711                 .swappiness = vm_swappiness,
1712                 .order = order,
1713                 .mem_cgroup = NULL,
1714                 .isolate_pages = isolate_pages_global,
1715         };
1716         /*
1717          * temp_priority is used to remember the scanning priority at which
1718          * this zone was successfully refilled to free_pages == pages_high.
1719          */
1720         int temp_priority[MAX_NR_ZONES];
1721
1722 loop_again:
1723         total_scanned = 0;
1724         nr_reclaimed = 0;
1725         sc.may_writepage = !laptop_mode;
1726         count_vm_event(PAGEOUTRUN);
1727
1728         for (i = 0; i < pgdat->nr_zones; i++)
1729                 temp_priority[i] = DEF_PRIORITY;
1730
1731         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1732                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1733                 unsigned long lru_pages = 0;
1734
1735                 /* The swap token gets in the way of swapout... */
1736                 if (!priority)
1737                         disable_swap_token();
1738
1739                 all_zones_ok = 1;
1740
1741                 /*
1742                  * Scan in the highmem->dma direction for the highest
1743                  * zone which needs scanning
1744                  */
1745                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1746                         struct zone *zone = pgdat->node_zones + i;
1747
1748                         if (!populated_zone(zone))
1749                                 continue;
1750
1751                         if (zone_is_all_unreclaimable(zone) &&
1752                             priority != DEF_PRIORITY)
1753                                 continue;
1754
1755                         /*
1756                          * Do some background aging of the anon list, to give
1757                          * pages a chance to be referenced before reclaiming.
1758                          */
1759                         if (inactive_anon_is_low(zone))
1760                                 shrink_active_list(SWAP_CLUSTER_MAX, zone,
1761                                                         &sc, priority, 0);
1762
1763                         if (!zone_watermark_ok(zone, order, zone->pages_high,
1764                                                0, 0)) {
1765                                 end_zone = i;
1766                                 break;
1767                         }
1768                 }
1769                 if (i < 0)
1770                         goto out;
1771
1772                 for (i = 0; i <= end_zone; i++) {
1773                         struct zone *zone = pgdat->node_zones + i;
1774
1775                         lru_pages += zone_lru_pages(zone);
1776                 }
1777
1778                 /*
1779                  * Now scan the zone in the dma->highmem direction, stopping
1780                  * at the last zone which needs scanning.
1781                  *
1782                  * We do this because the page allocator works in the opposite
1783                  * direction.  This prevents the page allocator from allocating
1784                  * pages behind kswapd's direction of progress, which would
1785                  * cause too much scanning of the lower zones.
1786                  */
1787                 for (i = 0; i <= end_zone; i++) {
1788                         struct zone *zone = pgdat->node_zones + i;
1789                         int nr_slab;
1790
1791                         if (!populated_zone(zone))
1792                                 continue;
1793
1794                         if (zone_is_all_unreclaimable(zone) &&
1795                                         priority != DEF_PRIORITY)
1796                                 continue;
1797
1798                         if (!zone_watermark_ok(zone, order, zone->pages_high,
1799                                                end_zone, 0))
1800                                 all_zones_ok = 0;
1801                         temp_priority[i] = priority;
1802                         sc.nr_scanned = 0;
1803                         note_zone_scanning_priority(zone, priority);
1804                         /*
1805                          * We put equal pressure on every zone, unless one
1806                          * zone has way too many pages free already.
1807                          */
1808                         if (!zone_watermark_ok(zone, order, 8*zone->pages_high,
1809                                                 end_zone, 0))
1810                                 nr_reclaimed += shrink_zone(priority, zone, &sc);
1811                         reclaim_state->reclaimed_slab = 0;
1812                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1813                                                 lru_pages);
1814                         nr_reclaimed += reclaim_state->reclaimed_slab;
1815                         total_scanned += sc.nr_scanned;
1816                         if (zone_is_all_unreclaimable(zone))
1817                                 continue;
1818                         if (nr_slab == 0 && zone->pages_scanned >=
1819                                                 (zone_lru_pages(zone) * 6))
1820                                         zone_set_flag(zone,
1821                                                       ZONE_ALL_UNRECLAIMABLE);
1822                         /*
1823                          * If we've done a decent amount of scanning and
1824                          * the reclaim ratio is low, start doing writepage
1825                          * even in laptop mode
1826                          */
1827                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1828                             total_scanned > nr_reclaimed + nr_reclaimed / 2)
1829                                 sc.may_writepage = 1;
1830                 }
1831                 if (all_zones_ok)
1832                         break;          /* kswapd: all done */
1833                 /*
1834                  * OK, kswapd is getting into trouble.  Take a nap, then take
1835                  * another pass across the zones.
1836                  */
1837                 if (total_scanned && priority < DEF_PRIORITY - 2)
1838                         congestion_wait(WRITE, HZ/10);
1839
1840                 /*
1841                  * We do this so kswapd doesn't build up large priorities for
1842                  * example when it is freeing in parallel with allocators. It
1843                  * matches the direct reclaim path behaviour in terms of impact
1844                  * on zone->*_priority.
1845                  */
1846                 if (nr_reclaimed >= SWAP_CLUSTER_MAX)
1847                         break;
1848         }
1849 out:
1850         /*
1851          * Note within each zone the priority level at which this zone was
1852          * brought into a happy state.  So that the next thread which scans this
1853          * zone will start out at that priority level.
1854          */
1855         for (i = 0; i < pgdat->nr_zones; i++) {
1856                 struct zone *zone = pgdat->node_zones + i;
1857
1858                 zone->prev_priority = temp_priority[i];
1859         }
1860         if (!all_zones_ok) {
1861                 cond_resched();
1862
1863                 try_to_freeze();
1864
1865                 goto loop_again;
1866         }
1867
1868         return nr_reclaimed;
1869 }
1870
1871 /*
1872  * The background pageout daemon, started as a kernel thread
1873  * from the init process.
1874  *
1875  * This basically trickles out pages so that we have _some_
1876  * free memory available even if there is no other activity
1877  * that frees anything up. This is needed for things like routing
1878  * etc, where we otherwise might have all activity going on in
1879  * asynchronous contexts that cannot page things out.
1880  *
1881  * If there are applications that are active memory-allocators
1882  * (most normal use), this basically shouldn't matter.
1883  */
1884 static int kswapd(void *p)
1885 {
1886         unsigned long order;
1887         pg_data_t *pgdat = (pg_data_t*)p;
1888         struct task_struct *tsk = current;
1889         DEFINE_WAIT(wait);
1890         struct reclaim_state reclaim_state = {
1891                 .reclaimed_slab = 0,
1892         };
1893         node_to_cpumask_ptr(cpumask, pgdat->node_id);
1894
1895         if (!cpus_empty(*cpumask))
1896                 set_cpus_allowed_ptr(tsk, cpumask);
1897         current->reclaim_state = &reclaim_state;
1898
1899         /*
1900          * Tell the memory management that we're a "memory allocator",
1901          * and that if we need more memory we should get access to it
1902          * regardless (see "__alloc_pages()"). "kswapd" should
1903          * never get caught in the normal page freeing logic.
1904          *
1905          * (Kswapd normally doesn't need memory anyway, but sometimes
1906          * you need a small amount of memory in order to be able to
1907          * page out something else, and this flag essentially protects
1908          * us from recursively trying to free more memory as we're
1909          * trying to free the first piece of memory in the first place).
1910          */
1911         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
1912         set_freezable();
1913
1914         order = 0;
1915         for ( ; ; ) {
1916                 unsigned long new_order;
1917
1918                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1919                 new_order = pgdat->kswapd_max_order;
1920                 pgdat->kswapd_max_order = 0;
1921                 if (order < new_order) {
1922                         /*
1923                          * Don't sleep if someone wants a larger 'order'
1924                          * allocation
1925                          */
1926                         order = new_order;
1927                 } else {
1928                         if (!freezing(current))
1929                                 schedule();
1930
1931                         order = pgdat->kswapd_max_order;
1932                 }
1933                 finish_wait(&pgdat->kswapd_wait, &wait);
1934
1935                 if (!try_to_freeze()) {
1936                         /* We can speed up thawing tasks if we don't call
1937                          * balance_pgdat after returning from the refrigerator
1938                          */
1939                         balance_pgdat(pgdat, order);
1940                 }
1941         }
1942         return 0;
1943 }
1944
1945 /*
1946  * A zone is low on free memory, so wake its kswapd task to service it.
1947  */
1948 void wakeup_kswapd(struct zone *zone, int order)
1949 {
1950         pg_data_t *pgdat;
1951
1952         if (!populated_zone(zone))
1953                 return;
1954
1955         pgdat = zone->zone_pgdat;
1956         if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0))
1957                 return;
1958         if (pgdat->kswapd_max_order < order)
1959                 pgdat->kswapd_max_order = order;
1960         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1961                 return;
1962         if (!waitqueue_active(&pgdat->kswapd_wait))
1963                 return;
1964         wake_up_interruptible(&pgdat->kswapd_wait);
1965 }
1966
1967 unsigned long global_lru_pages(void)
1968 {
1969         return global_page_state(NR_ACTIVE_ANON)
1970                 + global_page_state(NR_ACTIVE_FILE)
1971                 + global_page_state(NR_INACTIVE_ANON)
1972                 + global_page_state(NR_INACTIVE_FILE);
1973 }
1974
1975 #ifdef CONFIG_PM
1976 /*
1977  * Helper function for shrink_all_memory().  Tries to reclaim 'nr_pages' pages
1978  * from LRU lists system-wide, for given pass and priority, and returns the
1979  * number of reclaimed pages
1980  *
1981  * For pass > 3 we also try to shrink the LRU lists that contain a few pages
1982  */
1983 static unsigned long shrink_all_zones(unsigned long nr_pages, int prio,
1984                                       int pass, struct scan_control *sc)
1985 {
1986         struct zone *zone;
1987         unsigned long nr_to_scan, ret = 0;
1988         enum lru_list l;
1989
1990         for_each_zone(zone) {
1991
1992                 if (!populated_zone(zone))
1993                         continue;
1994
1995                 if (zone_is_all_unreclaimable(zone) && prio != DEF_PRIORITY)
1996                         continue;
1997
1998                 for_each_evictable_lru(l) {
1999                         /* For pass = 0, we don't shrink the active list */
2000                         if (pass == 0 &&
2001                                 (l == LRU_ACTIVE || l == LRU_ACTIVE_FILE))
2002                                 continue;
2003
2004                         zone->lru[l].nr_scan +=
2005                                 (zone_page_state(zone, NR_LRU_BASE + l)
2006                                                                 >> prio) + 1;
2007                         if (zone->lru[l].nr_scan >= nr_pages || pass > 3) {
2008                                 zone->lru[l].nr_scan = 0;
2009                                 nr_to_scan = min(nr_pages,
2010                                         zone_page_state(zone,
2011                                                         NR_LRU_BASE + l));
2012                                 ret += shrink_list(l, nr_to_scan, zone,
2013                                                                 sc, prio);
2014                                 if (ret >= nr_pages)
2015                                         return ret;
2016                         }
2017                 }
2018         }
2019
2020         return ret;
2021 }
2022
2023 /*
2024  * Try to free `nr_pages' of memory, system-wide, and return the number of
2025  * freed pages.
2026  *
2027  * Rather than trying to age LRUs the aim is to preserve the overall
2028  * LRU order by reclaiming preferentially
2029  * inactive > active > active referenced > active mapped
2030  */
2031 unsigned long shrink_all_memory(unsigned long nr_pages)
2032 {
2033         unsigned long lru_pages, nr_slab;
2034         unsigned long ret = 0;
2035         int pass;
2036         struct reclaim_state reclaim_state;
2037         struct scan_control sc = {
2038                 .gfp_mask = GFP_KERNEL,
2039                 .may_swap = 0,
2040                 .swap_cluster_max = nr_pages,
2041                 .may_writepage = 1,
2042                 .swappiness = vm_swappiness,
2043                 .isolate_pages = isolate_pages_global,
2044         };
2045
2046         current->reclaim_state = &reclaim_state;
2047
2048         lru_pages = global_lru_pages();
2049         nr_slab = global_page_state(NR_SLAB_RECLAIMABLE);
2050         /* If slab caches are huge, it's better to hit them first */
2051         while (nr_slab >= lru_pages) {
2052                 reclaim_state.reclaimed_slab = 0;
2053                 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
2054                 if (!reclaim_state.reclaimed_slab)
2055                         break;
2056
2057                 ret += reclaim_state.reclaimed_slab;
2058                 if (ret >= nr_pages)
2059                         goto out;
2060
2061                 nr_slab -= reclaim_state.reclaimed_slab;
2062         }
2063
2064         /*
2065          * We try to shrink LRUs in 5 passes:
2066          * 0 = Reclaim from inactive_list only
2067          * 1 = Reclaim from active list but don't reclaim mapped
2068          * 2 = 2nd pass of type 1
2069          * 3 = Reclaim mapped (normal reclaim)
2070          * 4 = 2nd pass of type 3
2071          */
2072         for (pass = 0; pass < 5; pass++) {
2073                 int prio;
2074
2075                 /* Force reclaiming mapped pages in the passes #3 and #4 */
2076                 if (pass > 2) {
2077                         sc.may_swap = 1;
2078                         sc.swappiness = 100;
2079                 }
2080
2081                 for (prio = DEF_PRIORITY; prio >= 0; prio--) {
2082                         unsigned long nr_to_scan = nr_pages - ret;
2083
2084                         sc.nr_scanned = 0;
2085                         ret += shrink_all_zones(nr_to_scan, prio, pass, &sc);
2086                         if (ret >= nr_pages)
2087                                 goto out;
2088
2089                         reclaim_state.reclaimed_slab = 0;
2090                         shrink_slab(sc.nr_scanned, sc.gfp_mask,
2091                                         global_lru_pages());
2092                         ret += reclaim_state.reclaimed_slab;
2093                         if (ret >= nr_pages)
2094                                 goto out;
2095
2096                         if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
2097                                 congestion_wait(WRITE, HZ / 10);
2098                 }
2099         }
2100
2101         /*
2102          * If ret = 0, we could not shrink LRUs, but there may be something
2103          * in slab caches
2104          */
2105         if (!ret) {
2106                 do {
2107                         reclaim_state.reclaimed_slab = 0;
2108                         shrink_slab(nr_pages, sc.gfp_mask, global_lru_pages());
2109                         ret += reclaim_state.reclaimed_slab;
2110                 } while (ret < nr_pages && reclaim_state.reclaimed_slab > 0);
2111         }
2112
2113 out:
2114         current->reclaim_state = NULL;
2115
2116         return ret;
2117 }
2118 #endif
2119
2120 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2121    not required for correctness.  So if the last cpu in a node goes
2122    away, we get changed to run anywhere: as the first one comes back,
2123    restore their cpu bindings. */
2124 static int __devinit cpu_callback(struct notifier_block *nfb,
2125                                   unsigned long action, void *hcpu)
2126 {
2127         int nid;
2128
2129         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2130                 for_each_node_state(nid, N_HIGH_MEMORY) {
2131                         pg_data_t *pgdat = NODE_DATA(nid);
2132                         node_to_cpumask_ptr(mask, pgdat->node_id);
2133
2134                         if (any_online_cpu(*mask) < nr_cpu_ids)
2135                                 /* One of our CPUs online: restore mask */
2136                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2137                 }
2138         }
2139         return NOTIFY_OK;
2140 }
2141
2142 /*
2143  * This kswapd start function will be called by init and node-hot-add.
2144  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2145  */
2146 int kswapd_run(int nid)
2147 {
2148         pg_data_t *pgdat = NODE_DATA(nid);
2149         int ret = 0;
2150
2151         if (pgdat->kswapd)
2152                 return 0;
2153
2154         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2155         if (IS_ERR(pgdat->kswapd)) {
2156                 /* failure at boot is fatal */
2157                 BUG_ON(system_state == SYSTEM_BOOTING);
2158                 printk("Failed to start kswapd on node %d\n",nid);
2159                 ret = -1;
2160         }
2161         return ret;
2162 }
2163
2164 static int __init kswapd_init(void)
2165 {
2166         int nid;
2167
2168         swap_setup();
2169         for_each_node_state(nid, N_HIGH_MEMORY)
2170                 kswapd_run(nid);
2171         hotcpu_notifier(cpu_callback, 0);
2172         return 0;
2173 }
2174
2175 module_init(kswapd_init)
2176
2177 #ifdef CONFIG_NUMA
2178 /*
2179  * Zone reclaim mode
2180  *
2181  * If non-zero call zone_reclaim when the number of free pages falls below
2182  * the watermarks.
2183  */
2184 int zone_reclaim_mode __read_mostly;
2185
2186 #define RECLAIM_OFF 0
2187 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
2188 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
2189 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
2190
2191 /*
2192  * Priority for ZONE_RECLAIM. This determines the fraction of pages
2193  * of a node considered for each zone_reclaim. 4 scans 1/16th of
2194  * a zone.
2195  */
2196 #define ZONE_RECLAIM_PRIORITY 4
2197
2198 /*
2199  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2200  * occur.
2201  */
2202 int sysctl_min_unmapped_ratio = 1;
2203
2204 /*
2205  * If the number of slab pages in a zone grows beyond this percentage then
2206  * slab reclaim needs to occur.
2207  */
2208 int sysctl_min_slab_ratio = 5;
2209
2210 /*
2211  * Try to free up some pages from this zone through reclaim.
2212  */
2213 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2214 {
2215         /* Minimum pages needed in order to stay on node */
2216         const unsigned long nr_pages = 1 << order;
2217         struct task_struct *p = current;
2218         struct reclaim_state reclaim_state;
2219         int priority;
2220         unsigned long nr_reclaimed = 0;
2221         struct scan_control sc = {
2222                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2223                 .may_swap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2224                 .swap_cluster_max = max_t(unsigned long, nr_pages,
2225                                         SWAP_CLUSTER_MAX),
2226                 .gfp_mask = gfp_mask,
2227                 .swappiness = vm_swappiness,
2228                 .isolate_pages = isolate_pages_global,
2229         };
2230         unsigned long slab_reclaimable;
2231
2232         disable_swap_token();
2233         cond_resched();
2234         /*
2235          * We need to be able to allocate from the reserves for RECLAIM_SWAP
2236          * and we also need to be able to write out pages for RECLAIM_WRITE
2237          * and RECLAIM_SWAP.
2238          */
2239         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2240         reclaim_state.reclaimed_slab = 0;
2241         p->reclaim_state = &reclaim_state;
2242
2243         if (zone_page_state(zone, NR_FILE_PAGES) -
2244                 zone_page_state(zone, NR_FILE_MAPPED) >
2245                 zone->min_unmapped_pages) {
2246                 /*
2247                  * Free memory by calling shrink zone with increasing
2248                  * priorities until we have enough memory freed.
2249                  */
2250                 priority = ZONE_RECLAIM_PRIORITY;
2251                 do {
2252                         note_zone_scanning_priority(zone, priority);
2253                         nr_reclaimed += shrink_zone(priority, zone, &sc);
2254                         priority--;
2255                 } while (priority >= 0 && nr_reclaimed < nr_pages);
2256         }
2257
2258         slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2259         if (slab_reclaimable > zone->min_slab_pages) {
2260                 /*
2261                  * shrink_slab() does not currently allow us to determine how
2262                  * many pages were freed in this zone. So we take the current
2263                  * number of slab pages and shake the slab until it is reduced
2264                  * by the same nr_pages that we used for reclaiming unmapped
2265                  * pages.
2266                  *
2267                  * Note that shrink_slab will free memory on all zones and may
2268                  * take a long time.
2269                  */
2270                 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
2271                         zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
2272                                 slab_reclaimable - nr_pages)
2273                         ;
2274
2275                 /*
2276                  * Update nr_reclaimed by the number of slab pages we
2277                  * reclaimed from this zone.
2278                  */
2279                 nr_reclaimed += slab_reclaimable -
2280                         zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2281         }
2282
2283         p->reclaim_state = NULL;
2284         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
2285         return nr_reclaimed >= nr_pages;
2286 }
2287
2288 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2289 {
2290         int node_id;
2291         int ret;
2292
2293         /*
2294          * Zone reclaim reclaims unmapped file backed pages and
2295          * slab pages if we are over the defined limits.
2296          *
2297          * A small portion of unmapped file backed pages is needed for
2298          * file I/O otherwise pages read by file I/O will be immediately
2299          * thrown out if the zone is overallocated. So we do not reclaim
2300          * if less than a specified percentage of the zone is used by
2301          * unmapped file backed pages.
2302          */
2303         if (zone_page_state(zone, NR_FILE_PAGES) -
2304             zone_page_state(zone, NR_FILE_MAPPED) <= zone->min_unmapped_pages
2305             && zone_page_state(zone, NR_SLAB_RECLAIMABLE)
2306                         <= zone->min_slab_pages)
2307                 return 0;
2308
2309         if (zone_is_all_unreclaimable(zone))
2310                 return 0;
2311
2312         /*
2313          * Do not scan if the allocation should not be delayed.
2314          */
2315         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
2316                         return 0;
2317
2318         /*
2319          * Only run zone reclaim on the local zone or on zones that do not
2320          * have associated processors. This will favor the local processor
2321          * over remote processors and spread off node memory allocations
2322          * as wide as possible.
2323          */
2324         node_id = zone_to_nid(zone);
2325         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
2326                 return 0;
2327
2328         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2329                 return 0;
2330         ret = __zone_reclaim(zone, gfp_mask, order);
2331         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2332
2333         return ret;
2334 }
2335 #endif
2336
2337 #ifdef CONFIG_UNEVICTABLE_LRU
2338 /*
2339  * page_evictable - test whether a page is evictable
2340  * @page: the page to test
2341  * @vma: the VMA in which the page is or will be mapped, may be NULL
2342  *
2343  * Test whether page is evictable--i.e., should be placed on active/inactive
2344  * lists vs unevictable list.  The vma argument is !NULL when called from the
2345  * fault path to determine how to instantate a new page.
2346  *
2347  * Reasons page might not be evictable:
2348  * (1) page's mapping marked unevictable
2349  * (2) page is part of an mlocked VMA
2350  *
2351  */
2352 int page_evictable(struct page *page, struct vm_area_struct *vma)
2353 {
2354
2355         if (mapping_unevictable(page_mapping(page)))
2356                 return 0;
2357
2358         if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2359                 return 0;
2360
2361         return 1;
2362 }
2363
2364 static void show_page_path(struct page *page)
2365 {
2366         char buf[256];
2367         if (page_is_file_cache(page)) {
2368                 struct address_space *mapping = page->mapping;
2369                 struct dentry *dentry;
2370                 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
2371
2372                 spin_lock(&mapping->i_mmap_lock);
2373                 dentry = d_find_alias(mapping->host);
2374                 printk(KERN_INFO "rescued: %s %lu\n",
2375                        dentry_path(dentry, buf, 256), pgoff);
2376                 spin_unlock(&mapping->i_mmap_lock);
2377         } else {
2378 #if defined(CONFIG_MM_OWNER) && defined(CONFIG_MMU)
2379                 struct anon_vma *anon_vma;
2380                 struct vm_area_struct *vma;
2381
2382                 anon_vma = page_lock_anon_vma(page);
2383                 if (!anon_vma)
2384                         return;
2385
2386                 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
2387                         printk(KERN_INFO "rescued: anon %s\n",
2388                                vma->vm_mm->owner->comm);
2389                         break;
2390                 }
2391                 page_unlock_anon_vma(anon_vma);
2392 #endif
2393         }
2394 }
2395
2396
2397 /**
2398  * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2399  * @page: page to check evictability and move to appropriate lru list
2400  * @zone: zone page is in
2401  *
2402  * Checks a page for evictability and moves the page to the appropriate
2403  * zone lru list.
2404  *
2405  * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2406  * have PageUnevictable set.
2407  */
2408 static void check_move_unevictable_page(struct page *page, struct zone *zone)
2409 {
2410         VM_BUG_ON(PageActive(page));
2411
2412 retry:
2413         ClearPageUnevictable(page);
2414         if (page_evictable(page, NULL)) {
2415                 enum lru_list l = LRU_INACTIVE_ANON + page_is_file_cache(page);
2416
2417                 show_page_path(page);
2418
2419                 __dec_zone_state(zone, NR_UNEVICTABLE);
2420                 list_move(&page->lru, &zone->lru[l].list);
2421                 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
2422                 __count_vm_event(UNEVICTABLE_PGRESCUED);
2423         } else {
2424                 /*
2425                  * rotate unevictable list
2426                  */
2427                 SetPageUnevictable(page);
2428                 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
2429                 if (page_evictable(page, NULL))
2430                         goto retry;
2431         }
2432 }
2433
2434 /**
2435  * scan_mapping_unevictable_pages - scan an address space for evictable pages
2436  * @mapping: struct address_space to scan for evictable pages
2437  *
2438  * Scan all pages in mapping.  Check unevictable pages for
2439  * evictability and move them to the appropriate zone lru list.
2440  */
2441 void scan_mapping_unevictable_pages(struct address_space *mapping)
2442 {
2443         pgoff_t next = 0;
2444         pgoff_t end   = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
2445                          PAGE_CACHE_SHIFT;
2446         struct zone *zone;
2447         struct pagevec pvec;
2448
2449         if (mapping->nrpages == 0)
2450                 return;
2451
2452         pagevec_init(&pvec, 0);
2453         while (next < end &&
2454                 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
2455                 int i;
2456                 int pg_scanned = 0;
2457
2458                 zone = NULL;
2459
2460                 for (i = 0; i < pagevec_count(&pvec); i++) {
2461                         struct page *page = pvec.pages[i];
2462                         pgoff_t page_index = page->index;
2463                         struct zone *pagezone = page_zone(page);
2464
2465                         pg_scanned++;
2466                         if (page_index > next)
2467                                 next = page_index;
2468                         next++;
2469
2470                         if (pagezone != zone) {
2471                                 if (zone)
2472                                         spin_unlock_irq(&zone->lru_lock);
2473                                 zone = pagezone;
2474                                 spin_lock_irq(&zone->lru_lock);
2475                         }
2476
2477                         if (PageLRU(page) && PageUnevictable(page))
2478                                 check_move_unevictable_page(page, zone);
2479                 }
2480                 if (zone)
2481                         spin_unlock_irq(&zone->lru_lock);
2482                 pagevec_release(&pvec);
2483
2484                 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
2485         }
2486
2487 }
2488
2489 /**
2490  * scan_zone_unevictable_pages - check unevictable list for evictable pages
2491  * @zone - zone of which to scan the unevictable list
2492  *
2493  * Scan @zone's unevictable LRU lists to check for pages that have become
2494  * evictable.  Move those that have to @zone's inactive list where they
2495  * become candidates for reclaim, unless shrink_inactive_zone() decides
2496  * to reactivate them.  Pages that are still unevictable are rotated
2497  * back onto @zone's unevictable list.
2498  */
2499 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
2500 void scan_zone_unevictable_pages(struct zone *zone)
2501 {
2502         struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
2503         unsigned long scan;
2504         unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
2505
2506         while (nr_to_scan > 0) {
2507                 unsigned long batch_size = min(nr_to_scan,
2508                                                 SCAN_UNEVICTABLE_BATCH_SIZE);
2509
2510                 spin_lock_irq(&zone->lru_lock);
2511                 for (scan = 0;  scan < batch_size; scan++) {
2512                         struct page *page = lru_to_page(l_unevictable);
2513
2514                         if (!trylock_page(page))
2515                                 continue;
2516
2517                         prefetchw_prev_lru_page(page, l_unevictable, flags);
2518
2519                         if (likely(PageLRU(page) && PageUnevictable(page)))
2520                                 check_move_unevictable_page(page, zone);
2521
2522                         unlock_page(page);
2523                 }
2524                 spin_unlock_irq(&zone->lru_lock);
2525
2526                 nr_to_scan -= batch_size;
2527         }
2528 }
2529
2530
2531 /**
2532  * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
2533  *
2534  * A really big hammer:  scan all zones' unevictable LRU lists to check for
2535  * pages that have become evictable.  Move those back to the zones'
2536  * inactive list where they become candidates for reclaim.
2537  * This occurs when, e.g., we have unswappable pages on the unevictable lists,
2538  * and we add swap to the system.  As such, it runs in the context of a task
2539  * that has possibly/probably made some previously unevictable pages
2540  * evictable.
2541  */
2542 void scan_all_zones_unevictable_pages(void)
2543 {
2544         struct zone *zone;
2545
2546         for_each_zone(zone) {
2547                 scan_zone_unevictable_pages(zone);
2548         }
2549 }
2550
2551 /*
2552  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
2553  * all nodes' unevictable lists for evictable pages
2554  */
2555 unsigned long scan_unevictable_pages;
2556
2557 int scan_unevictable_handler(struct ctl_table *table, int write,
2558                            struct file *file, void __user *buffer,
2559                            size_t *length, loff_t *ppos)
2560 {
2561         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
2562
2563         if (write && *(unsigned long *)table->data)
2564                 scan_all_zones_unevictable_pages();
2565
2566         scan_unevictable_pages = 0;
2567         return 0;
2568 }
2569
2570 /*
2571  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
2572  * a specified node's per zone unevictable lists for evictable pages.
2573  */
2574
2575 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
2576                                           struct sysdev_attribute *attr,
2577                                           char *buf)
2578 {
2579         return sprintf(buf, "0\n");     /* always zero; should fit... */
2580 }
2581
2582 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
2583                                            struct sysdev_attribute *attr,
2584                                         const char *buf, size_t count)
2585 {
2586         struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
2587         struct zone *zone;
2588         unsigned long res;
2589         unsigned long req = strict_strtoul(buf, 10, &res);
2590
2591         if (!req)
2592                 return 1;       /* zero is no-op */
2593
2594         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2595                 if (!populated_zone(zone))
2596                         continue;
2597                 scan_zone_unevictable_pages(zone);
2598         }
2599         return 1;
2600 }
2601
2602
2603 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
2604                         read_scan_unevictable_node,
2605                         write_scan_unevictable_node);
2606
2607 int scan_unevictable_register_node(struct node *node)
2608 {
2609         return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
2610 }
2611
2612 void scan_unevictable_unregister_node(struct node *node)
2613 {
2614         sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
2615 }
2616
2617 #endif