memcg: reclaim shouldn't change zone->recent_rotated statistics
[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                 /*
621                  * Anonymous process memory has backing store?
622                  * Try to allocate it some swap space here.
623                  */
624                 if (PageAnon(page) && !PageSwapCache(page)) {
625                         if (!(sc->gfp_mask & __GFP_IO))
626                                 goto keep_locked;
627                         if (!add_to_swap(page))
628                                 goto activate_locked;
629                         may_enter_fs = 1;
630                 }
631
632                 mapping = page_mapping(page);
633
634                 /*
635                  * The page is mapped into the page tables of one or more
636                  * processes. Try to unmap it here.
637                  */
638                 if (page_mapped(page) && mapping) {
639                         switch (try_to_unmap(page, 0)) {
640                         case SWAP_FAIL:
641                                 goto activate_locked;
642                         case SWAP_AGAIN:
643                                 goto keep_locked;
644                         case SWAP_MLOCK:
645                                 goto cull_mlocked;
646                         case SWAP_SUCCESS:
647                                 ; /* try to free the page below */
648                         }
649                 }
650
651                 if (PageDirty(page)) {
652                         if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && referenced)
653                                 goto keep_locked;
654                         if (!may_enter_fs)
655                                 goto keep_locked;
656                         if (!sc->may_writepage)
657                                 goto keep_locked;
658
659                         /* Page is dirty, try to write it out here */
660                         switch (pageout(page, mapping, sync_writeback)) {
661                         case PAGE_KEEP:
662                                 goto keep_locked;
663                         case PAGE_ACTIVATE:
664                                 goto activate_locked;
665                         case PAGE_SUCCESS:
666                                 if (PageWriteback(page) || PageDirty(page))
667                                         goto keep;
668                                 /*
669                                  * A synchronous write - probably a ramdisk.  Go
670                                  * ahead and try to reclaim the page.
671                                  */
672                                 if (!trylock_page(page))
673                                         goto keep;
674                                 if (PageDirty(page) || PageWriteback(page))
675                                         goto keep_locked;
676                                 mapping = page_mapping(page);
677                         case PAGE_CLEAN:
678                                 ; /* try to free the page below */
679                         }
680                 }
681
682                 /*
683                  * If the page has buffers, try to free the buffer mappings
684                  * associated with this page. If we succeed we try to free
685                  * the page as well.
686                  *
687                  * We do this even if the page is PageDirty().
688                  * try_to_release_page() does not perform I/O, but it is
689                  * possible for a page to have PageDirty set, but it is actually
690                  * clean (all its buffers are clean).  This happens if the
691                  * buffers were written out directly, with submit_bh(). ext3
692                  * will do this, as well as the blockdev mapping.
693                  * try_to_release_page() will discover that cleanness and will
694                  * drop the buffers and mark the page clean - it can be freed.
695                  *
696                  * Rarely, pages can have buffers and no ->mapping.  These are
697                  * the pages which were not successfully invalidated in
698                  * truncate_complete_page().  We try to drop those buffers here
699                  * and if that worked, and the page is no longer mapped into
700                  * process address space (page_count == 1) it can be freed.
701                  * Otherwise, leave the page on the LRU so it is swappable.
702                  */
703                 if (PagePrivate(page)) {
704                         if (!try_to_release_page(page, sc->gfp_mask))
705                                 goto activate_locked;
706                         if (!mapping && page_count(page) == 1) {
707                                 unlock_page(page);
708                                 if (put_page_testzero(page))
709                                         goto free_it;
710                                 else {
711                                         /*
712                                          * rare race with speculative reference.
713                                          * the speculative reference will free
714                                          * this page shortly, so we may
715                                          * increment nr_reclaimed here (and
716                                          * leave it off the LRU).
717                                          */
718                                         nr_reclaimed++;
719                                         continue;
720                                 }
721                         }
722                 }
723
724                 if (!mapping || !__remove_mapping(mapping, page))
725                         goto keep_locked;
726
727                 /*
728                  * At this point, we have no other references and there is
729                  * no way to pick any more up (removed from LRU, removed
730                  * from pagecache). Can use non-atomic bitops now (and
731                  * we obviously don't have to worry about waking up a process
732                  * waiting on the page lock, because there are no references.
733                  */
734                 __clear_page_locked(page);
735 free_it:
736                 nr_reclaimed++;
737                 if (!pagevec_add(&freed_pvec, page)) {
738                         __pagevec_free(&freed_pvec);
739                         pagevec_reinit(&freed_pvec);
740                 }
741                 continue;
742
743 cull_mlocked:
744                 if (PageSwapCache(page))
745                         try_to_free_swap(page);
746                 unlock_page(page);
747                 putback_lru_page(page);
748                 continue;
749
750 activate_locked:
751                 /* Not a candidate for swapping, so reclaim swap space. */
752                 if (PageSwapCache(page) && vm_swap_full())
753                         try_to_free_swap(page);
754                 VM_BUG_ON(PageActive(page));
755                 SetPageActive(page);
756                 pgactivate++;
757 keep_locked:
758                 unlock_page(page);
759 keep:
760                 list_add(&page->lru, &ret_pages);
761                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
762         }
763         list_splice(&ret_pages, page_list);
764         if (pagevec_count(&freed_pvec))
765                 __pagevec_free(&freed_pvec);
766         count_vm_events(PGACTIVATE, pgactivate);
767         return nr_reclaimed;
768 }
769
770 /* LRU Isolation modes. */
771 #define ISOLATE_INACTIVE 0      /* Isolate inactive pages. */
772 #define ISOLATE_ACTIVE 1        /* Isolate active pages. */
773 #define ISOLATE_BOTH 2          /* Isolate both active and inactive pages. */
774
775 /*
776  * Attempt to remove the specified page from its LRU.  Only take this page
777  * if it is of the appropriate PageActive status.  Pages which are being
778  * freed elsewhere are also ignored.
779  *
780  * page:        page to consider
781  * mode:        one of the LRU isolation modes defined above
782  *
783  * returns 0 on success, -ve errno on failure.
784  */
785 int __isolate_lru_page(struct page *page, int mode, int file)
786 {
787         int ret = -EINVAL;
788
789         /* Only take pages on the LRU. */
790         if (!PageLRU(page))
791                 return ret;
792
793         /*
794          * When checking the active state, we need to be sure we are
795          * dealing with comparible boolean values.  Take the logical not
796          * of each.
797          */
798         if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
799                 return ret;
800
801         if (mode != ISOLATE_BOTH && (!page_is_file_cache(page) != !file))
802                 return ret;
803
804         /*
805          * When this function is being called for lumpy reclaim, we
806          * initially look into all LRU pages, active, inactive and
807          * unevictable; only give shrink_page_list evictable pages.
808          */
809         if (PageUnevictable(page))
810                 return ret;
811
812         ret = -EBUSY;
813         if (likely(get_page_unless_zero(page))) {
814                 /*
815                  * Be careful not to clear PageLRU until after we're
816                  * sure the page is not being freed elsewhere -- the
817                  * page release code relies on it.
818                  */
819                 ClearPageLRU(page);
820                 ret = 0;
821         }
822
823         return ret;
824 }
825
826 /*
827  * zone->lru_lock is heavily contended.  Some of the functions that
828  * shrink the lists perform better by taking out a batch of pages
829  * and working on them outside the LRU lock.
830  *
831  * For pagecache intensive workloads, this function is the hottest
832  * spot in the kernel (apart from copy_*_user functions).
833  *
834  * Appropriate locks must be held before calling this function.
835  *
836  * @nr_to_scan: The number of pages to look through on the list.
837  * @src:        The LRU list to pull pages off.
838  * @dst:        The temp list to put pages on to.
839  * @scanned:    The number of pages that were scanned.
840  * @order:      The caller's attempted allocation order
841  * @mode:       One of the LRU isolation modes
842  * @file:       True [1] if isolating file [!anon] pages
843  *
844  * returns how many pages were moved onto *@dst.
845  */
846 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
847                 struct list_head *src, struct list_head *dst,
848                 unsigned long *scanned, int order, int mode, int file)
849 {
850         unsigned long nr_taken = 0;
851         unsigned long scan;
852
853         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
854                 struct page *page;
855                 unsigned long pfn;
856                 unsigned long end_pfn;
857                 unsigned long page_pfn;
858                 int zone_id;
859
860                 page = lru_to_page(src);
861                 prefetchw_prev_lru_page(page, src, flags);
862
863                 VM_BUG_ON(!PageLRU(page));
864
865                 switch (__isolate_lru_page(page, mode, file)) {
866                 case 0:
867                         list_move(&page->lru, dst);
868                         nr_taken++;
869                         break;
870
871                 case -EBUSY:
872                         /* else it is being freed elsewhere */
873                         list_move(&page->lru, src);
874                         continue;
875
876                 default:
877                         BUG();
878                 }
879
880                 if (!order)
881                         continue;
882
883                 /*
884                  * Attempt to take all pages in the order aligned region
885                  * surrounding the tag page.  Only take those pages of
886                  * the same active state as that tag page.  We may safely
887                  * round the target page pfn down to the requested order
888                  * as the mem_map is guarenteed valid out to MAX_ORDER,
889                  * where that page is in a different zone we will detect
890                  * it from its zone id and abort this block scan.
891                  */
892                 zone_id = page_zone_id(page);
893                 page_pfn = page_to_pfn(page);
894                 pfn = page_pfn & ~((1 << order) - 1);
895                 end_pfn = pfn + (1 << order);
896                 for (; pfn < end_pfn; pfn++) {
897                         struct page *cursor_page;
898
899                         /* The target page is in the block, ignore it. */
900                         if (unlikely(pfn == page_pfn))
901                                 continue;
902
903                         /* Avoid holes within the zone. */
904                         if (unlikely(!pfn_valid_within(pfn)))
905                                 break;
906
907                         cursor_page = pfn_to_page(pfn);
908
909                         /* Check that we have not crossed a zone boundary. */
910                         if (unlikely(page_zone_id(cursor_page) != zone_id))
911                                 continue;
912                         switch (__isolate_lru_page(cursor_page, mode, file)) {
913                         case 0:
914                                 list_move(&cursor_page->lru, dst);
915                                 nr_taken++;
916                                 scan++;
917                                 break;
918
919                         case -EBUSY:
920                                 /* else it is being freed elsewhere */
921                                 list_move(&cursor_page->lru, src);
922                         default:
923                                 break;  /* ! on LRU or wrong list */
924                         }
925                 }
926         }
927
928         *scanned = scan;
929         return nr_taken;
930 }
931
932 static unsigned long isolate_pages_global(unsigned long nr,
933                                         struct list_head *dst,
934                                         unsigned long *scanned, int order,
935                                         int mode, struct zone *z,
936                                         struct mem_cgroup *mem_cont,
937                                         int active, int file)
938 {
939         int lru = LRU_BASE;
940         if (active)
941                 lru += LRU_ACTIVE;
942         if (file)
943                 lru += LRU_FILE;
944         return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
945                                                                 mode, !!file);
946 }
947
948 /*
949  * clear_active_flags() is a helper for shrink_active_list(), clearing
950  * any active bits from the pages in the list.
951  */
952 static unsigned long clear_active_flags(struct list_head *page_list,
953                                         unsigned int *count)
954 {
955         int nr_active = 0;
956         int lru;
957         struct page *page;
958
959         list_for_each_entry(page, page_list, lru) {
960                 lru = page_is_file_cache(page);
961                 if (PageActive(page)) {
962                         lru += LRU_ACTIVE;
963                         ClearPageActive(page);
964                         nr_active++;
965                 }
966                 count[lru]++;
967         }
968
969         return nr_active;
970 }
971
972 /**
973  * isolate_lru_page - tries to isolate a page from its LRU list
974  * @page: page to isolate from its LRU list
975  *
976  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
977  * vmstat statistic corresponding to whatever LRU list the page was on.
978  *
979  * Returns 0 if the page was removed from an LRU list.
980  * Returns -EBUSY if the page was not on an LRU list.
981  *
982  * The returned page will have PageLRU() cleared.  If it was found on
983  * the active list, it will have PageActive set.  If it was found on
984  * the unevictable list, it will have the PageUnevictable bit set. That flag
985  * may need to be cleared by the caller before letting the page go.
986  *
987  * The vmstat statistic corresponding to the list on which the page was
988  * found will be decremented.
989  *
990  * Restrictions:
991  * (1) Must be called with an elevated refcount on the page. This is a
992  *     fundamentnal difference from isolate_lru_pages (which is called
993  *     without a stable reference).
994  * (2) the lru_lock must not be held.
995  * (3) interrupts must be enabled.
996  */
997 int isolate_lru_page(struct page *page)
998 {
999         int ret = -EBUSY;
1000
1001         if (PageLRU(page)) {
1002                 struct zone *zone = page_zone(page);
1003
1004                 spin_lock_irq(&zone->lru_lock);
1005                 if (PageLRU(page) && get_page_unless_zero(page)) {
1006                         int lru = page_lru(page);
1007                         ret = 0;
1008                         ClearPageLRU(page);
1009
1010                         del_page_from_lru_list(zone, page, lru);
1011                 }
1012                 spin_unlock_irq(&zone->lru_lock);
1013         }
1014         return ret;
1015 }
1016
1017 /*
1018  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1019  * of reclaimed pages
1020  */
1021 static unsigned long shrink_inactive_list(unsigned long max_scan,
1022                         struct zone *zone, struct scan_control *sc,
1023                         int priority, int file)
1024 {
1025         LIST_HEAD(page_list);
1026         struct pagevec pvec;
1027         unsigned long nr_scanned = 0;
1028         unsigned long nr_reclaimed = 0;
1029
1030         pagevec_init(&pvec, 1);
1031
1032         lru_add_drain();
1033         spin_lock_irq(&zone->lru_lock);
1034         do {
1035                 struct page *page;
1036                 unsigned long nr_taken;
1037                 unsigned long nr_scan;
1038                 unsigned long nr_freed;
1039                 unsigned long nr_active;
1040                 unsigned int count[NR_LRU_LISTS] = { 0, };
1041                 int mode = ISOLATE_INACTIVE;
1042
1043                 /*
1044                  * If we need a large contiguous chunk of memory, or have
1045                  * trouble getting a small set of contiguous pages, we
1046                  * will reclaim both active and inactive pages.
1047                  *
1048                  * We use the same threshold as pageout congestion_wait below.
1049                  */
1050                 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1051                         mode = ISOLATE_BOTH;
1052                 else if (sc->order && priority < DEF_PRIORITY - 2)
1053                         mode = ISOLATE_BOTH;
1054
1055                 nr_taken = sc->isolate_pages(sc->swap_cluster_max,
1056                              &page_list, &nr_scan, sc->order, mode,
1057                                 zone, sc->mem_cgroup, 0, file);
1058                 nr_active = clear_active_flags(&page_list, count);
1059                 __count_vm_events(PGDEACTIVATE, nr_active);
1060
1061                 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1062                                                 -count[LRU_ACTIVE_FILE]);
1063                 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1064                                                 -count[LRU_INACTIVE_FILE]);
1065                 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1066                                                 -count[LRU_ACTIVE_ANON]);
1067                 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1068                                                 -count[LRU_INACTIVE_ANON]);
1069
1070                 if (scan_global_lru(sc)) {
1071                         zone->pages_scanned += nr_scan;
1072                         zone->recent_scanned[0] += count[LRU_INACTIVE_ANON];
1073                         zone->recent_scanned[0] += count[LRU_ACTIVE_ANON];
1074                         zone->recent_scanned[1] += count[LRU_INACTIVE_FILE];
1075                         zone->recent_scanned[1] += count[LRU_ACTIVE_FILE];
1076                 }
1077                 spin_unlock_irq(&zone->lru_lock);
1078
1079                 nr_scanned += nr_scan;
1080                 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
1081
1082                 /*
1083                  * If we are direct reclaiming for contiguous pages and we do
1084                  * not reclaim everything in the list, try again and wait
1085                  * for IO to complete. This will stall high-order allocations
1086                  * but that should be acceptable to the caller
1087                  */
1088                 if (nr_freed < nr_taken && !current_is_kswapd() &&
1089                                         sc->order > PAGE_ALLOC_COSTLY_ORDER) {
1090                         congestion_wait(WRITE, HZ/10);
1091
1092                         /*
1093                          * The attempt at page out may have made some
1094                          * of the pages active, mark them inactive again.
1095                          */
1096                         nr_active = clear_active_flags(&page_list, count);
1097                         count_vm_events(PGDEACTIVATE, nr_active);
1098
1099                         nr_freed += shrink_page_list(&page_list, sc,
1100                                                         PAGEOUT_IO_SYNC);
1101                 }
1102
1103                 nr_reclaimed += nr_freed;
1104                 local_irq_disable();
1105                 if (current_is_kswapd()) {
1106                         __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scan);
1107                         __count_vm_events(KSWAPD_STEAL, nr_freed);
1108                 } else if (scan_global_lru(sc))
1109                         __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scan);
1110
1111                 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
1112
1113                 if (nr_taken == 0)
1114                         goto done;
1115
1116                 spin_lock(&zone->lru_lock);
1117                 /*
1118                  * Put back any unfreeable pages.
1119                  */
1120                 while (!list_empty(&page_list)) {
1121                         int lru;
1122                         page = lru_to_page(&page_list);
1123                         VM_BUG_ON(PageLRU(page));
1124                         list_del(&page->lru);
1125                         if (unlikely(!page_evictable(page, NULL))) {
1126                                 spin_unlock_irq(&zone->lru_lock);
1127                                 putback_lru_page(page);
1128                                 spin_lock_irq(&zone->lru_lock);
1129                                 continue;
1130                         }
1131                         SetPageLRU(page);
1132                         lru = page_lru(page);
1133                         add_page_to_lru_list(zone, page, lru);
1134                         mem_cgroup_move_lists(page, lru);
1135                         if (PageActive(page) && scan_global_lru(sc)) {
1136                                 int file = !!page_is_file_cache(page);
1137                                 zone->recent_rotated[file]++;
1138                         }
1139                         if (!pagevec_add(&pvec, page)) {
1140                                 spin_unlock_irq(&zone->lru_lock);
1141                                 __pagevec_release(&pvec);
1142                                 spin_lock_irq(&zone->lru_lock);
1143                         }
1144                 }
1145         } while (nr_scanned < max_scan);
1146         spin_unlock(&zone->lru_lock);
1147 done:
1148         local_irq_enable();
1149         pagevec_release(&pvec);
1150         return nr_reclaimed;
1151 }
1152
1153 /*
1154  * We are about to scan this zone at a certain priority level.  If that priority
1155  * level is smaller (ie: more urgent) than the previous priority, then note
1156  * that priority level within the zone.  This is done so that when the next
1157  * process comes in to scan this zone, it will immediately start out at this
1158  * priority level rather than having to build up its own scanning priority.
1159  * Here, this priority affects only the reclaim-mapped threshold.
1160  */
1161 static inline void note_zone_scanning_priority(struct zone *zone, int priority)
1162 {
1163         if (priority < zone->prev_priority)
1164                 zone->prev_priority = priority;
1165 }
1166
1167 static inline int zone_is_near_oom(struct zone *zone)
1168 {
1169         return zone->pages_scanned >= (zone_lru_pages(zone) * 3);
1170 }
1171
1172 /*
1173  * This moves pages from the active list to the inactive list.
1174  *
1175  * We move them the other way if the page is referenced by one or more
1176  * processes, from rmap.
1177  *
1178  * If the pages are mostly unmapped, the processing is fast and it is
1179  * appropriate to hold zone->lru_lock across the whole operation.  But if
1180  * the pages are mapped, the processing is slow (page_referenced()) so we
1181  * should drop zone->lru_lock around each page.  It's impossible to balance
1182  * this, so instead we remove the pages from the LRU while processing them.
1183  * It is safe to rely on PG_active against the non-LRU pages in here because
1184  * nobody will play with that bit on a non-LRU page.
1185  *
1186  * The downside is that we have to touch page->_count against each page.
1187  * But we had to alter page->flags anyway.
1188  */
1189
1190
1191 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1192                         struct scan_control *sc, int priority, int file)
1193 {
1194         unsigned long pgmoved;
1195         int pgdeactivate = 0;
1196         unsigned long pgscanned;
1197         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1198         LIST_HEAD(l_inactive);
1199         struct page *page;
1200         struct pagevec pvec;
1201         enum lru_list lru;
1202
1203         lru_add_drain();
1204         spin_lock_irq(&zone->lru_lock);
1205         pgmoved = sc->isolate_pages(nr_pages, &l_hold, &pgscanned, sc->order,
1206                                         ISOLATE_ACTIVE, zone,
1207                                         sc->mem_cgroup, 1, file);
1208         /*
1209          * zone->pages_scanned is used for detect zone's oom
1210          * mem_cgroup remembers nr_scan by itself.
1211          */
1212         if (scan_global_lru(sc)) {
1213                 zone->pages_scanned += pgscanned;
1214                 zone->recent_scanned[!!file] += pgmoved;
1215         }
1216
1217         if (file)
1218                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -pgmoved);
1219         else
1220                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -pgmoved);
1221         spin_unlock_irq(&zone->lru_lock);
1222
1223         pgmoved = 0;
1224         while (!list_empty(&l_hold)) {
1225                 cond_resched();
1226                 page = lru_to_page(&l_hold);
1227                 list_del(&page->lru);
1228
1229                 if (unlikely(!page_evictable(page, NULL))) {
1230                         putback_lru_page(page);
1231                         continue;
1232                 }
1233
1234                 /* page_referenced clears PageReferenced */
1235                 if (page_mapping_inuse(page) &&
1236                     page_referenced(page, 0, sc->mem_cgroup))
1237                         pgmoved++;
1238
1239                 list_add(&page->lru, &l_inactive);
1240         }
1241
1242         spin_lock_irq(&zone->lru_lock);
1243         /*
1244          * Count referenced pages from currently used mappings as
1245          * rotated, even though they are moved to the inactive list.
1246          * This helps balance scan pressure between file and anonymous
1247          * pages in get_scan_ratio.
1248          */
1249         if (scan_global_lru(sc))
1250                 zone->recent_rotated[!!file] += pgmoved;
1251
1252         /*
1253          * Move the pages to the [file or anon] inactive list.
1254          */
1255         pagevec_init(&pvec, 1);
1256
1257         pgmoved = 0;
1258         lru = LRU_BASE + file * LRU_FILE;
1259         while (!list_empty(&l_inactive)) {
1260                 page = lru_to_page(&l_inactive);
1261                 prefetchw_prev_lru_page(page, &l_inactive, flags);
1262                 VM_BUG_ON(PageLRU(page));
1263                 SetPageLRU(page);
1264                 VM_BUG_ON(!PageActive(page));
1265                 ClearPageActive(page);
1266
1267                 list_move(&page->lru, &zone->lru[lru].list);
1268                 mem_cgroup_move_lists(page, lru);
1269                 pgmoved++;
1270                 if (!pagevec_add(&pvec, page)) {
1271                         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1272                         spin_unlock_irq(&zone->lru_lock);
1273                         pgdeactivate += pgmoved;
1274                         pgmoved = 0;
1275                         if (buffer_heads_over_limit)
1276                                 pagevec_strip(&pvec);
1277                         __pagevec_release(&pvec);
1278                         spin_lock_irq(&zone->lru_lock);
1279                 }
1280         }
1281         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1282         pgdeactivate += pgmoved;
1283         if (buffer_heads_over_limit) {
1284                 spin_unlock_irq(&zone->lru_lock);
1285                 pagevec_strip(&pvec);
1286                 spin_lock_irq(&zone->lru_lock);
1287         }
1288         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1289         __count_vm_events(PGDEACTIVATE, pgdeactivate);
1290         spin_unlock_irq(&zone->lru_lock);
1291         if (vm_swap_full())
1292                 pagevec_swap_free(&pvec);
1293
1294         pagevec_release(&pvec);
1295 }
1296
1297 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1298         struct zone *zone, struct scan_control *sc, int priority)
1299 {
1300         int file = is_file_lru(lru);
1301
1302         if (lru == LRU_ACTIVE_FILE) {
1303                 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1304                 return 0;
1305         }
1306
1307         if (lru == LRU_ACTIVE_ANON &&
1308             (!scan_global_lru(sc) || inactive_anon_is_low(zone))) {
1309                 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1310                 return 0;
1311         }
1312         return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1313 }
1314
1315 /*
1316  * Determine how aggressively the anon and file LRU lists should be
1317  * scanned.  The relative value of each set of LRU lists is determined
1318  * by looking at the fraction of the pages scanned we did rotate back
1319  * onto the active list instead of evict.
1320  *
1321  * percent[0] specifies how much pressure to put on ram/swap backed
1322  * memory, while percent[1] determines pressure on the file LRUs.
1323  */
1324 static void get_scan_ratio(struct zone *zone, struct scan_control *sc,
1325                                         unsigned long *percent)
1326 {
1327         unsigned long anon, file, free;
1328         unsigned long anon_prio, file_prio;
1329         unsigned long ap, fp;
1330
1331         /* If we have no swap space, do not bother scanning anon pages. */
1332         if (nr_swap_pages <= 0) {
1333                 percent[0] = 0;
1334                 percent[1] = 100;
1335                 return;
1336         }
1337
1338         anon  = zone_page_state(zone, NR_ACTIVE_ANON) +
1339                 zone_page_state(zone, NR_INACTIVE_ANON);
1340         file  = zone_page_state(zone, NR_ACTIVE_FILE) +
1341                 zone_page_state(zone, NR_INACTIVE_FILE);
1342         free  = zone_page_state(zone, NR_FREE_PAGES);
1343
1344         /* If we have very few page cache pages, force-scan anon pages. */
1345         if (unlikely(file + free <= zone->pages_high)) {
1346                 percent[0] = 100;
1347                 percent[1] = 0;
1348                 return;
1349         }
1350
1351         /*
1352          * OK, so we have swap space and a fair amount of page cache
1353          * pages.  We use the recently rotated / recently scanned
1354          * ratios to determine how valuable each cache is.
1355          *
1356          * Because workloads change over time (and to avoid overflow)
1357          * we keep these statistics as a floating average, which ends
1358          * up weighing recent references more than old ones.
1359          *
1360          * anon in [0], file in [1]
1361          */
1362         if (unlikely(zone->recent_scanned[0] > anon / 4)) {
1363                 spin_lock_irq(&zone->lru_lock);
1364                 zone->recent_scanned[0] /= 2;
1365                 zone->recent_rotated[0] /= 2;
1366                 spin_unlock_irq(&zone->lru_lock);
1367         }
1368
1369         if (unlikely(zone->recent_scanned[1] > file / 4)) {
1370                 spin_lock_irq(&zone->lru_lock);
1371                 zone->recent_scanned[1] /= 2;
1372                 zone->recent_rotated[1] /= 2;
1373                 spin_unlock_irq(&zone->lru_lock);
1374         }
1375
1376         /*
1377          * With swappiness at 100, anonymous and file have the same priority.
1378          * This scanning priority is essentially the inverse of IO cost.
1379          */
1380         anon_prio = sc->swappiness;
1381         file_prio = 200 - sc->swappiness;
1382
1383         /*
1384          * The amount of pressure on anon vs file pages is inversely
1385          * proportional to the fraction of recently scanned pages on
1386          * each list that were recently referenced and in active use.
1387          */
1388         ap = (anon_prio + 1) * (zone->recent_scanned[0] + 1);
1389         ap /= zone->recent_rotated[0] + 1;
1390
1391         fp = (file_prio + 1) * (zone->recent_scanned[1] + 1);
1392         fp /= zone->recent_rotated[1] + 1;
1393
1394         /* Normalize to percentages */
1395         percent[0] = 100 * ap / (ap + fp + 1);
1396         percent[1] = 100 - percent[0];
1397 }
1398
1399
1400 /*
1401  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1402  */
1403 static unsigned long shrink_zone(int priority, struct zone *zone,
1404                                 struct scan_control *sc)
1405 {
1406         unsigned long nr[NR_LRU_LISTS];
1407         unsigned long nr_to_scan;
1408         unsigned long nr_reclaimed = 0;
1409         unsigned long percent[2];       /* anon @ 0; file @ 1 */
1410         enum lru_list l;
1411
1412         get_scan_ratio(zone, sc, percent);
1413
1414         for_each_evictable_lru(l) {
1415                 if (scan_global_lru(sc)) {
1416                         int file = is_file_lru(l);
1417                         int scan;
1418
1419                         scan = zone_page_state(zone, NR_LRU_BASE + l);
1420                         if (priority) {
1421                                 scan >>= priority;
1422                                 scan = (scan * percent[file]) / 100;
1423                         }
1424                         zone->lru[l].nr_scan += scan;
1425                         nr[l] = zone->lru[l].nr_scan;
1426                         if (nr[l] >= sc->swap_cluster_max)
1427                                 zone->lru[l].nr_scan = 0;
1428                         else
1429                                 nr[l] = 0;
1430                 } else {
1431                         /*
1432                          * This reclaim occurs not because zone memory shortage
1433                          * but because memory controller hits its limit.
1434                          * Don't modify zone reclaim related data.
1435                          */
1436                         nr[l] = mem_cgroup_calc_reclaim(sc->mem_cgroup, zone,
1437                                                                 priority, l);
1438                 }
1439         }
1440
1441         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1442                                         nr[LRU_INACTIVE_FILE]) {
1443                 for_each_evictable_lru(l) {
1444                         if (nr[l]) {
1445                                 nr_to_scan = min(nr[l],
1446                                         (unsigned long)sc->swap_cluster_max);
1447                                 nr[l] -= nr_to_scan;
1448
1449                                 nr_reclaimed += shrink_list(l, nr_to_scan,
1450                                                         zone, sc, priority);
1451                         }
1452                 }
1453         }
1454
1455         /*
1456          * Even if we did not try to evict anon pages at all, we want to
1457          * rebalance the anon lru active/inactive ratio.
1458          */
1459         if (!scan_global_lru(sc) || inactive_anon_is_low(zone))
1460                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1461         else if (!scan_global_lru(sc))
1462                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1463
1464         throttle_vm_writeout(sc->gfp_mask);
1465         return nr_reclaimed;
1466 }
1467
1468 /*
1469  * This is the direct reclaim path, for page-allocating processes.  We only
1470  * try to reclaim pages from zones which will satisfy the caller's allocation
1471  * request.
1472  *
1473  * We reclaim from a zone even if that zone is over pages_high.  Because:
1474  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1475  *    allocation or
1476  * b) The zones may be over pages_high but they must go *over* pages_high to
1477  *    satisfy the `incremental min' zone defense algorithm.
1478  *
1479  * Returns the number of reclaimed pages.
1480  *
1481  * If a zone is deemed to be full of pinned pages then just give it a light
1482  * scan then give up on it.
1483  */
1484 static unsigned long shrink_zones(int priority, struct zonelist *zonelist,
1485                                         struct scan_control *sc)
1486 {
1487         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1488         unsigned long nr_reclaimed = 0;
1489         struct zoneref *z;
1490         struct zone *zone;
1491
1492         sc->all_unreclaimable = 1;
1493         for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1494                 if (!populated_zone(zone))
1495                         continue;
1496                 /*
1497                  * Take care memory controller reclaiming has small influence
1498                  * to global LRU.
1499                  */
1500                 if (scan_global_lru(sc)) {
1501                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1502                                 continue;
1503                         note_zone_scanning_priority(zone, priority);
1504
1505                         if (zone_is_all_unreclaimable(zone) &&
1506                                                 priority != DEF_PRIORITY)
1507                                 continue;       /* Let kswapd poll it */
1508                         sc->all_unreclaimable = 0;
1509                 } else {
1510                         /*
1511                          * Ignore cpuset limitation here. We just want to reduce
1512                          * # of used pages by us regardless of memory shortage.
1513                          */
1514                         sc->all_unreclaimable = 0;
1515                         mem_cgroup_note_reclaim_priority(sc->mem_cgroup,
1516                                                         priority);
1517                 }
1518
1519                 nr_reclaimed += shrink_zone(priority, zone, sc);
1520         }
1521
1522         return nr_reclaimed;
1523 }
1524
1525 /*
1526  * This is the main entry point to direct page reclaim.
1527  *
1528  * If a full scan of the inactive list fails to free enough memory then we
1529  * are "out of memory" and something needs to be killed.
1530  *
1531  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1532  * high - the zone may be full of dirty or under-writeback pages, which this
1533  * caller can't do much about.  We kick pdflush and take explicit naps in the
1534  * hope that some of these pages can be written.  But if the allocating task
1535  * holds filesystem locks which prevent writeout this might not work, and the
1536  * allocation attempt will fail.
1537  *
1538  * returns:     0, if no pages reclaimed
1539  *              else, the number of pages reclaimed
1540  */
1541 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
1542                                         struct scan_control *sc)
1543 {
1544         int priority;
1545         unsigned long ret = 0;
1546         unsigned long total_scanned = 0;
1547         unsigned long nr_reclaimed = 0;
1548         struct reclaim_state *reclaim_state = current->reclaim_state;
1549         unsigned long lru_pages = 0;
1550         struct zoneref *z;
1551         struct zone *zone;
1552         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1553
1554         delayacct_freepages_start();
1555
1556         if (scan_global_lru(sc))
1557                 count_vm_event(ALLOCSTALL);
1558         /*
1559          * mem_cgroup will not do shrink_slab.
1560          */
1561         if (scan_global_lru(sc)) {
1562                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1563
1564                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1565                                 continue;
1566
1567                         lru_pages += zone_lru_pages(zone);
1568                 }
1569         }
1570
1571         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1572                 sc->nr_scanned = 0;
1573                 if (!priority)
1574                         disable_swap_token();
1575                 nr_reclaimed += shrink_zones(priority, zonelist, sc);
1576                 /*
1577                  * Don't shrink slabs when reclaiming memory from
1578                  * over limit cgroups
1579                  */
1580                 if (scan_global_lru(sc)) {
1581                         shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
1582                         if (reclaim_state) {
1583                                 nr_reclaimed += reclaim_state->reclaimed_slab;
1584                                 reclaim_state->reclaimed_slab = 0;
1585                         }
1586                 }
1587                 total_scanned += sc->nr_scanned;
1588                 if (nr_reclaimed >= sc->swap_cluster_max) {
1589                         ret = nr_reclaimed;
1590                         goto out;
1591                 }
1592
1593                 /*
1594                  * Try to write back as many pages as we just scanned.  This
1595                  * tends to cause slow streaming writers to write data to the
1596                  * disk smoothly, at the dirtying rate, which is nice.   But
1597                  * that's undesirable in laptop mode, where we *want* lumpy
1598                  * writeout.  So in laptop mode, write out the whole world.
1599                  */
1600                 if (total_scanned > sc->swap_cluster_max +
1601                                         sc->swap_cluster_max / 2) {
1602                         wakeup_pdflush(laptop_mode ? 0 : total_scanned);
1603                         sc->may_writepage = 1;
1604                 }
1605
1606                 /* Take a nap, wait for some writeback to complete */
1607                 if (sc->nr_scanned && priority < DEF_PRIORITY - 2)
1608                         congestion_wait(WRITE, HZ/10);
1609         }
1610         /* top priority shrink_zones still had more to do? don't OOM, then */
1611         if (!sc->all_unreclaimable && scan_global_lru(sc))
1612                 ret = nr_reclaimed;
1613 out:
1614         /*
1615          * Now that we've scanned all the zones at this priority level, note
1616          * that level within the zone so that the next thread which performs
1617          * scanning of this zone will immediately start out at this priority
1618          * level.  This affects only the decision whether or not to bring
1619          * mapped pages onto the inactive list.
1620          */
1621         if (priority < 0)
1622                 priority = 0;
1623
1624         if (scan_global_lru(sc)) {
1625                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1626
1627                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1628                                 continue;
1629
1630                         zone->prev_priority = priority;
1631                 }
1632         } else
1633                 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority);
1634
1635         delayacct_freepages_end();
1636
1637         return ret;
1638 }
1639
1640 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
1641                                                                 gfp_t gfp_mask)
1642 {
1643         struct scan_control sc = {
1644                 .gfp_mask = gfp_mask,
1645                 .may_writepage = !laptop_mode,
1646                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1647                 .may_swap = 1,
1648                 .swappiness = vm_swappiness,
1649                 .order = order,
1650                 .mem_cgroup = NULL,
1651                 .isolate_pages = isolate_pages_global,
1652         };
1653
1654         return do_try_to_free_pages(zonelist, &sc);
1655 }
1656
1657 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
1658
1659 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
1660                                                 gfp_t gfp_mask)
1661 {
1662         struct scan_control sc = {
1663                 .may_writepage = !laptop_mode,
1664                 .may_swap = 1,
1665                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1666                 .swappiness = vm_swappiness,
1667                 .order = 0,
1668                 .mem_cgroup = mem_cont,
1669                 .isolate_pages = mem_cgroup_isolate_pages,
1670         };
1671         struct zonelist *zonelist;
1672
1673         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1674                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1675         zonelist = NODE_DATA(numa_node_id())->node_zonelists;
1676         return do_try_to_free_pages(zonelist, &sc);
1677 }
1678 #endif
1679
1680 /*
1681  * For kswapd, balance_pgdat() will work across all this node's zones until
1682  * they are all at pages_high.
1683  *
1684  * Returns the number of pages which were actually freed.
1685  *
1686  * There is special handling here for zones which are full of pinned pages.
1687  * This can happen if the pages are all mlocked, or if they are all used by
1688  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1689  * What we do is to detect the case where all pages in the zone have been
1690  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1691  * dead and from now on, only perform a short scan.  Basically we're polling
1692  * the zone for when the problem goes away.
1693  *
1694  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
1695  * zones which have free_pages > pages_high, but once a zone is found to have
1696  * free_pages <= pages_high, we scan that zone and the lower zones regardless
1697  * of the number of free pages in the lower zones.  This interoperates with
1698  * the page allocator fallback scheme to ensure that aging of pages is balanced
1699  * across the zones.
1700  */
1701 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
1702 {
1703         int all_zones_ok;
1704         int priority;
1705         int i;
1706         unsigned long total_scanned;
1707         unsigned long nr_reclaimed;
1708         struct reclaim_state *reclaim_state = current->reclaim_state;
1709         struct scan_control sc = {
1710                 .gfp_mask = GFP_KERNEL,
1711                 .may_swap = 1,
1712                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1713                 .swappiness = vm_swappiness,
1714                 .order = order,
1715                 .mem_cgroup = NULL,
1716                 .isolate_pages = isolate_pages_global,
1717         };
1718         /*
1719          * temp_priority is used to remember the scanning priority at which
1720          * this zone was successfully refilled to free_pages == pages_high.
1721          */
1722         int temp_priority[MAX_NR_ZONES];
1723
1724 loop_again:
1725         total_scanned = 0;
1726         nr_reclaimed = 0;
1727         sc.may_writepage = !laptop_mode;
1728         count_vm_event(PAGEOUTRUN);
1729
1730         for (i = 0; i < pgdat->nr_zones; i++)
1731                 temp_priority[i] = DEF_PRIORITY;
1732
1733         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1734                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1735                 unsigned long lru_pages = 0;
1736
1737                 /* The swap token gets in the way of swapout... */
1738                 if (!priority)
1739                         disable_swap_token();
1740
1741                 all_zones_ok = 1;
1742
1743                 /*
1744                  * Scan in the highmem->dma direction for the highest
1745                  * zone which needs scanning
1746                  */
1747                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1748                         struct zone *zone = pgdat->node_zones + i;
1749
1750                         if (!populated_zone(zone))
1751                                 continue;
1752
1753                         if (zone_is_all_unreclaimable(zone) &&
1754                             priority != DEF_PRIORITY)
1755                                 continue;
1756
1757                         /*
1758                          * Do some background aging of the anon list, to give
1759                          * pages a chance to be referenced before reclaiming.
1760                          */
1761                         if (inactive_anon_is_low(zone))
1762                                 shrink_active_list(SWAP_CLUSTER_MAX, zone,
1763                                                         &sc, priority, 0);
1764
1765                         if (!zone_watermark_ok(zone, order, zone->pages_high,
1766                                                0, 0)) {
1767                                 end_zone = i;
1768                                 break;
1769                         }
1770                 }
1771                 if (i < 0)
1772                         goto out;
1773
1774                 for (i = 0; i <= end_zone; i++) {
1775                         struct zone *zone = pgdat->node_zones + i;
1776
1777                         lru_pages += zone_lru_pages(zone);
1778                 }
1779
1780                 /*
1781                  * Now scan the zone in the dma->highmem direction, stopping
1782                  * at the last zone which needs scanning.
1783                  *
1784                  * We do this because the page allocator works in the opposite
1785                  * direction.  This prevents the page allocator from allocating
1786                  * pages behind kswapd's direction of progress, which would
1787                  * cause too much scanning of the lower zones.
1788                  */
1789                 for (i = 0; i <= end_zone; i++) {
1790                         struct zone *zone = pgdat->node_zones + i;
1791                         int nr_slab;
1792
1793                         if (!populated_zone(zone))
1794                                 continue;
1795
1796                         if (zone_is_all_unreclaimable(zone) &&
1797                                         priority != DEF_PRIORITY)
1798                                 continue;
1799
1800                         if (!zone_watermark_ok(zone, order, zone->pages_high,
1801                                                end_zone, 0))
1802                                 all_zones_ok = 0;
1803                         temp_priority[i] = priority;
1804                         sc.nr_scanned = 0;
1805                         note_zone_scanning_priority(zone, priority);
1806                         /*
1807                          * We put equal pressure on every zone, unless one
1808                          * zone has way too many pages free already.
1809                          */
1810                         if (!zone_watermark_ok(zone, order, 8*zone->pages_high,
1811                                                 end_zone, 0))
1812                                 nr_reclaimed += shrink_zone(priority, zone, &sc);
1813                         reclaim_state->reclaimed_slab = 0;
1814                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1815                                                 lru_pages);
1816                         nr_reclaimed += reclaim_state->reclaimed_slab;
1817                         total_scanned += sc.nr_scanned;
1818                         if (zone_is_all_unreclaimable(zone))
1819                                 continue;
1820                         if (nr_slab == 0 && zone->pages_scanned >=
1821                                                 (zone_lru_pages(zone) * 6))
1822                                         zone_set_flag(zone,
1823                                                       ZONE_ALL_UNRECLAIMABLE);
1824                         /*
1825                          * If we've done a decent amount of scanning and
1826                          * the reclaim ratio is low, start doing writepage
1827                          * even in laptop mode
1828                          */
1829                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1830                             total_scanned > nr_reclaimed + nr_reclaimed / 2)
1831                                 sc.may_writepage = 1;
1832                 }
1833                 if (all_zones_ok)
1834                         break;          /* kswapd: all done */
1835                 /*
1836                  * OK, kswapd is getting into trouble.  Take a nap, then take
1837                  * another pass across the zones.
1838                  */
1839                 if (total_scanned && priority < DEF_PRIORITY - 2)
1840                         congestion_wait(WRITE, HZ/10);
1841
1842                 /*
1843                  * We do this so kswapd doesn't build up large priorities for
1844                  * example when it is freeing in parallel with allocators. It
1845                  * matches the direct reclaim path behaviour in terms of impact
1846                  * on zone->*_priority.
1847                  */
1848                 if (nr_reclaimed >= SWAP_CLUSTER_MAX)
1849                         break;
1850         }
1851 out:
1852         /*
1853          * Note within each zone the priority level at which this zone was
1854          * brought into a happy state.  So that the next thread which scans this
1855          * zone will start out at that priority level.
1856          */
1857         for (i = 0; i < pgdat->nr_zones; i++) {
1858                 struct zone *zone = pgdat->node_zones + i;
1859
1860                 zone->prev_priority = temp_priority[i];
1861         }
1862         if (!all_zones_ok) {
1863                 cond_resched();
1864
1865                 try_to_freeze();
1866
1867                 goto loop_again;
1868         }
1869
1870         return nr_reclaimed;
1871 }
1872
1873 /*
1874  * The background pageout daemon, started as a kernel thread
1875  * from the init process.
1876  *
1877  * This basically trickles out pages so that we have _some_
1878  * free memory available even if there is no other activity
1879  * that frees anything up. This is needed for things like routing
1880  * etc, where we otherwise might have all activity going on in
1881  * asynchronous contexts that cannot page things out.
1882  *
1883  * If there are applications that are active memory-allocators
1884  * (most normal use), this basically shouldn't matter.
1885  */
1886 static int kswapd(void *p)
1887 {
1888         unsigned long order;
1889         pg_data_t *pgdat = (pg_data_t*)p;
1890         struct task_struct *tsk = current;
1891         DEFINE_WAIT(wait);
1892         struct reclaim_state reclaim_state = {
1893                 .reclaimed_slab = 0,
1894         };
1895         node_to_cpumask_ptr(cpumask, pgdat->node_id);
1896
1897         if (!cpumask_empty(cpumask))
1898                 set_cpus_allowed_ptr(tsk, cpumask);
1899         current->reclaim_state = &reclaim_state;
1900
1901         /*
1902          * Tell the memory management that we're a "memory allocator",
1903          * and that if we need more memory we should get access to it
1904          * regardless (see "__alloc_pages()"). "kswapd" should
1905          * never get caught in the normal page freeing logic.
1906          *
1907          * (Kswapd normally doesn't need memory anyway, but sometimes
1908          * you need a small amount of memory in order to be able to
1909          * page out something else, and this flag essentially protects
1910          * us from recursively trying to free more memory as we're
1911          * trying to free the first piece of memory in the first place).
1912          */
1913         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
1914         set_freezable();
1915
1916         order = 0;
1917         for ( ; ; ) {
1918                 unsigned long new_order;
1919
1920                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1921                 new_order = pgdat->kswapd_max_order;
1922                 pgdat->kswapd_max_order = 0;
1923                 if (order < new_order) {
1924                         /*
1925                          * Don't sleep if someone wants a larger 'order'
1926                          * allocation
1927                          */
1928                         order = new_order;
1929                 } else {
1930                         if (!freezing(current))
1931                                 schedule();
1932
1933                         order = pgdat->kswapd_max_order;
1934                 }
1935                 finish_wait(&pgdat->kswapd_wait, &wait);
1936
1937                 if (!try_to_freeze()) {
1938                         /* We can speed up thawing tasks if we don't call
1939                          * balance_pgdat after returning from the refrigerator
1940                          */
1941                         balance_pgdat(pgdat, order);
1942                 }
1943         }
1944         return 0;
1945 }
1946
1947 /*
1948  * A zone is low on free memory, so wake its kswapd task to service it.
1949  */
1950 void wakeup_kswapd(struct zone *zone, int order)
1951 {
1952         pg_data_t *pgdat;
1953
1954         if (!populated_zone(zone))
1955                 return;
1956
1957         pgdat = zone->zone_pgdat;
1958         if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0))
1959                 return;
1960         if (pgdat->kswapd_max_order < order)
1961                 pgdat->kswapd_max_order = order;
1962         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1963                 return;
1964         if (!waitqueue_active(&pgdat->kswapd_wait))
1965                 return;
1966         wake_up_interruptible(&pgdat->kswapd_wait);
1967 }
1968
1969 unsigned long global_lru_pages(void)
1970 {
1971         return global_page_state(NR_ACTIVE_ANON)
1972                 + global_page_state(NR_ACTIVE_FILE)
1973                 + global_page_state(NR_INACTIVE_ANON)
1974                 + global_page_state(NR_INACTIVE_FILE);
1975 }
1976
1977 #ifdef CONFIG_PM
1978 /*
1979  * Helper function for shrink_all_memory().  Tries to reclaim 'nr_pages' pages
1980  * from LRU lists system-wide, for given pass and priority, and returns the
1981  * number of reclaimed pages
1982  *
1983  * For pass > 3 we also try to shrink the LRU lists that contain a few pages
1984  */
1985 static unsigned long shrink_all_zones(unsigned long nr_pages, int prio,
1986                                       int pass, struct scan_control *sc)
1987 {
1988         struct zone *zone;
1989         unsigned long nr_to_scan, ret = 0;
1990         enum lru_list l;
1991
1992         for_each_zone(zone) {
1993
1994                 if (!populated_zone(zone))
1995                         continue;
1996
1997                 if (zone_is_all_unreclaimable(zone) && prio != DEF_PRIORITY)
1998                         continue;
1999
2000                 for_each_evictable_lru(l) {
2001                         /* For pass = 0, we don't shrink the active list */
2002                         if (pass == 0 &&
2003                                 (l == LRU_ACTIVE || l == LRU_ACTIVE_FILE))
2004                                 continue;
2005
2006                         zone->lru[l].nr_scan +=
2007                                 (zone_page_state(zone, NR_LRU_BASE + l)
2008                                                                 >> prio) + 1;
2009                         if (zone->lru[l].nr_scan >= nr_pages || pass > 3) {
2010                                 zone->lru[l].nr_scan = 0;
2011                                 nr_to_scan = min(nr_pages,
2012                                         zone_page_state(zone,
2013                                                         NR_LRU_BASE + l));
2014                                 ret += shrink_list(l, nr_to_scan, zone,
2015                                                                 sc, prio);
2016                                 if (ret >= nr_pages)
2017                                         return ret;
2018                         }
2019                 }
2020         }
2021
2022         return ret;
2023 }
2024
2025 /*
2026  * Try to free `nr_pages' of memory, system-wide, and return the number of
2027  * freed pages.
2028  *
2029  * Rather than trying to age LRUs the aim is to preserve the overall
2030  * LRU order by reclaiming preferentially
2031  * inactive > active > active referenced > active mapped
2032  */
2033 unsigned long shrink_all_memory(unsigned long nr_pages)
2034 {
2035         unsigned long lru_pages, nr_slab;
2036         unsigned long ret = 0;
2037         int pass;
2038         struct reclaim_state reclaim_state;
2039         struct scan_control sc = {
2040                 .gfp_mask = GFP_KERNEL,
2041                 .may_swap = 0,
2042                 .swap_cluster_max = nr_pages,
2043                 .may_writepage = 1,
2044                 .swappiness = vm_swappiness,
2045                 .isolate_pages = isolate_pages_global,
2046         };
2047
2048         current->reclaim_state = &reclaim_state;
2049
2050         lru_pages = global_lru_pages();
2051         nr_slab = global_page_state(NR_SLAB_RECLAIMABLE);
2052         /* If slab caches are huge, it's better to hit them first */
2053         while (nr_slab >= lru_pages) {
2054                 reclaim_state.reclaimed_slab = 0;
2055                 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
2056                 if (!reclaim_state.reclaimed_slab)
2057                         break;
2058
2059                 ret += reclaim_state.reclaimed_slab;
2060                 if (ret >= nr_pages)
2061                         goto out;
2062
2063                 nr_slab -= reclaim_state.reclaimed_slab;
2064         }
2065
2066         /*
2067          * We try to shrink LRUs in 5 passes:
2068          * 0 = Reclaim from inactive_list only
2069          * 1 = Reclaim from active list but don't reclaim mapped
2070          * 2 = 2nd pass of type 1
2071          * 3 = Reclaim mapped (normal reclaim)
2072          * 4 = 2nd pass of type 3
2073          */
2074         for (pass = 0; pass < 5; pass++) {
2075                 int prio;
2076
2077                 /* Force reclaiming mapped pages in the passes #3 and #4 */
2078                 if (pass > 2) {
2079                         sc.may_swap = 1;
2080                         sc.swappiness = 100;
2081                 }
2082
2083                 for (prio = DEF_PRIORITY; prio >= 0; prio--) {
2084                         unsigned long nr_to_scan = nr_pages - ret;
2085
2086                         sc.nr_scanned = 0;
2087                         ret += shrink_all_zones(nr_to_scan, prio, pass, &sc);
2088                         if (ret >= nr_pages)
2089                                 goto out;
2090
2091                         reclaim_state.reclaimed_slab = 0;
2092                         shrink_slab(sc.nr_scanned, sc.gfp_mask,
2093                                         global_lru_pages());
2094                         ret += reclaim_state.reclaimed_slab;
2095                         if (ret >= nr_pages)
2096                                 goto out;
2097
2098                         if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
2099                                 congestion_wait(WRITE, HZ / 10);
2100                 }
2101         }
2102
2103         /*
2104          * If ret = 0, we could not shrink LRUs, but there may be something
2105          * in slab caches
2106          */
2107         if (!ret) {
2108                 do {
2109                         reclaim_state.reclaimed_slab = 0;
2110                         shrink_slab(nr_pages, sc.gfp_mask, global_lru_pages());
2111                         ret += reclaim_state.reclaimed_slab;
2112                 } while (ret < nr_pages && reclaim_state.reclaimed_slab > 0);
2113         }
2114
2115 out:
2116         current->reclaim_state = NULL;
2117
2118         return ret;
2119 }
2120 #endif
2121
2122 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2123    not required for correctness.  So if the last cpu in a node goes
2124    away, we get changed to run anywhere: as the first one comes back,
2125    restore their cpu bindings. */
2126 static int __devinit cpu_callback(struct notifier_block *nfb,
2127                                   unsigned long action, void *hcpu)
2128 {
2129         int nid;
2130
2131         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2132                 for_each_node_state(nid, N_HIGH_MEMORY) {
2133                         pg_data_t *pgdat = NODE_DATA(nid);
2134                         node_to_cpumask_ptr(mask, pgdat->node_id);
2135
2136                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2137                                 /* One of our CPUs online: restore mask */
2138                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2139                 }
2140         }
2141         return NOTIFY_OK;
2142 }
2143
2144 /*
2145  * This kswapd start function will be called by init and node-hot-add.
2146  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2147  */
2148 int kswapd_run(int nid)
2149 {
2150         pg_data_t *pgdat = NODE_DATA(nid);
2151         int ret = 0;
2152
2153         if (pgdat->kswapd)
2154                 return 0;
2155
2156         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2157         if (IS_ERR(pgdat->kswapd)) {
2158                 /* failure at boot is fatal */
2159                 BUG_ON(system_state == SYSTEM_BOOTING);
2160                 printk("Failed to start kswapd on node %d\n",nid);
2161                 ret = -1;
2162         }
2163         return ret;
2164 }
2165
2166 static int __init kswapd_init(void)
2167 {
2168         int nid;
2169
2170         swap_setup();
2171         for_each_node_state(nid, N_HIGH_MEMORY)
2172                 kswapd_run(nid);
2173         hotcpu_notifier(cpu_callback, 0);
2174         return 0;
2175 }
2176
2177 module_init(kswapd_init)
2178
2179 #ifdef CONFIG_NUMA
2180 /*
2181  * Zone reclaim mode
2182  *
2183  * If non-zero call zone_reclaim when the number of free pages falls below
2184  * the watermarks.
2185  */
2186 int zone_reclaim_mode __read_mostly;
2187
2188 #define RECLAIM_OFF 0
2189 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
2190 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
2191 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
2192
2193 /*
2194  * Priority for ZONE_RECLAIM. This determines the fraction of pages
2195  * of a node considered for each zone_reclaim. 4 scans 1/16th of
2196  * a zone.
2197  */
2198 #define ZONE_RECLAIM_PRIORITY 4
2199
2200 /*
2201  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2202  * occur.
2203  */
2204 int sysctl_min_unmapped_ratio = 1;
2205
2206 /*
2207  * If the number of slab pages in a zone grows beyond this percentage then
2208  * slab reclaim needs to occur.
2209  */
2210 int sysctl_min_slab_ratio = 5;
2211
2212 /*
2213  * Try to free up some pages from this zone through reclaim.
2214  */
2215 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2216 {
2217         /* Minimum pages needed in order to stay on node */
2218         const unsigned long nr_pages = 1 << order;
2219         struct task_struct *p = current;
2220         struct reclaim_state reclaim_state;
2221         int priority;
2222         unsigned long nr_reclaimed = 0;
2223         struct scan_control sc = {
2224                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2225                 .may_swap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2226                 .swap_cluster_max = max_t(unsigned long, nr_pages,
2227                                         SWAP_CLUSTER_MAX),
2228                 .gfp_mask = gfp_mask,
2229                 .swappiness = vm_swappiness,
2230                 .isolate_pages = isolate_pages_global,
2231         };
2232         unsigned long slab_reclaimable;
2233
2234         disable_swap_token();
2235         cond_resched();
2236         /*
2237          * We need to be able to allocate from the reserves for RECLAIM_SWAP
2238          * and we also need to be able to write out pages for RECLAIM_WRITE
2239          * and RECLAIM_SWAP.
2240          */
2241         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2242         reclaim_state.reclaimed_slab = 0;
2243         p->reclaim_state = &reclaim_state;
2244
2245         if (zone_page_state(zone, NR_FILE_PAGES) -
2246                 zone_page_state(zone, NR_FILE_MAPPED) >
2247                 zone->min_unmapped_pages) {
2248                 /*
2249                  * Free memory by calling shrink zone with increasing
2250                  * priorities until we have enough memory freed.
2251                  */
2252                 priority = ZONE_RECLAIM_PRIORITY;
2253                 do {
2254                         note_zone_scanning_priority(zone, priority);
2255                         nr_reclaimed += shrink_zone(priority, zone, &sc);
2256                         priority--;
2257                 } while (priority >= 0 && nr_reclaimed < nr_pages);
2258         }
2259
2260         slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2261         if (slab_reclaimable > zone->min_slab_pages) {
2262                 /*
2263                  * shrink_slab() does not currently allow us to determine how
2264                  * many pages were freed in this zone. So we take the current
2265                  * number of slab pages and shake the slab until it is reduced
2266                  * by the same nr_pages that we used for reclaiming unmapped
2267                  * pages.
2268                  *
2269                  * Note that shrink_slab will free memory on all zones and may
2270                  * take a long time.
2271                  */
2272                 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
2273                         zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
2274                                 slab_reclaimable - nr_pages)
2275                         ;
2276
2277                 /*
2278                  * Update nr_reclaimed by the number of slab pages we
2279                  * reclaimed from this zone.
2280                  */
2281                 nr_reclaimed += slab_reclaimable -
2282                         zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2283         }
2284
2285         p->reclaim_state = NULL;
2286         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
2287         return nr_reclaimed >= nr_pages;
2288 }
2289
2290 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2291 {
2292         int node_id;
2293         int ret;
2294
2295         /*
2296          * Zone reclaim reclaims unmapped file backed pages and
2297          * slab pages if we are over the defined limits.
2298          *
2299          * A small portion of unmapped file backed pages is needed for
2300          * file I/O otherwise pages read by file I/O will be immediately
2301          * thrown out if the zone is overallocated. So we do not reclaim
2302          * if less than a specified percentage of the zone is used by
2303          * unmapped file backed pages.
2304          */
2305         if (zone_page_state(zone, NR_FILE_PAGES) -
2306             zone_page_state(zone, NR_FILE_MAPPED) <= zone->min_unmapped_pages
2307             && zone_page_state(zone, NR_SLAB_RECLAIMABLE)
2308                         <= zone->min_slab_pages)
2309                 return 0;
2310
2311         if (zone_is_all_unreclaimable(zone))
2312                 return 0;
2313
2314         /*
2315          * Do not scan if the allocation should not be delayed.
2316          */
2317         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
2318                         return 0;
2319
2320         /*
2321          * Only run zone reclaim on the local zone or on zones that do not
2322          * have associated processors. This will favor the local processor
2323          * over remote processors and spread off node memory allocations
2324          * as wide as possible.
2325          */
2326         node_id = zone_to_nid(zone);
2327         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
2328                 return 0;
2329
2330         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2331                 return 0;
2332         ret = __zone_reclaim(zone, gfp_mask, order);
2333         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2334
2335         return ret;
2336 }
2337 #endif
2338
2339 #ifdef CONFIG_UNEVICTABLE_LRU
2340 /*
2341  * page_evictable - test whether a page is evictable
2342  * @page: the page to test
2343  * @vma: the VMA in which the page is or will be mapped, may be NULL
2344  *
2345  * Test whether page is evictable--i.e., should be placed on active/inactive
2346  * lists vs unevictable list.  The vma argument is !NULL when called from the
2347  * fault path to determine how to instantate a new page.
2348  *
2349  * Reasons page might not be evictable:
2350  * (1) page's mapping marked unevictable
2351  * (2) page is part of an mlocked VMA
2352  *
2353  */
2354 int page_evictable(struct page *page, struct vm_area_struct *vma)
2355 {
2356
2357         if (mapping_unevictable(page_mapping(page)))
2358                 return 0;
2359
2360         if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2361                 return 0;
2362
2363         return 1;
2364 }
2365
2366 /**
2367  * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2368  * @page: page to check evictability and move to appropriate lru list
2369  * @zone: zone page is in
2370  *
2371  * Checks a page for evictability and moves the page to the appropriate
2372  * zone lru list.
2373  *
2374  * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2375  * have PageUnevictable set.
2376  */
2377 static void check_move_unevictable_page(struct page *page, struct zone *zone)
2378 {
2379         VM_BUG_ON(PageActive(page));
2380
2381 retry:
2382         ClearPageUnevictable(page);
2383         if (page_evictable(page, NULL)) {
2384                 enum lru_list l = LRU_INACTIVE_ANON + page_is_file_cache(page);
2385
2386                 __dec_zone_state(zone, NR_UNEVICTABLE);
2387                 list_move(&page->lru, &zone->lru[l].list);
2388                 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
2389                 __count_vm_event(UNEVICTABLE_PGRESCUED);
2390         } else {
2391                 /*
2392                  * rotate unevictable list
2393                  */
2394                 SetPageUnevictable(page);
2395                 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
2396                 if (page_evictable(page, NULL))
2397                         goto retry;
2398         }
2399 }
2400
2401 /**
2402  * scan_mapping_unevictable_pages - scan an address space for evictable pages
2403  * @mapping: struct address_space to scan for evictable pages
2404  *
2405  * Scan all pages in mapping.  Check unevictable pages for
2406  * evictability and move them to the appropriate zone lru list.
2407  */
2408 void scan_mapping_unevictable_pages(struct address_space *mapping)
2409 {
2410         pgoff_t next = 0;
2411         pgoff_t end   = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
2412                          PAGE_CACHE_SHIFT;
2413         struct zone *zone;
2414         struct pagevec pvec;
2415
2416         if (mapping->nrpages == 0)
2417                 return;
2418
2419         pagevec_init(&pvec, 0);
2420         while (next < end &&
2421                 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
2422                 int i;
2423                 int pg_scanned = 0;
2424
2425                 zone = NULL;
2426
2427                 for (i = 0; i < pagevec_count(&pvec); i++) {
2428                         struct page *page = pvec.pages[i];
2429                         pgoff_t page_index = page->index;
2430                         struct zone *pagezone = page_zone(page);
2431
2432                         pg_scanned++;
2433                         if (page_index > next)
2434                                 next = page_index;
2435                         next++;
2436
2437                         if (pagezone != zone) {
2438                                 if (zone)
2439                                         spin_unlock_irq(&zone->lru_lock);
2440                                 zone = pagezone;
2441                                 spin_lock_irq(&zone->lru_lock);
2442                         }
2443
2444                         if (PageLRU(page) && PageUnevictable(page))
2445                                 check_move_unevictable_page(page, zone);
2446                 }
2447                 if (zone)
2448                         spin_unlock_irq(&zone->lru_lock);
2449                 pagevec_release(&pvec);
2450
2451                 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
2452         }
2453
2454 }
2455
2456 /**
2457  * scan_zone_unevictable_pages - check unevictable list for evictable pages
2458  * @zone - zone of which to scan the unevictable list
2459  *
2460  * Scan @zone's unevictable LRU lists to check for pages that have become
2461  * evictable.  Move those that have to @zone's inactive list where they
2462  * become candidates for reclaim, unless shrink_inactive_zone() decides
2463  * to reactivate them.  Pages that are still unevictable are rotated
2464  * back onto @zone's unevictable list.
2465  */
2466 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
2467 void scan_zone_unevictable_pages(struct zone *zone)
2468 {
2469         struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
2470         unsigned long scan;
2471         unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
2472
2473         while (nr_to_scan > 0) {
2474                 unsigned long batch_size = min(nr_to_scan,
2475                                                 SCAN_UNEVICTABLE_BATCH_SIZE);
2476
2477                 spin_lock_irq(&zone->lru_lock);
2478                 for (scan = 0;  scan < batch_size; scan++) {
2479                         struct page *page = lru_to_page(l_unevictable);
2480
2481                         if (!trylock_page(page))
2482                                 continue;
2483
2484                         prefetchw_prev_lru_page(page, l_unevictable, flags);
2485
2486                         if (likely(PageLRU(page) && PageUnevictable(page)))
2487                                 check_move_unevictable_page(page, zone);
2488
2489                         unlock_page(page);
2490                 }
2491                 spin_unlock_irq(&zone->lru_lock);
2492
2493                 nr_to_scan -= batch_size;
2494         }
2495 }
2496
2497
2498 /**
2499  * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
2500  *
2501  * A really big hammer:  scan all zones' unevictable LRU lists to check for
2502  * pages that have become evictable.  Move those back to the zones'
2503  * inactive list where they become candidates for reclaim.
2504  * This occurs when, e.g., we have unswappable pages on the unevictable lists,
2505  * and we add swap to the system.  As such, it runs in the context of a task
2506  * that has possibly/probably made some previously unevictable pages
2507  * evictable.
2508  */
2509 void scan_all_zones_unevictable_pages(void)
2510 {
2511         struct zone *zone;
2512
2513         for_each_zone(zone) {
2514                 scan_zone_unevictable_pages(zone);
2515         }
2516 }
2517
2518 /*
2519  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
2520  * all nodes' unevictable lists for evictable pages
2521  */
2522 unsigned long scan_unevictable_pages;
2523
2524 int scan_unevictable_handler(struct ctl_table *table, int write,
2525                            struct file *file, void __user *buffer,
2526                            size_t *length, loff_t *ppos)
2527 {
2528         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
2529
2530         if (write && *(unsigned long *)table->data)
2531                 scan_all_zones_unevictable_pages();
2532
2533         scan_unevictable_pages = 0;
2534         return 0;
2535 }
2536
2537 /*
2538  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
2539  * a specified node's per zone unevictable lists for evictable pages.
2540  */
2541
2542 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
2543                                           struct sysdev_attribute *attr,
2544                                           char *buf)
2545 {
2546         return sprintf(buf, "0\n");     /* always zero; should fit... */
2547 }
2548
2549 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
2550                                            struct sysdev_attribute *attr,
2551                                         const char *buf, size_t count)
2552 {
2553         struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
2554         struct zone *zone;
2555         unsigned long res;
2556         unsigned long req = strict_strtoul(buf, 10, &res);
2557
2558         if (!req)
2559                 return 1;       /* zero is no-op */
2560
2561         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2562                 if (!populated_zone(zone))
2563                         continue;
2564                 scan_zone_unevictable_pages(zone);
2565         }
2566         return 1;
2567 }
2568
2569
2570 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
2571                         read_scan_unevictable_node,
2572                         write_scan_unevictable_node);
2573
2574 int scan_unevictable_register_node(struct node *node)
2575 {
2576         return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
2577 }
2578
2579 void scan_unevictable_unregister_node(struct node *node)
2580 {
2581         sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
2582 }
2583
2584 #endif